Ad 0767826

AMCP-706-192
JC PAMPHLET WICP 706-192
vy^K*-"
7
<7fe7P-b
ENGINEERING DESIGN
HANDBOOK
to*.:; . ' ?=' vsmm
REDSTONE ^Ci.\;:i;/iCir:-r.:::{;f-.TiON Cf.NT
COMPUTER AIDED DESIGN

^ REDSTONE SC.ENTIRCINFOBMA^
5 0510 00078496 4
MECHANICAL SYSTEMS
L
nn f.*" >-*"".- <?"" '' * w
."^.S!
y&i *.<> k~i^-- ' -i-v. '.;;A,
EADQUARTERS, U.S. ARMY MATERIEL COMMAND JULY 1973

HEADQUARTERS
US ARMY MATERIEL COMMAND
5001 EISENHOWER AVENUE
ALEXANDRIA, VA 22304
15 July 1973
AMC PAMPHLET
NO. 706-192
ENGINEERING DESIGN HANDBOOK

COMPUTER AIDED DESIGN
OF MECHANICALSYSTEMS
TABLE OF CONTENTS
Paragraph Page
LIST OF ILLUSTRATIONS ix
LIST OF TABLES xiii
PREFACE XV
CHAPTER 1. ELEMENTS OF COMPUTER AIDED DESIGN
1- -1 Synthesis vs Analysis in Engineering Design 1-1

1- -2 The Philosophy of System Engineering 1-2
1- -3 Computer Aided Design in the
Mechanical Sciences 1-4
1- -4 Mathematical Preliminaries 1-S
1- -5 Illustrative Military Computer Aided
Design Problems 1-8
1- -5 1 Optimal Design of Structures 1-8
1- -5.2 Application of the Steepest Descent Method
in Interactive Computer Aided Design 1-13
1- -5 3 Design of Artillery Recoil Mechanisms 1-15
References 1-17
AMCP 70G-192
TABLE OF CONTENTS (CorVt.)
Paragraph Page
CHAPTER 2. FINITE DIMENSIONAL

UNCONSTRAINED OPTIMIZATION
2-1 Introduction 2-1

22 Necessary Conditions for Extrema 22
23 One-dimensional Minimization 26
23.1 Quadratic Interpolation 26
23.2 Fibonacci Search (or Golden Section Search) .... 27
24 The Method of Steepest Descent (or Gradient) . . . 28
2-5 A Generalized Newton Method 2-11
26 Methods of Conjugate Directions 212
2-6.1 The Conjugate Gradient Method 2-14
2-6.2 The Method of Fletcher and Powell 2-16
26.3 A Conjugate Direction Method
Without Derivatives 217
27 Comparison of the Various Methods 218
27.1 Method of Steepest Descent 218
2-7.1.1 Cost Function^ (x) 2-18
27.1.2 Cost Function/2(x) 219
2-7.1.3 Cost Function/3 (x) 2-19
27.2 Generalized Newton Method 219
2-7.2.1 Cost Function/j (x) 2-19
27.2.2 Cost Function f2(x) 220
2-7.3 Conjugate Gradient Method 2-20
2-7.3.1 Cost Function/! (x) 2-20
2-7.3.2 Cost Function/2(x) 2-21
2-7.4 Fletcher-Powell Method 2-21
27.5 Conjugate Directions Without Derivatives 222
28 An Application of Unconstrained
Optimization to Structural Analysis 224
References 225
AMCP 70G-192
Paragraph Page
CHAPTER 3. LINEAR PROGRAMMING
31 Introduction 31
3_2 Properties of Linear Programs 32
3_3 The Simplex Algorithm 38
33.1 Determination of a Basic Feasible Point 38
3-3.2 Solution of LP 3-9
3 3.3 The Degenerate Case 310
3_4 Minimum Weight Truss Design 3 12
35 An Application of Linear Programming
to Analysis 3 14
References 3 17
CHAPTER 4. NONLINEAR PROGRAMMING AND

FINITE DIMENSIONAL OPTIMAL DESIGN
41 Introduction to the Theory of

Nonlinear Programming (NLP) 41
4- -1.1 Nonlinear Programming Problems 41
4--1.2 Global Theory 44
4- -1.3 Local Theory 45
4- -2 Theory of Finite Dimensional Optimal Design .... 48
4- -2.1 Finite Dimensional Optimal Design
Problems 48
4- -2.2 Local Theory 410
4- -3 Sequentially Unconstrained Minimization
Techniques (SUMT) 4-11
4- -3.1 Interior Method 4-11
4- -3.2 Exterior Method 414
4- -3.3 Mixed Interior-Exterior Method 416
4- -3.4 Determination of an Interior Point 418
4- -4 Steepest Descent Methods for NLP 4-19
4- -4.1 The Direction of Steepest Descent 420
4- -4.2 Step Size Determination 423
4- -4.2.1 Rosen's Method for Linear Constraints 423
4- -4.2.2 Fixed Step With Variable Weighting 4-24
4- -4.2.3 Steepest Descent With Constraint
Tolerances 425
4-4.3 A Steepest Descent Method With
Constraint Error Compensation 425
111
AMCP 706-192
TABLE OF CONTENTS (Con't.)
Paragraph rage
45 Steepest Descent Solution of the Finite
Dimensional Optimal Design Problem 427
45.1 An Approximation of the Problem OD 428
45.2 Solution of the Approximate Problem 430
45.3 Steepest Descent Algorithm 433
45.4 Use of the Computational Algorithm 434
References 435
CHAPTER 5. FINITE DIMENSIONAL

OPTIMAL STRUCTURAL DESIGN
51 Introduction 51
5 1.1 Lightweight vs Structural Performance
Trade-offs 5-2
5 1.2 Weapon Development Problems Associated
With Lightweight Requirements 52
5 1.2.1 Aircraft Armament 52
5-1.2.2 Gun Barrel Design 5-3
5-1.2.3 Towed Artillery 5-3
5-1.2.4 Other Weapon Problems 5-4
5 1.3 Plan for Technique Development 54
52 Elements of the Elastic Structural
Design Problem 54
5-2.1 The Optimality Criterion 5-5
52.2 Stress and Displacement Due to
Static Loading 55
52.3 Natural Frequency and Buckling 56
5-2.4 Method of Solution 5-7
53 Steepest Descent Programming for
Optimal Structural Design 57
53.1 Linearized Cost and Constraint Functions 58
53.2 Steepest Descent Algorithm for
Optimal Structural Design 5 10
53.3 Computational Considerations 510
54 Optimization of Special Purpose Structures 5 11
5-4.1 A Minimum Weight Column 5-11
54.2 A Minimum Weight Vibrating Beam 514
54.3 A Minimum Weight Portal Frame With a
Natural Frequency Constraint 516
54.4 A Minimum Weight Frame With Multiple
Failure Criteria 5 18
AMCP70G-192
Paragraph Page
54.5 A Minimum Weight Plate With Frequency

Constraints 520
55 General Treatment of Truss Design 522
5_5.1 Special Problem Formulation 522
55.1.1 Frequency Constraints 524
5-5.1.2 Stress Constraints 5-25
5-5.1.3 Buckling Constraints 5-25
5 5.1.4 Displacement Constraints 526
5 5.1.5 Bounds on Design Variables 526
5-5.2 Multiple Loading Condtions 5-28
5 5.3 Example Problems 528
56 A General Treatment of Plane Frame Design 536
56.1 Problem Formulation 537
56.2 Stress Constraint Calculations 539
56.3 Example Problems 546
57 Interactive Computing in Structural
Optimization 551
57.1 The Interactive Approach 551
57.2 Interactive Structural Design Using
Sensitivity Data 5-52
57.3 Example Problems 5 54
57.4 Interactive Computing Conclusions 561
References 562
CHAPTER 6. THE CALCULUS OF VARIATIONS AND

OPTIMAL PROCESS THEORY
6-1 Introduction 6-1

6-2 The Fundamental Problem of the Calculus
of Variations 6-4
6-2.1 Necessary Conditions for the Fundamental
Problem 6-5
6-2.2 Special Cases and Examples 6 8
6-2.3 Variational Notation and Second-order
Conditions 6-10
6-2.4 Direct Methods 6-13
6-2.4.1 The Ritz Method 6-14
6-2.4.2 Method of Finite Differences 6-15
6-3 A Problem of Bolza 6-16
6-3 1 Statement of the Problem 6-16
6-3.2 6-17
AMCP 706192
Pa e
Paragraph 8
63.3 Necessary Conditions for the Bolza

Problem 6-18
63.4 Application of the Bolza Problem 622
64 Problems of Optimal Design and Control 627
64.1 Design Variable Inequality Constraints 628
64.2 State Variable Inequality Constraints 631
64.3 Application of the Theory of Optimal
Design 633
65 Methods of Satisfying Necessary Conditions 640
65.1 Initial Value Methods (or Shooting
Techniques) 640
6-5.2 A Generalized Newton Method 6-42
References 645
CHAPTER 7. OPTIMAL STRUCTURAL DESIGN

BY THE INDIRECT METHOD
7- -1 Introduction 71
7- -1.1 The Class of Problems Considered 71
7- -1.2 Historical Development 71
7- -1.3 Methods Employed 72
7- -2 A Minimum Weight Column 72
7- -3 A Minimum Weight Structure With
Angular Deflection Requirements 75
7- -3.1 Statement of the Problem 75
7- -3.2 Tower With One Design Variable 77
7- -3.2.1 Method 1. Tower With Base Rigidly
Fastened to the Earth 78
7-3.2.2 Method 2. Tower With Base Pinned to
Earth and With Top Supported by
Guy Lines 712
73.3 Tower With Two Design Variables 713
7-3.3.1 Method 1. Tower With Base Rigidly
Fastened to the Earth 714
7-3.3.2 Method 2. Tower With Base Pinned to
Earth and With Top Supported by
Guy Lines 7-14
73.4 Discussion of Results 715
74 Minimum Weight Design of Beams With
Inequality Constraints on Stress and
Deflection 7-16
AMCP 706-192
Paragraph Page
74.1 Statement of the Problem 716

74.2 Necessary Conditions for the Beam
Design Problem 719
74.3 Statically Indeterminate Problems 723
7_4.4 Solutions of the Equations of Theorem 7-1 724
74.5 Beams With Rectangular Cross Section
of Variable Depth 7-25
7-4.5 1 A Problem Which Can Be Solved
Analytically 7-29
74.5.2 Simply Supported Beam With Positive
Distributed Load 7-32
7_4.5.3 A Problem of a More General Type 7-38
7-4.5.4 Conclusions 7-43
References 744
CHAPTER 8. METHODS OF STEEPEST DESCENT

FOR OPTIMAL DESIGN PROBLEMS
3 1 Introduction 81
32 A Steepest Descent Method for the Basic
Optimal DesignProblem 82
3-2.1 The Problem Considered 8-2
32.2 Effects of Small Changes in Design
Variables and Parameters 82
32.3 A Steepest Descent Approach 85
33 A Steepest Descent Method for a General
Optimal Design Problem 815
3-3.1 The Problem Considered 8-15
33.2 The Effect of Small Changes in Design
3-3.3 A Steepest Descent Computational
Algorithm 8-23
3-4 Steepest Descent Programming for a Class of Systems
Described by Partial Differential Equations 825
3-4. 1 The Class of Problems Considered 825
3-4. 2 Effect of Small Changes in Design
84.3 A Steepest Descent Computational
Algorithm 8-29
Vii
AMCP 706-192
Paragraph Page
85 Optimal Design of an Artillery Recoil

Mechanism 835
85.1 Formulation of the Problem 837
8-5.2 Equations of Motion for the XM164
Howitzer 8-39
85.3 Steepest Descent Formulation 843
85.3.1 Determination of the Adjoint Equations 844
85.3.2 Determination of the Boundary Conditions
for the Adjoint Equations 845
8 5.3.3 Computation of Design Improvements 847
85.4 Results and Conclusions 848
References 849
CHAPTER 9. APPLICATION OF STEEPEST DESCENT

METHODS TO OPTIMAL STRUCTURAL DESIGN
91 Introduction 91
9-2 Steepest Descent Method for Optimal
Structural Design 92
9-3 A Minimum Weight Column 9-8
9_4 A Minimum Weight Vibrating Beam 99
95 A Minimum Weight Vibrating Frame 910
96 A Minimum Weight Frame With Multiple
Failure Criteria 914
9-7 A Minimum Weight Vibrating Plate 9-16
References 918
APPENDIX A CONVEXITY A-l
APPENDIX B ANALYSIS OF BEAM-TYPE

STRUCTURES B-l
INDEX i_ i
Mil
AMCP 706-192
LIST OF ILLUSTRATIONS
Hgr. N3. Title Bage
11 A System Engineering Model 12

12 Structural Requirement 18
13 Conceptual Designs 19
1 4 Uniform Initial Design 1 11
15 Direction of Steepest Descent 1 12
16 Tower With Base Rigidly Fastened
to the Earth 1-12
17 Tower With Base Simply Supported and
Top Supported With Guy Lines 1 13
18 Sensitivity to Design Variations 1 14
19 Sensitivity to Two Performance Indicators 1 14
1-10 Howitzer. Towed. 105mm, XM164 1-16
1 11 Traditional Recoil Design Goal 1 16
1 12 Recoil Distribution in Time 1 16
1-13 Sensitivity to Gun Hop 1-17
1 14 Optimum Recoil Curve 1 17
2-1 f{x) = (x - 2)2 2-2
2-2 A Cost Function 2-3
23 Function of Single Variable 27
2-4 Interval Partition 2-7
2-5 Descent Steps 2-10
31 Graphical Solution of Example 3-1 32
3-2 Polyhedral Constraint Set 3-5
33 Admissible Joints for Bridge Truss 3 14
3 4 Optimum Bridge Trusses 3 14
35 Boundary Condition for Example 3-4 3 16
4 1 Graphical Solution of Example 4-1 43
42 First-order Constraint Qualification 46
43 Penalty Functions 411
5-1 Column 5-12
5-2 Column Element 5-12
53 Profiles of Optimal Columns 5 14
5-4 Stepped Beam 5-14
55 Typical Element 5 15
56 Profile of Optimum Beam 5 16
5-7 Portal Frame 5-16
58 Typical Elements 5 16
59 Optimum Portal Frame for w = 3000
rad/sec 5 18
IX
AMCP 70G-192
LIST OF ILLUSTRATIONS (Con't.)
Fig. No. Title Page
5 10 Frame With Side Loading 5 18

5_11 Typical Elements 5-19
5 _ 12 Profile of Optimal Frame With Multiple
Criteria (q = 25 lb/in.) 5-20
5-13 Rectangular Plate 5-20
514 Collocation Points 5-21
5-15 Optimal Design Variable h(x, y)
for Vibrating Plate 522
516 Description of a Truss 523
5 17 A Truss Element 5-23
518 Four-bar Truss (Example 5-1) 529
519 Iteration vs Weight Curves for Example 5-1.
Four-bar Truss 529
5_20 Transmission Tower (Example 5-2) 531
521 Iteration vs Weight Curves for Example 5-2,
Transmission Tower 533
5-22 47-Bar Plane Truss (Example 5-3) 5-34
47-Bar Plane Truss 5-36
5_24 Description of a Frame 538
5-25 A Frame Element 5-38
526 Simple Portal Frame (Example 5-4) 546
Simple Portal Frame 547
5-28 One-bay, Two-story Frame (Example 5-5J 5-48
529 Iteration vs Weight Curves for Example 5-5;
One-bay, Two-story Frame 549
5 30 Two-bay, Six-story Frame (Example 5-6) 549
53 1 Iteration vs Weight Curves for Example 5-6;
Two-bay, Six-story Frame. With Stress
Constraints Only 551
5 32 Iteration vs Weight Curves for Example 5-6;
Two-bay, Six-story Frame. With All
Constraints 551
5 33 Vector Change in Design Space 552
5_34 Three-bar Truss 5-53
535 Display of Design Sensitivity Data 5 53
5-36 Local Optima 5-54
5-37 Trusses (Example 5-7) 5-55
AMCP 706-192
Eg. No. Eife Bags

Three-bar Truss. With Stress Constraints
Only 5-55
Three-bar Truss. With All Constraints 556
5_40 Transmission Tower (Example 5-8) 558
Transmission Tower. With Stress Constraints
Only 5-62
Transmission Tower. With All Constraints 562
6-1 Shortest Path 6-1
62 Curve for Minimum Time 62
63 Examples of Continuous Functions 63
6-4 A Neighborhood of x (t) 6-4
65 Perturbation from Optimum 65
66 Graphical Proof of Lemma 6-1 66
67 Minimizing Sequence 69
6-8 Particle in Motion 6-22
6-9 Orbit Transfer 6-25
6-10 Thrust Program 6-25
6-11 Ground Vehicle 6-33
6 12 Extremal Arcs 636
6-13 Extremal Arcs With Straight Section 6-36
6 14 Bounded Brachistochrone 638
615 Bounded Brachistochrone Solution 640
71 Column Under Consideration 72
7-2 Profiles of Optimal Columns 7-6
73 Tower Considered 7-6
7-4 Loading of Tower 78
75 Tower With Base Rigidly Fastened to the
Earth 7-12
7-6 Tower With Guy Lines 7-12
77 Tower With Base Simply Supported and
Top Supported With Guy Lines 7-13
78 Beam Loaded in a General Way 7 17
79 Rectangular Cross Section 725
7-10 Simple Cantilever Beam 7-29
711 Cantilever Beam of Minimum Weight 732
AMCP 706-192
Fig. No. Title Page
712 Simply Supported Beam With Positive

Distributed Load 7-32
713 Subdivision of the Beam With Distributed
Load 7-33
7-14 Beam With an Inflection Point 7-38
715 Subdivision of the Beam 7 39
7_ 16 Volume vs Deflection Requirement 743
7_17 Profile of Optimal Beam for A = 0.16 7-43
7_18 Profile of Optimal Beam for A = 0.15 7-43
8-1 Howitzer. Towed. 105 mm. XM164 8-36
8-2 Recoil Force for a Rigid Mount 8-37
8-3 Time Intervals 8-37
8-4 Schematic of XM164 105 mm Towed
Howitzer Dynamic Model 839
85 Recoil Time Interval 845
8-6 Optimal Rod Force 8-48
8-7 Optimal Control Rod Design 8-48
9 1 Simply Supported Beam 92
92 Simply Supported Vibrating Beam 99
9-3 Profile of Optimal Beam 9-11
9_4 Portal Frame 9-11
9_5 Profile of Optimum Frame 913
96 Laterally Loaded Frame 914
9-7 Free Bodies 9-14
98 Profile of Minimum Weight Frame 917
9-9 Simply Supported Plate 9-17
9-10 Contours of Optimum Plate 9-18
A 1 Examples; Convex Case and Nonconvex Case A 1
A-2 Graph of/(x) = x2 in R[ A-l
B 1 Basic Beam Element B-1
AMCP 706-192
LIST OF TABLES
Table No. Title Page
1_1 Weights of Towers 1 13

2 1 Steepest Descent Method Iterative Data
for Cost Function/j (x) 2-19
22 Steepest Descent Method Iterative Data
for Cost Function/2(x) 2-19
23 Steepest Descent Method Iterative Data
for Cost Function/3 (x) 2-19
24 Generalized Newton Method Iterative Data
for Cost Function/j (x) . , 2-20
for Cost Function/2 (x) 220
for Cost Function/3 (x) , 2-20
27 Conjugate Gradient Method Iterative Data
for Cost Function/j (x) 221
for Cost Function/2 (x) 2-21
for Cost Function/3 (x) 221
2-10 Fletcher-Powell Method - Iterative Data
for Cost Function/j (x) 222
213 Conjugate Directions Without Derivatives
Method Iterative Data for Cost
Function/j (x) 2-23
Function/2 (x) 223
Function/3 (x) 2-23
51 Comparison of Uniform and Optimal Columns ... 513
52 Cross-sectional Areas of Optimum Columns 5 14
53 Comparison of Optimum Beams 5 16
5_4 Material Properties for Aluminum 517
55 Comparison of Uniform and Optimal Frames
for Aluminum 5 18
AMCP 706-192
LIST OF TABLES (Con't.)
Table No. Hue Page
56 Optimal Design Variable bi for

Vibrating Frame 5 18
57 Optimal Design Variable b{ for
Static Frame 5-20
58 Volume of Optimum Frame 520
59 Four-bar Truss (Example 5-1) 530
510 Transmission Tower (Example 5-2) 532
5 11 47-bar Plane Truss (Example 5-3) 535
5-12 Simple Portal Frame (Example 5-4) 5-47
513 One-bay, Two-story Frame (Example 5-5) 548
5 14 Two-bay. Six-story Frame (Example 5-6) 550
5 15 Optimum Three-member Trusses (Example 5-7) .. 557
516 Design Information for Transmission Tower
(Example 5-8) 5-59
517 Optimum Transmission Towers With Stress
Constraints Only (Example 5-8) 560
518 Optimum Transmission Towers With All
Constraints (Example 5-8) 561
71 Results for Column Problem 75
7-2 Constants 7-7
7-3(A) Weights of Simply Supported Towers.
One Design Variable 711
73(B) Weights of Guy-line Supported Towers.
One Design Variable 711
7-3(C) Weights of Towers 7-11
74 Results for Simply Supported Beam With
q(t) = t 7-37
75 Results for Beam With S-shaped Deflection
Curve 7-42
91 Comparison of Optimal Beams 911
92 Weight of Optimum Frames 913
93 Volume of Optimum Frame 916
xiv
AMCP 706-192
PREFACE
The Engineering Design Handbooks of the U. S. Army Materiel Command

are a coordinated series of handbooks containing basic information and
fundamental data useful in the design and development of Army materiel
and systems.
This text treats a broad class of optimal design problems through use of a
consistent set of computational techniques ideally suited for computer
application to mechanical design problems. No attempt has been made to be
exhaustive in the treatment of optimization techniques or the full range of
mechanical applications. Rather, the class of problems treated is concisely
formulated (in Chapters 4 and following) in terms of design and state
variables that occur in mechanical design. A steepest-descent approach
which has served as a workhorse, reliable technique in fields such as
aerodynamic system design, control theory, and nonlinear programming is
developed here for mechanical system design.
Extensive application of design optimization techniques is made in the

field of structural design, as well as in a limited number of specific weapon
design problems. The examples are presented in considerable detail, as they
are encountered in practice, to provide the practicing engineer with insight
into use of the methods for his class of problems. A consistent design
philosophy is maintained throughout the text to assist the designer in
extrapolating the methods to classes of problems that are only similar
mathematically to the examples treated here.
The text is structured so that it can be understood and used by practicing

engineers with a good background in calculus and matrix theory. Computa-
tional algorithms are stated in considerable detail so that they can be
effectively implemented by junior engineers, with only problem formulation
and general supervision provided by a senior project engineer. As with
virtually all computer aided design techniques, some computing art is
required for effective implementation of these techniques. The detailed
treatment of structural applications in Chapters 5,7, and 9 should provide
insight into this computational art. References are given to more advanced
literature for proofs of theorems and extensions of methods to other classes
of problems.
AMCP 706-192
The Handbook was written by Dr. Edward J. Haug, Jr. of the U. S. Army
Weapons Command. It is based on lecture materials used by him in a
two-semester graduate sequence on "Optimization of Structural Systems",
taught at the University of Iowa since 1968. Examples treated in the text are
derived primarily from Dr. Haug's research, Dr. Jasbir Arora's University of
Iowa dissertation, and the work of Messrs. Tom Streeter and Stephen Newell
of the U.S. Army Weapons Command.
The Engineering Design Handbooks fall into two basic categories, those
approved for release and sale, and those classified for security reasons. The
Army Materiel Command policy is to release these Engineering Design
Handbooks to other DOD activities and their contractors and other
Government agencies in accordance with current Army Regulation 70-3 1,
dated 9 September 1966. It will be noted that the majority of these
Handbooks can be obtained from the National Technical Information
Service (NTIS). Procedures for acquiring these Handbooks follow:
a. Activities within AMC, DOD agencies and Government agencies other

than DOD having need for the Handbooks should direct their request on an
official form to:
Commander
Letterkenny Army Depot
ATTN: AMXLE-ATD
Chambers burg, PA 17201
b. Contractors and universities must forward their requests to:
National Technical Information Service

Department of Commerce
Springfield, VA 22151
(Requests for classified documents must be sent, with appropriate "Need to

Know" justification, to Letterkenny Army Depot.)
Comments and suggestions on this Handbook are welcome and should be

addressed to:
US Army Materiel Command

ATTN: AMCRD-TV
5001 Eisenhower Avenue
Alexandria, VA 22304
DA Forms 2028 (Recommended Changes to Publications), which are

available through normal publications supply channels, may be used for
comments/suggestions.
XVI
AMCP 708-192
CHAPTER 1
ELEMENTS OF COMPUTER AIDED DESIGN
1-1 SYNTHESIS VS ANALYSIS IN ENGI- sign changes are made and the structure is re-
NEERINGDESIGN analyzed. This process continues until the
designer is satisfied with his design. This has
Engineering is defined (Ref. 1) as "the art been the principal use of the computer in the
or science of making practical application of design process.
the knowledge of pure sciences such as
physics, chemistry, biology, etc.". Although In general, then, before the designer can
broad, this definition implies that the job of assure himself that he has the best system, he
engineering is to synthesize, or put together, must be capable of analyzing all candidates.
useful systems by applying knowledge and In the past half century, outstanding advances
methods derived from the "pure" sciences. in engineering analysis have been made. The
The meaning of "practical" in the given digital computer has allowed the engineer to
definition should be interpreted as best, or quantitatively analyze the behavior of systems
optimal; i.e., the job of engineering design is that were examined only qualitatively in the
to develop the best possible system for the past. The mechanical sciences, particularly,
given application, consistent with the re- have benefited from this boom in analysis
sources allocated to the development phase. capability. Structural analysis, stress analysis,
The purpose of this handbook is to present a analysis of mechanisms, and heat transfer
class of methods that allow for efficient use analysis, just to name a few, have made
of the computer in the design process. spectacular advances in the past twenty years.
Since the computer can be viewed simply Until the early 1960's, and even to the
as a device to handle large quantities of data present day to a lesser extent, the attention of
and perform simple algebraic operations and engineering research has been focused pri-
logic rapidly, it is important to look first into marily on developing analysis capability. Dur-
the role of calculation in design. The usual ing this period of emphasis on analysis,
approach to design is to conceive of a inadequate attention was paid to developing a
candidate system and then test it to see if it synthesis, or design, capability that is able to
works. Great strides have been made with efficiently use the newly developed analysis
digital computers in the past two decades to methods. In some of the mechanical sciences,
allow for numerical analysis as a test of the this problem is particularly acute. In struc-
idea, or concept, rather than previous cut-and- tural mechanics, for example, it is possible to
try techniques. For example, in structural analyze a structure under a given loading to
design one chooses the configuration and obtain accurate values for stress, displace-
member sizes, and then tests the structure by ment, and even natural frequency. It is not
analyzing its response to given loads. If the clear, however, how a structure should be laid
structure does not behave as desired, then de- out and proportioned to efficiently utilize
1-1
AMCP 706-192
material in order to meet strength require- A key challenge to developers of practical

ments. A more difficult problem is the pro- computer aids to designers is to take maxi-
portioning of a structure so as to efficiently mum advantage of human judgment in the
limit displacement and meet constraints on design process. The potential of interactive
natural frequency and buckling. For a review computation and design information display
of the state of optimal structural design is only now in a developing stage and holds
through 1967, see Ref. 2. promise for significant improvement of the
value of the computer in design.
It appears that the analysis capability
needed for computer aided design is available. 1-2 THE PHILOSOPHY OF SYSTEM ENGI-
The next problem to be addressed, then, is NEERING
the matter of what is meant by best, or
optimum. The idea of best enters very natu- In the middle 1950's a formalized approach
rally into engineering design efforts. In to the development of large-scale, man-made
profit-motivated industries as well as in systems began to appear in the literature, see
Government laboratories, the objective is to Refs. 3, 4, 5. This approach, which has
maximize some return function while satisfy- features common to most problem solving
ing constraints such as resource allocation, processes, was given the name "system engi-
performance requirements, and human limita- neering" and is the very essence of computer
tions. aided design. A feature which sets system
engineering and computer aided design off
Once some return function or measure of from most of the logical problem solving
value is chosen and constraints are identified, schemes is the explicit inclusion of key
the system designer would like to have some considerations peculiar to engineering design
optimal design methodology that is capable of of systems. A second important feature of
aiding him in the determination of the best, system engineering is the attention paid to
or practically best, system. It must be empha- quantitative description of the system and its
sized at this point that the search is not for an behavior.
automatic optimization technique that can
solve any design problem fed to it. Rather, The basic idea in system engineering is to
the need is for an optimal design methodol- begin with a statement of system require-
ogy that can aid the engineer in the implements and objectives, and move in an orga-
mentation of his concepts and guide him in a nized way toward an optimum system. A
direction which, if continued indefinitely, process which illustrates the approach is
would yield a mathematical optimum. shown in Fig. 1-1.
I
"XT
ES
Examine
" " * *L Complete
yp.fv
leeasn ana ^
Function
Analysis
_ Conceptual
Design * Optimal
Design
Description
Inputs Output
Objective
Figure 1-1. A System Engineering Model
1-2
AMCP 708-192
The purpose of this text is not to give a to describe analytically. Conceptual design, as
detailed treatment of system engineering, but its name implies, is the identification of the
rather to present aspects of the theory of various concepts or basic system configura-
computer aided design, with emphasis on tions that might meet the system objectives.
optimal design. The simplified model of a It is desirable in this step to leave the
system engineering process shows that opti- concepts as general as possible so as not to
mal design is a part of system engineering, eliminate candidate systems that might be
but, indeed, by no means the dominant part. very effective. For example, if the function to
The purpose of this paragraph is to discuss the be performed is to propel a vehicle over the
interface of optimal design with the remaining surface of the earth, conceptual designs might
essential elements of system engineering. include wheels, tracks, legs, air cushion, etc.
System engineering begins with the identifi- It is important at this time to identify
cation of a need by a potential user of the ranges of values of parameters describing the
system to be developed. It is often the case system so that, for any parameter in this
that the user knows that he needs a system to range of values, the system will perform the
do a job, but he may have difficulty in stating functions identified in the previous step, i.e.,
his needs and objectives quantitatively. It the set of parameters describing admissible
then becomes the joint responsibility of the systems is identified. It is at this time that the
system engineer and user to quantify system experienced designer can be extremely valu-
objectives so that a meaningful set of objec- able in reflecting state-of-the-art capabilities
tives may be established for the development of technologies involved in the system devel-
to follow. opment.
Once the needs and objectives for a system The optimal design step has as its objective
are identified, it is necessary to define func- the choice of the undetermined parameters
tions that must be performed by the system identified in the previous step. These param-
and any subsystems that are required. This eters must be in the ranges defined by
process is called function analysis, and its technological limitations and system func-
purpose is to pick out functions or operations tions. The criterion for choosing system
that must be performed in order to accom- parameters is maximization of system worth
plish the mission required of the system being or value. It should be emphasized that a
developed. These functions then become mathematically precise optimum may be im-
lower level objectives for the development of possible to attain and must therefore serve
subsystems. Identification of functions tends only as a goal. Methods for choosing system
to be qualitative in nature. However, once a parameters should, however, have the prop-
function or operation is identified, it must be erty that if an optimum does exist, then given
described in quantitative terms, if at all enough patience and computer time, that
possible. For example, if a function must optimum should be approached as a limit.
occur quickly, the time allowed should be
specified. What appears to be the final step in the
system engineering model of Fig. 1-1, Descrip-
The next step shown in Fig. 1-1 is one of tion, is, in reality, probably just an inter-
the most difficult functions in system engi- mediate step. Unless the system design pro-
neering and certainly the most difficult step cedure has been unusually effective, the sys-
1-3
AMCP 706-192
tern decided upon will probably not satisfy it applies to the mechanical sciences. There
the user. More likely, it will probably not are peculiarities of mechanical design, as
satisfy the system engineering team. Having opposed to classical control system design,
the results of one pass through the system which require specialized treatment. Further,
engineering process, the user can probably the mathematics involved in mechanical sys-
remember some constraints which he forgot tem design is quite different from the math-
to specify and which the optimum system ematics of control theory. These distinctions
violates. The designer probably also will see are highlighted throughout the text.
concepts that he did not see before. Much as
the user, he too will remember technological In the chapters that follow, optimal control
constraints which he forgot to specify and theory is interpreted as treating feedback
which the optimum system violates. Finally, controllers; i.e., an optimal control system has
the sponsoring activity will undoubtedly de- active elements that sense errors in output,
cide that it will be all right to decrease the due to fluctuations in inputs, and adjust
measure of system value a small amount if it system controls so as to maximize some
will save some money. measure of system performance. Optimal de-
sign, on the other hand, is taken as the
The next step in the procedure is for each problem of choosing system elements or
member of the team to take a deep breath, parameters describing these elements, which
sigh and go back to work, armed with his are fixed for the life of the elements, so that
hard earned new knowledge. It is for this the system is optimum in some sense. In
purpose that all the feedback paths in the control literature this is called open loop
model of Fig. 1-1 are shown. This iterative control. The principal difference in the two
procedure is then continued until the sponsor- problems is that the variables chosen in the
ing activity decides that the system developed optimal design problem are fixed for the life
is what it really needs. This will probably be of the system, whereas variables in a feedback
another human decision, rather than a pro- control device are to be adjusted according to
grammed mathematical one. inputs as the system operates. Mathematical-
ly, the difference in the two results is that the
The remaining chapters will be devoted to control law describes how the system vari-
the problem of computer aided and optimal ables should be adjusted as a function of the
design. If the design methods presented later state of the system, whereas an optimum
are to be of maximum value to the reader, he design is simply a set of parameters describing
must have a clear picture of how these system elements and will not be changed
methods fit into the larger problem of system during the life of the system. This distinction
engineering. For further literature on the is not uniform in the control literature but is
basic ideas involved in system engineering, see used here to identify the class of problems
Refs. 3, 4, and 5. treated.
1-3 COMPUTER AIDED DESIGN IN THE In most literature on control problems,

MECHANICAL SCIENCES sequential systems are treated, i.e., operations
of the system progress one after another as if
The theory of computer aided and optimal they were occurring in time in a pre-arranged
design is developed in subsequent chapters as order. Many optimal design problems are not
1-4
AMCP 706-192
of this kind. For example, in designing a techniques available for design optimization.
structure one must be concerned with stresses Since the subject of this handbook is com-
due to applied loads. These stresses are puter aids to design, the practical distinction
interpreted as the state of the structural is made here. For a unifying mathematical
system. They are determined by a boundary- treatment, the reader is referred to Ref. 7.
value problem that cannot be interpreted as a
sequential process (initial-value problem). In Finally, it is important to realize that
some design problems it is possible to define engineering design optimization and engineer-
auxiliary variables so that the governing equa- ing analysis are fundamentally different in
tions form an initial-value problem with addi- nature. In analysis, one is generally assured
tional constraints. This procedure, however, that a solution exists and numerical methods
generally complicates the problem unneces- are generally stable. In optimal design, on the
sarily. For this reason the problems in other hand, existence of even a nominal
succeeding chapters are formulated as bound- design satisfying objectives is not assured,
ary as opposed to initial-value problems. much less existence of an optimal design.
Moreover, even when an optimum exists,
In order to illustrate the use of the meth- numerical methods for its solution are often
ods presented, applications are made pri- quite sensitive to initial estimates and require
marily in optimal structural design. Applica- much computational art for iterative con-
tions are chosen to illustrate the use of the vergence. These properties will be observed
methods on problems having a number of over and over in this handbook when example
design variables which might be found in problems are treated.
engineering applications. Further, since many
of the methods are relatively new, it is It is important that the reader take a
anticipated that improvements in computa- mathematical outlook when doing computer
tional efficiency may be realized in specific aided design and optimization. A purely
problems if advantage is taken of special intuitive approach can lead to erroneous
features of the class of problems treated. results that may not be apparent until some-
one happens onto a nominal design which is
It is appropriate to highlight a significant vastly superior to a "supposed" optimum
computational distinction between two design.
classes of design problem. The reader may
note that Chapters 2 through 5 of this text 1-4 MATHEMATICAL PRELIMINARIES
deal with problems in which system design
and performance are specified by a finite The level of mathematical background re-
number of parameters (real numbers). Chap- quired for an understanding of the methods
ters 6 through 9, on the other hand, deal with of optimal design presented in the following
systems that are described by functions on chapters is a course in advanced calculus and
some given space or time domain. Mathemati- the ability to use matrix notation. Since
cally, these problems are called finite and engineers often require results of rather deep
infinite dimensional, respectively. Optimiza- mathematical analyses to solve real-world
tion theory for these two classes of problems problems, several results have been accepted
can be put in the same form, but there are with references given to proofs. The purpose
very real differences in the computational of this paragraph is to present notation and
1-5
AMCP 706-192
some basic mathematical ideas used through- Two vectors are called orthogonal if their
out the text. inner product is zero.
Since most real-world problems involve The idea of convergence of a sequence

several parameters, it is convenient to utilize { x' (in R" with norm (Eq. 1-2) is much like
vector notation. For example, rather than convergence of real numbers. That is, j^ x' =
writing x x..., xn repeatedly, these n variables x if for any E > 0 there is anN > 0 suchthat
are collected into a column vector \\x' x || < E for all i> N. An important
property of sets in optimization theory is
closedness. A subset D of R" is called closed
(1-1) if every sequence in D which converges has its
limit inD.
Unless otherwise noted, all vector variables Just as the idea of collecting n real numbers
will be column vectors. A vector of the form into a vector in R", it is helpful to define a
(Eq. 1-1) may be interpreted as a point in vector functiong(x) for x&R" as
n-dimensional real space, R" . The spaced" is
simply the collection of all n-vectors of real
numbers. For example, the real line is A1 and g(x)- (1-4)
the plane is R2.
It will often be convenient to deal with a Such a function is called continuous at x if

collection of points in the space R". A for any E > 0 there is a 6 > 0 such that \\g(x)
collection of points D in R" will be called a g(x) || < if \\'x x \\n < 6.The subscripts
set, or a subset of R". A point x inR" which m and n on the norms denote the dimension
is in D will be denoted xeD. This will be the of the space on which the norm is defined.
extent of set notation required in later chap-
ters. It will often be desirable to deal with a set
of functions which satisfy
In R" there is a well defined idea of length
of a vector. This analog of length in the real gAx)<0,i=\, m (1-5)
world will be denoted
1/2 In this case it is convenient to define inequal-
(1-2) ity among vectors as
g(x)<0 (1-6)
and is called a norm on". There are many
norms defined on" but Eq. 1-2 will be
where inequality is taken componentwise, i.e.,
sufficient for the purposes of this text. Along
Eq. 1-6 is defined to mean the same thing as
with the idea of norm on Rn goes the concept
Eq. 1-5.
of dot product or inner product. The inner
product of two elementsx andy oiRn is One of the most useful notations in the
(1-3) following chapters is the idea of the derivative
x, y > =x'y
i -1 of a vector function with respect to its vector
1-6
AMCP 706-192
variable. This notation is I: a* /(*)

+...+
dgjx) bgj (x)
dx m x n (1-7)
X.
Oi -xji.-.iyn-xjn
where i is a row index and / is a column index.

If fix) is a real valued function of x&R", this 1 L 9^+1/(g)
notation is (fc + i)i / +. ..+/ = fc + i ai:. ...a+
dfjx) d/(x),...,a/(x) (1-8)

dx OX. dx
(yi-xi)71--- (y -*)'"
The derivative of a real valued function is
often called the gradient of that function and
is denoted For proof of this theorem see Ref. 6, page 56.
dfjx) (1-9)
V/(x)s In many places in the following chapters,
dx
Taylor's Theorem will be used to obtain an
approximate expression for a function at a
The gradient is one of the few standard point sufficiently near a point where the
symbols which denotes a row vector rather function is known. The most common ap-
than a column vector. Likewise, for a real proximation is the one obtained by deleting
valued function the matrix of second deriva- second and higher order terms. For example,
tives may be defined as the matrix if \\x ~y\\ is small,
d2f(x) d2f{x)
= 72f(x) = (1-10)
f(y) -fix) =* (y-x) (1-12)
dx2' OX; OXj
dx
An important theorem in the analysis of where by Eq. 1-11 the error in Eq. 1-12 is
functions appearing in optimal design prob- at most a constant times \\y x\\2 if f(x) has
lems is Taylor's Theorem. bounded second order derivatives. The left
side of Eq. 1-12 is generally denoted by
Taylor's Theorem: Let the real valued 8f(x), where y x is denoted Sx. In this
function fix) have k + 1 continuous deriva- notation,
tives in Rn . Then for xeR", there is a point
= ax + (1 u)y with 0 < a < 1, such that
;
/<*) dx Sx.
' (1-13)
" ofix)
/O0 =/(*)+ 2 -^-t-ty-x,) (1-11) Eq. 1-13 may be thought of as a total
; = 1 dx- ' '
differential. Even for vector functions g(x),
1 n n d2f(x) Eq. 1-13 holds for each component so if
+
2 2 . ' : ' (yt -x;.)0,.-x.)
2 /'=1 1=1 3jC; 9X;
8g(x)=[d8l(x),.. ., 6gm(x)]T, then
1-7
AMCP 706-192
dg(x) structural design.

&g(x) = 5x. (1-14)
dx
As a specific example, let us consider a
In later work, g(x) will often be a function design problem whereby a highly directional
of x&R" and zsRp. In this case, Eq. l-14is transmission device, or perhaps a gun, is to be
mounted on a tower or gun mount that is
d8(X Z)
"su/; - ' - -dgiX'ZUz (1-15) required to support the device at some given
dx ~" dz distance away from the basic supporting
structure, such as the earth. A schematic of
lere
the problem is shown in Fig. 1-2. The basic
dg(x,z) dgt{x,z)
dx dx-
m X n
and
dgjx.z)
dz dz,
_I m X p
Figure 1-2. Structural Requirement
Most of the common notation used in later
chapters now has been defined. Special nota-
tion and results required locally for some problem is to design a structure that supports
development will be defined and used there. the device under consideration and which is as
light as possible for purposes of transporta-
1-5 ILLUSTRATIVE MILITARY COM- tion and erection on the battlefield, or per-
PUTER AIDED DESIGN PROBLEMS haps mounting on a helicopter. A basic design
requirement for this structure is that the
In this paragraph two illustrative military device mounted on the top shall not have an
optimal design problems are formulated, and angular deflection of more than 6 radians, in
computer aided design techniques are out- order to hit the receiver or target. The loading
lined for their solution. The treatment here is that is to be considered is a wind load of up
only for the purpose of introducing concepts. to a given velocity, which would cause angular
These examples are treated in more depth in deflection of the top of the tower.
Chapters 7 and 8.
The needs and objectives in this design
1-5.1 OPTIMAL DESIGN OF STRUCTURES problem are well established, so no additional
inputs need be considered at the present time.
The optimization technique described in Further, the requirement that the tower
this paragraph was initially developed for support the device with only a given allowable
application to minimum weight structural angular deflection is the only basic function
design problems. For this reason, and to give required of the tower; thus the function
an engineering feel for application of the analysis block of Fig. 1-1 is also complete.
technique, the method will be presented along The next stage, and one that is quite difficult
with examples from the field of optimal to describe analytically, is that of arriving at
AMCP 708-192
conceptual towers which might perform the Figs. 1-3(A) and (B), involve rigidly fixing the
given mission. tower at its base to the fundamental support-
ing structure. In both towers, variable spacing
Four different conceptual designs are as a function of height is allowed between
shown in Fig. 1-3. The first two concepts, vertical members of the structure. In addition,
*-Z6
Figure 1-3. Conceptual Designs

one of the concepts allows for varying the ent area and spacing as desired. Three are
area of the main structural members as a shown for convenience in the figure.
function of height. The second set of con-
cepts, Figs. 1-3(C) and (D), involves towers In each of the conceptual towers of Fig.
that are pinned at their base to the supporting 1-3, the variables b{ through b3 describe the
structure and that are supported by guy wires variable spacing of the members of the tower.
at the top of the structure. Likewise, in both In two of the concepts, Figs. 1-3(B) and (D),
of these concepts, variable spacing of the 64 through bb specify the variable areas in
main vertical members is allowed. In the the construction of the main vertical member.
second concept, variation of area along the These variables serve as design parameters, in
length of the tower is also allowed. It should that the designer can choose these variables
be noted that the conceptual designs in Fig. and completely specify the design of the
1-3 can have as many subsections with differ- tower.
1-9
AMCP 706-192
In addition to the design variables, a main ture, it is required that the design variables be
part of the design problem is the behavior of bounded uniformly away from zero. This is
the structure under wind load, since one of given formally by the inequality
the major constraints on behavior of the
structure is that the angular deflection of the b >b
i io> 0,i=\,. ,m. (1-19)
top of the tower not exceed an angle 6. For
this reason, the angular deflection of each of The fundamental constraint in the present
the joints must be determined, along with problem is that the angular deflection at the
lateral deflection due to lateral wind loading. top of the tower not exceed the angle 0. This
This is a relatively routine analysis problem is expressed analytically by the inequality
when one uses the techniques of finite ele-
ment structural analysis. Not shown in Fig. u,u e. (i-20)
1-3, but required in the construction, are
cross members which maintain spacing of the The final step in formulation of an optimal
main vertical members. In order to state the design problem is to identify the cost func-
optimal design problem mathematically, first tion to be minimized. In the present case, the
define vectors of design variables b( and state cost function is structural weight J and is
variables z, given by an expression of the form
b=[b,,b2,. ..bm]T
(1-16) / = 7.2 cibi (1-21)
Z = [Z,, Zi, -z
n]
Using finite element structural analysis tech- where y is material density and c( are weight-
niques, define the stiffness matrix as ing factors representing lengths of structural
elements and weight requirements for lateral
A(b)=laij(b)]nxn (1-17) stiffners.
where the dependence of stiffness on the We now have an optimal structural design
design variables is explicitly shown. Using this problem that is well formulated from a
matrix, the structural response is given by the mathematical point of view. The objective is
following matrix equation to find the design variables bt through b
that satisfy constraint Eqs. 1-19 and 1-20, and
A(b)z=q (1-18) which minimize the structural weight as given
by Eq. 1-21. The technique used to solve this
where q is the wind loading matrix. problem, and in fact a large class of optimal
system design problems, is based on a very
Now that the relationship between the simple idea of engineering design. The idea of
design variables and the structural response is the technique is to allow small variations in
specified by Eq. 1-18, the next step in some nominal design, and analyze the effect
formulating an optimal design problem is the of these variations on the equations of the
identification of constraints. In order to problem and the cost function associated with
prevent dimensions or structural areas from the problem. As a result of allowing only
going to zero, resulting in an unstable struc- small design changes, certain approximations
1-10
AMCP 708-192
may be made that allow the best change in 66,. >*,-*,- (1-24)
design variables to be determined in order to
decrease the cost function of the problem as Or, if the angular deflection constraint is
much as possible, while still not violating violated, for example,
constraints of the design problem. For
example, one might choose as an initial zi > e (1-25)
estimate of the optimal design a uniform
tower as shown in Fig. 1-4. The estimated then, to correct the constraint error it is
design variable in this case is denoted by b^\ required that
8z, < e - z i. (1-26)
Finally, the change in structural weight due to

the change in design 6b is given by
5/ = 7/i c.bb.. (1-27)
The object of the new problem is to

/////////////////// determine 6b so as to minimize the linearized
cost function of Eq. 1-27, subject to con-
straint Eqs. 1-24 and 1-26. Due to the special
nature of this problem, the optimum change
Figure 1-4. Uniform initial Design
8b can be determined in closed form. For a
detailed derivation of this optimum perturba-
Let 8b be a small change in the design tion, the reader is referred to Chapter 5. For
variable 6(0). Any change in the design discussion here, the results of this calculation
variable will result in a change in the struc- will be denoted by
tural response, denoted by 8z. The nature of
the structural analysis problem guarantees 8b = rjB + C (1-28)
that small 6b yields small 8z. Further, a
Taylor series approximation of terms appear- where the vectors B and C depend on 6(0),
ing in Eq. 1-18 yields constraint errors, and equations of the prob-
lem. The parameter 77 is an undetermined
parameter that plays the role of a step size,
A(b^)8z+-^-(A(b)z\ m)8b = 0. (1-22) when viewed in the geometry of design
variable space. For example, if there are only
ff an inequality constraint is violated, such two design parameters bt and b2, the direc-
as tion of the desired change is shown by B in
Fig. 1-5, and r\ is a step size along that
bs < b, (1-23) direction. In the terminology of optimization
theory, B is known as the direction of
then in order to correct the constraint error it steepest descent. It is analogous to the direc-
is required that tion one would go downhill in order to reduce
1-11
AMCP 706-192
descent method is based on requesting a

Optimum
certain reduction in the cost function due to
the changed 6b. This request, then, deter-
mines the step size 77 and one can calculate bb
from Eq. 1-28. This 6b is the best change in
the estimated design variable M\ This best
change is then added to the initial estimate to
obtain a new estimate that corresponds to a
Figure 1-5. Direction of Steepest Descent structure of less weight and that still satisfies
the constraints of the problem, i.e.,
his altitude as rapidly as possible. It is clear 6(l)=ft(0)+fife. (1-29)

that on normal hills, as in most design cost
functions, the direction of steepest descent This process is repeated as many times as
changes, depending on the location on that required to obtain convergence to the mini-
hill. For this reason, the direction of steepest- mum weight structure.
descent does not generally pass through the
optimum point as shown in Fig. 1-5. The optimum towers for each of the four
basic configurations chosen are shown in Figs.
There are many techniques for choosing 1-6 and 1-7, with a table of results being given
the step size r\. The one used in the steepest in Table 1-1. These results were obtained
(A) One Design Variable (B) Two Design Variable
Figure 7-6. Tower With Base Rigidly Fastened to the Earth
1-12
AMCP 707-192
using a finite element model with approxi-

mately forty elements so that the resulting
structure has an essentially continuous distri-
bution of material and spacing. The weights
shown in Table 1-1, corresponding to no
design variables are simply the weights of the
optimum towers having uniform members and
no variation in spacing. Note that there is a
significant reduction in structural weight for
the tapered optimum towers over uniform
towers. Extensive examples of this kind are
presented in Chapters 5, 7, and 9.
1-5.2 APPLICATION OF THE STEEPEST (A) One Design (B) Two Design
DESCENT METHOD IN INTERAC- Variable Variable
TIVE COMPUTER AIDED DESIGN Figure 1-7. Tower With Base Simply Supported
and Top Supported With Guy Lines
Very often in design problems, it is not
practical to specify a unique cost function to Consider, for example, the problem treated
be minimized, hence the formal optimization in par. 1-5.1. The initial estimate of the
problem described in par. 1-5.1 does not optimum tow er was taken as a uniform tower.
apply directly. The fact that the vector B in The components of the vector 6b can be
Eq. 1-28 is a direction of steepest descent, projected on a cathode ray tube, along with a
however, is extremely valuable information to picture of the structure as shown in Fig. 1-8.
a designer. The application of this informa- The algebraic sign of the components of 8b,
tion to a structural design problem, using corresponding to each of the design variables,
interactive graphics, is a technique which is an indication of the effect a change in that
shows considerable promise in design. design variable will have on the cost function
TABLE 1-1
WEIGHTS OF TOWERS
Guy-line Guy-line Guy-line

Cantilevered Cantilevered Cantilevered Supported Supported Supported
Number of
Design
Variables
Best Weight W = 2440.6 lb W = 2111.4 W = 1827.9 W= 1563.99 W= 1356.6 W= 1265.71

Height h 63.7 in. h_ =91.4 h
max = 80 2
- h = 46 ,h =46.5, , ,h = 36.55
Cross-sec-
tional area of
member A = 7.96 lb A = 6.97 A = 10.03 A = 3.84 A = 4.434 A
max = 495
1-13
AMCP 706-192
86, on the other hand, the designer has trend

information that he can use in altering the
distribution of material in a structure, he can
better use his experience in making design
86, improvements. This capability can be invalu-
662 able to large-scale structural designers. It
* * includes the effect of individual design vari-
66=
able changes on overall structural value, while
663 taking into account the effect of that design
-
change on all design constraints.
56.
In real-world structural design problems,

/77777777777777777 the designer must design his structure for
Figure 1-8. Sensitivity to Design Variations more than simply light weight. He must be
concerned with structural vibration and
buckling characteristics, since these are major
of interest. For example, if 5b 1 were positive,
sources of structural failure. Often, as in par.
this would indicate that an increase in the
1-5.1, it is possible to determine design
dimension A, will decrease the structural
perturbations that have a desirable effect on
weight. On the other hand if the algebraic sign
such structural properties as natural fre-
of 5b! were negative, then an increase in 5b 1
quency and weight simultaneously. Both of
would increase the structural weight. Like-
wise, the algebraic signs of 5b4 through these factors can then be displayed on a
cathode ray tube as shown in Fig. 1-9. In this
5b6 indicate the effect that a change in these
element areas will have on structural weight. case, 5bl indicates the direction in which the
These data give the designer valuable informa- design variable should be changed to reduce
tion, according to which he should change his structural weight, and 5b2 indicates the direc-
nominal design to improve the structure, tion in which the variable should be changed
while still satisfying all the essential con- to increase natural frequency. This informa-
straints. tion can then be used by experienced design
861
Traditionally, in structural design by graph-

sb*
ics, the designer puts areas and dimensions
into a structural analysis routine and then
66* , 662
requests a stress calculation, the results of * 4 4
which are shown on the screen of a cathode

ray tube. This technique has been used by 5
Lockheed-Georgia in the design of the C5A.
56
*5' S"l
S
While this technique has been quite useful in 3
* *
structural design, it is extremely difficult for 661 , 662

6 6
the designer with only stress information to 86,
determine how he should change just one
element in the structure to reduce overall / / / 77- 77-
structural weight. The difficulty comes in the Figure 1-9. Sensitivity to Two Performance
interplay between structural constraints. If, Indicators
1-14
AMCP 706-192
personnel in making design changes that will tires. A photograph of the first prototype of
have desirable effects on overall aircraft struc- this weapon is shown in Fig. 1-10.
tural properties, for example. This is extreme-
ly important in large-scale structural design The recoil mechanism for this weapon was
due to the difficulty in determining the effect designed according to traditional recoil mech-
of changes in an individual design parameter anism design goals. Namely, the objective in
on several different structural properties. the design was for a constant retarding force
Computation of these data and interactive which is transmitted by the recoil mechanism
aspects of the technique are discussed in to the undercarriage, as shown in Fig. 1-11. A
Chapter 5. recoil mechanism was designed which de-
livered approximately this recoil force R(t) as
This design technique is feasible from a a function of time.
computational point of view in that very little
additional computer time is required to When the weapon was built and fired, a
generate sensitivity information from stress nearly constant recoil occurred, as desired;
and vibration analyses that are required. While but, at high angles of fire, the weapon
most structural optimization work has been exhibited unacceptable dynamic response.
done in the batch mode, it is shown in During firing, the tires of the weapon com-
Chapter 5 that utilization of the steepest- pressed and after firing and the subsequent
descent technique with interactive graphics is release of the recoil forces, the weapon
a much more practical way to design struc- rebounded off the ground approximately 6 in.
tures, particularly in cases where several This unacceptable behavior required a re-
measures of structural performance are im- design cycle for the recoil mechanism with a
portant. design goal of minimizing the dynamic re-
sponse, or hop, of the weapon after firing.
Development and display of sensitivity
information in design is a form of information It was determined that the peak recoil
transfer to design personnel. This technique force could be allowed to reach 22,000 lb
depends on the availability of interactive without damaging the support structure. The
graphics software and hardware, which are optimization problem is then to determine
currently being developed. the recoil force R(f) as a function time such
that
1-5.3 DESIGN OF ARTILLERY RECOIL R(t)< 22,000 (1-30)

MECHANISMS
and the peak dynamic response, denoted by
As an application of this same optimization
technique to a weapon design problem, cer- / = max [h(t)] (1-31)
t
tain aspects of the design of a lightweight
artillery piece will now be outlined. The is as small as possible, where h (t) is the height
requirement was stated for a lightweight of the tires off the ground at any time t.
artillery piece that can be fired with very Graphically, this problem is to determine a
short implacement time. For this reason it recoil force which lies beneath the 22,000-lb
was determined that the weapon must be level in Fig. 1-12, and which minimizes the
capable of being fired while it is resting on its peak dynamic response of the weapon. In this
1-15
AMCP 706-192
Figure \-10. Howitzer, Towed, 105mm, XM 164
problem, the dynamic response h (t) is deter- The same philosophy of small design
mined by the second order differential equa- changes about some nominal estimate, as in
tions of motion of the artillery piece. the structural design problem of par. 1-5.1,
was employed in this case. A sensitivity
R(t)
22,000
20,000 -
Time, sec Time, sec
Figure 1-11. Traditional Recoil Design Goal Figure 1-12. Recoil Distribution in Time
1-16
AMCP 706-192
function is first determined, which indicates 22,000
the desirable direction of change in the

nominal design variable. For example, taking
the previously designed constant retarding
force as the nominal base line, a sensitivity Pi OS
function is determined as shown in Fig. 1-13.

Time, sec
ff a constant multiple of this function is
added to the retarding force, a reduction will Figure 1-14. Optimum Recoil Curve
occur in peak dynamic response and other
constraints of the problem will continue to be
satisfied. The dotted curve in Fig. 1-13 shows This sensitivity information could easily be
the altered design, which gives better charac- displayed on the screen of a cathode ray tube
teristics than the original design estimate. and could be used by design personnel in
determining desirable changes in the recoil
design. Even in this relatively simple problem
it was not clear in what way the design should
R/t) be altered to obtain improved response of the
22,000 artillery piece. This particular problem was
20,000 solved in the batch mode by doing many
R<> (tj
small step iterations of the kind previously
'3 - described until convergence to an optimum
o -
a! as was obtained. The optimum recoil force curve
Sensitivity
' Function
is shown in Fig. 1-14 and resulted in a peak
Time, sec dynamic response of less than 0.5 in. Detailed
solution of this problem is presented in
Figure 1-13. Sensitivity to Gun Hop Chapter 8, par. 8-5.
REFERENCES
1. The American College Dictionary, Ran- Systems, John Wiley & Sons, New York,
dom House, New York, 964. 1962.
2. C. Y. Sheu and W. Prager, "Recent Devel- 5. M. K. Starr, Product Design and Decision
opments in Optimal Structural Design", Theory, Prentice-Hall, Englewood Cliffs,
Appl. Mech. Rev., Vol. 21, No. 10, Octo- New Jersey, 1963.
ber 1968, pp. 985-992.
6. C. Goffman, Calculus of Several Variables,
3. H. H. Goode and R. E. Machol, System Harper and Row, New York, 1965.
Engineering, McGraw-Hill, New York,
1957. 7. D. G. Luenberger, Optimization By Vector
Space Methods, John Wiley & Sons, New
4. W. Gosling, The Design of Engineering York, 1969.
1-17
AMCP706192
CHAPTER 2
FINITE DIMENSIONAL UNCONSTRAINED OPTIMIZATION
2-1 INTRODUCTION must be able t hoose, out of a collection of

objects which satisfy the restrictions of the
In many engineering design problems cer- preceding paragraph, that one which is
tain information which helps to prescribe the "best". More specifically, out of all design
object being designed is specified. However, a parameters in the admissible set D, the de-
certain number of parameters called design signer must pick that one, x, which describes
parameters are left open to the designer's the "best" system. This discussion has still
choice. These parameters must uniquely not given the meaning of "best". An effective
determine the object if the optimal design way of defining "best" is to give a real valued
problem is to be meaningful. In the discussion function whose domain of definition is the
which follows, the design parameters will be admissible set/), say f(x). "Best", then, may
denoted by xl * or in vector notation be taken as the minimum or maximum of fix)
simply as x = (xl,,xn)T. for x in D. If the function/(x) is a cost of the
system being designed, then it is to be
In virtually all design problems there are minimized. If, on the other hand, f(x) is a
restrictions on the object being designed. return or profit, it is to be maximized.
These may include the performance required,
physical limitations such as size, weight, The cost or return function will be defined
resource limitations, and organizational poli- in each optimal design problem. As a result,
cy. These restrictions or constraints generally very little can be said about its nature in
will involve the design parameters so that the general. It is clear, however, that maximizing
range of values of design parameters may be a real valued function r(x) is equivalent to
restricted. If the vector of design parameters minimizing r(x). Therefore, optimal design
(hereafter called the design parameter) is problems may always be put into J form
viewed as an element of real Euclidean space which may be interpreted as minimization of
R" , then the effect of the listed restrictions is a cost function. For convenience this will be
to confine the designer's choice of design done in the following development.
parameters to a subset D of R" called the
admissible set of design parameters. The Example 2-1: As a hypothetical optimal
nature of this set will be determined by the design problem let the scalar x be the design
nature of the requirements placed on the parameter and f(x) = (x 2)2 be the cost
system being designed. This aspect of the function. In Fig. 2-1 the cost function is
optimal design problem will be treated ex- plotted versus x. It is clear that the minimum
tensively in later chapters. cost of zero occurs at x = 2.
When one speaks of optimal design, he Example 2-1 is included here as an aid to
2-1
AMCP 706-192
intuition in more complex problems. Even

when x is an n-vector, one can think of
plotting the cost function above the x-hyper-
plane to obtain the cost surface. The optimal
design problem is then to find the lowest
point on this surface.
Even though real-world optimal design

problems invariably have constraints placed
on the design parameter, the methods pre- Figure2-1. ffx)=(x- 2)1
sented in this chapter will ignore constraints.
There are two reasons for considering this eter which is admissible, say x(0). This choice
simplified problem in some detail. First, it of design parameter will probably not put him
may happen that the design parameter x that at the lowest point on the cost surface.
minimizes fix) lies in the interior of the Rather than discarding this nonoptimal point
admissible set D. In this case the constraints and picking another trial point at random he
play no part in locating X. Second, even might attempt to find a second point x(1)
though the point x may push some constraint which is closer to the lowest point of the cost
to its limit and lie on the boundary of D, surface. The designer's view of the cost
there are iterative methods for finding x surface is limited to only a small area due to
which require minimization of an auxiliary the local nature of mathematical tests which
cost function, subject to no constraints at he may perform. Using only this local
each iteration. Methods which take con- information, he chooses a strategy which
straints into account are presented in Chap- insures that he makes a move to a new point
ters 3 and 4. x^ which is lower than xM. The direct
methods presented in pars. 2-3 to 2-7 arejust
Two basically different methods of solving a mathematical implementation of these
unconstrained minimization problems are pre- elementary ideas.
sented in this chapter. The first method,
called the indirect method, is based on de- 2-2 NECESSARY CONDITIONS FOR EX-
rived properties of the cost function at its TREMA
minimum; i.e., if one pictures himself as being
at the lowest point of the cost surface (x = 2 As described in par. 2-1, the approach
in Fig. 2-1), he may notice that the surface is taken in the indirect method is to assume/(x)
required to have certain special properties has a minimum at x and then derive condi-
there. He may then use these special prop- tions which f{x) must satisfy there. These
erties to locate the lowest point on any such conditions may then be used to find the
surface. This intuitive idea is made rigorous in minimum point of any real valued function
par. 2-2. f(x). They are valuable in giving the designer
an insight into the minimization portion of an
The second method of solving optimization optimal design problem, even when he is using
problems is more direct in nature and is direct computational methods to solve the
appealing from an engineering point of view. problem. Before these ideas may be devel-
The designer initially chooses a design param- oped, several definitions are required.
2-2
AMCP 706-192
Definition 2-1: A real valued function/(x)

defined on a subset D of R" has an absolute
minimum at x in D if
fix) </(*) (2-1)
for all x in D. The function g(x) has an

absolute maximum at x if g(x) has an
absolute minimum there. The minimum is
called strict if only strict inequalities hold in
Eq. 2-lforx#X.
Figure 2-2. A Cost Function
Note that fix) can have a strict absolute
minimum at only one point in D whereas it
fix). Without making some assumptions as to
could have an absolute minimum at several
the regularity of fix) it is difficult to verify
distinct points in D provided it has the same
the required inequalities. Consider the case of
value at all these points.
a function fix) of the real variable x which
has two continuous derivatives. The Taylor
Definition 2-2: A function fix) defined on
formula is
a subset D of R" has a relative minimum
(maximum) at x if there exists an E > 0 so
fix+h)=fix)+f'ix)h
that f(x) has an absolute minimum (maxi-
mum) in a subset of D whose points satisfy
+-^%x + 6h)h2 (2-2)
I*.- I< e, / = 1 .... n.

where 0< 6 < 1. Since f"(x + Oh) is bounded
Verbally, this definition says that/(x) has a for A in a closed bounded set, it is clear that if
relative minimum at x if it has an absolute f'{x) + 0 then for small enough h the linear
minimum in a sufficiently small neighborhood term in h dominates the squared term so that
of x. It is clear that if fix) has an absolute fix +h) may be made both larger and smaller
minimum at x, then it also has a relative than f(x) through choice of the appropriate
minimum there. The converse is not neces- sign of h. Therefore, in order for/(x)to have
sarily true. a relative minimum or maximum at x it is
necessary that /'(3c) = 0. It follows directly
Example 2-2: Locate all relative and abso- from Eq. 2-2 that if fix) = 0, then/"(x) > 0
lute maxima and minima of fix) on 0 < x < (< 0) is a sufficient condition for/(x) to have
3, where fix) is given graphically in Fig. 2-2. a relative minimum (maximum) at X
The function fix) has a strict absolute In case x is in R", results analogous to
maximum at x = 1, absolute minima (not those just obtained are given in Theorem 2-1.
strict) at x = 0 and 2, relative maxima at x = 1
and 3, and relative minima atx = 0 and 2. Theorem 2-1: Necessary Condition: Let
fix) be defined on a subset D ofR" and have
In Definitions 2-1 and 2-2 no continuity or a continuous derivative in a neighborhood of
differentiability requirements were placed on a point x which is in the interior of D. If fix)
2-3
AMCP 706-192
has a relative minimum at x then A = (a.)nX is positive definite if and

only if the determinate of each of the
V/(x) = 0 (2-3) matrices A m , formed from the first m rows
and first m columns of A, is positive, m =
Sufficient Condition: Let f(x) have two con- 1,...,.
tinuous derivatives in a neighborhood of x
and let Eq. 2-3 hold. Then if the matrix Example 2-3: Obtain explicit necessary and
sufficient conditions for fixi,x2) to be a
a2/ minimum and a maximum at x, where fix!,
V2Ax) = (2-4)
x2) has two continuous derivatives inZ) and x
dXj dx.
is an interior point of D.
is positive definite, fix) has a relative mini-
mum at X As necessary conditions for either a mini-
mum or a maximum, Eq. 2-3 demands
For convenience in later discussions, Defi-
nition 2-3 is made.
Definition 2-3: A point at which Eq. 2-3 A sufficient condition for x to be a

holds is called a stationary point of f(x). minimum point for/(.x) is that in addition to
the above equations, the matrices
It is imperative that the reader be aware of
Ai^fXlXl(x)^nd
the hypothesis of Theorem 2-1 which requires
x to be in the interior of the region D. The
A2 =
theorem does not apply if x is on the fxlXl&fxlXi&>
boundary of D. Example 2-2 illustrates this
requirement graphically. Points x = 0 and x =
3 of Fig. 2-2 yield a relative minimum and a
have positive determinates, i.e..
relative maximum, respectively, but neither
point is stationary (i.e., neither satisfies Eq.
2-3). The same example also illustrates the fXlXlix)>0^dfXiXiix)fX2X2ixl
need for verification of the differentiability
properties of fix). Even though x = 1 yields -WXlx2ix)]2 > 0.
an absolute maximum of fix) and is in the
interior of D ,it is not a stationary point since The function fix) has a relative maximum
fix) does not have a continuous derivative at x if the functiong(x) = fix) has a relative
there. This example illustrates the need to minimum there. Therefore, in addition to
faithfully verify all the hypotheses before fx (3c) = fx (3c) = 0 sufficient conditions
Theorem 2-1 is employed. for Hx) to have a relative minimum at x are
In order to verify the sufficiency condition **,*, <*>> 0and

^^, &gxx )
11 11 2 2
of Theorem 2-1, one must have a verifiable
test for positive dsfiniteness of a matrix. -tg,, (*)]2> 0.
Probably the most useful test is the following
(Ref. 2, page 103): A symmetric matrix For a relative maximum oi fix) at x then
2-4
AMCP 706-192
sufficient conditions are < 1, is also in D. A real valued function f{x)

defined on a convex set D is called a convex
/ (je)< Oand/, (*)/, (?) function onD if for any two pointsy and z in
D
[/, (?)]2>o.
fly + 6(z-y)] <f(y) + 6[f(z)-f(y)],
Thus far in this paragraph only properties Q< d a 1.
of fix) precisely at the minimum point have
been investigated. If the designer viewed the That is, f(x) is convex onfl if the straight
graph of fix) versus x to be a surface, then line segment f(y) + 8[/(z) -fiy)] is above the
Theorem 2-1 tells him what the surface will graph of f{x) on the line segment^ +0(z y)
look like when he finds its lowest point. in D, 0 < 8 < 1. For a more detailed
Theorem 2-1, however, does not tell him that discussion of convex functions, see Appendix
a lowest point exists. In order to solve his A.
optimization problems, the designer would
like to have tools which allow him to stand Theorem 2-3 gives the designer valuable
back from the cost surface and learn some- information about the global properties of the
thing about its global properties. Two theo- cost function. It is proved in detail in Chapter
rems are now stated which give him a better
overall view of the optimization problem.
Theorem 2-3: Let fix) be a convex func-
Theorem 2-2: If fix) is continuous on a tion defined on a convex set D inRn . Then a
closed and bounded subset D ofRn ,then/(x) relative minimum of fix) on D is also an
has an absolute minimum in D. absolute minimum of/(x) onD.
This theorem does not hold, in general, if This theorem is of obvious value to the
any of the hypotheses are deleted. For ex- designer. It assures him that if his design
ample, consider the function/(x) = x onD = problem satisfies the hypotheses of Theorem
(x\ 0< x < 1). D is not closed and/O) does 2-3 and if he has found a relative minimum
not have an absolute minimum in D. If then he is through; he has also found the
Z)={x|0<x< 1} then D is closed and fix) absolute minimum.
has an absolute minimum at x = 0.
Computational methods for finding ex-
Note: The hypotheses of Theorem 2-2 may be trema based on the theorems of this para-
weakened by demanding that fix) be only graph generally involve the solution of non-
lower semi-continuous rather than continu- linear algebraic equations. In particular, Eq.
ous. For proof, see Ref. 1, page 58. 2-3, which is in general nonlinear, can be
solved by a numerical method to locate all
Theorem 2-3 depends on the concept of admissible interior extrema. Methods for solv-
convexity. ing such equations are given in Ref. 3,
Chapter 2. It generally has been found,
Definition 2-4: A subset D of R" is called a however, that direct methods for finding
convex set if whenever x andj are inD, then extrema are superior to the solution of Eq.
the straight line segment x + 6iy x), 0< 8 2-3. For this reason no computational
2-5
AMCP 708-192
methods based on the indirect method will be The object here is to treat more general
presented here. functions, but it is possible to make a
quadratic approximation to / [xM + ots]
It will be the purpose of the remainder of which will hold near the minimum point.
this chapter to present methods that the Then, the minimum point of the approximat-
designer may use to locate interior relative ing function, which may be easily found, is an
minima. Relative minima on the boundary of approximation of the true minimum point.
the admissible region will be treated in Chap-
ters 3 and 4. The quadratic approximation of / [xM +
asj is constructed by passing a quadratic
2-3 ONE-DIMENSIONAL MINIMIZATION curve in a through three computed values of
the function. The distance between the three
In the direct minimization methods to trial points will be 6 > 0, where 6 is initially
follow, a multidimensional minimization chosen to be a small fraction of the expected
problem will be reduced to a sequence of one- range of a. It is known, however, that if the
dimensional minimization problems; i.e., the starting point of the process is quite far from
problem of determining a scalar a so that a the minimum point then the minimum point
given function^(a) will be a minimum. of the approximating function may not be
near the true minimum point. To prevent
In the problem of minimizing f(x) for x in making large, inaccurate steps in this case, a
R", all the methods of solution presented in maximum allowable step size A is chosen
this chapter are based on successive improve- before the process begins. A reasonable choice
ments in certain directions; i.e., at a point x(,) of A is 50% of the expected range of a.
one finds a direction, s, in which f(x) de-
creases. The object is now to move along the The following algorithm implements the
vector x(,) + as, by adjusting a, a > 0, until procedure described:
f(x) is as small as possible. The resulting point
is then called x('+1 \ and the entire process is Step 1. Define a0 = Oand/ = 1.
repeated. It is clear that the intermediate
problem of determining a so as to minimize Step 2. Compute
/(x(,) +as) is one-dimensional. This paragraph
will be devoted to presentation of methods h =/[x('>+(/-' -6)s
for solving the one-dimensional problem.
/o=/[*(0+/-1 s]
2-3.1 QUADRATIC INTERPOLATION
h =/[JC(0+(a/-1 +6)s].
(,)
If the function /Qc + as) of the scalar
variable a x^ and the unit vector s are Step 3. A quadratic polynominal in a
fixed were quadratic in a,then the value of a1'1 = z is fitted through ( 6, /,),
a which minimizes the function could be (0, /o), (6, fi). Its minimum is
found by setting
2
-~2(/1-2/0+/2ylt/l lu
+ f2 = 0. If this quantity is zero,

^-(f[x^+as])=0.
da x then the approximation is a
2-6
AMCP 708-192
straight line with minimum at ? = a = (k 1)5. If the initial step 6 was too
A , depending on which of/j and small, many steps will have to be made before
f2 is smaller. the minimum point is located.
Step 4. Define In Fibonacci search the same basic proce-

dure is followed except that if, after a given
da. = min(|zJ,A)-sgn(zm) step, the functional value has decreased, then
the next step size is taken as 1.618 times the
and previous step size. In this way if the minimum
point is a long way from a = 0, the Fibonacci
ai = al'x + da technique will isolate it much more rapidly
than the previous method with constant step
Step 5. If | da | is less than a specified size. Note that there is a penalty, in that the
tolerance, the process is stopped interval which contains the minimum point
and a* is taken as the minimum may have length much greater than 26. This is
value of a. Otherwise, replace /' illustrated in Fig. 2-3.
withj + 1 and return to Step 2.
w
f& +]
2-3.2 FIBONACCI SEARCH (OR GOLDEN
SECTION SEARCH)
The Fibonacci search technique is a

method based on isolating a relative minimum
in an interval and successively decreasing the
5 2.6186 5.2326 9.6666
size of the interval. The process thus gives
successively better estimates for the location Figure 2-3. Function of Single Variable
of the minimum point. For a proof that the
method converges very rapidly the reader is
referred to Ref. 4, page 236. Here, only the Once the minimum point is restricted to
basic ideas behind the method will be given, some interval, this interval is broken up into
and an iterative algorithm stated. three subintervals by inserting points located
a distance of 0.382 times the length of the
Starting at a = 0 one might evaluate/[x*1' interval from each end. A test is then per-
+ as] at a = 6 and check to see if the formed to see which subinterval the minimum
functional value is smaller than at a = 0. If it point lies in. For a given subinterval the
is, one might then evaluate the function at a = partitioning is shown in Fig. 2-4.
26 and compare with the value of a = 6.
Again if a decrease occurs, one moves on to a
= 36, etc. The process will terminate when 0.382 (a-aj 0.382 (a - a)
u *
/ [*<'> +(fc + l)5s] > f [xu) + kSs]. It is -* >
then known that (k 1)5 < a < (k + 1)5
contains the minimum point and a more
accurate result, if required, may be obtained
by reducing 6 and repeating the process from Figure 2-4. Interval Partition
2-7
AMCP 706-192
The search process is terminated when the /-i

Note that aa = 2 6(1.618)* so/[x(l) +
minimum point is isolated in a sufficiently k=o
small subinterval. otas] is known.
The Fibonacci search method has the prop- Step 3. Compare/ [JC(0 +aas] and/ [xu)
erty of being best in a certain sense among all + ubs] and go to Step 4, 5, or 6.
search techniques which isolate a in an
interval. A measure of the effectiveness of any Step 4. If/ [x(,) + <y] < / [x(,) +abs] ,
then a2 < a*'-1 < <xb. By the choice
such technique is the ratio of the length of
the largest interval in which a may lie after n of a and ab , the new points s =
steps to the length of the original interval a,' and a'u = a,boahave a'. = a .
which contained a. It is shown in Ref. 4, page Compute now / [JC(,) + a'as]
253, that if/ [x(,) + as] has a unique relative where a'a = ae + 0.382 (a'u - a't).
minimum as a function of a,then Fibonacci Go to Step 7.
search minimizes the number of interval
Step 5. If/ [JC<'> +aes] > f [x + abs],
partitions.
then aa s a(,) < a,u . Similar to the
procedure in Step 4, put a'g = aa
The Fibonacci search technique may now
and a' = au so that a.'a = otb.
be given in the form of a computational
algorithm: Compute/[.x^'' + a.'bs] where ot'b =
0^+0.618(0^-^).
Step 1. First an upper bound must be Go to Step 7. -
found for a.,au. It is clear that 0 is
a lower bound, ae. For a chosen Step 6. If/ [x(,) +aas] = / [x(0 + <V]>
small step size 5 in a,let j be the put ae = aa and aj, = 6 -
smallest integer such that Return to Step 2.
Step 7. If a^ a is suitably small, put

/ \x(i) + [ X S(1.618)*]s
I k=o I + a
(') = -j-(a'u 's) and stop.
(,
>f JC '> + [ 2 6(1.618)*]s Otherwise, delete the primes on a'q,
fc = o
afl, c^, and c/ and return to Step
Then upper and lower bounds on <xu> are 3.
a = 2 5(1.618)* 2-4 THE METHOD OF STEEPEST DE-

*=o SCENT (OR GRADIENT)
a0 =;S 5(1.618)*.
fc = o The simplest and probably the best known
of the direct methods of minimization is the
Step 2. Compute/ [x(,) +
<xbs], where Method of Steepest Descent (or Gradient).
This method is based on the fact that if the
aa =ac +0.382(au -as) cost surface is smooth, then its tangent plane
is a good approximation to the surface near
a6 =ac +0.618(au - a8). the point of tangency. The philosophy of the
2-8
AMCP 706-192
Method of Steepest Descent is apparent in its as a better estimate of the minimum point
title. One wishes to change x(,) by an incre- and the process is continued until 7/ [x(y)] =
ment dx in such a way that /(x), x = x(,) + 0 oxdx is sufficiently small. This method may
dx, is reduced as much as possible for a given be given in compact form as the steepest
length of increment. The direction of the descent algorithm:
increment dx is called the direction of steep-
est descent. Step 1. Make the best engineering estimate
x'0) of the minimum point.
The direction of steepest descent is given
by Theorem 2-4. Step 2. Compute V/[x(0] and define a
normalized gradient s =
Theorem 2-4: Let f(x) be differentiable in II ?/[*"> Ill vW'l-Find^
R". The direction of steepest descent at a a(') which minimizes / [x(i) + as ]
point x is (where i is the number of iterations
completed). If V/ [x(0J = 0, ter-
dx = -aV/r(x) (2-4)
minate the process and x(z) is a
relative minimum point.
where a > 0 is a scalar factor.
The proof of Theorem 2-4 illustrates clear- Step 3. Put x(i+1) = x(0 - a.0)s. If 11 (0
ly that the direction of steepest ascent is and IIV / [x(,+ 1 *] || are less than
predetermined limits, terminate
dx = aVfT (x) (2-5) the process and let x(,+ 1) be the
approximation to the minimum
for a > 0. The reader should note carefully point. Otherwise return to Step 2.
that Eqs. 2-4 and 2-5 give only the direction
in the design parameter spaced" which yields It is interesting to note that successive,
the maximum rate of change of f{x). Since directions of steepest descent are orthogonal
the factor a is not determined explicitly, the to one-another in this algorithmi.e.,
size of step is not specified. V/[x (' + 1)] V/r[x(,) =0]. To see this, recall
that a(,) is chosen so that /[jc(i) as] is a
In order to start the steepest descent minimum in a. The necessary condition of
iterative technique, the designer makes the Theorem 2-1 then requires
best estimate of the design parameter avail- df 1 9/
able, x(0). The gradient V/[x(0>] is then 0= 2 ix (1 + 1)1
da 7/U (0, / = 1 ox-
computed at x(0) and a new point x(1) is
-0)1 =
1
determined by
dx, V/ " [x('> ]
7
r(D = r(0) x(0) V/r[x(0)] V/[x'0+1) 7/r[x(')]
where a<0) > 0 is chosen using methods of which was to be shown.

par. 2-3 so that / [x(0) - aV/r(x(0)) ] is a
minimum as a function of a. If V/ [x(0)] = In the case where x =1 x' , Fig. 2-5 is a
Othen/[x(1)] </[x(0)] , so x( l} is taken view of the design variable space. The closed
2-9
AMCP 708-192
(3) If x is the only point in S for which

V/(3c) = 0, thenx(') converges to x.
Several things which Theorem 2-5 does not

say are worthy of note. First, the theorem
does not guarantee that the sequence of
points xw generated by the Method of
Steepest Descent will converge. Further, un-
less the assumption of (3) holds, the sequence
need not converge to an absolute minimum; it
may converge to a relative minimum.
Figure 2-5. Descent Steps
curves in this figure are lines of constant f(x), The choice of the initial estimate x'0-* can
have a great deal to do with the limit point of
A relatively general convergence theorem the algorithm if it does converge. If it is not
pertaining to this algorithm will now be known beforehand that a unique relative
stated. The proof of this theorem may be minimum exists, it is general practice to start
found in Ref. 5, page 80. the iterative process at several initial esti-
mates. If the sequence x{,) converges to the
Theorem 2-5: Let f(x) be a continuous same point x each time, then one is led to
function defined on R" and let x(0) be any believe that he has indeed found an absolute
point such that the closed set minimum. Logic such as this can cause sleep-
less nights, however, particularly if a decision
S= \x\ f(x)<f[xw] j involving considerable resources and perhaps
even one's job depends on the outcome. For
is bounded, and Ax) is twice continuously this reason, the importance of at least making
differentiable on S. Let the matrix of second a serious attempt to apply theorems such as
derivatives of fix), those of par. 2-2 cannot be overemphasized.
Theorem 2-3, for example, if properly
92/(x)
H(x)-- applied, may prevent many anxious moments.
dxfix,
satisfy the condition In spite of the simplicity of the Method of
Steepest Descent, it has several severe restric-
I yTHy I < MyTy tions which are discussed in Ref. 5, page 159.
These are:
for some M, every y mRn , and every xinS.
Then for the sequence [x^ ] generated by 1. Even though convergence may be
the steepest descent algorithm: guaranteed by Theorem 2-5, an infinite num-
ber of iterations may be required for the
(1) A subsequence x 'm converges to a minimization of even a positive definite qua-
point 3c in.? for which V/(x) = 0. dratic form.
m
(2) / [x 'l decreases monotonically to 2. Each iteration is calculated indepen-
dently of the others so that no information is
2-10
AMCP 706-192
stored which might be used to accelerate matrix is required

convergence.
92/0c)
Hix) = W2fix)
3. The rate of convergence depends strong- dx.dx..
X n
ly on properties of the cost function. If the
ratio of the largest and smallest eigenvalues of Note that it is implicitly assumed here that
the matrix of second derivatives is large, the fix) has two derivatives. By Taylor's formula,
steepest descent algorithm generates short
zig-zagging moves. Convergence is, therefore, f[x(0) + Ax] ~/[x(0)] + V/[x(0)]x
very slow.
+ ÂxTH[x<-0)]Ax (2-6)
For an extensive treatment of modifica-
tions of the steepest descent method, which where Ax is a change inx(0). In case/(x) is
prevents certain of these difficulties, see Ref. locally convex convex in a neighborhood of
4, Chapter 7. Several methods, presented in x(0)- Theorem A-3 shows that H U(0)] is
the next three paragraphs, do not suffer so positive semi-definite. If, in addition,
severely from the problemsjust described. H [x(0)] is positive definite, then it has an
inverse. Further,/ [x(0) +Ax] in Eq. 2-6 is
2-5 A GENERALIZED NEWTON METHOD convex in Ax so Theorem 2-3 insures the
existence of a unique minimum point of the
In the Steepest Descent Method of par. 2-4, quadratic function in Eq. 2-6. By Theorem
only first-order derivatives that determine the 2-1, this unique minimum point is determined
tangent plane of the cost surface are used to by
represent the behavior of this surface. One
would expect that if second derivatives of the V/r[x(0)] +//U(0)] Ax = 0
cost function were available, then a quadratic
function could be constructed as an approxi- or
mation to the surface. The quadratic approxi-
l
mation should allow for much better approxi- Ax -H [*] V/r[*(0)] , (2-7)
mation of the minimum point of the cost
function. and the new estimate of the minimum point is
Xd)=x(0)+Ax
The idea of this method is to first use a

second-degree Taylor formula as an approxi- Since Eq. 2-6 is just an approximation, JC(1 >
mation to fix). K fix) is convex, or just con- will probably not be the precise minimum
vex near a minimum point then the minimum point of fix). Realizing that evaluation of
point of the quadratic should be near the Hix) requires computation of n (n + l)/2
minimum point of fix). The minimum point second derivatives of fix), one might be
of the quadratic approximation is then deter- tempted to improve the estimate for the
mined analytically and is taken as a good minimum point before recalculating all these
approximation of the minimum point of fix). derivatives.
In order to utilize Taylor's formula in- An easy way of improving the estimate of
cluding second degree terms, the following the minimum point is to change the length of
2-11
AMCP 708-192
the step Ax without altering its direction. The H(x) may not exist unless H{x) is
scalar a ~ 1 will be determined by methods of strictly positive definite.
par. 2-3 so as to minimize / [x(0) +a&x].
2. In nonconvex problems an iteration
This procedure may now be put down in does not necessarily decrease /[x(,)]
the form of a computational algorithm called when the current iterate x(r) is not near
Generalized Newton Method: the minimum point.
Step 1. Make an engineering estimate x(0) 3. For many engineering problems, H(x)
of the minimum point of f(x). will be extremely messy if not im-
possible to compute efficiently.
Step 2. Compute
-)#-V0] V/r[x(,)] , Even in nonconvex minimization problems

x(i+i) =XV)
the Generalized Newton Method may be used
where a = a*'' is chosen which in conjunction with a Steepest Descent Meth-
minimizes od to form an extremely effective tool. The
Steepest Descent Method has the property of
/{*<''> -a//"1 [x<'>] V/r[x(0]} making good progress even though only a
poor estimate of the minimum point is
as a function of a. Here, the index available. As a relative minimum is ap-
i is the number of iterations com- proached, however, the rate of convergence of
pleted. the Steepest Descent Method decreases. At
this point, however, the cost function should
Step 3. If IIV/ [x< 1II and II x(i+1 > - *<''> || be convex since a minimum point is nearby.
are sufficiently small, terminate Therefore, the Generalized Newton Method
the process and take x(l + x) as the may be employed for rapid convergence to
minimum point of f{x). Otherwise, the relative minimum point.
return to Step 2.
2-6 METHODS OF CONJUGATE DIREC-
The Generalized Newton Method presented TIONS
in this paragraph has been called the best for
minimizing convex cost functions when In par. 2-4 it is pointed out that the
second derivatives are available (see Ref. 5, Method of Steepest Descent had rather poor
page 162). Even in the case in which the cost convergence properties in many problems
function is nonconvex, if the starting point because it uses only first-order approxi-
x(0) is near enough to a relative minimum mations (involving only first-order deriva-
point so that the cost function is convex at tives). Further, the Steepest Descent Method
x(0), then good convergence may still be is not a learning process in that it does not
expected. store information from past iterations. The
first deficiency is corrected in par. 2-5 where
In spite of the advantages of this method, it a Generalized Newton Method employing
still has several shci-tcomings. second derivatives is presented. This method,
while having outstanding convergence prop-
1. Even if f(x) is convex, an inverse of erties, requires the computation of n(n+ l)/2
2-12
AMCP 706-192
second-order derivatives at each iteration (x is vectors S' are linearly independent. To see
in/?"). In most engineering design problems that this is true, form the linear combination
this is an extremely tedious, if not impossible,
task. Further, the Generalized Newton Meth-
2 a,S' = Q,
od is not a learning process. j'=0 '
The methods presented in this paragraph where the a. are scalars. Multiplying this sum
require the computation of only first deriva- T
on the left by S' A yields
tives. However, by making use of information
obtained from previous derivatives, con-
vergence is speeded as the minimum is ap- "z a.S'7AS' = a,SiTAS' = 0
i=o ' '
proached. In fact, as one of the methods
progresses, it develops an approximation to and since S' AS' = 0, a. = 0. Since ;' was
the matrix of second derivatives. In this arbitrary, a;. = 0,/' = 0, 1,.... n 1, and this is
respect the methods here have the desirable just the definition of linear independence.
features of both the Method of Steepest
Descent and the Generalized Newton Method. Consider now the problem of minimizing
the convex function
All Methods of Conjugate Directions are
based on the philosophy "If a method works f(x)=BTx+-xTAx (2-9)
well in minimizing all positive definite qua-
dratic forms, then it ought to work pretty where x is in./?", B is an n x 1 matrix and^l is
well on any smooth cost function." To be a symmetric, positive definite, n x n matrix.
more specific, Conjugate Gradient Methods The central idea of all methods based on
are guaranteed to minimize any positive conjugate directions is contained in Theorem
definite quadratic form in n iterations (the 2-6.
design parameter is in R"). Although this
ideal behavior will not carry over to general
Theorem 2-6: Let S, .... S*1-1 be nonzero
cost functions, since a convex cost function
vectors in R" which are conjugate with
often looks very much like a positive definite
respect to the positive definite matrix A.
quadratic form, similar behavior could be
Choose scalars X = A('\ i = 0, ..., n \,
expected. Experience has shown that this is
successively which minimize
the case.
/[x(0 +AS'] (2-10)

In order to be more precise, one makes
Definition 2-5.
where fix) is given in Eq. 2-9,
Definition 2-5: Let A be a symmetric
positive definite n x n matrix and S', i = 0,
(2-11)
1,..., n 1, be nonzero vectors in/?". TheS' fc = 0
are called conjugate with respect to A if
and x(0) is any point in/?". Thenx*") is the
SiT AS' = 0,i*j. (2-8) absolute minimum point off(x) over/?".
Since A is positive definite, the conjugate The two methods that follow are simply
2-13
AMCP 708-192
based on different ways of generating con- k+lT ASr

rfr+l - ,jt
v*H+ l
jugate directions. There are an unlimited r=0 SrTASr
number of ways to generate conjugate direc-
tions. Several ways are discussed in Ref. 6.
Many sets of vectors v. could be chosen to
generate conjugate directions. A natural
2-8.1 THE CONJUGATE GRADIENT
choice, however, is the set of gradient vectors
METHOD
of Ax), g' = Vfix^), where x(,) are defined
in Theorem 2-6. Define
Given any set of n linearly independent
vectors and a positive definite n x m matrix A
S=-g
a set of conjugate directions with respect to A
can be generated by a Gram-Schmidt ortho- ; + lT
k
5* + ! =_gk +1 + s g
ki
AS1
gonalization technique. Let v, ..., vn~' S' (2-12)
i=0
be linearly independent vectors and define S SiTASl
= v. Now put
Alternatively.
l
S' = v +a10S. k + iT
k pk+l ' ,A O!
^ + 1=-5* + 1 + Z i S< (2-13)
For A-conjugacy, it is required that i-O
S<jil'ASl
SÂS1 = 0=STA(vi +a10S) Since Ax) =~xTAx +BTx,

2
and
gk=VfT[x(k)] =AxM+B,
vl TAS
10 or from the proof of Theorem 2-6,
STAS
k-\
AssumingS1 ,..., 5* areA-conjugate, put gk =gi+1 +A S X<e)S8 (2-14)
2 = /+ 1
5* + i=v* + i+aJt + 1>05+...+a|t + 1^S*. Now,
For A-conjugacy it is required that k-l

gkTSi=gi+lTSi +SiTA
8 = 1+1
S* + 1 TA Sr = 0 = vk + 1 TA Sr + ak+, rSrT A Sr = 0,i< k (2-15)

due to A-conjugacy of the 5' and
where the second equality holds by S-con-
jugacy, so Vf[x<-k + 1)]Sk = 0,k=0,..., n - 1. (2-16)
J + 1T ASr FromEqs. 2-13 and 2-14

Tt + l.r ,r= 1,..., k.
rT r
S AS
S' + 2 -^ 5'
;'=0 5M S'
By induction, the resulting directions are
A-conjugate and 0,i< k. (2-17)
2-14
AMCP 706-192
yk+\T k + lT fe'+i
, ,+ 1 - >)
Thus, the g\ i = 0, 1, ..., n 1 are linearly gî'AS' 8^' g
independent and the S', i= 0,1, ..., n 1 are SiTASl g'V

/4-conjugate.
By Eq. 2-17, for i < k, the right side of the
The Conjugate Direction Method of Theo- above equation is zero. For / = k,
rem 2-6 may now be applied using the
conjugate gradients Sl. The result is called the gk+xTASk g k+\Tgk+l
Conjugate Gradient Method. In order to apply kT k
this method to nonquadratic problems, it is SkTASk gk gk
first necessary to eliminate explicit depen- Substituting this result into Eq. 2-12 yields
dence of the algorithm on the form of f(x).
(ek + \T *+i\
IS*. (2-20)
By definition,
gkTgk )
Eq. 2-20 now gives an algorithm for determin-

gi+i =Ax' +l
+B=A[x'+\ S>] +B(J) ing the conjugate directions, even without
knowledge of the matrix^ .
or
For a general function/(x),
si+1 = gJ + X(/)AS>. (2-18) gi = v/r[>>]
and the following algorithm for finding the

By Eq. 1!-16 unconstrained minimum of fix) is called the
Conjugate Gradient Method:
1
g/+lTS' = o=g'T S'+X^S' 'AS'
Step 1. Make an engineering estimate x(0)
Thus, of the minimum point and com-
pute
giTs>
X<" = - s = -v/r[*(0)]
S'TA S>
Step 2. For i = 0,1, ..., find a = a(0 which
Substituting for S' from Eq. 2-12 and using
minimizes/ [x^ + aS'].
Eq. 2-15, this is
Step 3. Compute
g'TS'
X0) (2-19)
S'TA S> x(+l) =xO)+aii)si
FromEqs. 2-18 and 2-19 Si+1 =- V/rU(/+1)] +'S'
SiTA S' where

AS' (gi+i -g').
_ 7/[x(i+1)]y/7U"+1)]
Now, V/[x(]V/rtx('>] '
2-15
AMCP70S-192
Step 4. Terminate the process if little insight into use of the method. For a
IIV/[*(''+1)]llandlU<!'+1) _x('>|| direct proof of convergence, etc., the reader is
are sufficiently small. Otherwise, referred to Ref. 7.
return to Step 2.
The computational algorithm is:
When this algorithm is applied to problems
in which f(x) is not of the form of Eq. 2-9, Step 1. Make an engineering estimate x(0)
convergence will not occur in n steps. of the minimum point and choose
Fletcher and Reeves recommend that after n a symmetric positive definite
steps the algorithm should be "restarted", i.e., matrix H(0).
(0)
x( + i) should be treated as x in the
algorithm. In a sense, the first few iterations Step 2. For i = 0,..., compute
of the algorithm build up information about
the curvature of the cost surface. After n S=-^>V/r[#].
iterations, this information is discarded and a
new estimate of curvature is built up during Step 3. Compute a = oS'^ which minimizes
(/ (i)
the next n iterations. This method then does /[JC > +aS ].
not accumulate information about curvature
of the cost surface over the entire iterative Step 4. Compute
process.
aU) =aV)S(i)
2-6.2 THE METHOD OF FLETCHER AND
POWELL x+i)=x(/)+0(fl
A second method based on a different set //0'+i)=Jtf(')+x(')+(0

of conjugate directions was suggested by
Davidon (Ref. 8) and modified by Fletcher where
and Powell (Ref. 9). This method is reported
to be one of the most powerful known ,(0 VfT[xV+D] _ V/r[*(0]
methods for general functions/(x), (Ref. 10).
A major reason for the success of this method
A =
is its capability to accumulate information
about the curvature of the cost surface during
the entire iterative process, even though only
_ HWyWyU)THl
first order derivatives of the cost function B (0 =
yU)T Hd) yd)
need to be computed.
The directions S^'\ generated by the al- Step 5. If || V/[x('+1>] || and ||x(,+ 1>
gorithm that follows, are conjugate if f(x) is Ai)v are sufficiently small, termi-
of the form of Eq. 2-9. This proof is given in nate the process. Otherwise return
Refs. 7 and 9. In Ref. 6 it is shown that the to Step 2.
method of Fletcher and Powell fits naturally
into a large class of conjugate direction Fletcher and Powell (Ref. 9) prove that this
methods. The derivation is tedious and lends algorithm has the following properties:
2-16
AMCP 706-192
1. The matrix H^ is positive definite for Step 1. Make an engineering estimate of

all i. This implies the method will the minimum point JT(0) of fix).
always converge to a stationary point Choose vectors s', j = 1,..., n, in
since the coordinate directions of Rn .
Step 2. Find a = c/, k = 1,..., n, which

da |a_0
minimize/[x'*"1' +oisk]
= - V/[jt(,'>]#(0V/r[.*('')] < 0 where
provided V/Uw ] = 0. This means that y =xw

(2)
/[JC ] may be decreased by choosing a
yk =yk-\ +aksk^ k = it_f n>
> 0if/[x(o] =0.
2. When this method is applied to the and i is the number of iterations

positive definite quadratic from Eq. 2-9, which have been completed. Note
//(,) converges to A'1. that in the one dimensional mini-
mization for ak, it is possible for
This method might be called a learning ak < 0.
process in that only first derivatives are ever
computed, but as the algorithm progresses an Step 3. Find the integer m, 1 < m < n for
approximation of the matrix of second deriva- which
tives is generated. Many experienced re-
searchers in the area of optimization methods f(ym-1)-f(ym)
laud this method as one of the best available.
is the largest and define
2-6.3 A CONJUGATE DIRECTION METH-
OD WITHOUT DERIVATIVES A = f(ym-1)-f(ym).
Occasionally in applications, one is faced Step 4. Define A =f(y), h =f(yn), and

n
with a problem in which computation of f3 =f(2y -y).
derivatives of the cost function is impossible
or at least prohibitive from a computational Step 5. If/3 > /\ or
point of view. There are many techniques for
solving this sort of problem given in Ref. 4. (fi ~2f2+fi)x(fl -f3 -A)2
An efficient technique, not presented in
common texts, was developed by Powell (Ref.
11) using conjugate directions.
put
A computational algorithm is presented
here without proof. For a proof that the c(l'+l) =yn _
algorithm generates conjugate directions the

reader is referred to Ref. 11. The computa- Terminate the process if ||JC('+1)
tional algorithm is: x(,)|l is sufficiently small. Other-
2-17
AMCP 706-192
wise return to Step 2 with the /2O1, x2, x3, x4) = (xy + 10x2)2
same set of s'\ j = 1,,.., n.
Step 6. If neither of the inequalities of (2-22)

Step 5 hold, defines = y " -y") and
find a = which minimizes + (x2 -2x3)4
f(yn +sis)
+ KK*! -x4)4
Put
and
xO+l) =yn +s
/3(x1(^x3)=x^ +2x; +:
(2-23)
Terminate the process if IIx 0+1)
x (0| is sufficiently small. Other-
wise return to Step 2 with the new
set of vectors ^ ..., xm~l, sm + l, The reader should verify that each of these
..., s", s functions has a strict absolute minimum
point. These points are (1,1), (0,0,0,0), and
For a discussion of use of the (0,0,0), respectively. Each iterative method
algorithm, see Ref. 11. will be started at points ( 1,1), (1,1,1,1),
and (1,1,1) for Eqs. 2-21, 2-22, and 2-23,
respectively. These functions will all be mini-
2-7 COMPARISON OF THE VARIOUS mized by each of the methods of pars. 2-4
METHODS through 2-6. The stopping criterion will be
that each component of the independent
During the development of the methods variable must be within 10" 2 of the known
presented in this chapter, theoretical advan- minimum point.
tages and disadvantages have been pointed
out. As a concrete test of these methods,
three functions will be minimized. Two of the Results will be presented in tabular form so
that a comparison of the behavior of each of
functions to be treated are terribly behaved
the methods may be made. For the sake of
and pose a meaningful test to any general
uniformity, each table will include the itera-
minimization technique. These functions re-
tion number i, the iterate x(,) = [x^, ....
semble a very deep valley at whose bottom (<)i T , and the value of the cost function.
the curvature in two orthogonal directions is
radically different. The third function is
quadratic and poses no serious obstacle to any 2-7.1 METHOD OF STEEPEST DESCENT
reasonable method. More specifically, these
functions are
2-7.1.1 COST FUNCTION: f^x) = 100(x2
/,(*,, x2)=100(x2 -x2)2 -x2)2 +(1 -x,)2.
(2-21)
+ (1 -x,)2 Exact solution: (1,1), /i(l,l) =0
2-18
AMCP 706-192
TABLE 2-1 2-7.1.3 COST FUNCTION: f3(x)=x] + 2x\

STEEPEST DESCENT METHOD -
+ 2x2 +2JC,X2 +2X2X3
ITERATIVE DATA FOR COST
FUNCTIONS U)
Exact solution: (0,0,0), /3(0) = 0.
fix 
TABLE 2-3
0 404.0 -1.0 -1.0
1 19.97 0.2576 -0.3743 STEEPEST DESCENT METHOD - ITERATIVE
2 0.8654 0.0707 0.00067 DATA FOR COST FUNCTION f3(x)
3 0.318 0.452 0.1910
4 0.3048 0.448 0.199
5 0.2929 0.472 0.211
fix11') U) m Si)
6 0.2828 0.4685 0.218
29 0.1752 0.5861 0.3373
0.1728 0.5846 0.3403 0 9.0 1.0 1.0 1.0
30
0.6739 0.4499 1 0.0714 0.2857 -0.0715 -0.0715
73 0.1081
0.1071 0.6729 0.4517 2 0.01311 0.1632 -0.153 0.0512
74
3 0.0088 0.1604 -0.114 0.065
4 0.00679 0.1245 -0.062 0.053
2-7.1.2 COST FUNCTION: f2(x) = (xt 5 0.00243 0.078 -0.0625 0.0204
+ 10x2)2 + 5(x3 - jc4)2 + (x2 - 2JC3)
4
+ 6 0.0018 0.073 -0.0476 0.02727
IOC*! -X4)4 7 0.00063 0.0218 -0.00305 0.0133
8 0.00006 0.014 -0.00956 0.0035
9 0.00005 0.011 -0.00686 0.00485
Exact solution: (0,0,0,0), /2(0) = 0 10 0.00003 0.004 -0.0036 0.0040
TABLE 2-2
It should be noted that the Steepest De-
STEEPEST DESCENT METHOD - ITERATIVE scent Method decreased the cost function
DATA FOR COST FUNCTION f2 (x) rapidly on the first iteration but in the first
flx(i)] 01 to (i) XA (I)
two problems failed to converge to the
*2
minimum point. That is typical behavior for
0 122.0 1.0 1.0 1.0 1.0 this method, particularly in problems for
1 16.43 0.9055 0.055 1.0 1.0 which the cost function has a long sharp
2 16.31 0.9019 0.023 0.9958 0.9581
5 16.03 0.8925 -0.0498 0.746
valley. It should be clear that blind use of the
0.969
6 15.06 0.886 -0.0756 0.923 0.463 Method of Steepest Descent can yield poor
9 12.25 0.641 -0.063 0.699 -0.156 results.
10 3.00 -1.048 0.0746 -0.1608 -0.9197
11 2.006 -1.039 0.1522 -0.258 -0.815
12 1.380 1.043 0.078 -0.298 -0.752 2-7.2 GENERALIZED NEWTON METHOD
13 1.188 1.033 0.1127 -0.3276 -0.7067
14 1.047 1.021 0.090 -0.362 -0.634 2-7.2.1 COST FUNCTION: /, (x) = 100(x2
15 1.041 -1.015 0.0949 -0.368 -0.619
16 1.040 -1.012 0.0960 -0.370 -0.611
-x\)2 +(1 -x,)2
38 1.039 -1.008 0.0967 -0.373 -0.603
74 1.039 -1.008 0.0968 -0.373 -0.6019 Exact solution: (1,1), ft (1,1) = 0
2-19
AMCP 708-192
TABLE 2 4 *Note: The trial starting point (1,1,1,1) was

a singular point for X72f2 so an alternate
starting point was chosen and the algorithm
GENERALIZED NEWTON METHOD - ITERATIVE
DATA FOR COST FUNCTION converged.
Mx)
W
/ Mx'"] *.'" x
x
7 2-7.2.3 COST FUNCTION: f3(x) = x] +2x
0 404.0 -1.0 -1.0 + 2x\ +2x!X2 +2*2*3
1 3.981 -0.9950 0.9869
2 3.403 -0.7919 0.5832 Exact solution: (0,0,0), /3(0) = 0.
3 2.588 -0.5248 0.2241
4 1.549 -0.1832 -0.5105
5 0.953 0.0887 -0.0271 TABLE 2-6
6 0.473 0.3642 0.1063
7 0.203 0.5955 0.3347 GENERALIZED NEWTON METHOD -
8 0.0531 0.8020 0.6315 ITERATIVE DATA FOR COST
9 0.0042 0.9536 0.9049 FUNCTION f3ix)
10 0.0002 0.9900 0.9810
11 2x 10 "6 1.0003 1.0007 f3[xu)] (/) (/I (/I
0 9.000 1.0 1.0 1.0

i 2xicr5 0.0015 0.0015 0.0015
2-7.2.2 COST FUNCTION: f2(x) = (x, +
\0x2)2 +50c3 -x4)2 + (*i - 2x3)4 + lOOcj
These results indicate that the Generalized
-x4y
Newton Method is indeed very powerful.
Even in the second cost function where the
Exact solution: (0,0,0,0), f2 (0)= 0
initial estimate caused a singularity in v2f2, a
second starting point yielded good results.
Similar behavior has been noted in the litera-
TABLE 2-5
ture, so one can expect to get good results
GENERALIZED NEWTON METHOD - ITERATIVE with this method. It must be remembered,
DATA FOR COST FUNCTION
however, that this method requires that sec-
f2M
ond derivatives of the cost function be com-
puted.
i f2\*m] *i
0) x2 in in x4 0)
0 137.0 1.0 1.0 1.0 2.0* 2-7.3 CONJUGATE GRADIENT METHOD

1 2.137 -0.3368 0.0175 0.3396 0.3249
2 0.0496 -0.0640 0.0250 0.1060 0.1229 2-7.3.1 COST FUNCTION: /,(x) = 100(x2
3 0.0025 -0.0591 0.0047 0.0627 0.0617
4 0.0007 -0.0236 0.0031 0.0263 0.0271 -x2)2 +(1 -JC,)2
5 0.00001 -0.0148 0.0014 0.0161 0.0160
6 lx ID"6 -0.0070 0.0007 0.0078 0.0079 Exact solution: (1,1), /, (1,1) = 0.
2-20
AMCP 706-192
TABLE 2-7 2-7.3.3 COST FUNCTION: f3{x) = x\2 +4.T..2

2x
+ 2Xj +2xlX2 +2*2*3
CONJUGATE GRADIENT METHOD-ITERATIVE
Mx)
Exact solution: (0,0,0), /3(0) = 0.
i fix'"] W x2
0 404.0 -1.0 -1.0 TABLE 2-9

9649. 0.1143 0.0102
0.0839
1 9649. 0.3258 0.0102 CONJUGATE GRADIENT METHOD -
22.19 0.5106 0.2360 ITERATIVE DATA FOR COST
2 22.19 0.5005 0.2482 FUNCTION f3(x)
0.5033 0.6307 0.3820
3 0.5033 0.6244 0.3882
0.2226 0.7267 0.5178
<;>! (/) (/) .w
4 0.2226 0.7227 0.5212 M* <i
0.001637 0,9919 0.9827
8 0 .001637 0.9842 0.9868 0 9.0 1.0 1.0 1.0
0 .000067 0.000013 0.999754 0.1181 0.3829 -0.2340 0.0744
11 0 .000067 0.999884 0.999768 0.0293 0.2571 -0.2285 0.1428
1 0 0 0 0
2-7.3.2 COST FUNCTION: f2{x) = (x, +

10x2)2 +5(x3 -x4)2 +(x2 _2x3)4 + 10(x1 The numerical results presented here indi-
cate that the Conjugate Gradient Method is
-x4)4
very effective even for the Rosenbrock func-
tion fi(x). The method requires approxi-
Exact solution: (0,0,0,0), /2(0) = 0.
mately the same amount of computation per
TABLE 2-8
step as the Steepest Descent Method but
shows spectacularly improved performance.
CONJUGATE GRADIENT METHOD - ITERATIVE
f7W
It should be noted, however, that con-
t [xM\ i,W *,|,) x M
x,U) vergence slows as the minimum point is
0 1220 1 0 1 0 10 1 0 approached. In fact, as shown in Table 2-8,
2925 27 0 9016 0 0346 0 9642 1000
2925 27 0 8632 0 01787 0 4158 0 9960
convergence to the required accuracy was not
1
36 55
192 2
0 85G1
0 8404
0 0101
0 0563
0 3642
0 3324
0 6573
0 4438
attained in one case.
192 2 0 8180 0 0860 0 3185 0 4540
12 11 0 7507 0 0685 0 2651 0 4787
3 023 06410 0 0850 0 2533 0 4104
2 26 29 0 3404 0 0281 0 2192 0 2075
x; 2'J 0 3378 0 0370 0 2079 0 2057
2-7.4 FLETCHER-POWELLMETHOD
1 G51 0 3331 Q 0319 0 1748 0 2130
0 1130 0 3159 0 0371 0 1473 0 1698
G 00531 0 0305 0 0029 0 0718 0 0717
2-7.4.1 COST FUNCTION: fx(x) = 100(*2

0 000751 0 0293 0 0031 0 0696 00714
0 000751 0 0294 0 0078 0 0695 00713
8
0 000556
0001091
0 0297
0 03519 0
0 0029
003519
0 068/
0 02338
000?7
0 02503
~x2)2+(l -*,)2
0 000517 0 035322 0 03498 0 0231 12 0 0231 19
0 0005 17 0035318 0 003530 0 073108 0 0231 19
1 K 10-7 0 035305 0 003528 0 023011 0 0731 15 Exact solution": (1,1), /, (1,1) = 0.
2-21
AMCP 706-192
TABLE 2-10 2-7.4.3 COST FUNCTION: f3(x)=x2l +2x22

+
2x\ +2XiX2 +2x2x3
FLETCHER-POWELL METHOD-ITERATIVE
DATA FOR COST FUNCTION fx (x)
Exact solution: (0,0,0), /3(0) = 0.
w
M* l x2
0 404.0 -1.0 - 1.0 TABLE 2-12

1 19.97 0.2570 -0.3746
2 0.7839 0.1 146 0.01249 FLETCHER-POWELL METHOD-
3 0.7570 0.1422 0.005683 ITERATIVE DATA FOR COST
4 0.7424 0.1727 0.005740 FUNCTION f3(x)
5 0,5377 0.3378 0.08262
6 0.4013 f3[*li)]
(/) (/) (/)
0.3689 0.1416
7 0.2968 0.4815 0.2151
8 0.2524 0.5616 0.2909 0 9.0 1.0 1.0 1.0
9 0.03621 0.8286 0.6784 1 0,05319 0.3830 -0.2340 0.07447
10 0.032 16 0.8207 0.6733 2 0.02857 0.2571 -0.2286 0.1429
11 0.02568 0.8536 0.7221 3 3x 10"13 2x icr7 -2x icr7 -3x 1CT7
12 0.01 162 0.9268 0.851 1
13 0.00437 0.9342 0.8733
14 0.00106 0.9760 0.9504 The Fletcher-Powell Method requires slight-
15 8x 10'^ 0.9982 0.9967
ly more computation than the Conjugate
Gradient Method. However, its convergence
properties are very good as the minimum
point is approached, in contrast to the be-
2-7.4.2 COST FUNCTION: f2(x) = (x, -
2 2 A havior of the Conjugate Gradient Method.
10JC2) + 5(x3 -xA) + (x2 -2x3) + 10(JC!
-*4)4 This method appears to have good prop-
erties in all ranges of the iterative process. It is
Exact solution: (0,0,0,0), f2 (0)= 0.
more stable than the Generalized Newton
Method in the early stages of computation
and converges more rapidly than the Gradient
and Conjugate Gradient Methods near the
TABLE 2-11
minimum point. In these respects it has the
FLETCHER-POWER METHOD-ITERATIVE DATA desirable properties of other methods without
FOR COST FUNCTION f2 U) having their undesirable properties.
x1'! (/')
0 122.0 1.0 1.0 1.0 1.0 2-7.5 CONJUGATE DIRECTIONS WITH-

1 14.4292 0.9017 0.03472 0.9642 1.0
2 2.3775 0.8630 0.4120 0.9960
OUT DERIVATIVES
-0.07820
3 0.6678 0.8430 -0.08740 0.3618 0.4986
4 0.3353 0.2087 -0.02560 0.3644 0.3305
5 0.05134 0.1117 0.006686 0.1883 0.1952
6 0.01059 0.07931 -0.009696 0.1532 0.1526
7 0.00067 0.02731 -0.0007003 0.06189 0.06276
2-7.5.1 COST FUNCTION:/,(*) = 100(x2
8 0.00016 0.02164 -0.002344 0.05417 0.05409 -x\)2 +(l-x,)2
9 8.3 x 10 "6 0.00267 -0,0000359 0.0191 0.0192
6
no 2.1x 10" 0.00148 -0.000 63 0.0172 0.0172
ii 10-7 -0.0057 0.00060 0.00341 0.00342 Exact solution: (1,1), /i(l,l) = 0.
2-22
AMCP 708-192
TABLE 2-13 TABLE 2-14 (Continued)
h [xW] v 1 li)
i x4 li)
x
CONJUGATE DIRECTIONS WITHOUT 2 *3 dt
DERIVATIVES METHOD-ITERATIVE DATA
FOR COST FUNCTIONS, M 0.4421 0.0469 0.0127 0.2799 0.4284
2 0.1415 0.0469 0.0127 0.2799 0.2423
li) li) 0.1418 0.0510 0.0127 0.2799 0.2423
fx\x Oh x2
0.1210 0.0510 -0.0015 0.2799 0.2423
0.0498 0.0510 -0.0015 0.1875 0.2423
0 404.0 -1.0 -1.0
0.0246 0.0510 -0.0015 0.1875 0.1749
100.1 0.0049 -1.0000
3 0.0082 0.0536 -0.0104 0.1291 0.1324
0.9902 0.0049 0.0000
4 0.0020 0.638 -0.0181 0.0794 0.0882
1 0.9902 0.0261 0.0211
5 0.0018 0.1322 -0.0147 0.0892 0.0940
0.9485 0.0261 0.0007
6 0.0010 0.0828 -0.0109 0.0580 0.0603
0.9402 0.0429 0.0174
7 0.0005 0.0412 -0.0057 0.0377 0.0322
2 0.7922 0.1287 -0.0016
8 0.0000 0.0078 -0.007 0.0058 0.0050
0.7022 0.1815 0.0509
0.6172 0.2147 0.0436
3 0.3958 0.4058 0.1440
4 0.2895 0.4785 0.2422 2-7.5.3 COST FUNCTION: &) = x] +\
5 0.2591 0.5308 0.3015
+ 2x2, +ill +2fi|
6 0.0770 0.7258 0.5225
7 0.0282 0.8564 0.7246
8 0.0125 0.8942 0.8033
9 0.01 19 0.91 16 0.8373
10 0.0116 0.9039 0.8218 Exact solution: (0,0,0), f3 (0) = 0.
11 0.0125 0.9469 0.9065
12 0.0042 1.0363 1.0792
13 0.0002 0.9886 0.9781
14 0.0002 1.0032 1.0079 TABLE 2-15
CONJUGATE DIRECTIONS WITHOUT DERIVATIVES

METHOD-ITERATIVE DATA FOR
2-7.5.2 COST FUNCTION: f2(x) = (*, COST FUNCTION f3(x)
-llijl + ^ . . x^ +$ti 2xi"| +10Oi -
W 0) 0) (i)
f3[x ]
Exact solution: (0,0,0,0), /2(0) = 0.
0 9.0 1.0 1.0 1.0
5.000 -1.0000 1.0000 1.0000
TABLE 2-14 3.000 -1.0000 0.0000 1.0000
1 1.000 -1.0000 0.0000 0.0000
CONJUGATE DIRECTIONS WITHOUT DERIVATIVES 0.000 0.0000 0.0000 0.0000
METHOD-ITERATIVEDATA FORCOST
FUNCTION f2 W
li) (i)
f2[x<"] x, (i) x4 li) The Conjugate Directions Without Deriva-
0 122.0 1.0 1.0 1.0 1.0 tives Method is not as efficient as some of the
109.1 0.2051 1.0000 1.0000 1.0000 methods that require computation of deriva-
18.45 0.2051 0.1140 1.0000 1.0000 tives. However, there are many problems in
7.667 0.2051 0.1140 0.4819 1.0000
which computation of derivatives is either
1 2.371 0.2051 0.1140 0.4819 0.4284
2.157 0.0469 0.1140 0.4819 0.4284 impossible or very difficult. In these prob-
1.075 0.0469 0.0127 0.4819 0.4284 lems, this method appears to be effective.
2-23
AMCP 706-192
2-8 AN APPLICATION OF UNCON- In this form, an iterative technique such as

STRAINED OPTIMIZATION TO the power (or iteration) method (Ref. 13,
STRUCTURAL ANALYSIS page 93) may be applied to obtain the largest
eigenvalue of the matrix K~lM and hence,
As pointed out earlier in this chapter,
the smallest eigenvalue of the original prob-
optimal design problems are seldom uncon-
lem. Even though the power method is
strained. There is, however, a large class of
efficient, this approach has the severe dis-
analysis problems which can be solved using
advantage of requiring that I'1 be com-
unconstrained optimization methods. In Ap-
puted.
pendix B, energy principles which govern
equilibrium, vibration, and stability of struc- A more promising approach to the above
tures are given. The condition for equilibrium eigenvalue problem utilizes the Rayleigh
is particularly direct since it requires that, in quotient (Ref. 13, page 83), i.e., the smallest
problems for which the strain energy is eigenvalue X, of Eq. 2-25 is given by
quadratic, the equilibrium state, x, minimizes min yTKy
V of Eq. B-18, Appendix B, 1 (2-27)
yÔ yTMy
T
V = TTX KX -xTF. (2-24) If the vector y is normalized by fixing one of
its elements, the resulting vector denoted y,
Even in some problems which are nonlinear then Eq. 2-27 reduces to
and the total potential energy is not qua-
min yTKy
dratic, the minimum energy principle applies. A1 ^ x^f.^ (2-28)
y y My
In view of the quadratic form of Eq. 2-24, The minimization Eq. 2-28 may now be
conjugate direction methods are indicated. solved by any of the methods of the present
Even for nonquadratic energy expressions, chapter. The method of conjugate directions
methods for conjugate directions appear to be has been recently applied to solve this class of
very efficient. For a much more detailed problems (Refs. 14, 15). It is interesting to
treatment of this class of equilibrium prob- note that this exact approach to the eigen-
lems, see Ref. 12. value problem was proposed by the inventor
of conjugate direction methods, M. R.
A second structural analysis problem for Hestenes, in 1955 (Ref. 16, page 93j. The
which unconstrained optimization methods technique was apparently not used in engi-
hold even more promise is the eigenvalue neering problems, however, until 1966.
problem, As shown in Appendix B, vibration
and buckling problems reduce to eigenvalue Iterative methods of the kind outlined in
problems of the kind this paragraph are particularly appropriate for
iterative optimal design techniques. As dis-
Ky = \My . (2-25) cussed in Chapter J, the most time consuming
task in iterative design methods is the re-
In this problem, the smallest eigenvalue X , of analysis of the system during each iteration;
the Eq. 2-25 is sought. One method of solving i.e., after the design variable is changed
this problem is to rewrite Eq. 2-25 as slightly, analysis for stresses, displacements,
and eigenvalues must be done even though it
1 (2-26)
K - 'My- is expected that these quantities will be very
2-24
AMCP 706-192
close to their values before the change in the minimization algorithm. In this way, rapid
design variables. By using an iterative tech- convergence to the new state of the system is
nique such as conjugate directions, the pre- attained. This approach has been applied with
vious state may be used as an estimate to start good success (Ref. 15).
REFERENCES
C. Goffman, Calculus of Several Vari- 10. M. J. Box,"A Comparison of Several

ables, Harper and Row, New York, 1965. Current Optimization Methods and the
Use of Transformations in Constrained
S. Perlis, Theory of Matrices, Addison- Problems", The Computer J., Vol. 9, p.
Wesley, Reading, Mass., 1952. 67, 1966.
3. T. L. Saaty and J. Bram, Nonlinear 11. M. J. D. Powell, "An Effective Method

Mathematics, McGraw-Hill, New York, for Finding the Minimum of Functions of
1964. Several Variables Without Use of Deriva-
tives", British Computer J., Vol. 7, pp.
4. D. J. Wilde and C. S. Beightler, Founda- 155-162, 1964.
tions of Optimization, Prentice-Hall,
Englewood Cliffs, New Jersey, 1967.
12. R. L. Fox and E. L. Stanton, "Develop-
ments in Structural Analysis by Direct
5. A. V. Fiacco and C. P. McCormick,
Energy Minimization," AIA A J., Vol. 6,
Nonlinear Programming. Sequential Un-
No. 6, June 1968,pp. 1036-1042.
constrained Minimization Techniques,
John Wiley & Sons, New York, 1968.
13. S. H. Crandall Engineering Analysis,
McGraw-Hill, New York, 1956.
6. J. C Pearson, On Variable Metric Meth-
ods of Minimization, Technical Paper
14. W. W. Bradbury and R. Fletcher, "New
RAC-TP-302, Research Analysis Corpora-
Iterative Methods for Solution of
tion, McLean, Virginia, February 1968.
Eigenproblem", Numerische Mathematik,
9, 1966,pp. 259-267.
7. R. K. Williamson, Direct Methods of
Parameter Optimization far Dynamical
Systems, Air Force Rept. No. SAMSO - 15. M P. Kapoor, Automated Optimum De-
TR - 69-432, July 15, 1969. sign of Structures Under Dynamic Re-
sponse Restrictions, Report No. 28,
8. W. C. Davidon, Variable Metric Method Division of Solid Mechanics Structures
far Minimization, Atomic Energy Com- and Mechanical Design, Case Western
mission R.E.D. Rept. ANL-5990, 1959. Reserve University, Cleveland, Ohio,
January 1969.
9. R. Fletcher and M. J. D. Powell, "A
Rapidly Convergent Descent Method for 16. M. R. Hestenes, "Iterative Computational
Minimization", The Computer J., Vol. 6, Methods", Comm. Pure Appi. Math., Vol.
p. 163, 1963. 8, 1955,pp. 85-96.
2-25
AMCP 706-192
CHAPTER 3
LINEAR PROGRAMMING
3-1 INTRODUCTION b. Power consumption must be belo,w a

specified level.
In the preceding chapter a function fix), x
in R", was minimized with no restrictions c. Capacitance of a proposed capacitor
placed on the location of the design variable must be within attainable limits.
x. Problems in the real world seldom reduce
to this form. In virtually all engineering design 3. Aerospace vehicle guidance:
problems, requirements are placed on the
object being designed, and these requirements a. Controller thrust must be within the
are stated in terms of equations involving the capability of the thruster.
design variable. More often, these require-
ments may be stated in terms of inequalities b. Total fuel consumption for a mission
involving the design variable. must be less than or equal to the vehicle's
storage capacity.
Examples of inequality constraints are
abundant in all areas of engineering design. c. Altitude must be greater than or
The following are examples: equal to zero.
1. Optimal structural design This list of typical inequality constraints

could be expanded many-fold. It is clear then
a. Stress mustbelessthanorequaltothe that the inequality constraint must play a
yield strength of the material. central role in any unified theory of design.
b. Buckling load must be greater than The class of problem considered in this
or equal to applied loads. chapter is very restricted. Only linear func-
tions are to be minimized subject to con-
c. Deflection of the structure must not straints which are linear in the design vari-
exceed specified limits. ables. In matrix notation this is, minimize
d. Natural frequency must lie within an fix) = CTx (3-1)

allowable range.
where C is an n x 1 matrix of constants. The
design variable x is required to satisfy
2. Optimal circuit design:
Ax <5
a. Voltage must remain within linear (3-2)
range of components. x> 0
3-1
AMCP 706-192
where A is an m x n matrix and B is an m x 1 As will be seen in the following paragraph,

matrix. The inequality, Eq. 3-2, is taken as this is typical of linear programming prob-
lems.
2 a....x < bp i= 1, ...,H .
/= l Before proceeding to the next paragraph, it
i.e., when one vector is less than or equal to is worthwhile to discuss the applicability of
another vector, each of the components of linear programming. The theory of linear
the vectors must satisfy this relation. programming arose out of studies of econom-
ic activities. In economics it is often the case
Example 3-1: Consider the problem of that behavior of an economic system is
minimizing predictable only in a rather crude way, so
frequently a linear relation among variables is
fix) = xi + 2*2 (3-3) as good a representation as can be expected.
subject to the constraints In engineering design, however, it is very

seldom that the behavior of an object or
2
*i
+ x2 .,1 process can be described by linear expres-
sions. One might be tempted, then, to com-
x
i > 0 | (3-4) pletely ignore linear programming. Even
W though it is not directly applicable to most
X
2 > 0. engineering design problems, however, linear
programming is still a very powerful tool.
The constraints, Eq. 34, are satisfied at all First, even though the computational pro-
points in the triangular region of Fig. 3-1. The cedures of linear programming do not carry
lines passing through this region are lines of over to the real nonlinear world, many facets
constant value of fix). It is clear that as the of the behavior of solutions are very similar in
line is translated downward, the value of fix) more general programming problems. The
decreases and that the lowest line that still engineer who has mastered linear pro-
contains points in the admissible region oc- gramming will go into the study of the much
curs for x, + 2x2 = 0. Since this line more complex nonlinear programming armed
intersects the admissible region only at (0,0), with a powerful tool intuition. Further, the
fix) takes on an absolute minimum at (0,0). solution of many nonlinear problems can be
reduced to the solution of a sequence of
linear programming problems. For a review of
some of these applications of linear pro-
gramming methods see Ref. 1.
3-2 PROPERTIES OF LINEAR PRO-

GRAMS
To formalize the discussion of the previous

paragraph, the following definition is made.
Figure 3-1. Graphical Solution of
Example 3-1 Definition 3-1: The linear programming
3-2
AMCP 706-192
problem is the problem of determining that x describe an admissible object or process, i.e.,
mR" which minimizes one which performs the required service but is
not necessarily optimal. In LP the constraint
B1 x (3-5) set is a polyhedron and, according to Def. 2-4,
this constraint set is convex. Further, accord-
and which satisfies ing to the same definition, the cost function
> LP f{x) for LP is convex. If the constraint set is
Ax> C (3-6) bounded and nonempty, it is necessarily also
closed and all the hypotheses of Theorems 2-2
x> 0 (3-7) and 2-3 are satisfied. One then concludes that
fix) has a strict absolute minimum in the
where C = 0 is an m x 1 matrix, A is an m x n constraint and that is has no other relative
matrix, B is an n x 1 matrix and the minima.
symbolism < () as applied to matrices means
that the relation less than or equal to (greater Further, if fix) had a minimum in the
than or equal to) holds for corresponding interior of the constraint set, the necessary
components of the matrices. condition of Theorem 2-1 implies
- 0, i = i],...,
df
It should be pointed out that Eqs. 3-5 -= c, -n
dx, '
through 3-7 do not explicitly cover all linear
optimization problems. For example, it may which contradicts Def. 3-1 of LP. Therefore,
be required to maximize a linear objective fix) cannot have a minimum point in the
function. Further, equality constraints may interior of the constraint set but must take on
be imposed and negative values of the x. may its minimum at the boundary. Weyl has
be allowed. However, all these variations on shown, in fact, that the solution must lie on
the linear programming problem may be put one of the vertices of the polyhedral con-
into the form of the problem previously straint set (Ref. 2).
considered. An objective function may be
maximized by minimizing its negative, equal- In spite of this elementary theory, it is
ity constraints are nothing more than a pair of possible that a linear programming problem
inequality constraints (i.e., y = 0 if and only if may not have a solution. This may happen for
y s 0 and y < 0), and a negative x. may two reasons. First, the constraint set may be
always be written as the difference between empty; and second, the constraint set may be
two new non-negative variables. There is unbounded and the cost function may be
therefore, no loss of generality in considering decreased without restriction. In order to
only the problem expressed by Eqs. 3-5 facilitate discussion of these difficulties,
through 3-7. Definition 3-3 is made.
Definition 3-2: The constraint set for the Definition 3-3: If the constraint set of LP
linear programming problem of Def. 3-1 is the is nonempty (empty), the problem is called
set of points mR" which satisfy Eqs. 3-6 and feasible (infeasible). If the constraint set is
3-7. unbounded and the cost function is not
The constraint set associated with a prob- bounded below, then the problem is called
lem is just the set of design variables which unbounded.
3-3
AMCP 706-192
The concept of the dual problem that will Theorem 3-2: Let LP and LPD both be
be used in constructing solutions of LP's will feasible. Then both have solutions x and yt
now be discussed. The dual problem will also respectively, andBTx = CTy.
play a major role in obtaining results for more
general optimization problems. The proof of Theorem 3-2 is involved and
does not yield a method of constructing
Definition 3-4: The linear programming solutions. It may be found in Ref. 3, page 44,
problem of maximizing or Ref. 4, page 118.
CTy (3-8) Since the solution of LP must lie on a

vertex of the polyhedral constraint set, it
iory mRm satisfying suffices to check at most a finite number of
> LPD points for the minimum. This procedure is
ATy o function cannot be decreased by moving
along an edge of the polyhedron that inter-
where the matrices^, B, and C are the same sects this vertex, then this vertex is the
as in LP, and are called the dual of LP. solution. If, however, the cost function de-
creases by moving along some edge, this
The results of Theorem 3-1 relating LP and policy is followed until a second vertex is
LPD are proved in Ref. 3, page 41, and Ref. reached and the cost function has been
4, page 118. reduced. Since there are only a finite number
of vertices and it is impossible to return to a
Theorem 3-1: Let x and y be in the previously occupied vertex, the process must
constraint sets of LP and LPD, respectively. terminate at the minimum over the constraint
Then set.
1. CTy < BTx. (3-10) In order to illustrate the argument pre-

sented in the preceding paragraph, consider
2. ff CTy = BTx thenxandy (3-11) Example 3-2.
are the solutions of LP and LPD, respectively.
Example 3-2: By moving along edges of
the constraint set, solve the LP
3. If LP (LPD) is unbounded, then LPD
(LP) is infeasible.
minimize/(*!, x2) = 2x i x2
4. If LP (LPD) is feasible and LPD (LP) is subject to

infeasible, then LP (LPD) in unbounded.
xy > 1
These results are useful in constructing
solutions of linear programming problems. x2 > 1
They are also used in providing Theorem 3-2
that is central to linear programming theory.
3-4
AMCP 706-192
Xi,x2 > 0. The only move admissible is toward (1/2,

1). A unit move in this direction is obtained
Solution: The polyhedral constraint set is from
shown in Fig. 3-2.
V?'
2
dx:
(1/2,1) 2.
which causes a change inf,

(1,1/2)
df = Vf(\M2)dx = + y/T-
v ^1=^Z> o.
2 2
*-xl
Figure 3-2. Polyhedral Constraint Set

Therefore, / may not be decreased in moving
from the vertex (1, 1/2) so this point is the
The vector solution of the problem.
The idea of moving from vertex to vertex is

-otVfT(Xl ,x2) = <x good for visualization but is poor for higher
dimensional problems. The same idea, how-
ever, can be implemented algebraically. In
whose direction as shown in Fig. 3-2 is the order to obtain relations which will be re-
direction of steepest descent of f(x). Starting quired for solution of LP, define slack vari-
at (0,0) a unit movement along the xt -axis ables w!,..., u so that
yields a change
Ax ~ C = u > 0. (3-12)
df = V/(0,0)dx = -2
The cost function of Eq. 3-5 will be denoted
and a unit movement along the x2 -axis yields by the variable
a change
w=BTx. (3-13)
df = V/(0,0)dx = - 1
The problem LP now takes the form
so both moves yield a decrease in/(x). Choose Ax - C - u = 0
the xx-axis and move to the first vertex (1,0).
The only movement possible is in the x> 0
+ x2 -direction from (1,0). A unit move in this LP'.
direction yields u> 0
df = Vf( 1,0) dx = - 1 w = BTx = minimum
which decreases f. Move in this direction to The solution of LP' is the same as the solution
the first vertex (1, 1/2). of LP.
3-5
AMCP 706-91
The information contained in Eqs. 3-12 using a method which is based largely on
and 3-13 is contained in the following matrix Theorem 3-3.
equation (called the simplex tableau):
Theorem 3-3: If in Eq. 3-15 b'. 0, i = 1,
an al2 ... aln _ct xt ux ..., n, and_c^ > 0, j = 1, ..., m, then the
solution of LP is
-c2
rt = 0, i = 1,..., n
(3-14)
S =
i -^J= 1,..., m
a
ml m2 w = 6.
bx b2 bn ! 0
It is clear from this theorem that any
method of choosing the variables sf and r,
Eq. 3-14 may be viewed asm + 1 equations
which wiU terminate with non-negative entries
involving the variables xt ,..., xn, ult... ,um,
in the last row and column, except perhaps
w. At present Eq. 3-14 may be interpreted as
for 6, will serve as a method of solving LP.
determining ux , ..., um , and w explicitly in
Before developing such a method, several
terms of xl,..., xn . It might be desirable to
definitions will be helpful.
determine some other combination of m + 1
of the variables in terms of the remaining n. Definition 3-5: In Eq. 3-15, the variables
Except in singular cases, this is possible. s.,j = 1 m, are called basic variables, while
the variables r., i = 1,..., n are called nonbasic
Assume now that m + 1 of the variables .?i, variables.
..., sm , and w have been determined explicitly
in terms of the remaining variables^ ,..., rn . Definition 3-6: The set of variables Sj,...,
Eq. 3-14 will then take the form s r r
m i >> will be called abasicpoint. If c.
[- < 0,;' = 1,..., m, in Eq. 3-15, then the basic
11 12
point will be called a basic feasible point.
21 22 2H - C2 r2 2
A certain geometric interpretation may
now be given for the nonbasic variables. In
(3-15) LP' it is clear that the boundary of the
constraint set of LP is obtained by setting
mi a'mi u
mn <-m various combinations of the variables xt, / = 1,
..., n and ;-, / = 1,..., m, equal to zero. In the
b\ b'2 b'n I spaced" of the design variable x, a vertex of
the polyhedral constraint set is obtained by
where primes denote coefficients obtained having n equality constraints among the xv i =
when the original set of equations is solved 1, ..., n, enforced. By the discussion, this
forsj,..., sm ,and w. occurs when r. = 0, i = 1,..., n. An edge of this
polyhedron is a line inR" obtained by setting
The solution of LP will be constructed r. = 0 for n 1 indices i. From Def 3-6 and
3-6
AMCP 706-192
Ey. 3-15, it is clear that a basic feasible point a-in a,kj

corresponds to a vertex of the polyhedral set. + ... + *kn
a
This is true since setting the nonbasic vari- U
ables of the basic feasible point equal to zero
L
yields admissible basic variables. Further, two i"k/
vertices lie on the same edge of the constraint

set if they have n 1 of their nonbasic
variables in common.
It is thus clear how the coefficients in Eq.
The process for interchanging the roles of a 3-15 change as the roles of a pair of variables
basic and a nonbasic variable thus becomes are interchanged. This process may be de-
the central tool for methods based on Theo- scribed concisely in the language of Definition
rem 3-3. Suppose it is desired to make s. a 3-7.
nonbasic variable and r, a basic variable. If a'.,
i= 0 then the ith equation from Eq. 3-15, Definition 3-7: The entry aL = 0, preced-
ing Eq. 3-16, is called the pivot of the
a rx +...+ai/rj+... +ainrn _c,=s; transformation. The transformation itself is
called a.pivot step.
may be solved for r. to obtain
The effect of the pivot step on the coef-

y-i
ficient matrix of Eq. 3-15 may be illustrated
V = r '"I j-i
' a.- a.. easily by the diagram
v v
V1
;+i
(3-16) r l a
a
H p (X
-> p P
(3-18)
7 V "J3
a!. r"'
fl J _p P
v
Using this expression for rjt r- may be

The diagram shown by Eq. 3-18 simply relates
eliminated from the left sides of the remain-
that in the coefficient matrix of Eq. 3-15 the
ing equations in Eq. 3-15. For k # / this yields
following changes occur. The pivot is replaced
by its inverse. All other elements in the same
'ii"kj
row as the pivot are multiplied by the
*k\
negative inverse of the pivot. All other ele-
ments in the same column as the pivot are
a
ij-xakj multiplied by the inverse of the pivot. All
"kj- /-I other elements in the matrix are decreased by
a
U
the product of the element in their column
aii+rakj and the row of the pivot, the element in their
*kj+1 a
'/, (3-17)
u row and the column of the pivot and the
inverse of the pivot.
3-7
AMCP 708-192
Example 3-3: given stage of the solution process and the

primes of Eq. 3-15 are dropped, i.e.,
Given ~ 2 1 -4~ ~rT Sl"
-Cl ?i S\
3 6 1 ?2 = 2
- 5 3 2_ _1_ _w_
(3-19)
a c
m 1 ~~ m rn Sm
interchange the role of rl and 2.
8 1 W
Solution: The new matrix relation is

Primes will now be used to denote the
2/3 -3 -14/3 coefficients that result from a pivot step
applied to Eq. 3-19. These new coefficients
1/3 - 2 - 1/3 are determined by applying Eq. 3-18.
-5/3 13 II/3JLI. 3-3.1 DETERMINATION OF A BASIC FEA-

SIBLE POINT
It is shown in Ref. 3, page 53, that this
pivoting transformation preserves the dual If some elements in the right-hand column
linear programming problem. of the matrix of Eq. 3-19 (other than 6) are
negative, then the present choice of variables
The pivoting transformation is an organized is not a basic feasible point. Let ck be the
tool which allows one to interchange basic negative entry nearest the bottom of the
and nonbasic variables. It remains only to column (again excluding 6). Since when r. =
obtain an algorithm which uses this tool and 0, / = 1, ..., n, sk = ck < 0, if there are
Theorem 3-3 to construct the solution of LP. admissible points in the constraint set of LP,
then it must be possible to increase sk by
3-3 THE SIMPLEX ALGORITHM increasing some r. from zero; i.e., there must
be some positive ak.. Choose/'0 so that ak, >
As was shown in par. 3-2, the solution of 0. This fixes the column index of the pivot.
the linear programming problem may be
reduced to the choice of pivot points. The To find an admissible row index con-
algorithm presented here will have two sider first that after the pivot step
phases. The first phase will consist of an
algorithm for obtaining a basic feasible point.
The second phase will operate only with basic !
o/o
feasible points and will successively reduce
the cost function until the hypotheses of It is clear then that candidates for the pivot
Theorem 3-3 are satisfied. a.'0^0
, must be limited to indices i for which
For convenience in the discussion which

follows, it is assumed that the choice of basic -> 0. (3-20)
and nonbasic variables has been made at a "To
3-8
AMCP 706-192
With this restriction in mind, consider the The process described may be given quite
values of c'. after the pivot step with i = i0. simply as the iterativeAlgorithm LP-A:
These are Step 1. Choose ck as the lowest negative

entry (with the exception of 6) in
'o '/o the right-hand column of the co-
c'. = c,+ (3-2 1)
% !o efficient matrix of Eq. 3-19.
In order to insure c't > 0, i > k, it is Step 2. Choose any positive element ak,
required that in the kth row of the matrix of Eq.
3-19.
.0.+'*"' > 0,i>k,i*L (3-22)
Step 3. Choose z'0 asinEq. 3-25
<o/o
ff a.,. > 0 this clearly holds. If au'/ o < 0,

J
Step 4. Perform the pivot step with pivot
i/o
however, the requirement, Eq. 3-22, may be a. .
'o/o
rewritten as
Step 5. If any cj < 0, i= 1,..., k, choose
C- ig
(3-23)
that one with largest index / and
L > , i > k, ; # /'0
return to Step 1. If - c. > 0,i= 1,
.... m, then a basic feasible solution
Further, for / = k, has been found and the process
may be terminated.
(3-24)
3-3.2 SOLUTION OF LP
since a n > 0. In par. 3-3.1 an algorithm is given for

K/O
finding a basic feasible point. Once this has

Inequalities, Eqs. 3-23 and 3-24, show thai been accomplished, the object is to find a
if iQ is chosen so that second algorithm which successively reduces
w.
^ = min [X I _fL > 0 \ (3-25) Since by Eq. 3-19, w = b^rx +... + bnr +
Vo i>k Wo ' aHo 6, it is clear that if b. < 0 for some; =/
/o
then w may be reduced by increasing r. from
then c\ > 0, i > k and c'k > ck. If ck zero. If a pivot step is performed which makes
is still negative, the process may be repeated. r./o a basic variable then w will be decreased,
Otherwise choose the next entry above ck The choice of the basic variable s. which is
'0
which is negative and repeat the process. to be made nonbasic must be made in such a
way that the point obtained after the pivot
If all the cjy i > k are nonzero, only a finite step is still a basic feasible point, i.e., so that
number of basic points are possible since the c. > 0, i = 1. ..., m. However, this is
process is monotone (nonrepeating). If there precisely the restriction which led to the
exists a point with c. > 0, i=\, ..., m, this choice of i0 in par. 3-3.1. Therefore, the same
process must find it. The degenerate case in procedure for choosing ('0 may be employed
which some c, = 0, / > k is discussed later. here.
3-9
AMCP 708-192
Since w'= w ci0bj0/ajo/0, the pivot step Viewed geometrically, the difficulty occurs
determined here guarantees w' < w provided because the path which successive basic points
all c. > 0, i = 1, ..., m. In this case, follow on the polygonal constraint boundary
therefore, only a finite number of pivot steps may form a closed loop. To prevent this
may be made, and the process must terminate behavior with only a small error in the final
at the solution of the linear programming solution an entry, c., which is zero, is
problem. Termination occurs when*' > 0,/ = replaced by an arbitrarily small parameter E >
1,..., n. Theorem 3-3 shows that this is the 0. The problem is not degenerate any longer
solution of the linear programming problem. and cycling cannot occur. Therefore, the
The degenerate case where some c( = 0 will be altered problem will proceed toward the
discussed par. 3-3.3. solution.
This process is given explicitly mAlgorithm Example 3-4: Use the simplex algorithm to
LP-B: solve the LP
Step 1. Choose any negative entry (except minimize 2Xi + 9x2+ x3

6) b, in the bottom row of the
coefficient matrix of Eq. 3-19. subject to
Step 2. Choose ;'0 according to Eq. 3-25 x1 + Ax 2+2*3 s. 5

withfc = 1.
3^! +x2 + 2x3 > 4
Step 3. Perform the pivot step with pivot
xl > 0
Step 4. If any bf < 0, j = 1,..., n, choose s> 0

*2
one bj < 0 and return to Step 1.
If bj > 0, / = 1, ..., n, then the Xi > 0.
solution of LP has been found.
First, LP is:
3-3.3 THE DEGENERATE CASE
minimize w where
In both pars. 3-3.1 and 3-3.2 the computa-
tional algorithms could have problems if some '14 2-5'
c. = 0. This situation is called degenerate since
when n constraints are made equalities by 3 1 2 - X7
putting r. = 0, / = 1,..., n, one has s. = c(. = 0
which means that still another constraint is an L2 0 1 0.
equality. The degeneracy arises from the fact
that in LP the n dimensional design variable x
= (Xj, ,xn) satisfies n +1 linear equalities.
Therefore, the n + 1 equations are not linearly subject to
independent.
xi > 0, /= 1,2, 3, w. > 0, /= 1, 2
3-10
AMCP 706-192
For the first pivot step in algorithm LP-A, pivot step leads to
k = 2. Choose j0 = 1 since a21 = 3 is the
largest element in the second row. i0 = 2 is - 1/6 2/3 -11/6 11/6" 1 ^3
the only choice available in Eq. 3-25 and
1/3 - 1/3 2/3 1/3 2 ^2
C2
- 4> 0.
1
a22 _ 17/6 -7/3 37/6 29/6_ *l W
=
The pivot is a22 1- This pivot step inter-
.1- =
changes u2 and x2' The result is
Put/0=2,
11 4 -6 11 *1 1
Cy C2
= -11/4, = 1,
3 1-24 "2 - x2 fl
12 a
12
L-25 9 - 17 36. *3 _w_ so i0 = 2 and a2 2 = 1/3 is the pivot. A pivot

step yields
I-
"1/2 -2 - 1/2 5/2" u, x{
Note that this basic point is already a basic
feasible point so that the process now trans- 1 - 3 2 1 = 2
2
fers to algorithm LP-B. Since b't is most
negative, choose;0 = l.Now, J/2 7 3/2 5/2_ 1
W
c C
i 2 _4_ .1-
Ml a2l
Since this is a basic feasible point and the
so z0 = 1. The pivot is then at t = 11. The
first three elements in the third row are
result of a pivot step is to interchange Xj and
positive, then the solution is immediate. The
M, . This results in the basic feasible point
nonbasic variables are zero,
Z. l/ll 4/11 -6/11 1 "l Xi
i = Xj =x2 =0
3/11 - 1/11 -4/11 1 2 = x2 and the basic variables take on the value
_25/ll - 1/11 -37/11 11J x3 w_ =

*3 5/2, u2 = 1, and w = 5/2.
j_,
Therefore, the solution to the original LP is
Choose /'0 = 3.
c, c2
x2 = 0
= 11/6, = 11/4,
13 Q23 x3=5/2.
=
so J'Q = 1. The pivot is a22 6/11 and a The minimum value of f(x') attained is 5/2.
3-11
AMCP 706-192
3-4 MINIMUM WEIGHT TRUSS DESIGN linearly independent equations in Eq. 3-26.
ff a is the maximum allowable stress (both
As will become apparent in subsequent
chapters, most optimal design problems are tensile and compressive) for the material from
nonlinear. Even the problems considered in which the truss is constructed, then stress
this paragraph appear at first glance to be constraints are
nonlinear. However, it is shown that the
problem can actually be solved as a linear I Sf I < oAf (3-27)
program. This will not be the case in general.
The class of problems and their solutions that Further, if p is the weight density of the
are discussed in this paragraph are taken from structural material, the total weight W of the
an outstanding paper by Dorn, Gomory, and truss which is to be minimized is
Greenberg (Ref. 5). Similar results have been
W-. p
reported more recently (Ref. 6). V/ (3-28)
The problem treated here is minimum where fi. is the length of the jth member.
weight design of plane trusses with constraints
on stress. The initial restrictions on the truss The problem of minimizing W of Eq. 3-28
include only the location of joints in the subject to the constraints of Eqs. 3-26 and
truss. The loads to be supported by the truss 3-27 is not the complete truss design problem.
are applied at joints. A member with non- In addition to the equilibrium conditions of
negative cross-sectional area is allowed to Eq. 3-26, a set of compatibility conditions
connect each pair of joints. If there are between displacements of the joints must be
joints, there may be M(M 0/2 members in satisfied. These compatibility conditions will
the truss. In general, then, statically indeter- be nonlinear in the variables S. and A,. In its
minate trusses are allowed. complete formuiation, then, the truss design
problem is not a linear programming problem.
Let A., j 1, .... n, denote the cross- It will be shown, however, that if the com-
sectional area of jth member and S. the load patibility conditions are ignored and the
in that member due to the external loads problem described by Eqs. 3-26, 3-27, and
applied to the truss; S. > 0 denotes tension. If 3-28 is solved, its solution satisfies the com-
m = 2;u,then equilibrium of the joints of the patibility conditions and is, therefore, the
truss is specified by the equations solution of the truss design problem.
Recalling that compatibility relations are

jZ 1 a St F, i=l, (3-26)
being ignored, it is required that
where F. are components of applied forces at
I a. 1 =aA;., /= 1,..., n. (3-29)
the joints, and a., are direction cosines of the
elements of the structure intersecting the jth
joint. All a., are zero if the jth element does This is true since if IS, I < aA- for some;, then
not intersect the point of application of F.. In A. could be reduced with an accompanying
order to satisfy three equilibrium equations reduction in W. The constraint, Eq. 3-27, is
for the applied loads (including reactions at therefore replaced by Eq. 3-29. The reader
supports), it is assumed there are m* = m 3 should note that this argument would not be
3-12
AMCP 708-192
valid if compatibility conditions were being LP', namely, find* to minimize

enforced, since a reduction in some A. may
result in a violation of a constraint not BTx (3-30)
involving,4. explicitly.
subject to
Since by Eq. 3-29,4- = \S-1, the optimiza-
tion problem is now to minimize Ax- C = 0 (3-31)
x> 0. (3-32)
W= 2 |S|,
a t =i /' /
This linear programming problem may now
subject to Eq. 3-26. In order to treat this be solved by the simplex method. Before the
problem as a linear programming problem, solution of the linear programming problem
define can be taken as the solution of the truss
design problem, however, it must be shown
Sj, ifSf> 0 that it satisfies the compatibility conditions.
s; = { It is clear that if the truss specified by the
0, if Sf < 0 linear programming problem is statically
determinate, it satisfies the compatibility con-
0, if 5) > 0 ditions trivially (i.e., there are no compatibil-
s ity conditions). For the analysis here, stat-
i=<
-Sjt if 5) < 0 ically determinate is taken to mean that the
member forces 5 are uniquely determined by
Now, the given loads and the equilibrium conditions
of Eq. 3-26.
srs;sr
As pointed out in Ref. 5, page 32, there
and will be m* possibility nonzero components of
x (basic variables) in the solution, correspond-
\Si\=S*+Sr ing to linearly independent columns of the
matrix A ; i.e., only m* of the Sy will possibly
Denote be nonzero. According to Eq. 3-27, then, only
m* of the areas may be nonzero. Further,
since the rank of A is m*, the member forces
* =(s? ,...s; ,s; ,...,s- > are uniquely determined. The resulting truss
CT =(Ft Fm), is, therefore, statically determinate and hence
is the solution of the original truss design
problem.
A -(aift -av)m x2
and It is pointed out (Ref. 5) that the simplex

method for solving many member truss design
BT =.ih,...,K,K...An) problems is relatively time-consuming. It is
proposed that the method be refined for this
In this notation, the problem is of the form class of problems to obtain a practical method
3-13
AMCP 708-192
of solving engineering design problems. Sever- subinterval, the member sizes are different. A
al examples are solved in considerable detail plot of W vs a and the forms of optimal
in Ref. 5; the results of one of these problems trusses are shown in Fig. 3-4.
will be discussed here.
A bridge truss is to be designed to span two

points, 1 and 13 of Fig. 3-3. Three vertical
24
<3V?t>
levels of joints are allowed with five horizon-
22
tal sets, a total of 15 points, as shown in Fig.
20
3-3. In the general case there could be 18
15(14)/2 = 105 members in the truss. Loads 16
on the floor of the truss are shown in Fig. 3-3. 14i
,2
n 9- 12- 15-
Figure 3-4. Optimum Bridge Trusses
11- 14-
The discussion here only touches on the
i_* 10- highlights of the very complete treatment of
H+-:
I t r the truss design problem in Ref. 5. The
interested reader is encouraged to study this
outstanding article in detail.
Figure 3-3. Admissible Joints for Bridge Truss
Before leaving the truss design problem, a
point of interest in the present results and in
In the solution presented in Ref. 5, it is the results obtained in future chapters may be
assumed that the truss is symmetric about the noted. In Fig. 3-4 it is clear that at two values
line of joints 7-8-9. This assumption reduces of a the form of the optimal truss changes
the number of variables to 57. Further, due to form drastically. Still, even though the
the assumed symmetry, there are only 14 topology of the structure is not continuous in
independent equilibrium conditions. There- a, the weight apparently is a continuous
fore, there will be only 14 members which function of a. The same sort of behavior
can be nonzero in the optimum truss. In the occurs in a beam design problem with con-
solution presented in Ref. 5 the problem is straints on deflection which is discussed in
made nondimensional by defining a = h/Z and par. 7-4. These problems might lead one to
= H/V, where h and 2 are the vertical and suspect that there is some basic mathematical
horizontal spacing, respectively, and H and V structure of the optimal structural design
are applied loads shown in Fig. 3-3. problem that has not been uncovered.
The solution presented in Ref. 5, page 45, 3-5 AN APPLICATION OF LINEAR PRO-
for a fixed value of ( = 1) shows that there GRAMMING TO ANALYSIS
are three subintervals of values of a on each
of which the truss has a constant geometrical . A major application of linear programming
form. For different values of a within a given in engineering design is, oddly enough, in
3-14
AMCP 706-192
nonlinear programming. It is seldom that a tial equations onXj o x o x2

realistic engineering design problem can be
formulated as an LP. Realistic problems are m
d'z
generally nonlinear when considered as a L[z]= 2 fl(jt) - (3-35)
i= o ' ax'
function of both state and design variables.
Several techniques of solution of nonlinear and the boundary operator is
programming problems are based on approxi-
mation of the nonlinear problem by a linear (3-36)
Biz] = Az(xy) +Bz(x2).
one, at least locally. These methods then
require that the approximating LP be solved.
In the case of partial differential equations,
This subject will be deferred until a discussion
of the general theory of nonlinear pro-
gramming has been given.
L[z] = 2 ajx).
la Km
aV1...a* "
A second application of linear pro-
gramming which is of concern to the engineer (3-37)
is in the solution of linear boundary-value
problems that arise in such fields as con- and the boundary operator is
tinuum mechanics. It should be emphasized
here that this application is not of an optimal Biz] = A(x)z(x),xonr . (3-38)
design nature, but rather falls in the field of
engineering analysis. The method to be discussed treats both the
partial and ordinary differential equations in
One of the important methods of solving the same way. Let 0-00, / = 1,.... k satisfy the
linear boundary-value problems is to approxi- homogeneous differential equation
mate the solution by a linear combination of
known functions. The question arises, "How L[<t>A =0,inJ2. (3-39)
should the coefficients be chosen so as to
obtain the 'best' approximation to the true Further, let 0O (x) be found such hat
solution?" "Best" may be defined in many
ways. A relatively new concept of "best" will L14>0] = Q{x),mSl. (3-40)
be discussed in this paragraph.
Since the operator L is linear the new-
The general linear boundary-value problem function
may be stated in operator notation as
L[z]=Q(x), xinft (3-33) 2=0+2 c-0.(x) (3-41)

u /= 1 ' '
B [z] =q(x), x onT (3-34)

satisfies the differential Eq. 3-33 regardless of
where Q, is the domain of the independent the value of the constants c The object is
variable X<ER" and T is its boundary. The now to find these constants so that z satisfies
dependent variable is a vector function of x, the boundary conditions of Eq.3-34 as closely
z(x) in Rm . In the case of ordinary differen- as possible.
3-15
AMCP 706-192
Define Example 3-4: Obtain an approximate solu-

tion of
\\B[z] -q \Bi[z] -<?,. |. (3-42)
_ d2z d^z_ 1 dz
In this notation, z will be the solution of the dxj ox\ x dx
boundary-value problem if and only if
ml={{xl,x2)\\xl | < 1, 0 < x2 < 1 } with
\\B[z{ _<?(*) ||=0 (3-43)
for all points x onF. z + on

-- = 0onr = {(*,,* )| |JC | = 1,
i J. L
The method to be treated here attempts to

minimize the error in Eq. 3-43 at a large x2 = 0 or x2 = 1 (3-49)
number of points xR, Q = 1,. . ., L, on T.
Define where n is the interior unit normal to T.
7= s II [!(**)] -?(**) (3-44) Put
The object now is to choose the constants c.

so as to minimize y. To see that this is a linear
programming problem, note that Eq. 3-44 is
equivalent to \ (3-50)
5,.[z(x8)] -<7,.(x*)< T (3-45)

03 =%x\ -24x\x\ + Zx\.
and
Note that these functions satisfy Eqs. 3-39

-B,[7(xe)] +<?,(xe)< y (3-46)
and 3-40.
for all i and Q.
The domain $2 and its boundary Y are
Note that Eqs. 3-45 and 3-46 are linear in shown in Fig. 3-5. Partial derivatives with
the c, and y. Since the 9 may be either respect to the interior normal are shown.
positive or negative, it is necessary to define X
2 dz _ dz
new constants cj > 0 and cj > 0 such that dn ~ dx2
c,=cj -cf (3-47) \

n
dz _ dz
dz _ dz * n n
. n dn~ dx
+
Now, the problem of choosing y, c , cT (all
non-negative) which satisfy Eqs. 3-45 and
an dx.
t
dz _ dz
1
3-46 and which minimize 7 is clearly a LP. dn dx
Further, it is just a restatement of the best Figure 3-5. Boundary Condition for
approximation criterion of Eq. 3-44. Example 3-4
3-16
AMCP 706-192
The procedure is now to form Similar inequalities in the c*, cT, and 7 will be
obtained at all other boundary points chosen.
*= \x\ ^c^c,{2x\ -x\) Under the requirements c* > 0, c~> 0, and
y > 0, the problem of minimizing 7 is then
+ c3(&x* -2Ax\x\ + 3*) solved.
Rabinowitz in Ref. 1, page 141, reports

and, with the aid of the expressions for dz/dn that an approximate solution obtained by the
in Fig. 3-5, compute z + dz/dn at L points above method is
around the boundary T. At a typical point,
e.g., (1, 1/2), Cl =-0.5571, c2 = 0.0764, c3 = 0.0024,
_ (E 1
z + = +c, -(1 1/4lc, 7 = 0.0053.
a 16
_(285/16)c3. This means that at all the boundary points xe,

I 7 + dz/dn I < 0.0053. A result called a
At this point it is required that maximum principle from the theory of
second-order elliptic partial differential equa-
1/16 + (c\ -c~) - (11/4) (c+2-c-) tions then implies
+
-(285/16)(C 3-C3)Q7 |Z(JC) -z(x) 1 < 0.0053, x in ft
and where z(x) is the true solution of Eqs. 3-48

and 3-49. This powerful result guarantees that
_1/16_(C; _c-) + (ll/4)(c+2-c2) the approximate solution z generated by
linear programming is within 0.0053 of the
-(285/16) (c+ -cp<7- true solution throughout SI.
REFERENCES
1. P. Rabinowitz, "Applications of Linear 4. T. L. Saaty, Mathematical Methods of

Programming to Numerical Analysis", Operations Research, McGraw-Hill, New
SIAM Review, Vol. 10, No. 2, p. 121, York, 1959.
April 1968.
5. W. S. Dorn, R. E. Gomory, and H. J.
2. H. Weyl, "The Elementary Theory of Greenberg, "Automatic Design of Optimal
Convex Polyhedra", in H. W. Kuhn and A. Structures", Journal deMecanique, Vol. 3,
W. Tucker, Contributions to the Theory of No. 1,March 1964, pp. 25-52.
Games, Annals of Mathematical Studies
No. 24, Princeton University Press, 1950.
6. W. R. Spillers and J. Farrell, "On the
3. G. Owen, Game Theory, Saunders, Analysis of Structural Design", J. Math.
Philadelphia, 1968. Anal. Appl., Vol. 25, 1969,pp. 285-295.
3-17
AMCP 708-192
CHAPTER 4
NONLINEAR PROGRAMMING AND FINITE

DIMENSIONAL OPTIMAL DESIGN
4-1 INTRODUCTION TO THE THEORY OF Definition 4-1: The first nonlinear pro-
NONLINEAR PROGRAMMING (NLP) gramming problem NLP, is: find xeR" to
As pointed out in the preceding chapter, minimize f(x) (4-1)

inequality constraints play a central role in
engineering design problems. The inequalities subject to \ NLP
treated in Chapter 3, however, are of a rather
special form, namely, they involve only linear g(x) < 0 (4-2)
functions of the variables of the problem. It is
a rare real-world design problem which can be
put into this form. In general, the inequality where g(x) =
constraints as well as the cost or return
function in real-world problems are nonlinear.
For this reason, a more general theory than
Unless otherwise specified, it will be
that presented in Chapter 3 is needed.
assumed that f(x) and g(x) are continuously
differentiable. Other than this differen-
The class of problems considered here is
tiability requirement, f(x) and g(x) are as
called nonlinear programming, or math-
general as required for a particular problem.
ematical programming. A vast amount of
literature has been devoted to this class of
A second form of nonlinear programming
problems in recent years. Several books on
problem, which may actually be included in
the subject which contain reviews of this
NLP, is given by Definition 4-2.
literature are Refs. 1,2, and 3. In view of this
extensive literature, the purpose of this para-
Definition 4-2: The second nonlinear pro-
graph is simply to state the nonlinear pro-
gramming problem NLP', is: find xeR" to
gramming problem and present some key
results needed in the study of methods of
optimal design.
minimize f(x) (4-3)
4-1.1 NONLINEAR PROGRAMMING

subject to
PROBLEMS
For convenience and clarity in the develop- g(x); 0, NLP' (4-4)

ment of methods of solution, the nonlinear
programming problem will be stated in two
forms. The first form is given by Definition
4-1. (4-5)
4-1
AMCP 706-192
solution with only very mild assumptions.

gl (x)
This result is a consequence of properties of
where g(x) =
R" . In the infinite dimensional optimization
Jm(x)_ problems of Chapter 6, the space of variables
and I -. lacks these properties so that no analogous
,(X) result is available.
h(x) =
Theorem 4-2 provides an easy test for
-V*>_ closedness of the constraint set.
Theorem 4-2: If the functions g(x) and

Unless otherwise specified, it will be
h(x) are continuous, then the sets D and D'
assumed that f(x), g(x), and h(x) are con-
are closed in Rn .
tinuously differentiable.
The boundedness hypothesis of Theorem
Very much as in the linear programming
4-1 may be more difficult to check, par-
problem, the points x which satisfy the
ticularly in complex problems. One must
constraints of NLP and NLP' are charac-
show that there exists a number a such that if
terized by Definition 4-3.
xeD or D1, then xTx < a.
Definition 4-3: The sets of points x&Rn
To see that NLP' can actually be included
that satisfy the constraints NLP and NLP' are
in NLP, define
called constraint sets. They are denoted
D = {x^R" | g(x) < 0 ) 8i + m (x) = hl(x)J= l,..-,p
and
for NLP, and
gi + m+p(x)=- ft.(x), i= 1,...,p.
D' = { x^R" | g(x) < 0 and h(x) = 0 }
Now, NLP' is equivalent to the NLP:
for NLP'.
minimize f(x)
For convenience, Theorem 2-2, which was
stated previously in Chapter 2, is given here
subject to
(Theorem 4-1) as it applies to nonlinear
programming problems.
00 * 0,
Theorem 4-1: If f(x) is continuous on D
gi (*)
(D'Jand this set is closed and bounded ini?" ,
where g(x) =
then NLP (NLP') has a solution which is an
absolute minimum of/Cx) mD (D') . g
m+lp(X>
This theorem is one of the most easily This is true since

obtained yet most powerful results in opti-
mization theory. It guarantees existence of a ,-(*)< 0, i=m + 1,..., m +2p
4-2
AMCP 706-192
is just linear programming problem, one must be

careful not to generalize too much. To
hjix)* 0,j = l,...,P illustrate some differences between linear and
nonlinear programming, two examples will
and now be treated.
-*/*)< 0,/= l...,p Example 4-1 :
which is equivalent to Minimize
h(x) = 0.
fM = (Xi -3)2 +(x2 -3)2
It should be clear that problems of maxi- subject to constraints

A
mizing fix) are put into the form NLP or
NLP' simply by defining fix) = fix).
xx < 0
Further, constraints of the form g(x) > Oare
transformed to the proper form simply by
x2 < 0
defining g(x) = g(x). These transforma-
tions involve no theoretical or practical diffi-
x, +x2 - 4<0.
culty. As will be seen in par. 4-2, even though
the transformation of NLP' into NLP involves
The constraint set is the shaded triangular
no theoretical difficulty, severe practical diffi- region in Fig. 4-1.
culties occur. The explicit characterization of
equality constraints in NLP' will be useful
later, when methods of constructing solutions
are discussed.
(4,0)
Comparing nonlinear programming prob-
lems with the unconstrained problems of
Chapter 2, one might conclude that the
nature of the cost function fix) will deter-
mine the location of the minimum point, with
only a check required to verify that con- (0, 0) (4,0) 1
straints are satisfied. Since the linear pro-
Figure 4-1. Graphical Solution of Example 4-1
gramming problem is a special case of the
nonlinear programming problem, the results
of Chapter 3 show vividly that this conclusion If the constraints are ignored,/(x) takes on
is false. In the linear problem, the cost its minimum at the point (3,3). Observing the
function plays only a minor role in the circles, which are plots of constant value
simplex algorithm and most of the computa- curves of fix), it is clear that the smallest
tional effort is expended operating on the value f{x) takes on in the shaded triangle is
constraint functions. /(2,2) = 2. This is, therefore, the solution of
the problem.
While results from the linear programming
problem yield valuable insight into the non- It should be noted that even though the
4-3
AMCP 708-192
solution occurred on the boundary of the increased complexity of nonlinear as opposed

constraint set, it did not occur at a corner as to the linear problems is not surprising since
it would have if the problem had been linear. increased complexity generally accompanies
this transition in all mathematical disciplines.
Example 4-2:
Minimize Due to the complexity of NLP and NLP',

methods of obtaining their solutions are
f(x)=(Xl _1)2 +(x2 -l)2 generally computational in nature. Moreover,
in many meaningful engineering problems,
subject to the constraints convergence proofs are not available so the
designer must depend heavily on his engineer-
X[ < 0 ing intuition. One must be extremely careful
in applying engineering intuition to certain
x2 < 0 aspects of optimization problems, however. In
most problems of engineering analysis, exis-
*i +x2 4< 0. tence and uniqueness of solutions are taken
for granted since these properties hold for
The constraint set is just the same as in the very general classes of problems such as linear
previous problem. The cost function, how- elasticity, dynamics, circuit theory, and struc-
ever, has been modified. tural analysis. Existence and uniqueness ques-
tions in optimization problems are, however,
If the constraints are ignored, ftx) takes on by no means trivial. For instance, before the
its minimum at (1,1). Since this combination designer commits himself to a design based on
of design variables satisfies the constraints, it an optimum obtained by a computational
is the solution of Example 4-2. The solution algorithm, he should seriously consider the
of this nonlinear programming problem, possibility that this optimum is only relative
therefore, occurs in the interior of the con- and an absolute optimum exists that will give
straint set. This behavior contrasts sharply much better results.
with that of linear programming problems
where the solution must occur on the bound- Due to the weakness of intuition in dealing
ary of the constraint set. with optimization problems and the inherent
complexity of these problems, the importance
These examples show conclusively that the of theoretical results concerning existence,
properties of NLP, and hence, also NLP', uniqueness, and necessary and sufficient con-
differ considerably from those of LP. ditions cannot be overemphasized. The re-
mainder of this paragraph and par. 4-2 are
Theoretical results and computational devoted to these questions, while pars. 4-3
methods for NLP and NLP' will also be more through 4-5 contain methods for obtaining
complex than those for the linear pro- solution of NLP and NLP'.
gramming problem. The reason for this is
clear. Strong use was made of linearity of the 4-1.2 GLOBAL THEORY
functions involved in the linear programming
problem, and this linearity is not present in In nonlinear programming problems one
the nonlinear programming problem. The often obtains a relative minimum of f(x) in
4-4
AMCP 708-192
the constraint set. The question arises, "Is this 4-1.3 LOCAL THEORY
relative minimum an absolute minimum?" In
general problems it is difficult to answer this Without convexity it is difficult to say
question. There is a class of problems, how- much about global properties of the solution
ever, in which this question is easily answered. of NLP or NPL'. Considerable theory is
This class is described by Definition 4-4. available, however, which characterizes local
minima. The approach in the local theory is
Definition 4-4: If D (D') is a convex set to suppose that/(x) has a relative minimum at
and fix) is convex onfl (D') then NLP (NLP') a point inD orD' and then find conditions on
is called a convex programming problem. fix), gix), and h(x) which must hold at this
point. In this way, many points inZ) and D'
Theorem A-l, Appendix A, guarantees that may be eliminated as candidates for a relative
if gj(x)> i= 1, , rn, are convex functions, extrema and perhaps relative extrema can
then the set D is convex. Since the equalities even be located using these conditions. Such
(Eq. 4-5) in NLP' define a surface inR" , it is conditions, therefore, are called "necessary".
clear that D' is the intersection of that In some problems it will be possible to obtain
surface with the set { xe^lg^x) =s 0, i = 1, a set of conditions that, if satisfied at a point,
..., m) . The surface is convex if and only if it guarantee that this point yields a relative
is a plane, or equivalently, if and only if each extremum. Conditions of this kind, of course,
hXx) is linear in x. Since by Theorem A-6, are called "sufficient".
Appendix A, the intersection of two convex
sets is convex, D' is convex if g((x), i = 1,..., As often happens in engineering, the engi-
m, are convex and hXx),j = 1,..., p are linear. neer needs a powerful result developed in
The class of problems NLP' which are convex mathematics to solve his problem. Proof of
is, therefore, quite restricted. this result, however, may be very complex
and, in fact, contribute very little to the
As will be clear from what follows, con- engineer's insight into his problems. This
vexity is a very desirable property. However, appears to be the case in many phases of
in the real world, many optimization prob- optimization theory, in particular, in the
lems are nonconvex. In spite of this fact, the study of necessary a.id sufficient conditions
study of convex problems is justified. Many in nonlinear programming. In the remainder
results which hold only in convex problems of this paragraph results will be borrowed
have led to constructive methods which are from mathematical developments.
effective for finding local extrema in noncon-
vex problems. Some of these methods would
Before meaningful results may be given for
probably never have been developed if only
NLP and NLP', the following conditions will
general nonconvex problems had been
be required of the constraint functions g(x)
treated.
and h(x).
One of the powerful results which follows
due to convexity is given in Theorem 4-3. Definition 4-5: (First-order constraint
qualification): Let x be a point in the
Theorem 4-3: A relative minimum in a constraint setD' (orZ> if there are no equality
convex programming problem is an absolute constraints) and let the functions gix) and
minimum. h(x) be differentiable at x. Then the first-
4-5
AMCP 706-192
order constraint qualification holds at x if In this result, and in fact, in the remainder
for any nonzero yeR" suchthat Vgj(x)y o 0 of this paragraph, the problem NLP' is de-
for each i with g.(x) = 0 and Vh(x)y = 0, scribed. It is clear, however, that putting/ = 0
then y is tangent to a differentiable arc in NLP' yields NLP. One of the principal
passing from x into the constraint set. results of nonlinear programming may now be
stated. For proof the reader is referred to Ref.
Geometrically, this definition says that if 1, page 20.
the vector y is a direction which, to first
order, appears to point from x into the Theorem 4-5: (Kuhn-Tucker Necessity
constraint set, then there is a curve with y as Theorem): Let the functions f(x), g(x), and
tangent which actually passes from x into h{x) be differentiable and let the constraint
the constraint set. The conditions W,(*0)j Q functions satisfy the first-order constraint
0 forg,U) = 0 and Vh(x) = 0 are just first qualifications at a point x mD' of NLP'. In
order perturbations of g{(x) and h(x) which order that x be a relative minimum for NLP'
indicate that a small move in the y-direction it is necessary that there exist multipliers
ought to do the right thing to gt(x) and veRm and we '' suchthat
h(x). This is illustrated in Fig. 4-2.
Vj T> 0, i = 1,..., m (4-6)
2 j,w = o VjgjCx) = 0, i= 1, ..., m (4-7)
and
V L(x,v,w) =0 (4-8)
where
L(x,v,w) =f(x) + vTg(x) + wTh(x) (4.9)
is called the Lagrangian.

Figure 4-2. First-order Constraint
Qualification In a sense, Theorem 4-5 is an existence
theorem. It asserts that if x yields a relative
minimum for NLP', then the multipliers v and
While all constraints do not satisfy the
w exist and that Eq. 4-8 is satisfied. Occa-
first-order constraint qualification, the follow-
sionally, one will run across an argument
ing theorem (Ref. 1, page 19) identifies a class
attempting to justify this theorem which
of constraints which do.
states that
Theorem 4-4: If g(x) and h(x) are differ-
fix) =f(x) + vTg(x) + wTh(x) =L(x,v,w)
entiable at x in D' and if the gradients
Vg;-(*), for i wifhg.(jc) = 0, and Vh^x0)
are linearly independent, ; = 1,. . ., p, then since v is defined by Eq. 4-7 and h = 0. It is
the first-order constraint qualification is satis- then claimed that since x yields a relative
fied. minimum for f(x) it must yield a relative
4-6
AMCP 708-192
minimum for L(x,v,w), so VL(x,v,w) = 0 and if for every nonzero yeR" such that
must hold. This argument is not valid. For a Vgj(x)y = 0 for v( > 0, Vgt(x)y < 0 ioig^x)
rigorous proof of Theorem 4-5 the reader is = 0 and v. = 0, and Vh(x)y = 0, it is true that
referred to Ref. 1.
y' V L{x,v,w)y > 0 (4-11)
Theorem 4-6 states additional conditions
which are required to hold if the functions then x yields an isolated relative minimum for
appearing in NLP' have two derivatives. NLP'.
Theorem 4-6: (Second-order Necessary It should be noted that there is a gap

Conditions): Let fix), g(x), and Mx) have between the sufficient conditions of Theorem
two continuous derivatives at a point x inD'. 4-7 and the necessary conditions of Theorem
Further, let the vectors \/gt(x), for all i with 4-6. Strict inequality is required in Eq. 4-11
gfx) = 0, and Vh(x) be linearly independent. for a larger set of vectors y that may yield
If x yields a relative minimum for NLP', then only equality in Eq. 4-10. It is doubtful that a
it is necessary that there exist v and w single, tractable set of conditions exist that
satisfying Eqs. 4-6, 4-7, and 4-8. Further, for are both necessary and sufficient for the
every y<= R" suchthat Vgffiy = Owheng.(J) general problem NLP'.
= 0, and Vh(x)y = 0, it is necessary that
There is one class of nonlinear pro-
yT S72L(x,v,w)y > 0 (4-10) gramming problems in which conditions may
be given that are both necessary and sufficient
For proof of this theorem, see Ref. l,page for an absolute extremum. This class is the
25. Note that the existence of v and w convex programming problem.
satisfying Eqs. 4-6, 4-7, and 4-9 is a conse-
quence of Theorem 4-5. Even though this Theorem 4-8: Let/(x) andg(-(x), i= 1,...,
theorem involves second-order conditions, it m, be continuously differentiable and convex,
still gives only necessary conditions. then necessary and sufficient conditions forx
to be an absolute minimum point of NLP are
A theorem which gives conditions which, if that there exists veRm such that
satisfied at some point, are sufficient to
guarantee that this point yields a relative
minimum for NLP' will now be stated. For g0t)< 0
proof of this theorem, see Ref. 1, page 30.
vigi(x)= 0, i= l,...,m
Theorem 4-7: (Second-order Sufficient
Conditions): Let f(x), g(x), and h(x) be twice vi > 0, i= 1,..., m
differentiable functions at a point x. If for
XED' there exist v and w satisfying and
v. > 0,i= 1,..., m

7/(3?) + v VgAx) = 0.
i=i '
Vjgjix) = 0,i= 1,..., m

The technical presentation of par. 4-1 ends
VL(x,v,w) = 0 with this satisfying result. Several comments
4-7
AMCP 706-192
are, however, appropriate at this point. The from the theorems stated in the preceding
analytic necessary and sufficient conditions of paragraph.
par. 4-1 could be used to construct solutions
of NLP by solving systems of nonlinear 4-2.1 FINITE DIMENSIONAL OPTIMAL
equations. This is particularly true of the DESIGN PROBLEMS
results of Theorem 4-8. If one reads the
current literature, however, he is led to the The class of problems to be treated in this
distinct conclusion that iterative methods paragraph is, in a sense, a special case of the
based on successive improvements are too nonlinear programming problem NLP'. How-
effective to bypass in favor of methods that ever, by developing a theory for the new class
require solution of complicated, nonlinear, of problems which takes advantage of its
algebraic equations. special features, a more efficient solution
algorithm may be obtained.
Even if the results of par. 4-1 are never
used by the designer to construct solutions of The general optimal design problem must
nonlinear programming problems, they are have several of the features of NLP'. Namely,
still very powerful tools. Verification of the it is required to have a cost (return) function
hypotheses of one of the theorems may mean which is to be minimized (maximized) and a
the difference between going onto the com- set of constraints that describe the perfor-
puter with the comforting knowledge that a mance demanded of the object being de-
unique solution exists as opposed to the signed. It is in the representation of con-
frustrating experience of having computer straints that the optimal design problem
print-out which may be meaningless. differs from NLP'.
4-2 THEORY OF FINITE DIMENSIONAL In most problems of design in the real-

OPTIMAL DESIGN world the object being designed is required to
behave according to some law of physics. This
The nonlinear programming problems of behavior is described analytically by a set of
par. 4-1 are quite general and may be applied variables called state variables. Further, there
to a variety of optimization problems. As is is a second set of variables that describe the
frequently the case with very general formula- object itself rather than its behavior. These
tions of problems, special features of some variables are called design variables since they
problems within the class being studied are are to be chosen by the designer so that the
not exploited. This appears to be the case object being designed performs its required
when general nonlinear programming theory function. It generally happens that the laws of
is applied to solve optimal design problems. physics that determine the state variables
Interpretation of certain of the variables and depend on the design variables so the two sets
constraints in the problem NLP', in the of variables are related.
context of optimal design, yields very effective
computational methods of solution. This To illustrate the difference between state
paragraph will be devoted to stating the finite and design variables, consider the following
dimensional optimal design problem, drawing design problems:
an analogy with NLP', and stating necessary
and sufficient conditions that follow directly 1. Find the coefficient of damping in an
4-8
AMCP 708-192
automobile shock absorber so that peak ac- optimal design problem (OD) is a problem of
celeration in the passenger compartment due determining beRk to
to road conditions is as small as possible.
minimize f{z,b) (4-12)
The coefficient of damping is the design
variable since it describes the object being subject to
designed, and its magnitude is to be fixed by OD
the designer. Acceleration on the other hand h(z,b) = 0 (4-13)
is a state variable since it describes the
behavior of the object being designed. <j>{z,b) < 0 (4-14)
Further, this state variable may be determined
by Newton's laws of motion. Note that the where
designer has no direct control over the state
ht(z,b)
variable. He may effect it only indirectly by
adjusting the design variable. This is typical of h(z,b)
state and design variables.
hn (z,b)
2. Determine the size of beams to be used ~<t>dz,b)

in a structure so that when a given set of loads
are applied stresses are within certain given <S>{z,b) (4-15)
limits, the deflection of certain points on the
structure is within given limits, and the J
structure is as light in weight as possible.
and all the functions of the problem are re-
Beam sizes are the design variables in this quired to have first-order derivatives. Further,
problem since they describe the structure it is required that the (n + k) vectors
being designed and they must be chosen by
the designer. Stress and deflection, however, 3, 9*i
(4-16)
are state variables that are determined by 9z ' db
equilibrium and force deflection relations.
Again, the designer has no direct control over are linearly independent for all i with <t>t(z,b)
stress and deflection. He may effect these = 0 and that the matrix
quantities only by varying the size of beams
in the structure. d_h_
(4-17)
dz
In most real-world design problems the
state and design variables are clearly identi- is nonsingular.
fied. In what follows, the state variable will be
an n-vector, ze", and the design variable will
be a k-vector, beRk. The basic elements of the The assumption that the matrix mr is
optimal design problem are described by nonsingular guarantees, by the implicit func-
Definition 4-6. tion theorem (Ref. 4, page 181), that for
given b there is a unique solution of Eq. 4-13
Definition 4-6: The finite dimensional forz. Further, the state variable z, determined
4-9
AMCP 706-192
from Eq. 4-13 as a function of b, is differ- linearly independent at z,b. Then there exist
entiable with respect to b. This fact will be multipliers AeJ?" and peRm, with > 0 such
needed later when constructive methods are that for
developed.
H=f(z,b) + \Th(z,b) +jiT<t>{z,b) (4-21)
4-2.2 LOCAL THEORY
dH
(z,b) = 0 (4-22)
Since it is very seldom that the state db
equations (Eq. 4-13) are linear in both z and
b, convexity of the constraint set and hence (z,6) = 0 (4-23)
the problem will be rare. For this reason, no oz
global results based on convexity will be
discussed. In case Eq. 4-13 is linear, however, and
global results may be obtained by applying
the Theorems 4-3 and 4-8. M/0/(z,6) = O, /=!,..., (4-24)
It is clear that if a new variable X<ER"

+
is The proof of this theorem may be con-
defined as structed by simply writing down the
necessary conditions of Theorem 4-5 in terms
of x and then separating the components of x
(4-18) asinEq. 4-18.
then the problem OD may be put into the In exactly the same way the second-order
form NLP'. According to Theorem 4-4, the necessary and sufficient conditions of
first-order constraint qualification will be Theorems 4-6 and 4-7, respectively, may
satisfied for OD (with xeR" + * as indepen- be stated for the problem OD. No essential
simplification of the statements of those
dent variable) if the row vectors
theorems occurs, however, so the theorems
are not restated here.
<iht bhi
i = 1, .... n (4-19)
8z~' db Theorem 4-9, just as Theorem 4-5, is
difficult to use in constructing solutions of
30, 30, OD. Considerable difficulty arises because one
-r-{z,b\ ~{z,b)
Oz db does not know which of the inequalities in
OD is an equality. For problems with a small
for; with 4>-(z,b) = 0 (4-20) number of inequality constraints this may not
be a difficult obstacle, particularly if the
are linearly independent. Theorem 4-5 may designer has a good intuitive idea of which
now be applied to the problem OD. constraints will be equalities. If, on the other
Theorem 4-9: (First-order Necessary Con- hand, there are a large number of inequality
ditions): Let all the functions appearing in constraints, then the number of combinations
OD be differentiable at a point z, b which of constraints which may be equalities is
satisfies Eqs. 4-13, 4-14, and 4-15. Further, large. It is, therefore, difficult to determine
let the vectors, (Eqs. 4-19, 4-20, and 4-21) be just which combinations will be equalities. An
4-10
AMCP 706-192
analytic solution is extremely difficult in this of Fiacco and McCormick (Ref. 1). Theoret-
case. ical results guaranteeing convergence are
presented here to indicate the level of the
Rather than attempt to use the necessary known theory of SUMT, rather than as a
conditions to construct candidate solutions, a complete treatment of the subject.
more direct approach will be followed. The
remainder of this chapter will be devoted to 4-3.1 INTERIOR METHOD
direct methods of solving NLP, NLP', and
OD. The interior SUMT is based on the idea of
using the constraint functions to erect a
4-3 SEQUENTIALLY UNCONSTRAINED
barrier at the boundary of the constraint set
MINIMIZATION TECHNIQUES (SUMT)
D of NLP by adding a penalty function to
fix) which approaches infinity as the bound-
A favorite method of solving difficult ary of D is approached from the interior.
problems, particularly among mathematicians,
Once the solution of the augmented problem
is to reduce a difficult problem to a sequence is obtained, the penalty function is altered so
of easy problems. Each of the easy problems as to effect fix) less in the interior of D.This
is solved and if the method is any good, the behavior is illustrated in Fig. 4-3.
sequence of solutions of easy problems will
converge to the solution of the difficult
f(x) + Penalty Function UK '
problem. As the title might imply, SUMT
follows just this pattern. It should be clear
I
that a central part of this method must be
results which guarantee convergence, at least / I
in cases where solutions are known to exist.
The method presented here essentially re-
duces NLP and NLP' to a sequence of Penalty ""^^^^
Function (1.) /
auxiliary problems which may be solved by \ /
the methods of Chapter 2. The cost function
of NLP or NLP' is augmented by a function ,ru
called a penalty function. The penalty func- (A)
tion is formed from the constraint functions
in such a way that as a parameter approaches
zero (or perhaps infinity) the unconstrained
minimum of the augmented cost function
converges to the solution of NLP or NLP'.
Two basically different ways of doing this are
presented here. Each has its computational
and theoretical advantages and disadvantages
that will be described later.
Due to the large body of theory concerning
SUMT, results will be presented in this para-
graph without proof. The reader is referred
for proofs and an extended discussion of
SUMT to the complete and well-written text Figure 4-3. Penalty Functions
4-11
AMCP 706-192
As illustrated in Fig. 4-3, when the penalty 1. I(x] is continuous and non-negative on
function is decreased on the interior of D, the the interior of the constraint set D and if
minimum of the second augmented cost | .^^ is any sequence of points in R"
functionx(2) is closer to the solutions than converging to x where g(x) = 0 for some/,
the minimum of the first augmented cost then^ /(**) = +
function J'. The idea, of course, is that the
sequence of points x(,) generated in this way 2. S(r) is continuous and if r{ > r2 > 0,
converges to x. then S(r1) > S(r2) 0 and if r. is a sequence
of numbers converging to zero, then j^ S(r.)
It should be clear why this approach is = 0.
discussed only for NLP and not NLP'. The
constraint set of NLP' can have no interior Probably the most common penalty func-
due to the equality constraints. It is possible tions I(x} and S(r) are
that NLP has no interior and in this case the
interior SUMT is not applicable. In what 1
I(x} = - S - (4-26)
follows, it is assumed that the constraint set D
of NLP has an interior.
and
The sequence of points xw which is to
converge to the minimum point is generated S(r)=r. (4-27)
by minimizing
/(x)+S(r.) /(*) (4-25)

Any pair of functions satisfying properties
No. 1 and No. 2 associated with Eq. 4-25,
without regard to constraints, where S(r.) I(x) however, is suitable. It may be to the
is continuous for x in the interior of D and
designer's advantage to choose another form
S(r() I(x) = + for any x such that g;(x) = 0
for any particular problem. For other suitable
for any 1 Q / < m. It is clear that if one begins
choices of penalty functions, see Ref. l,page
an iterative minimization technique of Chap-
68.
ter 2 at a point in the interior of D, then a
relative minimum point will be found which
must lie in the interior of D. Otherwise, the The algorithm for solving NLP by the
minimizing sequence would have had to climb interior point technique is given in Definition
over a portion of the auxiliary cost surface 4-7.
that is infinitely high and none of the
methods will do this.
Definition 4- 7: The interior point sequen-
In order to obtain the sequence of points tially unconstrained minimization algorithm is
x(!\ the parameter r. is allowed to approach given by the following:
zero. To insure that the sequence xM con-
verges to a relative minimum point, the Step 1. Define the function
functions I(x] and S(r) are required to have
the following properties: U{x.f)=f{x)+Sif) I(x), (4-28)
4-12
AMCP 706-192
where S(r) and I(x) satisfy prop- where x is not an isolated point of
erties No. 1 and No. 2. Choose r D, (4-32)
> 0 and x^0) in the interior of the
constraint setD. { r. } be a strictly decreasing
sequence which converges to zero. (4-33)
Step 2. Beginning atx(0) minimize U(x,rQ)
without regard to constraints to Then for xW> sufficiently near x
obtain x^\ Any of the methods andr. sufficiently small,
of Chapter 2 may be employed for
this purpose. (4-34)
Further,
Step 3. For i = 0,1,2,..., choose r - j > 0
such that r{ < r.. Beginning at .1^ S(rt) I[xu)] = 0 (4-35)
x^!) minimize U(x,r. * t) without
regard to constraints to obtain
x^l+ l\ where i is the iteration in-
dex. (4-36)
Step 4. As r.ôo.if ||xO'+i> _x(0 II and f{xu']\ is monotone decreasing (4-37)

|/[X('+ D]_/[^(0] | are suffi.
ciently small, terminate the process and
and take x1-' + 1> as the solution of
NLP. Otherwise return to Step 3. { / [x(,)]} is monotone increasing. (4-38)
In order to be sure that this algorithm will

lead to a solution of NLP, one would like to
have a result that as rt 0, a solution is For proof of this theorem see Ref. I, page
approached. Such a result is contained in 47.
Theorem 4-10.
It has been noted throughout the previous
Theorem 4-10: In the interior point development that if NLP is convex i.e.,
algorithm just given let: Ax), gl(x), . ..,gm(x) are convex - then
"nice" things happen: One of these "nice"
Ax), gl(x), . ., gm(x) be con- things is given in Theorem 4-11.
tinuous on the constraint set D, (4-29)
Theorem 4-11: If NLP is convex with a
S(r) and I(x) satisfy properties No. unique minimum point x, gx), j- 1, . .., m,
land No. 2, (4-30) are twice continuously differentiable, and if
Eqs. 4-29 through 4-33 hold, then x(0 gen-
The interior of D be nonempty, (4-31) erated by the given algorithm will converge to
the minimum point.
There be a relative minimum point
x inD such that/(x) < /(x) for all x It should be noted that Step 1 of the
^ x in some neighborhood of x, algorithm (Def. 4-7) required a point x(0) in
4-13
AMCP 708-192
the interior of the constraint set but no 4-3.2 EXTERIOR METHOD

method of obtaining such a point was given.
This question will be addressed later in this Unlike the interior method, starting points
paragraph. for the exterior SUMT are not required to be
in the constraint set of NLP. The basic idea in
Example 4-3: Solve the LP the exterior method is to add to the cost
function a penalty function that is positive
/(jCj, x2) = Xj + x2 = minimum for points outside the constraint set and zero
inside the constraint set. This, in effect,
discourages the minimum of the new
gi(xl ,x2) = - x1 < 0
augmented cost function from being too far
from the constraint set if the original cost
g2(x1,x2) = - x2 < 0
function f(x) is "well behaved" outside the
constraint set. It is clear that this approach
using the interior point SUMT.
may not be taken if/(x) is undefined or takes
on negatively infinite values outside the con-
Solution:
straint set. One very appealing aspect of the
exterior method is that it handles equality as
U(x,r) = Xj +x2 r well as inequality constraints without diffi-
x, x2 J culty, so that it can be used on NLP'.
The functions f(x), g1(x), and g2(x) are The penalty function employed for the
exterior method will have the form
convex and by Theorem A-5, Appendix
A, so are 1/gjfxând l/g2(x). Since r > 0,
P(t) E(x) (4-39)
U(x,r) is convex and thus has a unique
minimum. To find it, put w'here P{t)andE(x) are required to satisfy the
conditions:
dU
= 0= 1
bx. (x,)2 1. E(x) = 0 if x is in the constraint set, and
E(x) > 0 if A; is outside the constraint set.
dU
0=1- 2 2. P(t) is continuous and if t2 > tx > 0,
dx 2 (x,)
v
^2 then E(t2) > (/j) > 0. Further, if tj - +
then ._f F(r.) = +
so
,1/2
Probably the most common choice for P(l)
and E(x) is
,1/2
Pit) = t (4-40)
As Y -* 0, jCj -* 0 and x2 -* 0 so the solution and

of Example 4-3 is
,(*)= 2 IgAx) + \g,(x)\ (4-41)

(xltx2) = (0,0).
4-14
AMCP 706-192
and obtain an unconstrained minimum

point of
= 2 [/*,(*)]:
T(x,tl)=f(x)+P(t1)E(x)
where
denoted x(().
E = Ei +E,.
Step4. As r. ->, if II [x(/) - x(M)] II
and '|/[*(0] - flxu-1}] I are
The basic idea for the exterior method was
sufficiently small, terminate the
given by R. Courant in 1943 (Ref. 5). He
process and take JC(,) as the solu-
argued that if
tion of NLP. Otherwise, return to
Step 3.
T(x,t)=f(x)+P(t)E(x) (4-42)
Very much as in the case of the interior

were minimized without regard to constraints
method, Theorem 4-12 guarantees a certain
using tx and t2 with t2 > tv, then since the
measure of success.
augmented cost function is penalized more
when t2 is used than when tx is used, the
Theorem 4-12: In the exterior point al-
minimum point corresponding to t2 should
gorithm let:
be closer to the constraint set and hence,
closer to the minimum point of f(x) on the
f{x), g^ $, ..., gm (x) be continuous
constraint set.
for all x. (4-43)
An explicit algorithm for solving NLP or
Eix) and^(f) satisfy conditions No.
NLP' by this method is given in Definition
1 and No. 2 of Eq. 4-39. (4-44)
4-8.
Definition 4-8: The exterior point sequen- There be a relative minimum point
tially unconstrained minimization algorithm is x in that admissible domain!) such
given by the following: that f{x) < fix) for all x = x in
some neighborhood of x, where x is
Step 1. Make an engineering estimate x(0) not an isolated point of D. (4-45)
of the solution of NLP or NLP'.
Step 2. Choose tx > 0 and beginning at The sequence I tf \ is strictly in-

creasing to + . (4-46)
x(0) find an unconstrained mini-
mum point of
Then for x(0) sufficiently close to x, and t.
sufficiently large,
T(x,tl)=fix)+Pitl)E(x)
denoted x(1). lim

xW=x (4-47)
Step 3. Continue with i = 2, by choosing

lim
ti > t(i and starting from x(,'x) />(/,)[*<] = 0 (4-48)
4-15
AMCP 706-192
lim 4x2 +2?(3x2 - 1) = 0

/[*(/)] =/(*) (4-49)
and

lim
ru, /.] =/(S) (4-50)
2 + 3/
[f[xM]} is monotone decreasing (4-5ij
As
{ E [xw ]) is monotone decreasing. (4-52)
For proof of this theorem, see Ref. l,page 1 , x, +-.

t^-,x2 -*and ,2
57.
3 3
Very much as in the interior method, if the The solution is then

NLP or NLP' is convex, then convergence is 1 2
guaranteed by Theorem 4-13. (x1,x2) = [
3 '3
Theorem 4-13: If NLP or NLP' is convex

4-3.3 MIXED INTERIOR-EXTERIOR
with a unique minimum point, and if Eqs.
METHOD
4-43, 4-44, and 4-46 hold, then regardless of
the estimates x(0) and t{, the sequence x(!)
generated by the algorithm given by Theorem Both the interior and the exterior methods
4-12 will converge to the minimum point. presented in pars. 4-3.1 and 4-3.2 are not
applicable in certain kinds of problems. In
Example 4-4: Solve particular, the interior method cannot be used
if the interior of the constraint set is empty,
fix i ,x2) = x j + 2x1 =
minimum such as in the case with equality constraints.
The exterior method cannot be used if some
h(xl,x2) =*i +x\ 1=0 constraint function is not defined or is ill-
behaved outside the constraint. A combina-
by the exterior point SUMT. tion of the two methods will now be given
which allows the treatment of problems
T(x,t) = x\ +2x22 + t x
( i +x
2 -I)2 which may have both these undesirable fea-
tures and thus could not be treated by either
pure interior or exterior methods.
= 2JC, +2r(x, + x2 -1) = 0
dx1
For convenience, consider NLP'
dT
= 4x2 +2t(x1 + x2 - 1) = 0 minimize fix) (4-53)
ox2
subject to
Subtracting,
gt(x) <, 0,i= 1, ..., m (4-54)
4x2 2x\ =0,orx1 =2x2.
A/(x)=0,/= 1, ...,/> (4-55)

Then
4-16
AMCP 706-192
where the set of all points which satisfy the m denoted x(,).
inequalities Eq. 4-54, has an interior. As
might be expected, the constraints, Eq. 4-54, Step 4. As r. - 0 and tf -+, if \\xO) -
will be dealt with using an interior point x(,'-1)|| and|/[*<'>] -/[^'-^Jlare
penalty function and the constraints, Eq. sufficiently small, terminate the
4-55, will be dealt with using an exterior process and take x(,) as the solu-
point penalty function. tion of NLP'. Otherwise return to
Step 3.
The penalty function used here will be
As might be expected from a study of the
two methods which were combined to form
the mixed method, a convergence result is
where S(r)I(x), P(t), and E(x) satisfy condi- given by Theorem 4-14.
tions No. 1 and No. 2 preceding Eq. 4-26, and
No. 1 and No. 2 of Eq. 4-39. It is understood Theorem 4-14: In the mixed point al-
that I(x) is a function of only the constraint gorithm let:
functions in Eq. 4-54 and E(x) is a function
of only those in Eq. 4-55. A general minimiz- gi(x), ..., gm(x) be continuous on
ing algorithm for NLP is now given in the nonempty interior of their con-
Definition 4-9. straint set and Rx),hl(x),..., h Ax)
be continuous for allx. (4-58)
Definition 4-9: The mixed interior-exterior
sequentially unconstrained minimization al- 5(r), I(x), P(t), and E(x) satisfy
gorithm is given by the following: conditions No. 1 and No. 2 preced-
ing Eq. 4-26, and No. 1 and No. 2
Step 1. Make an engineering estimate x(0) of Eq. 4-39. (4-59)
of the solution of NLP'.
There exist a relative minimum
Step 2. Choose rx > 0 and tx > 0 and point x in the admissible domain D'
obtain an unconstrained minimum of Eqs. 4-54 and 4-55 combined,
of such that f(x) < f (x) for all x = x
V(x,r1,t1)=f(x)+S(.r1)I(x) in some neighborhood of x, where
x is not an isolated point of D'. (4-60)
+ P(ti)E(x), (4-56)
The sequence ( ri J be strictly de-
(1) creasing to 0 and { t(} be strictly
denoted x .
increasing to+. (4-61)
Step 3. Continue with i = 2, ... by choosing
r. < rr_2 and t{ > ti_1 and starting
from x('_1* finding an uncon-
Then for x^ sufficiently close to x, r.
strained minimum point of
sufficiently small, and ti sufficiently large,
V(x,ri,ti)=f(x)+S(.rj)I(x)
lim
+ P(tt) E(x) (4-57)
S(r.)I[x^] =0 (4-62)
4-17
AMCP 706-192
lim
P(t.)E[xw] =0 (4-63)
2A: 2
x2=&-
and
Since g(xlfx2) < 0 is satisfied at all times, x2
lim (
V[je '\ rt, tt] =/(*). (4-64) > 1. Taking the limit as r -* 0, then, x2 -* 1.
As t -> + oo, it is necessary that x, +x2 1

For proof, see Ref. l,page 60. -* 0 or Eq. 4-66 will be violated. Therefore, in
the limit JC! = 0.
Example 4-5: Solve
The minimum point is, therefore, (xltx2) =
f(Xi,x2)= JCj +x2 = minimum (0,1).
g(Xi,x2)= -x2 < 4-3.4 DETERMINATION OF AN INTERI-

OR POINT
h(xl,x2)=xl +x2 - 1=0.
In order to begin the interior point or the
Solution: mixed interior-exterior point algorithm, it is
necessary to have a point x(0) which satisfies
Since g(xlt x2) is unbounded at x2 = 0, it a certain set of inequalities, i.e., a point
must be treated by the interior method and interior to a given constraint set. Let this set
since h(xy,x2) = 0 prevents the constraint set of inequalities be
from having an interior, it must be treated by
the exterior method. From 8j{x) < 0,i= 1,.... m. (4-68)
V(x,r,t)= -Xi +x2 + ff there are other inequalities or equalities

9.nx2 which will be treated by the exterior point
method, they are ignored for now.
+ t{xl +x2 - l)2 (4-65)
Let >>(0) be a first estimate of an interior
av i point of the set defined by Eq. 4-68. Denote
= -1 +2t(x1 +x2 - 1)=0 (4-66)
dx-i the inequalities of Eq. 4-68 which are not
strictly satisfied by
OV r
x
1
d 2 x2SLn7 x2 N= ( i\gt\y] > 0}
+ 2t(Xl +x, - D=0 (4-67)

and the inequalities which are strictly satisfied
by
Subtracting Eq. 4-67 from Eq. 4-66,
K= {/|*,[y<>] < 0}.
2
x2 S.n x2
The object now is to move from j><0) to
points where successively more inequalities
4-18
AMCP 708-192
move from N into K; i.e., so that no con- derivative information to make successive
straint which was previously satisfied is vio- improvements in an estimated solution of
lated, but constraints which were previously NLP. A study of the problem NLP' will be
violated are satisfied. This may be accom- better included in the next paragraph.
plished by minimizing the unconstrained cost
function Geometrically, the method presented here
will first investigate the direction of most
rapid decrease in the cost function/(x). As
U(y,r.)= gt(y) + S(r.) 2 /[g 00)
I i<=N ' ' ieK ' seen in par. 2-4, this is yfT{x). This
direction is then projected onto the tangent
with respect to y, where S(r) and I(x) satisfy hyperplane to the boundary of the constraint
the four constraints No. 1 and No. 2 set at x. A small move in the resulting
preceding Eq. 4-26, and No. 1 and No. 2 of direction will then decrease f{x) and will not
Eq. 4-39 and r. is a strictly decreasing cause excessive violation of constraints. This
sequence. The result is denoted y^: process is repeated as long as f(x) may be
decreased.
As soon as JJ(;) is such that gt [j>(/) ] < 0 for
some i previously mN, that constraint func- Instead of basing the derivation of the
tion is switched to K. In this way, constraint method on a geometric argument, the work
functions fromiV may get to K but those mK will all be done analytically. The reason for
may never fall back to N. Once all the this is twofold. First, geometric ideas in
constraints in N are switched to K, the higher dimensions are not always as clear as
process is stopped and the resulting y^ is in those in two and three dimensions. Second,
the interior of the constraint set of Eq. 4-68. the analytical method used here will be
If the minimum of U(y,r) is found as r. -* 0 employed in deriving algorithms in con-
and there are still constraints in N, then the tinuous problems where geometric concepts
constraint set defined by Eq. 4-68 has no are much more difficult.
interior. In this case, NLP is infeasible (has no
solution) or certain of the constraints of Eq. Extensive use will be made of matrix
4-68 will have to be treated by exterior point calculus notation in this paragraph.
methods.
Recall that for g(x) = x^R"

4-4 STEEPEST DESCENT METHODS FOR
NLP
In Chapter 2 a gradient method is pre-

sented for finding the minimum of an uncon-
dx m X n
strained function. Such a direct method has
properties that make it attractive and worth Further, the symbol
developing for the solution of NLP. It is clear,
however, that due to constraints the gradient
method studied earlier does not apply directly 8x =
to NLP. It is the object of this paragraph to
5x
develop a method which uses only first
4-19
AMCP 706-192
will denote a change in x and a 6 in front of a i.e., the collection of indices of constraint
quantity which depends onx will denote the functions which are equalities at the point x.
first order change in that quantity due to the
change bx in x. For example, for the scalar The object is now to find a direction of
function/(JC), change bx, bxT bx = 1,suchthat bx = kbx for
sufficiently small k > 0 will decrease f(x)
without violating any constraints. The prob-
/to bx.
dx lem is then to find bx such that
Note that this first order change is just the

first term in a Taylor expansion, Ref. 4, page = ^ [*<>] si
ox
84, of f(x), so bf(x) is an accurate approxima-
tion of the change in /(x) only for small 8x. is minimum subject to
The method to be developed here resem-

bles an interior method in the sense of par. S&[*<>] = ' [X<>]5JC < 0,
' ox
4-3. Therefore, the method of generating an
interior point (one which satisfies all the i<=A[x<-)]
constraints) presented in par. 4-3 may be
utilized to obtain a starting point. It is and
assumed now that his has been done, and that
an estimate x(0) of the solution of NLP is bxTbx = 1.
available which satisfies
For further convenience, define the column
vector of constraint functions which are zero
g(xW) < 0. as
4-4.1 THE DIRECTION OF STEEPEST DE- gXx)

SCENT 8M =
I6^[X(']
If the point x^0) is in the interior of the
constraint set, then the gradient method of In this notation the problem is
par. 2-4 applies and the direction in which
x(0) should be altered is
minimize -[x^* ] bx (4-71)
ox
5x (4-69)
dx subject to
k > 0.
[*<>]6f< 0, (4-72)
ox
(0)
In the remaining case, the point x is on
the constraint boundary sog;[x(0)] = 0 for bxTbx = 1 (4-73)
some i. For convenience define the set
It is assumed that at points where several
AM = { /1 gtM = 0 } , (4-70) gAx) = 0, the gradients are linearly indepen-
4-20
AMCP 706-192
dent. This is enough to satisfy the first-order

g,[*(0)]
constraint qualification (Theorem 4-4) for (0)
|[x ] =
the constraint, (Eq. 4-72). Theorem 4-5, there-
fore, may be applied to obtain the necessary
conditions With this new g , Eq. 4-74 yields
_1
%- + X^ f; + 2X.W = 0, (4-74) 6x :
2X o
where the components, X; > 0, of X corre- 1

spond to gj with indices inA [x(0) ]. r/_9| 3F\ 31
dxx \ bx bx I bx j i<"i
Assume for the time being that 8gj = 0 for
all fey4[x(0)]. Then taking the transpose of
Eq. 4-74 and premultiplying by bg/bx yields (4-76)
bx
bg of + 91 ogT r + -. 8 ,-_n Note that XQ > 0 is required since if A [x(0) ]

r -+ r X + 2h0u 5x = 0, is empty, then Eq. 4-76 must reduce to the
dx dx ax ox A:
negative gradient direction.
or since 6# = 0, Putting
3L+9dg^ - = 0
bx bx ax ax
Since the gradients of g,[x(0)] for

fe/l[x(0)] are assumed linearly independent,
" l!"",'lr"<0)i
the coefficient matrix of h is nonsingular and bg (4-77)
bx
bg bgT \ l
bg bf
(4-75)
bx bx bx bx Eq. 4-76becomes
If all components of X are non-neg- 6 (0)

* = -^
2X0 ^^^
dx ]
ative, then the assumption that all
bg. = 0, fev4fx(0)] isvahdandSx which solves Substituting this into Eq. 4-73,
the problems of Eqs. 4-71, 4-72, and 4-73, is
obtained directly from Eqs. 4-74 and 4-75. 1 bf bfT
Z
On the other hand, if X;. < 0 for some (2X)
ô> bx bx
fe^4[x(0)], then this component of g is Solvingfor 1/2X0, 5x becomes
removed from g. Equivalently, /4[x(0)] is
1/2
redefined as i bf
8x lx]PTP [x<>]
{ bx
^^ bx
i[x(0>] = {i\ g.[x(0)] =0andX.> 0} r
XP 9/ [x (0)i (4-78)
bx
and g [x(0 * ] is redefined as
4-21
AMCP 706-192
Eq. 4-78 gives a unit vector 5x in the g2(xl,x2) = x1 - x2 + 2 < 0.

constrained direction of steepest descent at x
= JC(0). The problem is now one of deter- First,
mining just how big a step should be taken,
i.e., *i(2,2) = 0
Sx = kbx (4-79) g2(2,2) = - 2.
where k > 0 must be chosen. Therefore, /l[x(1)] = { l } . As required by

Step 2 of the Algorithm
Before the problem of step size is treated, 3/

^-(2, 2) = [- 1,4]
however, an algorithm may be stated for ox
determination of the direction of steepest 3
descent, namely: -^- (2, 2) = [1, - 13
dx
Step 1. Using the method of par. 4-3,
obtain an estimate of the solution By Step 3,
of NLP, x(0), which is in the
constraint set.
[1,- 11
Step 2. Let / > 0 denote the number of the
present iteration. Compute x [1,-1] > 0
,-[*(y)]> / = 1, -, m, and form the
set A[xW>]. Compute df/dx[x^]
and dg./dx[x^] for is A [x(l)]. soA=k. For Step 4
-1
Step 3. Compute X in Eq. 4-75. For all X.
< 0, delete i from A [x) ] to form
H= P[u 11 J _ r 11 [1,-1]
L-! J
at*].
1/2 1/2
Step 4. Compute P in Eq. 4-77 and 8x in x [1,- 1] =
Eq. 4-78. ff P = 0, then this is the 1/2 1/2
solution of NLP.
Finally,
Example 4-6: Compute the direction of 8x =

steepest descent at the point (2,2) for the 1/2
NLP 1/2 1/2 1/2 l/2~ - 1
[1,-4]
1/2 1/2 1/2 1/2 4
minimize f(xl, x2) = (x2)2 x1

1/2 1/2 3/2
Si ixl,x2) = x1 - x2 < 0 1/2 1/2 3 3/2
4-22
AMCP 706-192
4-4.2 STEP SIZE DETERMINATION X(/ +D = XU) + TJC
There are many techniques presented in the if df/dx[xa + l)]8x <0. The process in the
literature for determining the size of step to algorithm for the direction of steepest descent
be taken in the constrained direction of is repeated from Step 2, and a new step size k
steepest descent. Three of these techniques is computed as above. If, on the other hand,
are presented here. The first technique applies dfldx[xU+1)]5x > 0, then a relative mini-
to a specialized class of problems in which the mum has been bypassed. To locate this
constraint functions are linear. The second relative minimum, do a oneKmensional
and third methods apply to the general search in the direction 8x starting at x^ to
nonlinear problem. obtain x(J+1).
4-4.2.1 ROSEN'S METHOD FOR LINEAR This process may be summarized inRosen's
CONSTRAINTS A Igorithm :
Step 1. Compute
ff the constraint functions are linear, then
once the direction of steepest descent 8x is mm
found, it may be followed without leaving the T bi-Gfx^'
k = IG( 8X =0
constraint boundary until a constraint,-(*) = Gjbx
0, for i not in A [x^ ]. This algorithm,
therefore, can lead to rather long step sizes. Step 2. Compute
Constraints here are restricted to the form

^ [x"> +k8x] .
[*<>> +
T
dx
G{ x - b} < 0, i = 1, ... m
If
where Gf is an n x 1 matrix of constants. The
step size is to be determined so that k is as
small as possible and still [xU) +k8x]3x < 0
dx
G,.r[x(/) + k8x] - b. = 0
put
for some iA [x^ ]. Only those i need to be x(/+i)=x(/)+fc5x
considered for which G? bx > 0, since other-

wise this constraint can never go from strict and go to Step 4.
inequality to equality. The step size k, there-
Step 3. If
fore, is chosen as
7\f
mm [x< +k8x]8x > 0,
dx
j Gfbx > 0 G:TSx
I i4A[x<-')] then find A: so as to minimize
The point x(/+1 * therefore is given by /[x( +k8x] .
4-23
AMCP 706-192
Put x(/+1> x(/) + k8x and go to P=W -l

Step 4.
dx \dx dx I dx
Step4. If If [x(/+1)]-f[x(')]|and||x(>'+1)
xl;)|| are sufficiently small, ter-
(4-82)
minate and take x(;+1) as the
solution of NLP. Otherwise, return
and
to Step 2 of the constrained
steepest-descent algorithm. 1/2
tbf_
8x = ~
b bx
4-4.2.2 FIXED STEP WITH VARIABLE prWP^
ox bx
WEIGHTING
When the auxiliary problem, Eqs. 4-71 (4-83)

through 4-73, was formulated, it would have
been possible to ask for the step size directly Note that if the step size is made large,
rather than just the direction of steepest then considerable progress may be made
descent. In many cases, the behavior of the toward the minimum point. However, since
solution is much more sensitive to changes in the constraint functions are nonlinear, viola-
one variable than another. For stability of tions may occur at any iteration. After a new
calculation then, rather than asking for a point x(/+1) has been computed, the con-
direction 8x satisfying Eq. 4-73, the designer straint functions should be checked. If any
might request a change 8x in x(;) which constraints are violated in excess of fixed
satisfies tolerance, the method of par. 4-3 may be used
to move x(/+1) back into the constraint set.
8xTW8x = e2 (4-80)
The computational method is then de-
where W is a positive definite matrix (usually scribed in Algorithm for Steepest Descent
diagonal) and is a predetermined constant. With Fixed Step Size:
The elements of W are often chosen so that
expected changes in various components of x,
Step 1. Using the method of par. 4-3,
8x, will contribute approximately the same
obtain an interior estimate of the
magnitude to 8xTW8x. The matrix W, there-
solution of NLP, A:'0-1, which is in
fore, is chosen based on the designer's ex-
the constraint set. Further, choose
perience.
the weighting matrix W and step
sizeinEq. 4-80.
The analysis performed in obtaining the
direction of steepest descent follows with
Step 2. Let / denote the number of the
only minor changes. The only changes of present iteration. Computeg([x^]
interest, computationally, are and form the set^4 [x(/) ]. Compute
bf/dxlx^] and bg/bx[x">] for
\dx bx I iÂ [x"> ].
bx bx
(4-81) Step 3. Compute X in Eq. 4-81. For all X.
4-24
AMCP 706-192
< 0, delete i from/4 [x^] to form vided f [ - ( * )kbx ] is decreasing, and con-
A[xV>]. straints do not exceed the given error toler-
ances. When either fails to hold, the resulting
Step 4. Compute P in Eq. 4-82 and Sx in point is called x^+1'.
Eq. 4-83.
If the process is stopped because a con-
Step 5. Compute xu+1) = x(i) + Sx. If any straint is violated in excess of its given
constraints are violated excessively, tolerance, the method of par. 4-3 is used to
use the method of par. 4-3 to get obtain a new point in the constraint set and
from x*'+1' back into the con- the process is repeated until the minimum
straint set. point is located.
Step 6. ff |/[x(/+1)] -/[*<] | andll^'+1) This method should be most effective when
JC(/)|| are sufficiently small, ter- constraint functions are easily evaluated but
minate. Otherwise, return to Step derivatives are costly in computer time. The
2 (possibly with altered 2 and W). basic idea of the method is to prevent an
excessive number of calculations of the con-
strained direction of steepest descent.
4-4.2.3 STEEPEST DESCENT WITH CON-
STRAINT TOLERANCES
4-4.3 A STEEPEST DESCENT METHOD
In par. 4-4.2.2 it was noted that a step may WITH CONSTRAINT ERROR COM-
be made so large as to violate a constraint in PENSATION
excess of an admissible error. The method of
choosing step size presented here will prevent
In previous subparagraphs, steepest descent
this difficulty.
methods were given which at boundary points
generated steps parallel to a constraint bound-
Let reasonable tolerances e(. be assigned to
ary in a direction which decreased the cost
constraint functions gt(x). The object here is
to move in the constrained direction of function as rapidly as possible. Due to non-
steepest descent until some constraint func- linearity of the constraint functions, and the
tion g{(x) violates the tolerance g(.(x) > ep or finite step size, however, some constraints will
until a minimum of invariably be violated. It is the object in this
paragraph to present a new method motivated
f[xW+k8x] by the article (Ref. 6) which automatically
corrects for violation in constraints.
is reached.
Let A[xiJ>] = { i\ gt[x(i>] > 0 } be the
A uniform step size in k may be chosen and indices of constraint functions which are zero
steps taken, checking or are violated. As in the preceding develop-
ment of this paragraph first-order Taylor
gt[xW +kSx] approximations will be used to approximate
functions appearing in NLP. The linearized
for each ieA [x(/) ] and each step in k. The version of NLP at an approximation to the
multiplier k is increased monotonically pro- solution, x^\ is
4-25
AMCP 706-192
3/ ,., x. Assuming Eq. 4-85 is an equality, then the

minimizes/ = [xU)]8x (4-84)
dx necessary conditions reduce to only Eq. 4-88
and Eq. 4-85 as an equality. This system is
subject to linear and can be solved. The multiplier A can
then be determined and a check made to see
(4-85) whether all components are non-negative. If
dx any component is negative, say k, then the
assumption that Eq. 4-85 is an equality is
g,[x(/)]
where g and A g is taken as the violated and it may be concluded that the kth
desired change in g, i.e., the total change component of Eq. 4-85 should have been
taken at the designer's discretion. Usually, so allowed to be a strict inequality. The index k
long as the constraints are not violated exis then deleted from^4.
cessively, the full violation may be corrected;
i.e. Premultiplying Eq. 4-88 by dg/dx and using
Eq. 4-85 yields
A$ = -S,[*W)], ie A [*<]. (4-86)
d ML + 3jL 3|L x + 20Sg = 0.
In order that step size is not excessive, it is ax ax ax ax
required that
It is assumed, as usual, that the gradients of
8xT8x =2 (4-87) all constraint functions which are zero or
violated are linearly independent. Therefore,
where is small. Assuming Eq. 4-85 is an the coefficient matrix of A is nonsingular and
equality, necessary conditions for the
linearized problem are obtained by using 3| 3I7"
X = - ^~
Theorem 4-5. From dx dx
x [all 9/l + 2Ag 1. (4-89)

ax ax 1
+ 8xT8x Substituting Eq. 4-89 into Eq. 4-88 yields
and Theorem 4-5, it is necessary that A. > 0, 1 dgT I dg dgT

and SJC =- /-
20 dx \ dx dx dx dx
df dgT
(4-88)
ox ax
dg- I dg- df Ag
ax Wax -m-
dx
and
(4-90)
Sx-Agf) =0, ieA.
\ax This expression for 8x could now be substi-
tuted into Eq. 4-87 to find 0. To be more
This set of equations is nonlinear in A and general, however, put 1/(20) = 7 > 0 and define
4-26
AMCP 706-192
6x1 = summarized in the Steepest-descent Al-

gorithm With Constraint Error Compensation.
df*_
Step 1. Make an engineering estimate of
dx \dx dx I dx. dx
the solution of NLP'.
(4-91) Step 2. Let the iteration number be j > 0.

Compute gf{x^] and form
and A [xU) ] and g.
dx2 ai(^)1 M (4-92) Step 3. Compute dg/dxlx^^l and

9//9x[x(/)] and choose the de-
sired change Ag ing.
Using this new notation Eq. 4-90 becomes
Step 4. Compute Sx1 and bx2 in Eqs. 4-91
6x = 76X1 + 8x2. (4-93) and 4-92.
This representation of Sx has important prop- Step 5. Choose 7 by a suitable scheme.

erties given by Theorem 4-15. Calculate X in Eq. 4-89. If any
components A. are less than zero
Theorem 4-15: Sx1 and Sx2 of Eqs. 4-91 for gj[x^] which are close to
and 4-92 satisfy the conditions zero, remove these components
T
from g and return to Step 3. If all
1. 6x1 Sx2 =0 X; > 0, proceed.
2. * 5x2 A# Step 6. Form

ax
6x = 75*1 + 8x2
3.^8* =0
9x and
JCU+1) = XW + 8x.
< 0
9x
A method of choosing 7 still has not been Step 7. If l/[x0+1)] -/[*<] I and II Sx ||
given. This parameter is interpreted as a are sufficiently small, terminate
step-size and may be determined by one- the process. Otherwise return to
dimensional search or any other scheme Step 2.
chosen by the designer. In different applica-
tions, different methods have proved effec- 4-5 STEEPEST DESCENT SOLUTION OF
tive. No single scheme has been found that THE FINITE DIMENSIONAL OPTIMAL
seems best. The choice of 7 at this time DESIGN PROBLEM
constitutes an art as much as a science.
In this paragraph a steepest-descent method
The use of this method may now be of solution of the problem OD is developed.
4-27
AMCP 708-192
In many ways, the method of this paragraph used to solve Eq. 4-95 for z- The object now
is similar to the method of par. 4-4. Here, is to determine a change in M0), denoted bb,
however, a distinction is made between design such that
and state variables, and the two types of
variables are treated quite differently. ^D = M> + 8b (4-97)
The problem to be solved here isjust as in will be an "improved" design. The meaning of
par. 4-2: Choose beRk and zeR"to minimize "improved" will be made clear as the analysis
progresses. If the new design variable fc(1)
(4-94) were substituted into Eq. 4-95, this equation
could be solved for the corresponding new
subject to state variable z(1). Since the matrix
9fc/3z[z(\M* ] is nonsingular, the implicit
h(z, b) = 0 OD (4-95) function theorem, Ref. 4, page 181, guaran-
tees that if II 6b || is small, then z(1) -z(0)
and will be small. The change inz is denoted bz so
that
<t>(z, b) < 0 (4-96)
Z(D = z(0) + fe- (4-98)
T
where h(z,b) = [hl(z,b), ..., hn(z,b)] , and
<t>(z,b) = [<t>xiz,b), ..., <t>m(z,b)]T. The state 4-5.1 AN APPROXIMATION OF THE
equations, Eq. 4-95, are put into vector form PROBLEM OD
here in order to take advantage of the The basic idea in the approach to OD
compact matrix calculus notation. presented here is to construct an approxima-
tion of OD which can be solved to obtain an
The steepest descent algorithm for OD is improvement 6b in M0'. The approximate
developed here by first approximating the problem is obtained by making linear approxi-
nonlinear elements of OD by linear expres- mations to nonlinear functions in OD. Linear
sions in the various variables. The difference approximations to the changes in f(z,b),
between the method presented here and that h(z,b), and <pz,b) due to the small changes
of par. 4-4 lies in the treatment of the state bb in () and bz in z^0' are, by Taylor's
variable. In a sense, the state variable is a Formula, Ref. 7, page 56:
nuisance since it does not really describe the
system being designed. The algorithm pre- 5/[z(0)j(0)]^[z()j(0)]fc
sented here is obtained by first eliminating
the state variable from the linearized problem (4-99)
and then solving an explicit problem for an
optimum improvement in the design variable. bb
Very much as in par. 4-4, an engineering

estimate of the optimum design is made. It is
bh[z<0\b^]=~[z<\bW]bz
denoted by (0). Then the state equations, (4-100)
Eq. 4-95, are solved for the corresponding dh
state z(0). Any method of analysis may be + TT [*<>. ><>]*
4-28
AMCP 706-192
and The set A (possibly empty) simply contains

all the indices j of constraints that will be
required to satisfy Eq. 4-103. To make
50 [z<\M>]= [z<\6<>]Sz
maximum use of vector calculus notation,
(4-101) define the column matrix
301 z(0)p 6(0)]
+ _ir 6fc.
bbL J
<t>jU,b)
4>(z,b)- (4-105)
In the development that follows, the argu-
ments [z(0),6(0)] of all functions will be
understood unless otherwise explicitly noted. If the set A is empty, then 0 is defined as
The symbol 6 in front of a quantity simply zero; i.e., all the constraint functions whose
denotes the total differential of that quantity. indices are in A are placed in a column
matrix. In this way, the conditions, Eq.
4-103, may now be written.
SinceMz(0), (0)] = 0 and z(0) + bz is to
satisfy the equation /J[Z(0) +8z,Z>(0)+ bb] =
0, the linearized version of this condition is 30 30
-^5z+ 6bo A0, (4-106)
simply bz bb
where the column matrix A$ is defined as

bz+6 = 0. (4-102)
bz bb
A0y
A$- (4-107)
Eq. 4-102 is viewed as determining bz as a je A
function of bb. It is clear that Eq. 4-102 can
be solved for bz since the matrix bh/bz has
If A is empty, A0 is defined to be zero.
been assumed nonsingular.
The object of the following analysis will be

Inequality constraints, Eq. 4-96, will be
to choose bb so that/[z(0) +6z,6(0) +bb] is
treated in an approximate manner. The
as small as possible. If this nonlinear function
method employed here is to require that if
of bz and bb is replaced by its Taylor
0,[z(o),fr(o)] > 0,then
approximation, the problem is to choose bz
and bb to minimize
Stfy < A0;, (4-103)
where A<fy is the required change in the value 3/ bf

bf = ~bz+ bb. (4-108)
of <t>, due to the changes bz and 6b inz(0) and bz bb
The entire argument up to this point has

For convenience of notation, define the set been based on the fact that II 6b || will be
of indices small. In order to insure that this is the case,
it will be required that
A = ij\ 07.[z(O), 6(0>] > 0> . (4-104) bbTWbb < 2 (4-109)
4-29
AMCP 708-192
for small and W a positive definite matrix. dence on 8z is to solve Eq. 4-111 for 8z as a
The matrix W will be used in particular function of 8b. This, however, requires the
problems to assign weights to the various inversion of the matrix dh/dz. The preceding
components of 8b. This is often necessary approach of applying necessary conditions
when the components of b represent different was scuttled for just this reason, so another
physical quantities that may be of different method of eliminating 8z must be found.
orders of magnitude. Usually W is diagonal. Note that if the terms (df/dz) 8z and
(30/dz) 8z could be found in terms of 8b,
To summarize the approximate problem, then dependence on 8z would be eliminated.
8b and 8z are to be chosen to minimize This is the approach that will be taken here
and also in a later chapter on infinite dimen-
sional problems.
8z + 6b (4-1 10)
dz db
Define the column matrix XJ as the solu-
subject to the constraints tion of
dh , dh
8z+6b=0, (4-111)
dz db (4-1 14)
dz dz
8z + 8b < A< (4-1 12) and the matrix X^ as the solution of

bz db
Zhx$ = df_
and (4-115)
dz dz
8bTW8b< *. (4-113)
Note that X* is a matrix whose columns are
solutions of
4-5.2 SOLUTION OF THE APPROXIMATE
PROBLEM dhT ^ = 3^
(4-1 16)
dz dz
Necessary conditions of Theorem 4-9 could
now be applied directly to the approximate for j&A. Note that Eqs. 4-114 and 4-115
problem, Eqs. 4-1 lOthrough 4-113. If this require the repeated solution of equations
course of action is followed, however, an with the same matrix on the left and different
explicit inverse of dh/dz must be computed. right-hand sides. There are efficient computa-
Since the dimension n of this matrix is often tion codes which can construct all the solu-
quite high, this operation would be very tions simultaneously.
costly. Instead of applying necessary condi-
tions immediately, Eq. 4-111 will be used to To see how these newly defined matrices
eliminate the dependence of the remaining are helpful, compute the transpose of both
functions of the problem on 8z. Necessary sides of Eqs. 4-114 and 4-115 and multiply
conditions may then be easily applied for the through on the right by 8z to obtain
determination of 8b.
(4-1 17)
The obvious method of eliminating depen- dz dz
4-30
AMCP 706-192
and
Xîf^ is not empty
(4-1 18) (4-123)

X* 6z = 6z.
9z 9z * =0, if ^ is empty.
Note that the terms on the right side of these

It should be noted that if the limitation,
equations are exactly the ones which are to be
Eq. 4-121, on the size of II 6b || is not
eliminated from Eqs. 4-110 and 4-112. Fur-
enforced, then the problem, Eqs. 4-119 and
ther, the term (dh/dz) 5z that appears in both
4-120, is just a linear programming problem
left-hand sides can be obtained from Eq.
that may be solved by well-established tech-
4-111 as
niques of linear programming. This technique
is similar to that used inZoutendijk's method
Sz = - 8b. of feasible directions (Ref. 8). For a discus-
dz db
sion of this method the reader is referred to
the literature.
Using this relation, Eqs. 4-117 and 4-1 li
become
The necessary conditions of Theorem 4-9
may now be applied to this reduced problem.
\J 8b = -8z In order to apply the theorem and in later
db dz
calculations, it is required that the matrix *
have full row rank; i.e., that the rows of S*
and (columns of 2*) are linearly independent.
,T bh 90 Further, for use of the theorem it isrequired
- X* 8b = 8z. that the column vector W6b be linearly
db dz
independent of the columns of *. It may be
Substituting these relations into Eqs. 4-110 noted that these assumptions require that
and 4-112, the approximate problem be- there can be no more than k 1 constraint
comes: 8b is to be chosen to minimize functions which are zero or positive at any
iteration. This is true since the matrix 2* has
only k rows and since its columns must be
ZjT 6b (4-1 19)
linearly independent of *, there can be at
most k 1 remaining linearly independent
subject to the constraints
columns. These assumptions are reasonable
from a physical point of view. If 8* had rank
* 6b < A0 (4-120) k then the equation
8bTW8b. 2 (4-121) 8$ 6b = A0"
where
would uniquely determine 8b, and there
would be no optimization problem.
\J (4-122)
db db
The constraints, Eqs. 4-120 and 4-121, will
and be treated differently, so different multiplier
4-31
AMCP 706-192
notation in Theorem 4-9 will be used for 4-121 and the problem is again solved with
each. First, define the reduced number of constraints.
In any method of solution of the approxi-
jT + T T T
H = S. db il ^ 6b+v8b W8b. mate problem, no information is gained if v =
0. Therefore, in the following v > 0 will be
assumed.
Theorem 4-9 requires that
Solving Eq. 4-124 for 8b,
M
=0=ZJ +jiT$ + 2v8bTW (4-124)
dSb
Sb = -W-'i^+îl). (4-127)
2v
where r- > 0 and v > 0
It is now assumed (to be checked later) that
AJ.(fi*'6ft-A0i)=o,i&4 (4-125) Eq. 4-120 is an equality. Substituting for 6b
from Eq. 4-127 into the equality Eq. 4-120,
and
1 r
v{8bT W8b -42) = 0. (4-126) * IV"1 aJ + Z$ji) = A<.
2v
At this point, a computational difficulty Rewriting this equation,

arises, it is difficult to determine 5b from
Eqs. 4-124, 4-125, and 4-126 since it is not
known which of the Constraints. Eqs. 4-120 .it1 TV-' e* = - e*' w~" s.J -2vi4>.
and 4-3 21. will be equalities and which will be
strict inequalities. The question is. "Which of Since ^" is required to have full row rank
the inequalities, Eq. 4-120 or Eq. 4-121. will and W~s is nonsinguiar, the matrix
become strict inequalities?" This can be inter- 1 , if A is empty
preted geometrically as a question of leaving ,4-128)
00
the boundary and going into the interior of
I ^' W~: S*, if A is not empty
the constraint set defined by Eqs. 4-120 and
4-121. It has been the experience with this
;s nonsinguiar. Therefore,
technique that once a constraint, say <j>,(z, b).
becomes zero, then for several small steps Sb
it will remain zero. This observation has ied to M = -M~\ ffi*r W~l 2J + 2v &$). (4-129)
the following computational procedure. First,
all constraints, Eqs. 4-120 and 4-121, will be Note that in the unconstrained case when A is
assumed equalities and Sb is determined using empty, p = O since ^ =0 and A$ = 0.
Eqs. 4-124, 4-120, and 4-121. Then the
algebraic signs of the ,. and v are checked. If Substituting from Eq. 4-129 into Eq. 4-127
they are all non-negative, then this is the
desired solution of the problem. If, on the
Sb = - W'1 (l - Z*M-\ *rW",)fi/
other hand, some ji; or v are negative, then
the constraints corresponding to these multi-
pliers are removed from Eq. 4-120 or Eq. + W~l ^M'\ A<5. (4-130)
4-32
AMCP 706-192
This expression for 6b could now be 3. * 8b2 =A<?

substituted into 8bTW8b = 2 to solve for v.
However, in practice it seems just as realistic 4. -ZjT8bl < 0.
to choose v > 0 in an iterative process as to
choose . Once v > 0 has been chosen A may An obvious check on convergence is to
be evaluated in Eq. 4-129. If any components monitor 8b and the associated reduction in/,
are negative, the corresponding elements of $ 8f. When small 6b occur and essentially no
are removed and 6b is calculated using the improvement is made in f, the process is
new $ matrix. terminated. This test, however, leaves a great
deal to be desired since the choice of v can
yield very small steps 8b and falsely lead the
To aid in interpreting the meaning of terms
designer to believe that the iterative process is
in Eq. 4-130 for 8b, define
converging.
8b1 =W-lll-^M'^^T W-lJiJ (4-131) A much better test is to monitor the

constrained gradient 8bl. Since in an uncon-
and strained problem the gradient must approach
zero at a minimum, one might expect that
8b2 = W'1 2*M:\ A< (4-132) once A$ = 0, the constrained gradient 8b1
should approach zero. The real quantity
In this notation, \\8bl || could then serve as a convergence
check. Theorem 4-17 makes these ideas more
rigorous.
6b = 8bl +8b2. (4-133)
2v
Theorem 4-17: Let/(z, b),h{z,b), and0(z,
b) be continuously differentiable functions. If
The vector 8b1 may be interpreted as a the sequences [b^] and [z(;)] generated by
constrained gradient with 1/2P taken as a step the above algorithm converge to the solution,
size. The matrix which multiplies %J in Eq. z, b of the problem OD and if 0 = 0 for all
4-131 essentially projects the gradient iJ of sufficiently large /', then it is necessary that
the cost function onto a tangent plane to the 6b1 approaches zero as/ approaches00.
constraint set. The term 8b2 serves to drive
any errors in constraint functions to zero.
These interpretations are supported by Theo-
45.3 STEEPEST DESCENTALGORITHM
rem 4-16.
The iterative procedure developed in this

Theorem 4-16: The vectors 8b{ and 8b2 of paragraph may be summarized as follows:
Eqs. 4-131 and 4-132 have the following
properties: Step 1. Make an engineering estimate of
the optimum design variable, b(0).
1. 8b2TW8b1 =0
Step 2. In the /th iteration, / > 0, solve Eq.
7
2. e* 8b =0 1
4-95 for z(/) corresponding to b^'K
4-33
AMCP 708-192
Step 3. Form the vector of constraint the designer should be familiar with the
functions $ in Eq. 4-105 and solve method of obtaining the given algorithm. In
Eqs. 4-114 and 4-115 for XJ and this way, problems with peculiar features
often can be treated by altering the general
algorithm slightly.
Step 4. Compute 27 and * in Eqs. 4-122
and 4-123.
There are two steps in the algorithm of par.
4-5.3 which are not complete. They are Steps
Step 5. ChooseA0 in Eq. 4-107.
8 and 10. In Step 8, a parameter 7 is to be
chosen, but no analytical method of choosing
Step 6. ComputeM^ inEq. 4-108.
it is given. This is the classical difficulty with
steepest-descent methods. They give a direc-
Step 7. Compute 8b' and 8b2 in Eqs.
tion but, unfortunately, they do not allow
4-131 andEq. 4-132.
analytical determination of a step size (l/(2c)
in this case).
Step 8. Choose 7 > 0 and evaluate A in Eq.
4-129. If any components of are
negative, take the corresponding A simple technique for choosing 7 which
elements out of 4> and return to has worked well in a number of problems is
Step 3. given here as a candidate scheme. Since it is
the 8b1 component of 6b which tends to
Step 9. Compute reduce f, the step size determination will be
based on 8b1. The basic idea is to choose 7 in
order to obtain a certain percentage reduction
bV+D=bU) sb1 +5b2.
2v in f. Let A/ (a negative quantity) be the
desired reduction in / for a single iteration
Step 10. If |/[x(/+1)] - /[*<>>] | and (perhaps a 5% to 10% reduction). Since for
II Sb1 || are sufficiently small, ter- A<?~0,
minate. Otherwise, return to Step
2.
8f=-2jT 8b1 (4-134)
2v
4-5.4 USE OF THE COMPUTATIONAL AL-
7 is chosen as
GORITHM
The algorithm presented in par. 4-5.3 will

2jT8bl
certainly not solve all optimization problems. (4-134)
2A/ '
It is presented primarily to guide the designer
to the proper equations developed in par. 4-5
while he is solving a problem. Almost surely a In many problems 7 has been chosen accord-
complicated real-world optimal design prob- ing to Eq. 4-134 on the first iteration and
lem will have some feature which is not held constant throughout the iterative pro-
explicitly contained in the general formula- cess. In other problems convergence prop-
tion OD. In order to utilize a steepest-descent erties were improved if 7 is changed during
philosophy similar to the one developed here, the iterative process.
4-34
AMCP 706-192
REFERENCES
1. A.V. Fiacco and G.P. McCormick, Non- Solution of Problems of Equilibrium and
linear Programming: sequential Uncon- Vibrations", Bull. Amer. Math. Soc, 49,
strained Minimization Techniques, John 1943, pp. 1-23.
Wey & Sons, New York, 1968.
6. 1.0. Melts, "Nonlinear Programming
2. J. Abadie, Ed., Nonlinear Programming, Methods for Optimizing Dynamical Sys-
John Wiley & Sons, New York, 1967. tems in Function Space",Automation and
Remote Control, No. 1, January 1968,pp.
3. D.J. Wilde and C.S. Beightler, Foundations 68-73.
of Optimization, Prentice-Hall, Englewood
Cliffs, New Jersey, 1967. 7. C. Goffman, Calculus of Several Variables,
Harper and Row, New York, 1965.
4. W. Rudin, Principles of Mathematical
Analysis, McGraw-Hill, New York, 1953. 8. G. Zoutendijk, Methods of Feasible Direc-
tions, Elsevier Publishing Company,
5. R. Courant, "Variational Methods for the Amsterdam, 1960.
4-35
AMCP 706-192
CHAPTER 5
FINITE DIMENSIONAL OPTIMAL STRUCTURAL DESIGN
5-1 INTRODUCTION infantry equipment has presented weapon

system designers with a major challenge. In
Throughout this handbook, structural the case of air mobility, minimum equipment
optimization problems are chosen to illustrate weight is a necessary condition for maximum
the use of the design methods developed. helicopter payload. In infantry applications
There are two principal reasons for using equipment weight limits the soldier's fire-
structural problems for illustration. First, power and mobility.
there has been great emphasis on helicopter
and man portability of materiel, which places In seeking lightweight designs, one is
a premium on structural weight. Illustrative of tempted to simply use lightweight materials
Army concern with lightweight structures is and lower safety factors. It becomes apparent,
the theme of rhe 1970 Army Mechanics however, that structural weight reduction can
Conference, "Lightweight Structures" (Ref. significantly degrade system performance. For
33). example, when the weight of an artillery ~iece
is reduced by 30%, dynamic response ! ;e :c
A second key reason for highlighting struc- firing ehe weapon becomes much more evere.
tural optimization ;s its advanced state of In infantry .veapons. "he requirement c-
development, relative to other areas of the reduced weight has Sec designers :o ;<jhtei
mecnanical engineering sciences such as dy- weight oceratins mechanisms "nr nividua
namics of machinery and mechanisms. A few weapons, m lightweight rifles, 'or examp.e.
examples in these related areas are treated in bolts are much lighter than ;n crevicus weac-
this handbook, but development of com- ons and hence are more sensitive to changes in
putational techniques remains to be done. It friction tue to lust ;n< external t>ar~'cles
is felt that if the reacer develops a thorough man vere 'ne mcr SU;;::YC xhts hi &i, M\-<
understanding of structural optimization and ind Ml Rifles, '"here ;!";- manv e'.amrsies.
computational techniques, he will be in a some of which will be discusseu later in this
good position to address problems outside the handbook, of instances m vnicn simply re-
realm of structures. The fact that the math- ducing weight of subsystems causes problems
ematics of structural analysis parallels that of which did not occur in heavier designs.
related mechanical disciplines strengthens this
feeling. The lightweight objective, then, requires
that the developer take an overall system view
A cursory review of Army materiel needs and consider the interaction between weapon
convinces one that light weight is a require- weight and performance of the weapon sys-
ment for a majority of weapon systems being tem. As is true in virtually every design
developed by the Army. The high priority problem in which the limits of technology are
placed on air mobility as well as lightweight approached, the iightweight weapon design
AMCP 706-192
problem must be considered simultaneously result, the minimum weight design problem is
with all other aspects of system design. It is often stated with explicit constraints on
not practical to expect, therefore, that one structural deflection, natural frequency, buck-
will find lightweight structural design special- ling load, and strength. A central part of the
ists operating independently of designers con- design problem, then, is representation of
cerned with other aspects of the weapon weapon system performance requirements
development. A technology is needed which that have an impact on structural design. It is
will allow trade-offs concerning weapon- often required that in doing structural design,
weight to be integrated into the overall dynamic weapon performance must be ana-
weapon design process. lyzed to assure that the proper constraints are
included in the structural design problem.
The objective in this treatment has been to
formulate the minimum weight structural
5-1.2 WEAPON DEVELOPMENT PROB-
design problem with constraints realistically
LEMS ASSOCIATED WITH LIGHT-
reflecting the performance requirements of
WEIGHT REQUIREMENTS
the weapon system. A detailed formulation
and solution of this structural design problem
is presented in this Chapter as well as Chap- To further explore some of the trade-offs
between lightweight and weapon system per-
ters 7 and 9.
formance, several typical problems en-
5-1.1 LIGHTWEIGHT VS STRUCTURAL countered in weapon development will be
PERFORMANCE TRADE-OFFS discussed in this paragraph. The discussion
here is presented to highlight some typical
Normally, achieving a lightweight structure problems, not necessarily to identify all light-
requires a reduction in the amount of material weight structural design problems faced in
used. The consequence is an increase in weapon development.
structural flexibility that causes increased
deflections, decreased natural frequencies,
5-1.2.1 AIRCRAFT ARMAMENT
and decreased buckling loads. Consequently,
failure modes that were not previously criti-
Some of the most critical lightweight struc-
cal, may now become limiting factors in
tural development problems in weaponry
design. For example, in gun supporting struc-
today are in the field of aircraft armament.
tures, increased deflection often reduces This is due to the very high priority placed by
effectiveness of the weapon system by in- the Army on improved air mobility and the
creasing dispersion. There are many ways in need for minimum weight weapon systems to
which such changes in structural performance be carried by helicopters. The combination of
can have an impact on overall system lightweight structural requirements and the
behavior. extreme environment under which the struc-
ture must perform in helicopter application,
The only effective approach to minimum generates a very difficult class of minimum
weight structural design is to formulate the weight structural design problems. The weap-
structural design problem to include con- on developer's interest in structural design for
straints on performance which are dictated by aircraft armament lies primarily in the area of
functional use of the weapon system. As a weapon and weapon support structures.
5-2
AMCP 706-192
The key structural requirement in this gun barrel development, particularly for
application is accurate aiming of an automatic infantry automatic weapon application. With
weapon during firing. Dynamic response of a great deal of emphasis being placed on
the weapon support structure due to inputs lightweight infantry weapons, the barrel is a
from the weapon and from the airframe, natural component in which to seek weight
which vibrates due to aerodynamic inputs, reduction. This is particularly true for rapid
must be considered in the design problem. fire weapons in which heavy barrels have
The most difficult feature of the minimum traditionally been used to alleviate tempera-
weight structural design problem for aircraft ture problems. For a particular barrel con-
weapon applications is the variety of per- figuration, decreased mass tends to cause
formance and failure constraints which must elevated temperatures and stresses. To com-
be treated in the design process. Constraints plicate matters, material strengths are highly
must generally be placed on stresses arising in temperature dependent, making stress con-
the structure, angular deflection of the struc- straints difficult to handle. Another potential
ture at the gun mount, and natural frequency problem, as one tends toward optimality in
of the supporting structure. These constraints barrel design, is the possibility that material
generally appear in the form of inequalities. yield properties will become critically depen-
For example, stress is required to be less than dent upon strain rates and require their
or equal to the allowable stress for the explicit inclusion in the design process.
material. This kind of constraint is very
realistic, from an engineering point of view, Another problem, which can arise in re-
but makes the solution of the optimal design duced weight design, is barrel deflection with
problem rather difficult. resulting reduction in weapon accuracy.
Deflection constraints must, therefore, be
In addition to altering the geometry and considered.
distribution of material in the structure to
obtain desirable performance, it is also possi- The objective of the barrel design problem
ble to induce damping into the structure and is to choose barrel dimensions and structural
to use active feedback control devices to material to minimize barrel weight in the
reduce response. These two methods of re- presence of constraints on dollar cost, tem-
ducing dynamic response will require addi- perature, stress, and perhaps strain rate. The
tional weight on board the helicopter. There optimal design problem must then include
is a trade-off between design of the structure equations of state of stress and temperature as
and design of other means of obtaining a function of time, both depending on the
improved weapon system performance. These barrel design features.
trade-offs, then, require that we treat the
aircraft weapon design problem as a system
problem, explicitly accounting for the inter- 5-1.2.3 TOWED ARTILLERY
action between structoral behavior, damping,
and active feedback control. The principal objective in towed artillery
design is to provide support for a large-caliber
5-1.2.2 GUN BARREL DESIGN tube that will, upon firing, transmit momen-
tum to the earth without doing damage to the
A second area in which lightweight struc- support structure and without undue dynamic
tural design is of critical importance is that of response. The fundamentals of the design
5-3
AMCP 706-192
problem then lie in the field of mechanics some of the more complex lightweight struc-
and, in particular, are highly dependent upon tural design problems faced in weapon devel-
the weight distribution within the artillery opment. They are simplifications of the real
piece. problems but are difficult enough to illustrate
the need for research in development of
In traditional artillery design, the support design methods. In view of the current em-
structure is flexible but has been quite heavy phasis within the Army on air mobility and
and stiff in the past so that the flexibility of lightweight systems, new design methods are
the structure was a higher order effect. Also, required which are capable of solving these
heavier carriages reduced the severity of the and many more lightweight design problems.
dynamic response problem due to their higher
mass. Recent developments, such as the 5-1.3 PLAN FOR TECHNIQUE DEVELOP-
Ml02, 105 mm Howitzer, have resulted in a MENT
weapon that weighs approximately 3200 lb,
as compared to the older Ml01 which The remainder of this chapter will be
weighed 4500 lb. As a result of the reduced devoted to formulation and application of a
weight, problems have arisen in providing a method of structural optimization. As noted
firm support for the artillery piece on soil. at the beginning of par. 5-1, an in-depth
More recent design efforts, including the treatment of lightweight structural design
XM164, 105 mm Howitzer, and XM198, 155 provides insight into application of the gen-
mm Howitzer, have resulted in weapons eral methods of Chapter 4.
which are considerably iighter than their
predecessors. As a result of the reduced For a comprehensive review of structural
:tructurai weight of the weapon, dynamic Optimization through 1967, the reader is
.response in both of these weapons became referred to Refs. i ana 2. Several of the major
critical and had to be treated as a key design classes of optimal structural design problems
constraint in development of the recoil mech- are outlined in Ref. 2. Some of the key papers
anism. For a discussion of a particular prob- which have appeared in the literature since
iem, :he reader Is referred to the artillery i 967 are listed in Refs. 3 through i 8.
design example of par. 8-5.
Although these are primarily mechanical 5-2 ELEMENTS OF THE ELASTC STRUC-
system design problems, they have arisen due TURAL DESIGN PROBLEM
to the lightweight design criterion. For this
reason, when one considers lightweight struc- A class of optimal structural design prob-
tural design he must be willing to fit his lems in which the structure must remain
structural design problem into a larger system elastic is treated in this paragraph. The objec-
design program and clearly understand the tive of this paragraph is to show how the
interfaces arising between structural and other optimization methods of Chapter 4 can be
system performance characteristics. used to solve realistic optimal design prob-
lems. No attempt is made here to present a
5-1.2.4 OTHER WEAPON PROBLEMS complete theory of optimal structural design
that is capable of solving all problems.
The example problems cited in par. 5-1.2
are meant to illustrate the essential features of The reader should note that, even for the
5-d
AMCP 708-192
class of problems considered here, it is not ment, buckling loads, and natural frequency.
possible to blindly apply the techniques of The collection of all variables required to
Chapter 4. A certain amount of knowledge of describe this response due to applied load will
structural analysis is required before a reason- be denoted by the state variable vector z. The
able statement of the design problem and a manner in which z is related to the design
method of solution can be obtained. Even variables and applied loads will be discussed in
more important, the structural designer needs some detail later in this paragraph.
to have a thorough knowledge of the opti-
mization methods of Chapter 4 and their The cost of the structure must now be
development. As will be seen, in some cases it described as a real valued function of the
is required that parts of the design problem be design and behavior variables. In keeping with
interpreted in light of the derivation of the the preceding notation this function will be
optimization method. In this way the method denoted as
may be adapted for solution of a particular
class of problems. / = /(*, b, r> (5-1)
5-2.1 THE OPTIMALITY CRITERION where f is one or more eigenvalues such as

buckling load and natural frequency. Before a
The meaning of optimal or best must be meaningful discussion of treatment of the
clearly established in each problem of inter- structural design problem may be given, the
est. In order to have a problem which may be behavior of the structure due to loads and
solved by the previously developed optimiza- constraints on that behavior must be ana-
tion methods, a real valued measure of the lyzed.
cost of the structure (value of the structure)
must be chosen. Such measures as dollar cost
of the structure, weight of the structure, or 5-2.2 STRESS AND DISPLACEMENT DUE
dynamic response of the structure may be TO STATIC LOADING
chosen.
It is assumed for now that the structure of
Along with the choice of a cost function, interest is either made up of a finite number
the parameters, or design variables, that repre- of distinct interconnected members or that
sent all design alternatives must be chosen. large continuous members in the structure
These parameters will often be dimensions of have been approximated by a finite number
structural members, area of member cross of elements as in finite element techniques.
sections, or locations of joints in the struc- Further, it is assumed that the entire structure
ture. In keeping with the notation of the is described by a vector design variable b.
preceding chapter, these design variables will
be denoted as b., i = 1,..., m. For con- Let stresses at critical points in the struc-
venience of notation, these variables will be ture be denoted zx, ..., zr and displacements
put in the vector form b = [b1 . . ., bm ]T. required for the analysis and design of the
structure be denoted zr+1 , . . ., z . The
Invariably, the behavior of the structure behavior of the structure due to any given
under load will have to be considered in the load may then be specified by the vector state
design problem. The response of the structure variable z = [zy ,. .., zn ]T. Attention will
may include quantities such as stress, displace- be restricted here and in the remainder of this
5-5
AMCP 708-192
chapter to structures which obey Hook's law, These constraints can generally be written in
i.e., stress and displacement are determined the form
by linear equations. It is clear, however, that
the design variables play a large part in the 0(z, b, r) < 0 (5-4)
response of the structure to loads. The depen-
dence on the design variables enters these where Mz.b.S) = [*x(*.*,), .,4>,{zM)\T
linear equations through the coefficients. The The inequality constraints, Eq. 5-4, are
equations for z will be denoted required to be satisfied for each of the states
z' due to different applied loads P'.
A(b)z=P (5-2)
It is clear that the Eqs. 5-2 and constraints,
where P is a matrix of loads and Eqs. 5-4, fit into the formulation of the finite
dimensional optimal design problem of par.
A(b)=[alf(b)]nx (5-3) 4-5. Treatment of the restrictions imposed by
Eq. 5-4, however, must be delayed until
is a matrix whose elements depend on the similar restrictions due to other behavior
design variables. constraints are accounted for. The entire
problem will be treated in par. 5-3.
In this formulation of the problem, z and P
may be generalized state and load variables.
Eq. 5-2 may be obtained through direct 5-2.3 NATURAL FREQUENCY AND
application of equilibrium and compatibility BUCKLING
conditions or through application of varia-
tional criterion for equilibrium. In today's As pointed out in par. 5-1, the desire to
structural analysis technology, Eqs. 5-2 are obtain lightweight structures has led to
very likely to be obtained by finite element resonance problems and, likewise, buckling
methods (Refs. 19, 20). If the structural problems. It is necessary, then, that a mean-
analysis problem is properly formulated, the ingful optimal design methodology be capable
matrix A(b) is nonsingular and z may be of enforcing constraints on eigenvalues
obtained by solving Eq. 5-2. It is assumed that associated with the system response. The sort
the elements of the matrix A(b) are differ- of constraint considered here is
entiable with respect to b.
f> fo (5-5)
In most real-world structural design prob-
lems the structure is required to carry a whole where is buckling load or natural frequency
family of loads that occur at different times and f0 is a lower bound on that eigenvalue.
in the life of the structure. The treatment More general restrictions than those of Eq.
here will be limited to a finite number of 5-5 are included in the general constraint, Eq.
loads, denoted P', i = 1,..., s. Associated 5-4.
with each load is a state z' determined by Eq.
5-2. Much as in Eq. 5-2, the equations of
vibration or buckling may be written in the
Constraints on behavior of the structure form
due to each of the applied loads P may
include bounds on stresses and displacements. K(b)y = tM(b)y (5-6)
5-6
AMCP 706-192
where y = fri ,.. .,yn ]T is an eigenvector of this quotient is the smallest eigenvalue. A
which plays the role of a state variable, direct method of minimizing the Rayleigh
quotient is discussed in par. 2-8.
w) = [k..mnxn (5-7)
5-2.4 METHOD OF SOLUTION

is generally symmetric positive definite ma-
trix, and
In the preceding formulation of the opti-
mal design problem, the cost functions and
M(b) = [m{l{b)] x (5-8)
constraints associated with stress and displace-
ment can be put into the format of the
is generally also a symmetric positive definite
problem treated in par. 4-5. The constraints
matrix. Eq. 5-6 is often obtained through
associated with natural frequency and buck-
a finite element formulation of the structural
ling, however, are not of exactly the same
analysis problem (Refs. 19,20).
form. One difficulty is that the coefficient
matrix for the eigenvector y, K(b) M(b),
There are many methods for obtaining the must be singular at the solution. This clearly
eigenvalue and associated eigenvector in Eq. contradicts the assumption in par. 4-5 that
5-6. The first method requires that the inverse the state equations uniquely determine the
of K(b) be computed. Multiplying through state variable.
Eq. 5-6 by K~ 1(b),
This situation is a direct result of Murphy's
K-\b)M(b)y=- y (5-9) law "if anything can go wrong it will".
Actually, it is not realistic to expect that a
This problem is now in standard form and the mathematical formulation of the kind pre-
largest eigenvalue of K"l(b)M(b) is sought. sented in par. 4-5 should contain all real-
The power method of obtaining this eigen- world design problems. Already, an important
value is quite effective (Ref. 21). It is par- problem has been encountered which requires
ticularly effective when a good estimate of an understanding of the development of par.
the eigenvector is available. In the iterative 4-5 in order to include the new problem in
design technique, a good estimate is generally the steepest-descent algorithm. The eigenvalue
available from the previous iteration. The problem, fortunately, can be treated very
power method is, therefore, well suited for nicely by the steepest-descent technique.
use in iterative techniques. This method does Development of the method will be done in
have the severe disadvantage that K~l (b) par. 5-3.
must be computed for each new b.
5-3 STEEPEST DESCENT PROGRAMMING
A different method of finding the smallest FOR OPTIMAL STRUCTURAL DE-
eigenvalue and associated eigenvector of Eq. SIGN
5-6 without computing K~ l (b) is based on
the Rayleigh quotient as discussed in par. 2-8 In order to obtain a steepest-descent al-
and Ref. 23. The smallest eigenvalue of Eq. gorithm for the design problem with con-
5-6 is obtained by choosing a normalized straints on eigenvalues, it is necessary to go
vector y which minimizes the quotient back into the derivation of the algorithm of
yTK(b)y/[yTM(b)y]. The minimum value par. 4-5. The major effort required here will
5-7
AMCP 706-192
be the linearization of the structural design

problem to obtain an approximate problem of (5-15)
dz
the kind described by Eqs. 4-1 19 through
4-121.
for each 0, > 0. All this follows since h{z,b) in
the general formulation is simply A(b)z in the
5-3.1 LINEARIZED COST AND CON-
present problem so
STRAINT FUNCTIONS
dh d
Since the cost and constraint functions [A(b)z]
Tb ' Yb
depend on z, b, and f, the first order
perturbation in these functions due to small
and
changes 8z, 8b, and Sf in z, b, and f is
dh
Sf=5LSz + KSb+v6i (510) 7z=A
oz ob ob
with A symmetric so A T =A.

and
Thus, the explicit dependence of Eqs. 5-10
and 5-11 on 5z can be easily eliminated. It
50 = 8z + 8b + 5f . (5-11)
dz db 9? " remains to determine 6f in terms of 6b. This
problem has been addressed in a completely
The problem of writing the perturbed cost rigorous manner by Kato (Ref. 23). Explicit
and constraint functions explicitly in terms of expressions are given there under quite restric-
6b now reduces to obtaining explicit ex- tive hypotheses. A formal development will
pressions for the terms involving 8z and 5f. be given here which obtains the same result.
From Eqs. 4-117 and 4-1 18, and the It is assumed that the eigenvalues and
perturbed state equation we obtain, just as eigenvectors of
Eq. 4-1 19,
K(b)y = W(b)y (5-16)
9/ ,T 9
5z = -X/ --[A(b)z]8b (5-12)
dz db depend continuously onb and further, that to
first order, the following perturbation equa-
and tion is accurate
30
6z = -X* [A(b)z]8b (5-13) K(b)8y+ -i [K(b)y]8b = 8$M(b)y
z ob ob
(5-17)
J
where \ and X* are determined by + [M(b)y]8b + W(b)8y
ob
A\J =~ (5-14)
dz where y and f satisfy Eq. 5-16.
and If K(b) or M(b) is not symmetric, it is
5-8
AMCP 706-192
necessary to solve the adjoint eigenvalue tion that Eq. 5-17 holds is highly questionable
problem from an operator theoretic point of view.
Under reasonable assumptions on the finite
K1 (b)y = W1 Wy (5-18) dimensional eigenvalue problem treated here
however, Eq. 5-19 is shown to hold (Ref. 23);
that has the same eigenvalue f asEq. 5-16 but i.e., even though the justification given here is
a different eigenvector y. Rearranging and not mathematically rigorous, the result, Eq.
premultiplying by y T this is 5-19, holds for a large class of problems.
Defining
yT[K(b)-^M(b))8y+yT [K{b)y]8b
ob
-yTt^r [M(b)y]8b
ob
"=- >^''^
= yTS$M(b)y + | ~/lyTM(b)y] \
lid )r_
Since the first term is a scalar, x | [K(b)y] y
\ \ ob )
yT \Kyb) - $M(b)} 8y = 8yT [KT(b)

? j [M(b)y\\ yj (5-20)
- CMT fb).
and
Sincey is an eigenvector of Eq. 5-! 8,
-4- -\[A{b)z]} ^
db ] ob j
[KT',b) -mT(b)]y=
and this aquation becomes ' +i \[K(b)y]\ y

X \\ ob \
i
y< [K(b)y] -Sy> [M(b)y] \ -5
in ob f : i 3 1 7 _ 'l
' -?{W(b)y]\ y I
S 3
= &Sy M(b)y.
Assuming y T
M{b)y = 0 which will generally \x^l/[y^M(b)y]
ob I
be the case. or
, 0, if $ is empty.
T
6? = \ y [K(b)y]
\ 6b Eqs. 5-10 and 5-11 become
(5-19)
- KyT [M(b)y] j ob [yTM(b)y] 8J = J 6b (5-22)
and
Derivation of the perturbation formula, Eq.
5-19, has been strictly formal. The assump- 50 = *r5. (5-23)
5-9
AMCP 706-192
The linearized problem is now to minimize -M~* (*T W1 & + 2vA$). If any
bJ, Eq. 5-22, subject to constraints component of is negative, re-
move the corresponding row from
T $ and return to Step 3.
$ bb < A$
Step 8. Compute
where A0 is the desired correction in con-
straint error and
8b1 = W1 (/ - 9?M[\ S?' W1 nJ
T 2
bb Wbb< %
and
where W is positive definite and % is small.
This is precisely the same problem in par. 4-5, bb2 = W-l$M-];A A<
Eqs. 4-119 through 4-121, so the theoretical
results and steepest-descent algorithm of that and form
paragraph apply with proper interpretation.
bb^-^-bb1 +bb2.
5-3.2 STEEPEST DESCENT ALGORITHM 2v
FOR OPTIMAL STRUCTURAL DE-
SIGN Step 9. Compute
Step 1. Make an engineering estimate of fcO+i) =bU) +5b.
the optimum design variable 6(0).

Step 10. If all constraints are satisfied and
Step 2. For/ = 0, 1, ..., solve Eq. 5-2 for bb1 is sufficiently small, termi-
zu\ Eq. 5-6 for/'' and f0), and nate. Otherwise, return to Step 2
Eq. 5-9 for y (if k(b) or M(b) is and continue the process.
not symmetric) with b = b^.
All the properties of bb1 and bb2 derived
Step 3. Form 0 as in Eq. 4-105. Solve Eqs. in par. 4-5.2 hold in this case. Further, the
5-14 and 5-15 for \J and X*. discussion of that paragraph regarding such
things as choosing v also hold. The reader
Step 4. Compute 7 and 5 in Eqs. 5-20 should refer to that paragraph for detailed
and 5-21. discussions.
Step 5. Choose A$ as the desired reduction 5-3.3 COMPUTATIONAL

in constraint error. CONSIDERATIONS
Step 6. Compute Several comments on the computational art

used in solution of these problems are in
1, if $ is empty order. First, if a feasible design was chosen
M^ = initially, large steps could be taken until one
S* W'1^, elsewhere. or more constraints were violated, at which
time the step size was reduced. Second, it was
Step 7. Choose 7 > 0 and evaluate noted that as the optimum was approached,
5-10
AMCP 706-192
oscillation occurred. By monitoring the dot number of constraints with positive compo-
product, 8b^ * 6_1), oscillations were nents of M was always less than or equal to the
sensed when negative values of the dot prod- number of design variables of the problem.
uct occurred. Thus, step size, l/(2i>), was This procedure of adjusting the constraint set
divided by two when negative values of the has worked very well and has minimized the
dot product occurred on two successive itera- possibility of divergence of the algorithm.
tions. Finally, the most effective method of
adjusting step size was to monitor successive The method presented is relatively auto-
reductions in cost function after feasibility matic in the sense that, for the computer
had occurred. Once insignificant reductions program developed, the input data given is the
occurred, the step size was reduced to obtain only pertinent design information required
finer convergence. for solution of the problem. All the necessary
matrices and their derivatives are automati-
The Power method used to compute the cally generated in the computer. Any person
smallest eigenvalue performs quite well. At with a reasonable knowledge of FORTRAN
every iteration, the starting value for the language should be able to handle the pro-
eigenvector is taken from the previous itera- gramming without any difficulty. The method
tion which manifested a very rapid rate of is developed to meet simultaneously displace-
convergence. An accuracy of 0.1% in each ment, strength, and frequency requirements
component of the eigenvector was used to on the structure. The technique, therefore,
compute the new eigenvector. The stiffness can be made user oriented.
matrix for the structure was inverted by the
Gauss-Jordan elimination procedure.
5-4 OPTIMIZATION OF SPECIAL
PURPOSE STRUCTURES
Another comment that is appropriate here
concerns the sign check on the Lagrange
Several special purpose structural optimiza-
multiplier vector p, called for in Step 7 of the
tion problems are solved in this paragraph on
computational algorithm (par. 5-3.2). The
an ad-hoc basis to illustrate the method of
algebraic sign of each component of the
par. 5-3. Subsequent paragraphs will treat
Langrange multiple vector p was checked at
large scale problems in a more unified man-
each iteration. If some of the components
ner.
were negative, then the matrix $ and the
vector A$ were adjusted accordingly. This
procedure is particularly useful whenever 5-4.1 A MINIMUM WEIGHT COLUMN
there were redundant constraint violations. In
some cases, the number of constraints vio- A column is to be constructed by making
lated is more than the number of design its cross section piecewise uniform as shown
variables of the problem, yielding a singular in Fig. 5-1. The objective of the design
matrix coefficient of p. In such cases numeri- problem is to choose the element areas so that
cal noise yielded a solution such that some of the column will support a vertical load P0
the components of the vector p were always without buckling or yielding under compres-
negative, indicating that the corresponding sive load. For the purpose of the present
constraints would be strictly satisfied in the problem the geometric shape of each column
next iteration. In numerical examples, the element is fixed and symmetric about two
5-11
AMCP 706-192
where er maxv is the allowable stress of the

column material in compression.
In order to apply the optimization method

of par. 5-3, the equations which determine
the buckling load in terms of b = [b,,..., bk ]T
must be obtained. Using the generalized co-
ordinates shown in Fig. 5-2 and Eq. B-4,
Appendix B, the potential energy of the ith
element under the buckling load P is
fimnnniiii J7TTT PEt=-if KKbW -Pu'T'D'ibW (5-28)

Figure 5-1. Column
orthogonal axes so that the cross-sectional where

area bt of the ith element completely specifies
the element. With this assumption, if a is the
second moment of the cross section of unit
area, then is as shown in Fig. 5-2. The matrices K'{b)
and D'{b) are from Eqs. B-4 and B-8, Appen-
/. = a b:2 (5-24) dix B
In this problem, weight of the column is to

be minimized so that the cost function is
u\jZ\ui
J = yXibjLi (5-25)
where y is material density and Li is the

length of the ith element of the column.
There are two basic constraints that must

be satisfied in this design problem. First, to
insure that the buckling load P is not less than Figure 5-2. Column Element
the applied load ô > it is required that
,+i^Po-P<0. (5-26)
12 ~6L{ - 12 -6Lt
Second, in order to insure that the column -6L. 4i(2 6Lt 2L)
Eabf
material does not yield under the applied load (5-29)
'(*)=-
P0, it is necessary that - 12 6Lt 12 2Lt
-6L, 2L; 2Li AL]

(5-27)
i= l,...,k and
5-12
AMCP 708-192
T
3 1 3 1
fifi -dli-= [o,...,o,-p0/b;,o,...,oi,
5L. ~2" 20 db
1 i i= 1,..., k (5-33)
20 15 20~ 60
D\b)=2 (5-30) since 0. does not depend on P, i = 1,..., k and
3 1 3 1
20 si; 20 { 9
[K(b)y]\ y
db
i 1
20 60 "20 15
p (yTDy),
\7b[Dib)y]\Tyi
Summing the total potential energies of all
the elements from Eq. 5-28 and defining a if^+1 > 0
new variable
[0],if^+1 < 0. (5-34)
y= [yi.yi, -,âtl7-
The computations required in Eq. 5-34 are
= [u\, u\,u\,u\, ..., uk, uk]T messy but they can be programmed for
automatic computation.
the total potential energy PE of the column
may be written All expressions required for direct applica-
tion of the steepest descent algorithm of par.
PE=-yT K(b)y - P-yT D(b)y 5-3 are now available. Numerical results and
profiles of optimum columns are shown in
Tables 5-1 and 5-2, and Fig. 5-3. Numerical
where K{b) and .0(6) are made up of elements data for the example problems are E = 3.0 x
of K'(b) and D'(b) and are symmetric. Apply- 107 psi, a = 0.079577, amax = 20,000 psi,
ing the theorem of minimum total potential and L = 10.0 in. Computation in each case
energy given in Appendix B, the governing required approximately 0.1 sec per iteration
equations of buckling are
K(b)y=PD(b)y. (5-3 1) TABLE 5-1

COMPARISON OF UNIFORM
AND OPTIMAL COLUMNS
Eq. 5-31 is now in the form of Eq. 5-6, with Volume of Volume of
proper interpretation of notation. Optimal Uniform* Material
P, lb Column, in? Column, in? Savings, %
In order to implement the computational
500 0.806 0.923 12.7
algorithm of par. 5-3, the following vectors 1000 1.143 1.300 12.1
are required: 1500 1.411 1.600 11.8
2000 1.640 1.840 10.9
4000 2.412 2.600 7.2
,T dJ
= = t7Ll,7L2
3 '""7i*] (5 32)
" 'Lightest uniform column which will support load P.
5-13
AMCP 706-192
TABLE 5-2
CROSS-SECTIONAL AREAS OF OPTIMUM COLUMNS
Element
No. i P = 500 lb P= 1000 1b P= 1500 lb P= 2000 1b P = 4000 lb
1 0.1070 0.1499 0.1833 0.2106 0.2947

2 0.1055 0.1480 0.1809 0.2076 0.2875
3 0.1035 0.1442 0.1763 0.2023 0.2789
4 0.1000 0.1383 0.1691 0.1942 0.2683
5 0.0960 0.1303 0.1593 0.1831 0.2505
6 0.0831 0.1 198 0.1464 0.1683 0.2302
7 0.0738 0.1064 0.1299 0.1493 0.2020
8 0.0623 0.0892 0.1088 0.1250 0.2000
9 0.0477 0.0668 0.0812 0.1000 0.2000
10 0.0267 0.0500 0.0750 0.1000 0.2000
together uniform sections of beams as shown

in Fig. 5-4. The objective is to choose the
sections so that the beam is as light in weight
as possible and still satisfies constraints on
strength and natural frequency. Due to dy-
namic inputs to the beam, it is required that
the natural frequency of the beam be above a
> l\ >} > >

I given limit co0 to prevent oscillation prob-
lems.
P= 500 ]* = 100 D
As in the preceding column design prob-

lem, the cross-sectional geometry is chosen,
but all dimensions of the cross section may be
varied in the same proportion. Thus, if bi
denotes the area of the ith section, then the
second moment of the cross-sectional area is
ab. (5-35)
i I'I 1111 ri
P = 2000 P = 4000 P = 6974

where a is a constant of proportionality
Figure 5-3. Profiles of Optimal Columns' depending on the geometry of the cross
section. The problem at hand is to minimize
and 15 iterations to converge on an IBM

360-65.
5-4.2 A MINIMUM WEIGHT VIBRATING

BEAM
A beam is to be designed by piecing Figure 5-4. Stepped Beam
5-14
AMCP 706-192
weight, so the cost function is 12/. eify -12/, -6/./.
-6/Z, 4/.L, 6/./. 2/x;

J = p 2 biLi (5-36)
i=i K\b) =- 12/, 6/,( 12/, 6/,L,
where p is material density and L(. is the -6/^ 2/A2 6/,i, 4/,4
length of the ith section.
(5-40)
As a strength constraint, it is required that

Forming a single vector y that contains all
0. =b0 -bj < 0, i= 1,-. * (5-37) displacements and rotations for the beam, the
total kinetic and potential energies are
where b0 > 0 is chosen so that the beam will yTM(b)y/2 and yTK(b)y/2, respectively.
support a lateral load. The constraint on Lagrange's equations, Eq. B-17, are then
natural frequency can be written as
M(b)y(t) + K{b)y{t) = 0 (5-41)
0fc + 1 ="o - OJ < 0. (5-38)
By neglecting compression of the beam, For harmonic motion of the structure, y(t) =
deformation of a typical element is shown in y sin ut, where y isjust a constant vector, / is
Fig. 5-5. By Appendix B, the kinetic energy time, and w is natural frequency. Substituting
into Eq. 5-41 and defining f = co2, the
eigenvalue equation is
77 ').
K(b)y = W(b)y (5-42)
Figure 5- 5. Typical Element

The problem of minimizing J of Eq. 5-36,
subject to the constraints of Eqs. 5-37 and
5-38, and with state Eq. 5-42, is in the form
of an element is lM'(b)'/2, where, from Eq.
B-6 of the general problem of par. 5-3. The
steepest-descent computational algorithm of
756 -22L, 54 13/,,
that paragraph can be applied directly to this
problem.
22Li 4L'-13A, -3Lf
M\b) -
As a numerical example, the beam problem
54 -13L. 756 227,..
420 was solved with the data E = 3 x 107 psi, L =
10 in., a = 1.0, and p = 0.00208 lb-sec2 /in3.
137,, -3L) 227,, 4L*
The computational algorithm required about
(5-39) 0.6 sec per iteration on an IBM 360-65 system
and approximately 15 iterations to converge.
Results for a range of natural frequencies are
Likewise, the potential energy of the ith given in Table 5-3 and the profile of an
element is u'TK'(b)u'/2, where, from Eq. B-4 optimum beam is shown in Fig. 5-6.
5-15
AMCP 708-192
TABLE 5-3
COMPARISON OF OPTIMUM BEAMS
Volume of Optimum
Frequency Uniform Volume, Material
rad/sec Beam*, in? in? Savings, %
3600 0.935 0.897 4.06

4000 1.155 1.062 8.05
4400 1.397 1.259 9.74
4800 1.663 1.481 10.94
5200 1.951 1.727 11.48
5600 2.263 1.993 11.93
6000 2.598 2.283 12.12
10000 7.217 6.330 12.29
't
'Uniform beam of lowest volume having required /77777777 ////////
natural frequency.
Figure 5-7. Portal Frame
ment stiffness matrix from Appendix B is
12 -61, -12 ~6Lt

Figure 5-6. Profile of Optimum Beam
Eabf 6L. AL) 6L, 1L\
K(bt)
Lf 12 6L, 12 6L,.
5-4.3 A MINIMUM WEIGHT PORTAL
61. 2L) 6Lt 4L?
FRAME WITH A NATURAL FRE-
QUENCY CONSTRAINT
(5-43)
A portal frame as shown in Fig. 5-7 is to be where L. is the length of the ith member and
proportioned so that it weighs as little as the element deformation variables are shown
possible and has its fundamental frequency at in Fig. 5-8. The potential energy PE of the ith
least as large as a specified frequency coQ. element is
Each member of the planar frame is formed
from several uniform sections whose areas are
to be determined as design variables. As in the t
preceding problems, the cross-sectional geom-
etry is taken as fixed and all dimensions of
cross sections varied proportionally. The
second moment of the cross-sectional area y ')
about a centroidal axis is/; = ubf where b. is
the cross-sectional area of the ith element.
Neglecting strain energy due to axial defor- "l JU2

mation of the horizontal member, the ele- Figure 5-8. Typical Elements
5-16
AMCP 706-192
where y is the vector of all displacements and

PE, =- I/KÛ* (5-44)
rotation, and f = w2. The matrices A'(ft) and
Af(ft) are formed from element stiffness and
where ' = [Wj, 2> w
3> "4!' mass matrices as outlined in Appendix B.
Eq. 5-47 is in exactly the form of Eq. 5-16

Likewise, from Appendix B the kinetic and the matrix for this problem is simply
energy KE of the typical element is weight of the structure which is
KE, = j 'TM(bty&' J = p 2 b,L. (5-48)

i- 1
where u denotes time derivative of and where p is density of the structural material.
156 -221.. 54 13L. The constraints imposed on the problem

include lower limits on cross-sectional area
-22L. 4L? - 13Z,,. -3Z.J
PbiLi
m,)- 420 <t>i = b0 -b.Q 0, i= 1,..., * (5-49)
54 -131, 156 22L.
13/.. 3L? 22L. 4L? where

here b0Q > 0 and a lower limit on natural
frequency
(5-45) <0 (5-50)

t+i=?o-f
Taking into account the lateral rigid body where f 0 is the lowest allowable eigenvalue of
motion of Member 2, the total kinetic energy
Eq. 5-47, r0 = w2 .
of the structure is
The steepest-descent algorithm may now be
KE = X-ui Af (&,.)'' + - MI (5-46) applied directly. Data for the specific prob-
1 2 2
lems solved are given in Table 5-4. The results
for an aluminum portal frame are given in
where M is the mass of Member 2 and uA is
Tables 5-5 and 5-6, with a typical profile
the horizontal velocity of points .
shown in Fig. 5-9. The design variable bi
shows the distribution of material for a
Requiring harmonic motion with frequency
minimum weight frame whose frequency of
co, the displacement vector y{t) made up of
vibration must be greater than or equal to a
all displacements is
y(t) = y sin cof TABLE 5-4

MATERIAL PROPERTIES FOR ALUMINUM
where y is a constant vector. Applying
Lagrange's equations and eliminating time a, dimensionless 0.07958
dependence yields p, lb-sec2/in.4 2.616X104
E, lb/in.2 10.3x10*
/0.in.4 0.009825
K(b)y = $M(b)y (5-47) L, in. 10.0
5-17
AMCP 706-192
TABLE 5-5 The results for the design variables bi are

COMPARISON OF UNIFORM AND OPTIMAL
FRAMES FOR ALUMINUM
the same for Members 1 and 3, and Member 2
converges to the lower bound /0, so only the
Weight of Weight of results for Member 1 are reported. For all
Frequency, Uniform Optimal Weight frequencies the values for the b{ for Member 2
rad/sec Frame, lb Frame, lb Reduction, %
are equal to 0.350.
2000 3.748 1.729 53.9
3000 8.434 2.562 69.6 5-4.4 A MINIMUM WEIGHT FRAME WITH
4000 14.994 3.590 76.1 MULTIPLE FAILURE CRITERIA
5000 23.428 4.688 80.0
To illustrate the applicability of the

steepest-descent method for the minimum
TABLE 5-6 weight design of structures with stress, buck-
OPTIMAL DESIGN VARIABLE b|
ling, and displacement constraints, an ex-
FOR VIBRATING FRAME
co, rad/sec ample of a statically loaded frame problem is
b| 2000 3000 4000 5000 presented. Fig. 5-10 shows the geometrical
*>1
1.577 1.964 2.907 4.020
b. 0.883 1.604 2.484 3.321
b, 0.552 1.416 1.912 2.622
b, 0.374 0.866 1.290 1.725
b, 0.350 0.360 0.671 0.836
be 0.350 0.350 0.350 0.350
Figure 5-10. Frame With Side Loading
configuration of the frame that is considered.

All members are assumed to be of the same
length L. Member 1 is subjected to a lateral
loading q(i). Member 3 has a uniform cross-
i ri 111 f 11 sectional area which is prescribed and will not
be allowed to vary. The connections at points
A and B are frictionless pins.
Figure 5-9. Optimum Portal Frame for
co =3000 rad/sec
The finite element method is used to
obtain the elastic response of the system for a
specified value. It can be seen from Table 5-5 given set of design variables, i.e., the cross-
that a significant material saving is possible in sectional areas of the elements. As in the
comparison to the portal frame with members preceding problem, the geometry of each
of constant cross section. cross section is the same with all dimensions
5-18
AMCP 706-192
of cross section varying proportionally. Thus, where the matrix D is derived from the
I. = abj where b, is the cross-sectionalareaof shortening of Member 2 as in par. 5-4.1 and, as
the ith element. The stiffness matrix K{b^) of in the previous problem, K(b) is a stiffness
a typical element, Fig. 5-11, can be written as matrix.
in par. 5-4.3 with generalized displacements
defined by The cost function to be minimized in this
problem is the structural weight of Members 1
u' = [t/j, U2, M3, u4]T. and 2 which is simply
S~\ / = 7 2 btLt
where y is the weight density of the material.
c
The weight of the frame is to be minimized
h subject to the following constraints:
1. Stress constraints at the ith node of

Member 1:
Figure 511. Typical Elements

/-%ax1 <0, i= 1,..., (5-53)
From the fundamental beam theory, if R is
where of = Mcib^lUib^} is bending stress,
the horizontal force transmitted from the
Member 1 to 3, and assuming that Member 2 c(bt) = (b!.)1 >2 is half the depth of the beam
remains straight without buckling, then at point i, i = 1,2,. . ., m and 0max is the
neglecting compression of Member 2, the maximum allowable stress. The parameter is
a property of the cross-sectional geometry.
deflection at A is uA =RL3/(3EI3). From
the equilibrium conditions on the transverse
2. Deflection constraint:
forces and moments at the nodes of Member
1, the generalized displacement z, which is
K I
=U
A - As 0 (5-54)
made up of the element displacements u{ can +
be evaluated from the following matrix equa-

where uA is the horizontal deflection at the
tion
top of Member 1 and A is the maximum
A(b)z=F (5-51) allowable lateral deflection of the top of the
frame.
where F is a vector load and A{b) is a
symmetric matrix. In a similar manner, if y is 3. Buckling constraint:
the displacement vector containing all ele-
3EI3
ment deflections associated with Member 2,
= u
the buckling load P can be determined by 4>m + i -[l~ A-p* (5-55)
solving the eigenvalue problem
where the first term is just the load.fi carried
K(b)y=PD(b)y (5-52) by the column.
5-19
AMCP 706-192
4. Compressive stress constraint at the jth :y

node of Member 2:
*w+/+2 =
X ~amax2 < (5-56)
for bj in Member 2,
The steepest-descent algorithm can now be

applied directly. Given data are: A =4 in.; TTTTTTT
%aJ, = 85,000 psi; am = 5,000 psi;

cross-sectional area of Member 3 is 4 in?; I = Figure 5-12. Profile of Optimal Frame
100in.;= SxlO1 psi, a = 0.07958, and = With Multiple Failure
Criteria (q = 25 lb/in.)
0.5642. The resulting horizontal forces R that
correspond to increasing constant lateral loads
4, given in Table 5-7, are 285 lb, 409 lb, 458
lb, and 458 lb, respectively. For 4 = 20 and
25 lb/in., the displacement in Eq. 5-54 is an 54.5 A MINIMUM WEIGHT PLATE WITH
equality. For lower loads it is a strict in- FREQUENCY CONSTRAINTS
equality. This was determined automatically
by the algorithm. The results for different As a final numerical example in this para-
side loadings are given in Tables 5-7 and 5-8. graph consider the problem of minimum
A profile of an optimal frame is shown in Fig. weight design of the simply supported rectan-
gular plate shown in Fig. 5-13 subject to a
5-12. Computation in each case required
natural frequency constraint. The bending
approximately 0.5 sec per iteration and 15
equation for plates of variable thickness is
iterations to converge on an IBM 360-65.
given in Eq. 5-58. When the deflection
TABLE 5-7 W{x,y,t) is written in the form
OPTIMAL DESIGN VARIABLE bj FOR
STATIC FRAME
W(x, y, t) = w(x, y) cos wf (5-57)
Cross-SectionalArea b., in.
Member 1 Member 2 the governing equation becomes

Element q, lb/in. q, b/in.
No. i 10 15 20 25 10 15 20 25
1 1.49 2.03 2.76 4.51 0.226 0.271 0.286 0.286

2 0.86 1.20 1.76 3.34 0.366 0.438 0.464 0.464
3 0.40 0.47 0.48 2.10 0.407 0.487 0.516 0.516
4 0.43 0.53 0.37 0.43 0.366 0.438 0.464 0.464
5 0.43 0.53 0.52 0.43 0.226 0.271 0.286 0.286
TABLE 5-8
VOLUME OF OPTIMUM FRAME
q, lb/in.
10 15 20 25
Optimum
Volume, 504.2 533.4 558.0 656.0
in.3 Figure 5-13. Rectangular Plate
5-20
AMCP 706-192
>V4w + 2 V2w + 2 V2w

9y 3x dy oy
d2D 32w
+ V2Z)-V2w -(1 -v)
3x2 3y2
a2> a2w 32z?a2w

3x3.y 9;c9>' 3.y2 3x2
= h pco2 w (5-58)
where Figure 5-14. Collocation Points
Eh3(x,y)
D(x,y) = (5-59) In the steepest-descent algorithm, the cost
12(1 -v2)
function that is to be minimized is
fcCx,>0 is the thickness of the plate, which is
j = p(\A . 2 h(Xj,y) (5-61)
the design variable, and p is the density of
plate material.
where AA is the area of the grid squares.
When the function w(x,y) is represented in
the form The constraints imposed on the design are
w{x,y)= 2 A,, sm sin ho-Hxpy/)<0 (5-62)

m , = 1 a b
(5-60) and
?o - ? < 0 (5-63)
the eigenvalue problem can be solved approxi-
mately by numerical methods. The problem
where h0 > 0 and f0 > 0 are lower limits on
posed here is solved using a collocation
plate thickness and eigenvalue of Eq. 5-58,
technique, i.e., the differential equation is
respectively.
satisfied at discrete points in the region, Fig.
5-14.
The steepest-descent computational al-
gorithm applies in a direct way. It should be
The number of discrete points is chosen noted that the collocation method for ap-
equal to the number of terms in the truncated proximate solutions of the equations for
series of Eq. 5-60. The drivatives of the natural frequency yields nonsymmetric
function D(x,y) at the grid points are evalu- matrices K and M in Eq. 5-16. Thus in this
ated by the use of finite differences. For a formulation of the plate optimization prob-
given set of design variables, i.e., h(x,y), the lem, the adjoint eigenvalue problem, Eq. 5-18,
lowest eigenvalue, J = pco2, and the associated must be solved along with the original eigen-
eigenvector {Amn I , which plays the role of value problem. If finite element methods for
y, are determined. plate analysis had been used, symmetric
5-21
AMCP 708-192
matrices would have been obtained. In this 5-5 GENERAL TREATMENT OF TRUSS
example, as well as in the preceding, a DESIGN*
minimum effort was expended to make com-
putations efficient. The emphasis has been The theory presented in pars. 5-2 and 5-3
placed on getting results. A subsequent effort will now be applied to the case of general
will be devoted to making algorithms more plane and space trusses. These types of
efficient. structures are encountered quite frequently in
practical situations. Most common among
The minimum weight plate problem was these are buildings, transmission towers,
solved by the algorithm of par. 5-3 with bridges, cooling towers, aircraft structures,
E = 3.0xl07 psi, p = 7.43 xlCT4 lb-sec2/in.4 , and lightweight military structures. In all these
v = 0.30, and co0 = 1375 rad/sec. The uni- cases, it is desirable that the structure simul-
form plate with t = pw2 = 1400 was taken as taneously should meet strength, deflection,
the initial estimate to the optimization prob- and frequency requirements and be of mini-
lem. The dimensions of the plate are 10.0 in. mum weight. In this chapter, all these con-
by 10.0 in. and the value of h0 =0.1 in. The straints will be considered.
material is assumed to have a constant density
and so minimum weight is equivalent to 5-5.1 SPECIAL PROBLEM FORMULATION
minimum volume. The volume of the uniform
plate is 11.44 in.3 and the volume of the In the problems to be considered here,
optimal plate is 10.8 in.3 which is a 5.6% geometry of the truss is assumed to be
material savings. Fig. 5-15 shows 25 colloca- specified and the loads are applied only at the
tion points. The numbers in the network are joints. The objective function for the problem
is taken as the total weight or the volume of
a the truss, and the design variable for each
0 2 member is taken as its cross-sectional area.
0.124 0.116 0.100 0.100 0.100 The objective function of Eq. 5-1 in this case
0.116 0.103 0.100 0.100 0.100 is a linear function of m design variables and
0.100 0.100 0.100 0.104 0.111
may be written as
0.100 0.100 0.104 0.121 0.128
h 0.100 0.100 0.111 0.128 0.136
2
<' /= 21 PiLtbt (5-64)
where p. and (. are material density and

Figure 5-15. Optimal Design Variable h(x, y) length of member i, respectively.
for Vibrating Plate
The displacement method of structural
the values of the thickness function h(x,y) at analysis is used, and nodal displacements of
each nodal point which is located at the the truss are considered as basic state vari-
center of each square. Double symmetry of ables. Therefore, the jth component of the
the optimal plate thickness was observed state variable represents the jth displacement
about axes through the point (a/2, b/2). component of the truss. Fig. 5-16 shows a
simple scheme of designating joints, members,
*This paragraph is based on the dissertation of Dr. J. Arora,
and displacement components of a truss. Fig.
Ref. 34. 5-17 shows a bar element with sign conven-
5-22
AMCP 706-192
Kt
K(b) (5-68)
K
2\ v4
where m is the total number of elements in
Figure 5-16. Description of a Truss the truss and K. is the stiffness matrix for the
ith element of the truss. The stiffness matrix
for the ith element may be written as
1 I
K.. (5-69)
L
' i 1 1
where E{ is Young's Modulus of Elasticity of

the ith element. Substituting Eqs. 5-65 and
5-66 into Eq. 5-67, one obtains
/= ['K(jb)]z
= K(b)z (5-70)
Figure 5-17. A Truss Element
where
tion to be used on element forces and

K(b)=TK(b) (5-71)
deformations. Basic equations of the displace-
ment method for a truss may be written as
is the structure stiffness matrix, which is
identical to A(b) in Eq. 5-2. The mass matrix
u = z (5-65)
M(b) for the truss may also be computed in a
similar way, and it is given by
F=K(b)u (5-66)
M(b)=JM(b) (5-72)
and
where M(b) is formed from element mass
f=fF (5-67)
matrices and is given by
where u is the element deformationvector, F
M,
is the element force vector, f is the vector of
external loads applied to structural nodes, and
M(b) = (5-73)
is a rectangular transformation matrix,
which transforms the nodal displacement vec-
tor z to the element deformation vector u. 'M_
The matrix K (b) is composed of element
stiffness matrices and is given by Here, Mt is the mass matrix for the ith
5-23
AMCP 708-192
element of the russ and is given by Assembling the matrix C* of Eq. 5-21 is
more tedious. It requires a formation of
2 1 constraint set $ as in Eq. 4-105 which will be
PMt
M,=- (5-74) discussed in detail now. The constraint set $
1 2
may be divided into five subsets, namely,
frequency, stress, buckling and displacement
Any nonstructural mass that is attached to constraints, and lower and upper bounds on
the trass may also be added to the mass the design variables. Explanation of these
matrix of Eq. 5-72 and it may be written as subsets follows one by one and, for each
subset, matrices 9#/9?, 9#/9<?, and 9$/9> are
M(b)=TM(b)+M0 (5-75) computed.
where M, is a matrix consisting of nonstruc- 5-5.1.1 FREQUENCY CONSTRAINTS

tural masses. In the example problems to be
presented later, MQ is taken to be a null In the example problems, only one fre-
matrix. However, there is no particular dif- quency constraint is considered. However, if
ficulty in incorporating this matrix if it is not other frequency constraints are also present,
zero. Its inclusion will simply change the these may be treated in a similar way. Since
lowest natural frequency of the truss. The the design variable vector b is available at any
derivation of the given structural analysis iteration, the matrices K(b) and M{b) are
equations and matrices is well documented in computed from Eqs. 5-71 and 5-72, respec-
the literature (Refs. 20, 24). tively. The lowest eigenvalue f and the asso-
ciated eigenvector^ are then obtained from
In order to apply the algorithm of par. Eqs. 5-5 and 5-6, respectively. Premultiplying
5-3.2, two main matrices %J and 2* of Eqs. both sides of Eq. 5-6 by K'lib), one obtains
5-20 and 5-21 must be computed. They can
be assembled very easily once various other
matrices required in them have been com- K-HbW(b)y=jy = yy (5-77)
puted. In the class of problems treated here
fib) does not depend on the design variable, where y = 1/f. The power method is used to
so df(b)/db = 0. Also, one obtains from Eq. find the largest eigenvalue y of K'l(b)M{b).
5-64 This method of obtaining the largest eigen-
value is quite efficient in the present problem,
bJ since a very good approximation to the
db = (Piî. .PmîJ eigenvector at each iteration, except for the
first one, is available from the previous
9/ iteration. The lowest eigenvalue is then given
by f = 1/7- The frequency constraint may
now be written as
and 9//9f = 0; and from Eq. 5-14
\J = (0, , 0)T. Substitution of these values J>? (5-78)
into Eq. 5-20 yields
where f0 corresponds to a given frequency.
y=(Piii- >PmLm)T- (5-76) In terms of the notation used in Eq.
5-24
AMCP 706-192
5-4, Eq. 5-78 may be written as inequality (Eq. 5-82) holds in such cases. For
tension members, a. and a? are negative;
$,&, z, f) = ro - f < 0 (5-79) therefore, Eq. 5-82 is written as a? s a.. The
expressions that follow are written for the
where s is a number assigned to this con- case of compression members. For tension
straint. If this constraint is violated, then members similar expressions can be readily
A<?s = _(ro _f), fy/db =(0, ,0),3#,/3? written. Inequality Eq. 5-82 may be written
= - l,andd#s/3z =(0, , 0). as
5-5.1.2 STRESS CONSTRAINTS $,&, z, f) = ff, -oCt< 0 (5-83)
Since the matrix K{b) is available, Eq. 5-70 where s is an index assigned to this constraint.
can be solved for the unknown displacement In all subsequent constraint subsets, subscript
vector z. Element forces can then be com- s on $ will have the same meaning. If Eq. 5-83
puted from Eqs. 5-65 and 5-66. Substitution is violated, then A$s = - (a. - af), 9</3? = 0,
of Eq. 5-65 into Eq. 5-66 yields
3tf>s 3#. /1 *F, 1
F= Sz (5-80) = 0 for i=j, " =1 - F. ,
dbf db; \bt bbi b2. '''
where S = K(b). It may be noticed from Fig. 90s /dat doi

5-17 that two forces are specified for each and
dz \ dz!' ' bzn I
element but the primary force remains con-
stant throughout a bar element. Therefore,
where dojdz. and 9F./9ft/ may be computed
F'2 = F[, where superscript i denotes the
from Eq. 5-80.
element number. Dimensions of the matrix S
may be reduced from 2m X n to m X n by
5-5.1.3 BUCKLING CONSTRAINTS
using this relationship. Stresses in the mem-
bers may now be calculated as
Each compression member of the truss is
F
i
also checked for the Euler buckling load given
(5-81) by
Once the stress for each member becomes P?

*2V,-
(5-84)
known, it is checked against the critical stress. L2
A number of these stresses may be violated in
a particular iteration. The stress constraint for
where P? and I{ are the critical buckling load
the ith member may be written as
and moment of inertia of the ith member,
respectively. It is assumed that the moment of
a, < ac. (5-82) inertia of the cross section of a member can
be written as
where act is the critical stress for member i. It
should be noted that, in terms of the nota- h = v; (5-85)
tions used in Fig. 5-17, compressive stress in a
member is- taken as positive and accordingly where a. is a constant depending upon the
5-25
AMCP 706-192
cross-sectional geometry of the ith member. where za is the maximum allowable /th
This is a convenient way of expressing the component of displacement. If a particular
moment of inertia in terms of the cross-sec- component of displacement is positive, then
tional area of a member, because the constant Eq. 5-89 is written as z- z? < 0;and if it is
a. can be specified by the designer quite negative, then it is written aszi?- z- < 0. The
readily. Therefore, Eq. 5-84 may now be expressions that follow are written for the
written as case of positive displacement, and similar
expressions can also be written for the case of
negative displacement. In terms of the nota-
(5-86) tion of Eq. 5-4 the constraint for the positive
displacement may be written as
where 0. =n2Eiaj/Lf. Eq. 5-86 may be writ-
4>s(b,z, f) = zj-z < 0 (5-90)
ten in terms of the critical buckling stress a*
as
If thejth displacement component exceeds an
Lp=0,.b,.. (5-87) allowable limit, then
Now the buckling constraint for the ith A0s=-rz/.-zp) - =0,^ =0, ,0),
compression member may be written as
w)--o;-
and
i,{b.z,f)=ot -of Q 0. (5-88)
*r=(0'
If this buckling constraint is violated, then
All the displacement components are checked
A0s=_(a. -a),30,/3f =0,
and any other violation is treated in a similar
way.
s = 0 for i # /, 5
dh. db,. b, db< ui '' '
5-5.1.5 BOUNDS ON DESIGN VARIABLES
30, bo, 3a,

and It may be necessary to put upper and lower
dz dz, 3z bounds on each design variable. This con-
straint may be demanded by many practical,
where dFj/db, and da./3z. may again be architectural or structural considerations.
computed from Eq. 5-80. The buckling con- Moreover, a lower limit on each design vari-
straints on all other compression members are able is required in the algorithm in order to
treated in a similar way. avoid the attainment of unrealizable designs
such as negative areas. This constraint may be
5-5.1.4 DISPLACEMENT CONSTRAINTS written as
The displacement components are known b\ < bt Q bl (5-91)

at this stage; therefore, the constraints on
them may be written as where bL, is the lower and bu, is the upper
bound on the ith design variable. Inequality
I*, I z* (5-89) Eq. 5-91 may be split up into two parts:
5-26
AMCP 708-192
(1) Lower Bound o n Design Variables: Problem par. 4-3.1, the present formulation
handles it without any difficulty.
This constraint is written as bLt < bt or in
terms of notation of Eq. 5^ After all the constraints have been con-
sidered, the matrices d(j>/db, 90/3?, and
4>s(b,z,f) =-bLt-b < 0 (5-92) 90/9z are available and X* can be solved
from Eq. 5-15. This still does not allow
Violation of this constraint yields, the matrix * of Eq. 5-21 to be assem-
bled. The following matrices must also be
A0S=-O>'' -K), 3*f = 0, computed
0,-1, o,... , 0), and [K(b)z] (5-94)

db
db (ith)
d_
90, [K(b)y] (5-95)
^ =(0, . ..., 0) db
az
and
(2) Upper Bound on Design Variables:
[M(b)y] (5-96)
This constraint is very similar to the previ- db
ous one and in the notation of Eq. 5-4 it is
written as These matrices are assembled automatically
from the quantities such as K(b),M(b), z, and
4>s(b,z, S) = bt -but < 0 (5-93) y, which have previously been calculated in
the computer. The procedure of computing
ff the upper bound on any design variable is the matrix of Eq. 5-94 will be explained here;
violated, then the matrices of Eqs. 5-95 and 5-96 are
calculated in an exactly similar manner. Eq.
5-71 may be written as (see Appendix B):
A0S = -(bi-by),
9f
K(b) = . 2 ' K.'
z = I '
*i =(0,. . .,0, 1,0, . . .,0),and
db (ith)
where Kt is the only quantity which is a
function of b. Now, Eq. 5-94 can be written
as follows, by substituting the above expres-
ai,=(o,...,o).
dz sion for K(b):
9 9 r -
it may be noticed here that the cross-sectional 2 l K
area of any member of the truss may be
assigned a predetermined value by putting the
same upper and lower bound on it. This
situation may be encountered in practice due (=i db
(tfTk,ff)2
to various reasons, and as shown in Example
(5-97)
5-27
AMCP 708-192
It should be noted here that the summation constraints. This procedure of taking into
sign in Eq. 5-97 represents the summation of consideration all the loading conditions has
m matrices of dimension (n x m). The worked out quite satisfactorily in the example
quantity inside the differentiation sign is an problems.
n-dimensional vector whose components may
be dependent upon the design variable vector 5-5.3 EXAMPLE PROBLEMS
b. Therefore, the quantity inside the summa-
tion sign is an n x m matrix for each index i Several trusses are designed by applying the
However, in the present case, since Kt is a procedure presented in this paragraph. A
function of only bf, the computation of Eq. computer program, based on the algorithm
5-97 is greatly simplified. Consideration of stated previously, was written in FORTRAN
each i in Eq. 5-97 generates a (n x m) matrix IV. The computations were performed on the
whose only nonzero elements are in its ith University of Iowa IBM 360-65 computer.
column. Computation of Eq. 5-97 is per- The stiffness matrix or the structure was
formed quite readily and automatically in the inverted by the Gauss-Jordan elimination
computer. procedure, and the power method was used to
find the smallest eigenvalue.
All the information being available, the
matrix 2* of Eq. 5-21 may now be assembled Results for three typical trusses are pre-
and the algorithm of par. 5-3.2 may be used sented here. All these structures were de-
to solve actual problems. signed with stress, displacement, buckling,
and frequency constraints. Examples 4-1 and
5-5.2 MULTIPLE LOADING CONDITIONS 4-2, par. 4-1.1, are compared with results in
Ref. 25. These were designed with and with-
Most structures are designed to withstand a out frequency and buckling constraints in
multiple loading environment. This is quite order to compare the results with Ref. 25.
reasonable, because only a certain combina- Example 4-3, par. 4-3.1, is treated in Ref. 26,
tion of loads may act on the structure at a and it was also designed with only stress
particular time. This situation is handled in constraints in one case to compare results
the par. 5-5.1 formulation by expanding the with Ref. 26. All sample problems had lower
state variable vector z to include all states. limits on areas of the elements and Example
The element force vector / is also expanded 4-3 had upper limits. The program is general
accordingly. Formulation of displacement, enough to handle different lower and upper
stress, and buckling constraints must also take bounds on stresses in an element, elements of
into consideration all states of the system. different materials, and a different buckling
This is handled in the manner that follows. parameter a. for each element. The examples
While formulating a particular displacement follow:
constraint, the value of displacement for each
loading case is checked and each violation is 1. Example 5-1. Five-node Four-bar Truss
entered into the reduced constraint vector 0.
After this, the procedure of calculating the Fig. 5-18 shows the geometry and the
matrices d<j>/db, 90/3f, and 9<A/3z is the same dimensions of the truss. Input and output
as explained earlier. An exact same procedure information is given in Table 5-9. In order to
is followed in treating stress and buckling compare the results with those of Ref. 25, the
5-28
AMCP 706-192
to the fact that in the first case the frequency

constraint was violated by 0.154%, whereas in
the second case this violation was 0.143%.
Fig. 5-19 shows variation of the cost function
with respect to the number of iterations for
this problem. It may be noted that for
practical purposes, convergence was obtained
in six to eight iterations for all the cases.
60 in.
Figure 5-18. Four-bar Truss (Example 5-1) 2. Example 5-2. Transmission Tower
trass was first designed for stress constraints, Fig. 5-20 shows the geometry and the
Fig. 5-19(A), and second for stress and dimensions of the tower. This example also is
displacement constraints, Fig. 5-19(B). It treated in Ref. 25. In this problem, the
may be noted that the results presented here cross-sectional area of each member of the
are at least as good as those presented in Ref.
25.
The final design weight with only stress

constraints was 9.09 lb and computation time
was 1.820 sec for 12 iterations. The final
design weight reported in Ref. 25 was 9.09 lb
with a computation time of 4 sec for 5 cycles.
The final design weight, with stress and
displacement constraints, was 14.28 lb and
the computation time was 1.500 sec for 12
iterations. The final design weight reported in 4' 8 12
Ref. 25 was 14.30 lb with a computation time Iteration Number
of 10 sec for 4 cycles. It is difficult to make (A) With Stress Constraints Only
an exact comparison of the computation
times because the computer used here is 26 (K 1 1 1
different from that used in Ref. 25. The
/Starting Point 2 _
computation times reported in Ref. 25 are on 220
IBM 7094-11-7044 DCS Computer. .OptimumWeight = 113.77 lb
y / With All Constraints
The trass was also designed by including 140 /Cv > Starting Point 1
buckling and frequency constraints along with
other constraints. Two different starting
points were used in optimizing this trass. fini 1 h h li
Starting Point 1 was infeasible and Starting Iteration Number
Point 2 was feasible. The final design weight (B) With AII Constraints
beginning at Starting Point 1 was 113.48 lb
and at Starting Point 2 was 113.77 lb. The Figure 5-19. Iteration vs Weight Curves for
slight difference in the two weights was due Example 5-1, Four-bar Truss
5-29
AMCP 706-192
TABLE 5-9
FOUR-BAR TRUSS (EXAMPLE 5-1)
Design Information: For each member, Young's Modulus of Elasticity Ejt the specific weight p/r lower limit on
area of cross sectionbL. ,and the constant a(. are 104 kips/in.2, 0.10lb/in.3, 0.10in.2,and 1, respectively. There is
no upper limit on member size. The resonant frequency for the truss is 284.6 Hz. For Output 1, the stress limits on
each member are 25.0kips/in.2 and the displacement limitsat node five are, 0.0, 0.3in., and 0.4in. inthex,-,
x2-, andx3-directions, respectively. For Output 2,the stress limits for each member are 15.0kips/in.2, and the
displacement limitsat node five are 0.15 in. in all three directions. There are three loading conditionsfor the truss;
they are: in positivexj-, x2-, andx3- directions, 5, 0, 0; 0, 5, 0; and 0,0, 7.5 kip, respectively, applied at node five.
OUTPUT 1. With Stress and Displacement Constraints Only
With only stress constraints With displacement constraints, also

Time per iteration = 0.152 sec Time per iteration = 0.124 sec
Total time = 1.820 sec Total time = 1.500 sec
El. Starting Final El. Starting Final

No. Values, Values, No. Values, Values,
in.2 in.2 in,2 in.2
1 0.100 0.130 1 0.500 0.234

2 0.200 0.192 2 0.500 0.3L9
3 0.200 0.120 3 0.500 0.184
4 0.100 0.100 4 0.500 0.128
Weight, Weight,
10.19 9.09 34.86 14.28
lb lb
OUTPUT2. With All Constraints
Starting Point 1 Starting Point 2

Time per iteration = 0.147 sec Time per iteration = 0.172 sec
Total time = 4.710 sec Total time = 4.640 sec
El. Starting Final El. Starting Final

No. Values, Values, No. Values, Values,
in.2 in.2 in.2 in.2
1 1.000 0.543 1 2.0 0.559

2 1.000 1.961 2 4.000 1.883
3 1.000 3.635 3 8.000 3.703
4 1.000 0.479 4 1.000 0.468
Weight, Weight,
69.72 113.48 257.68 113.77
lb lb
5-30
AMCP 706-192
Figure 5-20. Transmission Tower (Example 5-2)
trass is treated as an unknown design variable, cycles. The values of final design variables
and the results obtained are given in Table compare quite well with those in Ref. 25. At
5-10. The tower was designed first with only the final design point all constraints were
stress constraints. The final design weight in satisfied within 0.006%.
this case was 91.13 lb with a computation
time of 38 sec for 12 iterations. The final The tower was also designed with stress and
design weight reported in Ref. 25 was 91.14 displacement constraints and, finally, with all
lb with a computation time of 9 sec for 5 the constraints included. The design weight in
5-31
AMCP 706-192
TABLE 5-10
TRANSMISSION TOWER (EXAMPLE 5-2)
Design Information: For each member of the structure, the modulus of elasticity ), the specific weight pi, the
constanta/, and the stress limits are 104 kips/in.2, 0.10lb/in.3, l.Oand 40.0kips/in.2, respectively. The lower
limit on the area of cross section of each member is O.lOin.2 for the case with stress constraints only and 0.01 in.2
for other cases. There is no upper limit on the member sizes. The resonant frequency for the structure is 173.92 Hz,
and the displacement limits are 0.35 in. on all nodes and in all directions. There are six loading conditions, and they
are as follows (all loads are in kips):
Direction of Load Direction of Load

Load Node Load Node
x x x
Cond. l 2 3 Cond. x, x2 *3
1 1.0 10.0 -5.0 1 0 10.0 -5.0

2 0 10.0 -5.0 2 - 1.0 10.0 -5.0
1 3 0.5 0 0 2 4 -0.5 0 0
6 0.5 0 0 5 -0.5 0 0
1 1.0 -10.0 -5.0 1 0 -10.0 -5.0

2 0 -10.0 -5.0 2 -1.0 -10.0 -5.0
3 3 0.5 0 0 4 4 -0.5 0 O
6 0.5 0 0 5 -0.5 0 0
1 0 20.0 -5.0 1 0 -20.0 -5.0

5 6
2 0 -20.0 -5.0 2 0 20.0 -5.0
output:
With Stress With Stress and Dis-

With All Constraints
Constraints Only placement Constraints
El. Starting Final Starting Final Starting Final

No. Values, in.2 Values, in.2 Values, in.2 Values, in.2 Values, in.2 Values, in.2
1 0.200 0.100 1.000 0.010 0.500 0.010

2 0.500 0.376 3.000 2.322 2.500 2.092
3 0.500 0.376 3.000 2.322 2.500 2.075
4 0.500 0.376 3.000 2.322 2.500 2.095
5 0.500 0.376 3.000 2.322 2.500 2.083
6 0.500 0.47i 3.000 2.768 2.500 2.357
7 0.500 0.471 3.000 2.768 2.500 2.354
8 0.500 0.471 3.000 2.768 2.500 2.350
9 0.500 0.471 3.000 2.768 2.500 2.335
10 0.200 0.100 1.000 0.010 0.500 0.035
11 0.200 O.IOO 1.000 0.010 0.500 0.035
12 0.200 0.1OO 1.000 0.010 0.500 0.087
13 0.200 0.100 1.000 0.010 0.500 0.084
14 0.200 0.100 2.000 0.690 1.500 1.113
15 0.200 0.100 2.000 0.690 1.500 1.113
16 0.200 0.100 2.000 0.690 1.500 1.112
17 0.200 0.100 2.000 0.690 1.500 1.112
5-32
AMCP 706-192
TABLE 5-10 (Cont'd.)
Output: (Cont'd.)
With Stress With Stress and Dis-

Constraints Only placement Constraints
El. Starting Final Starting Final Starting Final

No. Values, in.2 Values, in.' Values, in.' Values, in.' Values, in.' Values, in.'
18 0.500 0.277 2.000 1.524 2.000 2.056

19 0.500 0.277 2.000 1.524 2.000 2.058
20 0.500 0.277 2.000 1.524 2.000 2.046
21 0.500 0.277 2.000 1.524 2.000 2.058
22 0.500 0.380 3.000 2.733 3.000 2.822
23 0.500 0.380 3.000 2.733 3.000 2.808
24 0.500 0.380 3.000 2.733 3.000 2.803
25 0.500 0.380 3.000 2.733 3.000 2.785
Weight,
132.37 91.13 772.24 546.18 669.80 590.32
lb
the first case was 546.18 lb with a computa- optimized for a single loading condition. The
tion time of 47 sec for 17 iterations, and the design information and the results are shown
maximum constraint violation was 0.00011%. in Table 5-11. In order to compare the results
The comparable design weight reported in with Ref. 26, the truss was first optimized
Ref. 25 was 555.11 lb with a computation with stress constraints only. The final design
time of 24 sec for 7 cycles. This shows that, weight was 2,993.37 lb with a computation
when displacement constraints are also in-
cluded, the results obtained with the new
780
gradient projection method are slightly better
than those of Ref. 25. For a design with all
the constraints included, the final weight was /W:*h All Constraints
590.32 lb with a computation time of 129 sec Optimum Weight = 590.32 ,b
for 36 iterations, and the maximum violation
of constraint was 0.028%. Fig. 5-21 shows With Stress and Displacement
variation of the cost function with respect to I Constraints
the iteration number for the last two cases of

this problem. It may be noted that for
practical purposes, convergence was obtained
in only 6 iterations.
520
3. Example 5-3. 4 7-Bar Plane Truss 4 8 12
Iteration Number
The schematic diagram of the structure
with dimensions is shown in Fig. 5-22. This Figure 5-21. Iteration vs Weight Curves
for Example 5-2, Trans-
example is also treated in Ref. 26 where it is mission Tower
5-33
AMCP 706-192
15.68K -5@5ft-0in- 15.68K
6.72K 18 10 3 17 @) 13 14 6.72K
2 ft 6 in.
2 ft 6 in.
F/gure 5-22. 47-Bar Plane Truss (Example 5-3)
time of 115 sec for 17 iterations. At this 9th iteration for which the design weight was
point the stress in member 18 was violated by 2,998.88 lb; maximum constraint violation
0.24% and all other violations were less than was 0.10% for stress in 7th member and all
0.035%. Another feasible design occurred at other violations were less than 0.016%. The
5-34
AMCP 706-192
TABLE 5-11
47-BAR PLANETRUSS (EXAMPLE 5-3)
Design Information: For each member of the structure, the modulus of elasticity Er the specific weight p:, and
the constanta. are 3.Ox 104 kips/in.2, 0.284lb/in.3, and 1.0,respectively. The resonant frequency for the structure
is 16.0 Hz and the displacement limits are 1 in. on all nodesand in all directions. There is one loading condition for
the truss which is shownon Fig. 5-22 .Allowable stress in tension for all members is 21.28kips/in.2
output:
Final Area, in.2
Compression With Stress

El. Lower Area Upper Area Stress Limit, Initial Area, Constraints With All
No. Bound, in.2 Bound, in.2 kips/in.2 in.2 Only Constraints
1 3.570 9.620 14.56 5.690 3.570 7.537

2 3.570 9.620 14.56 5.690 3.570 5.771
3 3.570 9.620 14.56 5.690 3.570 3.570
4 3.570 9.620 14.56 5.690 3.570 4.473
5 3.570 9.620 14.56 5.690 3.752 6.505
6 3.570 9.620 14.56 5.690 3.570 6.124
7 3.570 9.620 14.56 5.690 4.212 7.777
8 3.570 9.620 14.56 5.690 5.217 9.529
9 1.930 2.940 15.90 2.210 1.930 1.930
10 1.930 2.940 15.90 2.210 1.930 1.930
11 1.930 2.940 15.90 2.210 2.205 2.199
12 1.930 2.940 15.90 2.210 2.205 2.205
13 1.930 2.940 15.90 2.210 1.930 2.940
14 1.930 2.940 15.90 2,210 1.930 2.940
15 1.930 2.940 15.90 2.210 2.205 2.119
16 1.930 2.940 15.90 2.210 2.205 2.205
17 1.360 2.190 15.46 2.100 1.417 2.136
18 1.360 2.190 15.46 2.100 1.815 1.630
19 1.360 2.190 15.46 2.100 1.360 1.360
20 1.360 2.190 15.46 2.100 1.360 1.360
21 0.376 0.376 3.36 0.376 0.376 0.376
22 0.376 0.376 3.36 0.376 0.376 0.376
23 0.376 0.376 3.36 0.376 0.376 0.376
24 0.376 0.376 3.36 0.376 0.376 0.376
25 1.360 2.190 12.32 2.100 1.360 1.455
26 1.360 2.190 12.32 2.100 1.360 1.451
27 1.360 2.190 12.32 2.100 1.360 2.137
28 1.360 2.190 12.32 2.100 1.360 1.360
29 1.360 2.190 12.32 2.100 1.360 1.492
30 1.360 2.190 12.32 2.100 1.360 1.428
31 2.940 6.040 17.47 3.850 2.940 3.774
32 2.940 6.040 17.47 3.850 2.940 2.940
33 2.940 6.040 17.47 3.850 2.940 2.940
34 2.940 6.040 17.47 3.850 2.940 5.592
35 2.940 6.040 17.47 3.850 2.940 3.582
36 2.940 6.040 17.47 3.850 2.940 2.940
37 0.940 1.320 4.93 1.200 0.940 0.940
5-35
AMCP 708-192
TABLE 5-11 (Cont'd.)
output: (Cont'd.)
Final Area in.2
Compression With Stress

El. Lower Area Upper Area Stress Limit, Initial Area, Constraints With All
No. Bound, in.2 Bound, in.2 kips/in.2 in.2 Only Constraints
38 0.940 1.320 4.93 1.200 0.940 0.940

39 0.940 1.320 4.93 1.200 0.940 0.940
40 2.940 6.040 10.75 3.500 2.940 2.940
41 2.940 6.040 10.75 3.500 2.940 2.940
42 2.940 6.040 10.75 3.500 2.940 2.940
43 2.940 6.040 10.75 3.500 2.940 2.940
44 2.940 6.040 10.75 3.500 2.940 2.940
45 2.940 6.040 10.75 3.500 2.940 2.940
46 2.940 6.040 10.75 3.500 2.940 2.940
47 2.940 6.040 10.75 3.500 2.940 2.940
Weight, lb 3910.30 2993.37 3771.0
final weight reported in Ref. 26 was 3,328.5 that for practical purposes, convergence oc-
lb which is considerably higher than the one curred in approximately 6 iterations.
reported herein. This may be attributed to the
fact that in Ref. 26 the members are divided
into eight groups so that there are only eight 4000
independent design variables, whereas in this

treatment, area of cross section of each
member of the truss is treated as an unknown 3600
design variable.
The truss also was designed by imposing all

the constraints. The starting point, stress
3200 -
V With Only Stress Constraints
VOptimum Weight = 2993.37 lb
limits, and upper and lower bounds on the

areas are same as those used in Ref. 26. It 2900
4 8
may be noted that members 21, 22, 23, and Iteration Number
24 had the same upper and lower bounds on
areas. The final design weight was 3,771.0 lb Figure 5-23. Iteration vs Weight Curves for
with a computation time of 166 sec for 24 Example 5-3, 47-Bar Plane Truss
iterations. The maximum violation of the
constraint was 0.27% on stress for member
11. Fig. 5-23 shows variation of the cost 5-6 A GENERAL TREATMENT OF PLANE
function with respect to the number of FRAME DESIGN*
iteration for both the cases. It may be noted
In this paragraph, an application of the
*This paragraph is based on the dissertation of Dr. J. Arora,
Ref. 34. gradient projection method to framed struc-
5-36
AMCP 706-192
tures will be presented. Rigid frames are z. 1 =cfir (5-99)

found quite frequently in practical situations,
including building and vehicle structures. In and
the present work, optimization of planar
framed structures using wide flange steel
sections is considered under the assumption
r
i =dfir (5-100)
of elastic, working stress analysis. The AISC where At is the area of cross section, Zf is the
Steel Construction Manual (Ref. 27) is used section modulus, r. is the least radius of
for the properties of these sections. The gyration of the ith element of rigid frame, and
constraints considered are stress, buckling, a, ct, and df are constants. These constants
displacement, natural frequency, and restric- can be found by plotting curves of area of
tions on design variables. cross section, section modulus, and the least
radius of gyration versus the moment of
5-6.1 PROBLEM FORMULATION inertia of various economical beam and
column sections. These curves have been
In the problems considered here, the drawn by Nakamura (Ref. 28) for wide flange
geometry of the frame is assumed to be sections of AISC Steel Construction Manual
specified, i.e., lengths of the members or the (Ref. 27), and the same values are used in this
joint coordinates are not treated as design handbook. This approach of obtaining con-
variables. Multiple loading conditions for the tinuous relationship for area of cross section,
structure are treated by the procedure ex- section modulus, least radius of gyration, and
plained in par. 5-4.2. The moment of inertia the moment of inertia has also been used by
for each element is treated as the design other researchers in their work (Refs. 29, 30,
variable; therefore, ft is a vector whose ith 31).
component b{ is the moment of inertia of the
ith element. In calculating weight or volume The objective function, Eq. 5-1, for this
of the structure, element direct stresses, ele- problem is again taken as the total weight of
ment bending stresses, area of cross section, the frame which may be written as
and the section modulus of each element
must be known. Also, in order to calculate
the allowable compressive stress for an ele- /= 2 pAJi^V P^pA,1'1 (5-101)
/=; ' ' ' i=i ' ' ' '
ment, its least radius of gyration r. must be
known. These quantities are required as con- The displacement method of structural analy-
tinuous functions, rather than discrete num- sis is used, and nodal displacements of the
bers, in the present formulation. Since the frame are considered as basic state variables.
moment of inertia of each element is its only Therefore, the jth component of the state
design variable, the quantities area of cross variable represents the jth displacement com-
section, section modulus, and the least radius ponent of the frame. Fig. 5-24 shows a simple
of gyration must be expressed in terms of the scheme for designating joints, members, and
moment of inertia of the element. These displacement components of a frame in the
relationships of the ith element are written as structure coordinate system. Fig. 5-25 shows
follows: a frame element in the member coordinate
system with the sign convention to be used on
A, =a1b1l/2 (5-98)
element forces and deformations. It may be
5-37
AMCP 706-192
z2 z5 Z
8 noted that FJ and FJ are direct forces on the
L
element, F[ and Fs' are shearing forces, and
a F'z and F'6 are the moments at the end of an
element. The structural analysis equations
developed in par. 5-4 and in Appendix B are
1 3 5
also used here. The element forces are com-
<
j
puted from Eq. 5-80 which may again be

0 written as follows:
Tl*~xl TTT rr
F = Sz (5-102)
Figure 5-24. Description of a Frame
where
"2
S = K(b) (5-103)
'F'Z2
U'
.w kips/in. Dimensions of the matrices K(b) and are
I I I f\ 4 t I \ t I I t I F'4' F'4 adjusted for the case of frames and F is a
M r
vector which consists of element forces for all
the elements of the frame. The stiffness and
>*3 ~/Ai mass matrices for the frame element are
different from the truss element. They are
Figure 5-25. A Frame Element given by the following matrices:
1/2 1/2
fii 0 0 a
fii 0 0
12/L) 6/1, 0 -niL) 6/L,
6/. 4 0 -6/Lj 2
K.= (5-104)
1/2 -1/2
aibi 0 0 aibi 0 0
0 12/L,2 - 6/1,. 0 12/1,. -6/i,
0 6/Lj 2 0 - 61L. 4
5-38
AMCP 706-192
140 0 0 70 0 0
0 156 22L 0 54 - 13L
pj.L,.a./i/2 0 22L 4L2 0 13L - 3Z,2

M.. =- (5-105)
420
70 0 0 140 0 0
0 54 13L 0 156 -22L
0 13L -3L2 0 -22L 4L2
Now, as before, the matrices HJ and 0 of * of Eq 5-21. It requires formation of the

Eqs. 5-20 and 5-21, respectively, must be constraint vector /dz, and d$/db. The treatment of
once various other matrices have been com- frequency, displacement, and design variable
puted. Let us first consider computation of constraints in the case of a frame is exactly
the matrix 2J of Eq. 5-20. The matrix f{b), the same as in the case of a truss, which is
which is computed from externally applied developed in par. 5-4. So, these features will
loads, is independent of the design variable not be explained here, except for the fact that
vector b if the self weight of the elements is any point where a displacement constraint
neglected. This implies that df(b)/db = 0. must be imposed is treated as a nodal point.
Also, from Eq. 5-101 one obtains Computation of matrices such as d/db
[K(b)z], 3/96 [K(b)y], and d/db [M(b)y] is
_&Z_ } , -1/2 also carried out in the way explained in par.
ab 2' 5-4. The only constraint that remains to be
considered is the stress constraint, which will
1/2
(5-106) be considered next.
^m mm m )
5-G.2 STRESS CONSTRAINT CALCULA-

bJ
= (0, 0), (5-107) TIONS
9z
Let s denote the subscript for this con-

and 3//d? = 0. Eq. 5-14 now yields, \J = (0,
straint. If one can compute 90s/9f and row-
0)r. Substituting these values into Eq.
vectors 90s/9z and 90s/96 for this constraint,
5-20. one obtains
then he can assemble the matrix fi* of Eq.
1/2 5-21.
ey=-(p, ,*,,"
The members of a framed structure are

f>mLm*> bm~ ) (5-108) subjected to direct as well as bending stresses.
Thus, the effect of combined stresses must be
Next, consider computation of the matrix considered in implementing the stress con-
5-39
AMCP 706-192
straints. It should be noted here that a clear kL

distinction is made between the elements and (5-114)
rC
the members of a frame. This distinction is
necessitated by the fact that a member must Fa - allowable compressive stress
often be divided into several elements for
structural analysis and implementation of E = Young's modulus
displacement constraints. On the other hand,
the compressive stress for all elements making On the cross section of axially loaded col-
up a member is the same. In the present work, umns when kL/r exceeds C,,
the members subjected to direct and bending
stresses are required to satisfy the AISC 1.49 x 105
specification (Ref. 27). The permissible stress, F = ,Ksi (Kip/in?). (5-115)
(kL/r)2
according to this Steel Construction Manual,
are:
4. Combined Stresses:
1. Tension:
a. Axial Compression and Bending:
F, =0.60F (5-109)
Members subjected to both axial com-
pression and bending stresses shall be pro-
2. Bending:
portioned to satisfy the following require-
ments:
Fb = 0.66 Fy (5-110)
(1) When//F, < 0.15,
where F is material yield stress, F, is
allowablec tensile
1 stress, and Fb is allowable
f f
bending stress. Js. +J < 1.0 (5-116)
a b
3. Compression:
(2) When/a/Fa > 0.15,
On the gross cross-sectional area of axially
loaded compression members, when kL/r, the cmfb
largest effective slenderness ratio of any un- < 1.0 (5-117)
braced segment, is less than C (\ -^-\F.
F'
2 and, in addition at points braced in the plane

(5-111) of bending,
F.S.
f
+. 1.0 (5-1 18)
where F.S. = factor of safety = 0.6Fy Fb
5/3 +(3/8)e -(I/8)e3 (5-112)
where
Cc =y/2*EIFy (5-113) F = axial stress that would be permitted
5-40
AMCP 706-192
if axial force alone existed under consideration. Ml /A/2 is posi-

tive when the member is bent in
compressive bending stress that single curvature and negative when it
would be permitted if bending mo- is bent in reverse curvature.
ment alone existed
(c) For compression members in frames
1.43X 1J5 braced against joint translation in
F'e = ,Ksi (5-119)
the plane of loading and subjected to
(*V>>2 transverse loading between their
supports, the value of Cm may be
(In Eq. 5-119, for F'e, L, is the actual
determined by rational analysis.
unbraced length (in.) in the plane of
bending, rb is the corresponding radius of However, in lieu of such analysis, the
gyration (in.), and k is the effective length following values may be used.
factor in the plane of bending.)
J_ For members whose ends are
fa = computed axial stress restrained:
fb - computed compressive bending Cm = 0.85

stress at the point under con-
sideration 2. For members whose ends are
unrestrained:
Cm - a coefficient whose value shall be
taken as follows: C = 1.0 (5-122)
(a) For compression members in frames
subject to joint translation (side- b. Axial Tension and Bending
sway):
Members subject to both axial tension and
C = 0.85. (5-120) bending stresses shall be proportioned to
satisfy the requirements of Eq. 5-118 where
(b) For restrained compression members fb and F. are taken, respectively, as the
in frames braced against joint trans- computed and permitted bending tensile
lation and not subject to transverse stress.
loading between their supports in
the plane of bending: Eqs. 5-116, 5-117 and 5-118 are known as
the interaction equations. These equations, of
course, are derived from the linear super-
Cm =0.6 + 0.4-^1-
position of the direct stre-1 under axial load
M2
alone and the bending stress under bending
but not less than 0.4, (.5-121) moment alone. The factor Cm/(1 faIF'e) is
used in Eq. 5-117 to account for the magnifi-
where M1/M2 is the ratio of the cation of the primary bending moment due to
smaller to larger moments at the the axial load. This factor depends upon the
ends of that portion of the member, type of loading and end conditions of the
unbraced in the plane of bending member. The value of the coefficient Cm can
5-41
AMCP 706-192
be derived for various types of loadings and given by Eq. 5-112 includes an allowance for
members, but the values recommended in both of these factors and is adjusted to
Eqs. 5-120, 5-121, and 5-122 are conservative account for their varying influence. For short
and are used in the present work. For a columns, Eq. 5-112 approaches the basic
detailed development and discussion of these safety factor in tension (1.67); and, at e = 1,
equations the reader is referred to Ref. 32. it becomes 15%higher (1.92), a value which is
then used in the case when kL/r exceeds Cc.
The allowable compressive stress formula, Eq. 5-112 is an approximation of a quarter
Eq. 5-115, is derived based on the basic sine wave between the two limits, the curve
theory of column buckling. It is obtained by used in the specification as best representing
dividing the Euler buckling stress by a factor the influence of the two factors. For a
of safety of 1.92. Therefore, F = ir2 7/1.92 detailed discussion of the these factors, the
x (kL/r)2 ] and, taking E = 3.0 x 104 Ksi, Fa reader is again referred to Ref. 32.
= 1.49 x 10s KkL/r)2. Eq. 5-115 is applicable
when the largest slenderness ratio kL/r is The effective length factor k for each
greater than or equal to Cc. Experiments have member of the frame is found from the
shown that when kL/r < Cc , the values of the differential equation
failure stress predicted by the Euler critical
stress formula are seldom attained (Ref. 32).
dly Py n
This is due to the presence of residual stresses T + =0, (5-123)
and other imperfections in fabrication of the dx2 El
members. Therefore, when kL/r < C , the where P is the buckling load, and / is the
values of the allowable stress F are found second moment of the cross-sectional area.
from Eq. 5-111 which is derived based on the
parabolic approximation of the curve-critical The solution of this equation is given by
stress Fa versus the slenderness ratio kL/r in
the range kL/r < Cc. This approximation is
chosen based on the experimental results P P
y{x)=Dx sin/ x + D2 cos /- x
obtained at Lehigh University (Ref. 32). The yj El TJ El
value of the constant C, is found by assuming (5-129)
that the Euler critical stress formula holds
until the critical stress isF12. Therefore, In rigid frames, two cases must be discussed:
(1) frames without sidesway, and (2) frames
a2E
F 2
OTCc = (kL/r)l = y/27r2E/F with sidesway. The transcedental equation
y' ~ {kL/r)\ that comes from Eq. 5-124, while satisfying
where L and r must be expressed in the same the boundary conditions for a member of the
units. frame without sidesway, is given by Ref. 32
1 + GB + G.GJn/k)2 -1 Or/*)
The factor of safety is used to account for (G.A
2
small imperfections of form and loading, and
variations of support and restraint conditions
x sin (ir/k) ~(GA +G)(n/k)2 + 2
from those assumed in computation, which
cause the true effective length to be different
from that calculated. The factor of safety x cos (ir/k) + 2 = 0 (5-125)
5-42
AMCP 706-192
where GA and Q are given by the following equations is used in finding the roots of Eqs.
equations: 5-125 and 5-128 in the present work.
*KJL The interaction Eqs. 5-116, 5-117, and

cA' cA
GAA = (5-126) 5-118 are implemented at the point of maxi-
2/
bA '/AbA mum bending moment for an element. If
there are no loads between the end points of
and an element, then the maximum bending mo-
ment is at one of the ends; otherwise the
ZIrjL
cB' cB actual point of maximum bending moment is
GDB = {5-12 7) found and the interaction equations are
Zhslhi implemented there. As an example, consider
the case of a uniformly distributed load on a
frame element (Fig. 5-25); the moment at a
The subscripts A and B refer to the two ends distance x from the left end is given by
of the member under consideration and the
subscripts c and b refer to the compressed and
restraint members Respectively. For GA , the M = F! F'2x (5-129)
summations extend, over all members that are
connected to joints and for G, the summa-
and
tions extend over all the members that are
connected to joint B. So, for the first case of dM
a frame without sidesway, the value of the (5-130)
effective length factor k must be found by dx
solving the transcedental Eq. 5-125 for each
or
member of the frame.
F'
(5-131)
For the second case, i.e., a rigid frame with
sidesway, the transcedental equation that
comes out of Eq. 5-124 while satisfying the Therefore, from Eq. 5-129
boundary condition for a member of the
frameis given by ^m=-^3+^2')2/(2H') (5-152;
[GAG, i-nlk)2 -36] sin(V^) Eq. 5-132 is used in computing the maximum
bending stress required in the interaction
- 6{GA + GB )(ir/k) cos (v/k) = 0, equations. Now, the implementation of the
interaction equations will be considered one
(5-12; by one and the vectors 9<s/9z and 90s/d6 will
be computed in each case. For the sake of
where GA and G, are given by Eqs. 5-126 simplicity, let ./V be the direct force on the
and 5-127, respectively. Thus, for the case of element, M max be the maximum bending
a frame with sidesway, Eq. 5-128 must be moment, kj be the effective length factor, and
solved for k for each member of the frame. L- be the length of the member to which the
The secant method of nonlinear algebraic ith element belongs.
5-43
AMCP 706-192
b. Interaction Equations: 30s i /bN N db?

bb AiFa \bb 2b.. bb
(1) Interaction Eq. 5-116:
Eq. 5-116, for the maximum stress in the N 3F

ith element, can be written as AtFl bb ZtFb
N Mmax
-1.0< 0. (5-133)
dM
max max bb}
*,= ZF rr) (5-139)
AiFa i bb 4b. bb
where Zj is the beam bending stiffness.

The value of dMm ax /bb can be found from
Eq. 5-136 or Eq. 5-132, which are as follows:
In case Eq. 5-133 is violated, then one must
compute A0s, 90s/9f, b$s/bz, and bQjbb.
Therefore, from Eq. 5-133 30Jbz = Oand ma* *Fi3 bF'6
or (5-140)
bb bb bb
N M_
A0 s = -' 1.0] < 0. (5-134)
U..F Z..F,
rb m^.=m+^li (5-141)
ab ab w ab
Differentiating Eq. 5-133 with respect to z,
The value of bFab required in Eq. 5-139
is found from Eq. 5-111 or Eq. 5-115. First, if
1 bN 1 Mf
b<t>,s = ^,.F 9z +- (5-135) kjLl/rj < Cc> then Eq. 5-111 gives the value of
a Z,.^ 3* F and
The value of bMmax/bz depends on the

bF
expression that defines M, ajc. If Mmax oc- V* V I
curs at an end of the element, then bb 4bf(F.S.) L'+8(F.S.)
Mmax
m/IV =F'3OTF'6
(5-136) b^
x(l-ej)[l-_j
bb
and
3
Vf e.+-
bM_ bFk bFl
(5-137) k. (F.S.) I ' 8 (F.S.)
bz bz bz
M
(1 ^ i '
If Mmax occurs at a point other than the 1 --e2 ] ' (5-142)
ends, then Eq. 5-132 gives its value, and 2 ' I bb
dM_ bF', F\ bF1, where

(5-138)
bz bz w bz
Again, differentiating Eq. 5-133 with respect (5-143)

rC
tob, i c'
5-44
AMCP 706-192
When k.Ljr.* Cc, Eq. 5-115 gives the bM_ bC

+[C-* + M max )I4>
/'^lX
value of F_ and hb bb
bF _FJ>bJ_ 2Fa dk. C,m Mmax bb..

(5-144) (1 +N/AiF'e)-ft-
ab lb, ab k. ab <^l $2 4b
Substituting the appropriate expressions in 1 bN 2N 3/c.l

Eq. 5-139, the value of bQjbb can be found. iTrT-ET-TT, (5-149)
A,F' ab k^Fl ab']
(2) Interaction Eq. 5-117:
where
Eq. 5-117 for the maximum stress in the
ith element can be written as follows: ^2 =
('-} (5-150;
N M
n max (3) Interaction Eq. 5-118:
0 = + - 1.0 0 (5-145)
AtFa
Next, consider Eq. 5-118, which may be
where written as follows for the maximum stress in
the ith element
*i = (5-146)
(' -#'F"
N . Mmax
r 7 + -1.0 0..(5-151)
0.6FyAi ZtFb
If this constraint is violated, then one must
compute A$s, 30S/3J, bifrjbz, and bfijdb. In
this case 3<_/3f = 0 and Therefore, 30s/3f = 0 and
,1 / " CM \ N
Ai=J
S +ggf-i.0\ (5-152;
\0.6FyAf ZtFb ]
Differentiating Eq. 5-145 with respect to z Differentiating Eq. 5-151 with respect to z
andb, one obtains and b, one obtains
_9^ ai, i M \ bMmax

bz \AiFa +
4>l ÂtF'J bz + (5-153)
bz 0.f>FyAt bz ZiFb bz
(5-148)
, bMmax bCn and
+ \C +Mmax
" bz bz
ai, _ 1 V TV bb,
1
and Hb'o.FyA, Ib~ ~ lb, bb
(5-154)
30s 1 JbN N Q N bFa 1
(Mm 3Mmax bb\
bb ~ A.F\X>~ 2b, bb ~' A.Fl db
5-45
AMCP 706-192
It may be noted in the previous equations 54 and 5-5 that follow also are treated in
that Ref. 28 and were first designed for only stress
constraints in order to compare results with
db! (ith) those of Ref. 28.
-(0,.....0,1,0, ,0) (5-155)
ab
1. Example 5-4. Simple Portal Frame
The value of the vectors such as dN/dz,
Fig. 5-26 shows the dimension of the
dNIdb, dF'2/dz, Fib can be found directly
frame. The moment of inertia for each ele-
from the Eq. 5-102. The vectors dCm/dz and
ment of the frame is treated as an unknown,
dCm/db are zero for cases prescribed by Eqs.
and the results obtained are shown in Table
5-120 and 5-122. For the case prescribed by
5-12. The frame was first designed with only
Eq. 5-121, they are computed using the chain
stress constraints. The final weight in this case
rule of differentiation. It remains to find
value of the vector dk./db. This vector can be was 3050.5 lb with a computation time of
computed by differentiating Eq. 5-125 or Eq. 3.74 sec for 13 cycles. At the final design
5-128 with respect to the design variable point, the maximum constraint violation was
vector b. However, due to the fact that both 0.012% for stress in element 2. Optimal
GA and GB arefunctionsof b, this computa- weight reported in Ref. 28 was 3206 lb,
tion is quite tedious and time consuming on which is higher by approximately 5%.
the computer. Another approach that may be
followed for computing akjdb is to use the The frame was also designed by including
method of finite differences; but this ap- all the constraints. The resonant frequency
proach is equally time consuming on the limit for the structure was 25.0 Hz and the
computer. Moreover, it has been observed in final weight obtained in this case was 3803.0
lb with a computation time of 14.60 sec for
the numerical computation that the value of
31 iterations. At the final design point, the
kt does not change appreciably from one
maximum constraint violation was 0.0073%
design cycle to another. Therefore, without
significant loss of accuracy, the value of k. in
a particular design cycle is treated as a
constant. However, at the start of each design
cycle, k values for all the members of the -240 in.-
frame are recomputed. Thus, following this

procedure, dkjdb = (0, , 0). Now, all the
necessary information is available to assemble
matrix A of Eq. 5-20.
180 in.
5-6.3 EXAMPLE PROBLEMS
Several rigid frames were optimized using ,. (D

the computer program based on the algorithm "777 -*-*i
rm777
of par. 5-3.1. All the problems were solved
with stress, displacement, frequency, and de- Figure 5-26. Simple Portal Frame (Example
sign variable constraints. Example problems 5-4)
5-46
AMCP 706-192
TABLE 5-12
SIMPLE PORTAL FRAME (EXAMPLE 5-4)
Design Information: For each element of the frame, the modulus of elasticity, the specific weight, and the yield
stress are 3 x 10* kips/in.2, 0.2836lb/in.3, and 36.0kips/in.2, respectively. The constants a,-, c- and dj are 0.58,
0.58,and 0.67,respectively. The lower limit on the moment of inertia of each element is 1.0 in.4 and there is no
upper limit. The resonant frequency is 25.0 Hz and the displacement limits are 0.5 in. at nodes2, 3, and 4 in both
xv and x2 -directions. There are three loading conditions for the frame; first is uniformly distributed load of 0.5
kip/in. on elements 2 and 3, second is a load of 45.0 kips inx-direction at node 2,and the third is a load of 45.0
kips inx,-direction at node 4.
With All the Constraints

Computation time = 14.60 sec
El. Starting Values, Final Values, El. Starting Values, Final Values,
No. in.4 in.4 No. in.4 in.4
1 1600.0 1091.4 1 1600.0 1995.5

2 1600.0 768.3 2 1600.0 860.3
3 1600.0 768.3 3 1600.0 860.3
4 1600.0 1091.3 4 1600.0 1995.5
Wt, lb 3947.7 3050.5 Wt, lb 3947.7 3803.0
for stress in element 2. Fig. 5-27 shows the in only 5 iterations in the first case and in 7
variation of the objective function as the iterations in the second case. However, in the
iterations progress. It may be noted that, for second case, the cost function continued to
practical purposes, convergence was obtained reduce for a few cycles beyond the 7th
iteration without correcting the constraints.
This was due to the fact that the step size for
4400 1 i r- the problem was too large.
Optimum Weight = 3803.0 lb
3900
2. Example 5-5. One-bay Two-story Frame
i With Only Stress Constraints The present example is also treated in Ref.
Optimum Weight = 3050.5 lb 28. Fig. 5-28 shows the dimensions and the
loading conditions for this structure. Input
and output information for this example is
given in Table 5-13. This frame was first
designed for stress constraints only. The final
0 4 8 12 16 18 weight in this case was 8292.0 lb with a
Iteration Number computation time of 21.47 sec for 32 itera-
tions. Maximum constraint violation at the
Figure 5-27. Iteration vs Weight Curves for
Example 5-4, Simple Portal design point in this case was 0.27 percent for
Frame stress in element 3. The comparable final
5-47
AMCP 706-192
weight reported in Ref. 28 was 8810 lb, The frame was also designed while en-
which is again higher by approximately 5.8%. forcing all constraints. The resonant frequen-
cy limit in this case was 15.0 Hz. The final
, 6K/ft weight obtained in this case was 9722.5 lb
/ 45K with a computation time of 48.84 sec for 32
45K uiuuuniirn]
3 0' 4 III iterations. Maximum constraint violation was
0.38 X 10"3 % for displacement of node 3 in
the x{ -direction. Fig. 5-29 shows variation of
180 in.
6K/ft the cost function with respect to iteration
JL
45K
mimnimii 45K number, and it may again be noted that
convergence was obtained in 8 cycles in both
III
the cases.
180 in.
3. Example 5-6. Two-bay Six-story Frame
0)
Figure 5-30 shows the geometry and
240 in. - dimensions of the frame. This frame has 21
joints, 30 members, and 54 degrees of free-
Figure 5-28. One-bay, Two-story Frame dom. The frame was designed for four loading
(Example 5-5) conditions, and the input and output informa-
TABLE 5-13
ONE-BAY, TWO-STORY FRAME (EXAMPLE 5-5)
Design Information: For each element of the frame, the modulus of elasticity, the specific weight, and the yield
stress are 3 x 104 kips/in.2, 0.2836lb/in.3, and 36.0kips/in.2, respectively. The Constantsa., c;, and d,.are 0.58,
0.58,and 0.67,respectively. The lower limit on the moment of inertia of each element is 1.0in.4 and there is no
upper limit. The resonantfrequency for the frame is 15.0Hz and the displacement limits are l.Oin.at nodes2, 3,4,
5, 6, and 7 in both x, -andx2 -directions. There are three loading conditionsfor the structure, and they are as shown
on Fig. 5-28.
With Only Stress Constraints With All the Constraints

Computation time = 21.47 sec Computationtime = 48.84 sec
1 6400.0 3264.8 1 6400.0 3794.0

2 6400.0 901.4 2 6400.0 1436.3
3 6400.0 801.5 3 6400.0 845.7
4 6400.0 801.5 4 6400.0 845.7
5 6400.0 2598.7 5 6400.0 4618.8
6 6400.0 2598.7 6 6400.0 4618.8
7 6400.0 901.4 7 6400.0 1436.3
8 6400.0 3267.4 8 6400.0 3794.0
Wt, lb 15790.8 8292.0 Wt, lb 15790.8 9722.5
5-48
AMCP 708-192
16000/) the first few iterations that reductions in the

values of the design variables for elements 23,
24, 25, 28, 29, and 30 were relatively smaller
14400 , With All Constraints
than those of other elements. This is due to
Optimum Weight = 9722.5 lb nature of the gradient of objective function
, With QiLy Stress Constraints
for this problem (Eq. 5-106). So, the values of
these design variables were reduced con-
a siderably at the 7th iteration. This is shown
1, by the vertical drop in the graph at the 7th
iteration on Fig. 5-31. In the second case,
where all the constraints were considered,
râ^ôKsô^Ki^
variation of the cost function with respect to
the iteration is shown in Figure 5-32. In this
case, the starting point was infeasible and the
4 8 convergence was obtained in 8 iterations.
Iteration Number
Figure 5-29. Iteration Number vs Weight

Curves for Example 5-5; One-
bay, Two-story Frame
X2
2 3
1
lion for the problem is given in Table 5-14.
3
1
4
2
5
t
144 in
The frame was first optimized by imposing
the stress constraint only. The optimum
5 6 \
6 7
weight in this case was 21706.6 lb with a \
8 9 10 144 in.
computation time of 8.32 min for 21 itera- ''
8 9
tions. At the final design point, the maximum 7
11 12
constraint violation was 0.025% for stress in 13 14 15 144 in.
element number 25. Next, the frame was
designed by imposing all the constraints. The
10
16
11
17
12 fI
optimum weight in this case was 24290.1 lb 18 19 20 144*1.
with a computation time of 8.7 min for 20 14 15 \
iterations. At the final design point, the
maximum constraint violation was 0.0072%
13
23
21
24
22
25
I
144 31.
for displacement in the xl -direction at node 16

17 18 \
1. All other violations were less than that. 26 27
30
\
28 29 144 in.
19 20 21 \
Fig. 5-31 shows variation of the cost mm i mm m ft
function with respect to the iteration number.
The starting point in this case was quite a
Li * 240 in. 240 in.
distance away from the optimum point.

Therefore, a larger step size was used in the Figure 5-30. Two-bay, Six-story Frame
first few iterations. Also, it was observed from (Example 5-6)
5-49
AMCP 706-192
TABLE 5-14
TWO-BAY, SIX-STORY FRAME (EXAMPLE5-6)
Design Information: For each element of the frame the modulus of elasticity, the specific weight, and the yield
stress are 3x 104 kips/in.2, 0.2836lb/in.3, and 36.0 kips/in.2, respectively. The constantsa^c., andd.are 0.58,
0.58,and 0.67 respectively. The lower and the upper limits on the moment of inertia of each elementare 394.5in.4
and 6699.Oin.4, respectively. The resonant frequency of the structure is taken as 4.0Hz and the displacement limits
are 2.Oin. at all nodes in bothx!- and x2-directions. There arefour loading conditionsfor the frame: (1) Uniform-
ly distributed load of -4.0 kips/ft on element 1, 7,11, 17,21, and 27,and 1.0kip/ft on elements2, 6,12,16,22,
and 26; (2)Uniformly distributed load of - 4.0kips/ft on elements 2, 6, 16, 22, and 26,and 1.0kip/ft on
elements 1, 7,11, 17, 21,and 27; (3)Uniformly distributed load of - 1.0kip/ft on elements 1, 2,6,7, 11, 12,16,
17, 21, 22, 26,and 27, and loads of 9.Okips each at nodes 1, 4, 7, 10, 13, and 16 in direction of the x-axis; (4)
Uniformly distributed load of - 1.0kip/ft on elements 1, 2,6,7,11, 12, 16,17,21,22, 26,and 27, and loads of
- 9.Okips each at nodes 3,6, 9,12,15, and 18 in direction of xt axis.
i/Vith Only Stress Constraints With All Constraints

~omputationtime = S.32 min Computationtime= 8.70 min
1 2400.0 450.6 1 394.5 473.8

2 2400.0 450.6 2 394.5 473.8
3 2400.0 498.6 3 394.5 467.2
4 2400.0 394.9 4 394.5 437.5
5 2400.0 498.6 5 394.5 467.2
6 2400.0 530.8 6 394.5 568.7
7 2400.0 530.8 7 394.5 569.1
8 2400.0 394.3 8 394.5 394.5
9 2400.0 397.1 9 394.5 608.5
10 2400.0 394.1 10 394.5 394.5
11 3200.0 481.8 11 450.0 787.0
12 3200.0 481.7 12 450.0 786.4
13 3200.0 425.3 13 400.0 412.7
14 3200.0 472.7 14 450.0 794.1
15 3200.0 425.3 15 400.0 412.6
16 4000.0 521.9 16 550.0 930.2
17 4000.0 521.7 17 550.0 930.0
IS 4000.0 468.3 18 550.0 561.9
19 4000.0 723.5 19 750.0 920.4
20 4000.0 467.5 20 550.0 561.8
2L 4800.0 699.1 21 600.0 1019.1
22 4800.0 699.1 22 600.0 1018.7
23 4800.0 646.3 23 700.0 693.4
24 4800.0 1044.5 24 1100.0 1197.0
25 4800.0 646.5 25 700.0 693.3
26 5600.0 666.4 26 600.0 868.5
27 5600.0 666.4 27 600.0 867.9
28 5600.0 1099.0 28 1200.0 1245.4
29 5600.0 1489.7 29 1600.0 1658.6
30 5600.0 1099.0 30 1200.0 1245.3
Wt, lb 54290.9 2L706.6 Wt, lb 21243.6 24290.1
5-50
AMCP70G-192
54000 5-7 INTERACTIVE COMPUTING IN

STRUCTURAL OPTIMIZATION
48000
5-7.1 THE INTERACTIVE APPROACH
Structural optimization techniques treated

thus far consist of methods which seek to
determine an optimum design, within a well-
defined mathematical structure, by purely
mathematical techniques. A second approach
With Only Stress Constraints
consists of providing the designer with an
24000 interactive computing tool with which he can
try nominal designs, get rapid analysis feed-
back, and alter his initial design based on his
18000
4 8 12 16 18 knowledge of structural behavior. Both meth-
Iteration Number ods have been used with varying degrees of
Figure 5-31. Iteration Number vs Weight Curves success on a variety of design problems. In
for Example 5-6; Two-bay, Six- general, the first approach has been used for
story Frame, With Stress Constraints
Only problems with well-defined optimality cri-
teria, such as minimum weight or maximum
stiffness. The second approach has been used
to aid designers in large scale structural design
problems, primarily airframe design, such as
the Air Force C-5 transport aircraft.
25200
The possibility of utilizing a combination
of these two methods for structural design has
24400 been the subject of a recent paper (Ref. 36).
This paragraph presents the specifics of appli-
With All Constraints cation of the steepest-descent technique with
23600
designer interaction. This hybrid approach is
appealing from a number of points of view.
First, the problem of topological design, i.e.,
22800
determination of optimum structural configu-
ration, has been addressed with very limited
22000 success from an analytical point of view.
Topological design, in practice, is done by
experienced structural designers, occasionally
21200t
0 4 8 12 16 1 with the aid of interactive computation.
Iteration Number Combined analytical and interactive com-
puting methods appear to be essential for this
important class of problems. A second prob-
Example 54>; Two-bay, Six- lem area arises due to the difficulty in
story Frame, With All Constraints formulating a single optimality condition and
5-51
AMCP 708-192
mathematically precise design constraints. algorithm. The result is a hybrid structural

Often, conflicting design constraints and ob- optimization technique.
jectives arise during design which require
experienced judgment and defy a priori math- Reconsidering the design improvement step
ematical formulation. Such problems appear of the optimization algorithm, one might
to require an interactive computing capability draw a vector picture in design space, as is
but should profit from analytical methods depicted in Fig. 5-33. Here, 776ft1 is the
that are used in automated structural optimi- direction which will yield the greatest re-
zation. duction in J subject to the required con-
straints, and 5ft2 is the design change required
Due to unavailability of a large scale, to give the desired constraint error correction.
interactive system, the computations for this While useful in this form, there is a better
study were simulated. Instructions were pre- display of these data for use by the experi-
pared and computations were run in the batch enced structural designer. The scalar com-
mode. Output data were then displayed and ponents of 8b1 and 5b2 tell the designer
analyzed just as they would be in the inter- whether he should increase or decrease his
active mode, and instructions for recomputa- individual design variables to obtain desirable
tion were given by the designer and the changes in overall structural response. Fur-
process repeated. The delay in designer inter- ther, relative importance of design variable
action is felt to degrade performance some- changes is given. For this reason, 6ft1 may be
what, over true interactive computing, since interpreted as a vector of design sensitivity
the designer tends to forget pertinent detailed coefficients that relate individual design
data during the time delay. For this reason, parameter changes to overall structural
the results of this study should provide a characteristics. It is extremely important to
conservative estimate of the designer's per- note, at this point, that these sensitivity
formance in a truly interactive mode. coefficients account for constraints implicitly;
i.e., the direction of change indicated in the
5-7.2 INTERACTIVE STRUCTURAL DE-
SIGN USING SENSITIVITY DATA
The steepest-descent optimization method

developed in this Chapter has been used to
solve a number of relatively large scale struc-
tural optimization problems with good
success. All these problems, however, have
been well formulated mathematically and
have involved structures with a predetermined
form. Difficulties have occurred when certain
structural elements tend toward zero cross
section. Further, no universal method has
been found to determine the best step-size 17
in the optimization algorithm. These and
other inherent difficulties in automated opti-
mization lead one to interject an experienced Figure 5-33. Vector Change in Design
designer into the computational, optimization Space
5-52
AMCP 708-192
design parameters will not cause significant designer a clear picture of the manner in
violation in specified performance constraints which he should change his design parameters
such as stress limits and deflection limits. to reduce total weight, subject to stress
constraints. He can now choose the desired
To illustrate these ideas, consider the reduction A/ in weight and take the resulting
simple structural design problem in Fig. 5-34. design change 8b, or if he wishes, he can input
modified design changes through an inter-
active computer terminal.
There are a number of other respects in

which this mode of designer interaction with
the computer alogrithm is beneficial. First, it
often happens in the automated use of the
algorithm that oscillation of admissible de-
signs occurs because too large a design im-
provement has been requested. Such oscilla-
tion can often be identified by the designer
after only a few iterations and the step size
can be reduced to prevent loss of computer
Figure 5-34. Three-bar Truss time, which can be significant in large scale
problems. Conversely, if an estimate quite far
from the optimum is chosen to initiate the
The cost function here is structural weight. If, algorithm, it often happens that the designer
for example, the stress in member 1 is at its chooses far too small a step size. The result is
allowable limit under one of the loads, then a very small improvement in the design which
the indicated changes in design (50 5, can be sensed by the designer and improved
-8b\, Sbi) will not increase the stress in before excessive computation time is ex-
member 1. To make the design sensitivity pended.
data of maximum use to the designer, con-
sider the graphical display in Fig. 5-35. In this
display, a. are the stresses in the various A second important benefit from designer
members. This display gives the experienced interaction with the algorithm arises due to
the occurrence of local minima and singu-
larities in the analytical formulation of the
design problem. The problem of local minima
is illustrated by Fig. 5-36. Virtually all opti-
mization methods seek local optima and do
not solve the global optimization problem. It
is easy for an optimization technique to get
hung up at point B and not get to point A,
which is the global minima, so the designer
must try different starting points to obtain
the global solution. This is a very time
Figure 5-35. Display of Design Sensitivity consuming and indefinite technique with very
Data few analytical aids to the designer. Part of the
5-53
AMCP 706-192
J(b, z, () changes in configuration as outlined. There

are no general optimization methods, to date,
which will remove a member during iterative
design. The reason is that as a member cross
section goes toward zero, as is required to
remove a member, the equations of structural
mechanics and stress constraints become
singular. This sort of behavior is typical when
the configuration of a system is changed and a
different set of equations is required to
describe the behavior. At the present time,
allowing the designer to make changes in
Figure 5-36. Local Optima configuration appears to be the most feasible
approach, which requires that he play an
active role in the iterative optimization al-
difficulty here arises because Figure 5-36 is gorithm.
the wrong display for the designer in that it
does not utilize his knowledge and experience 5-7.3 EXAMPLE PROBLEMS
with structures.
1. Example 5-7. A Three-member Truss
A much better approach for the designer is
to look at a display such as Figure 5-35. He As an illustrative example of the technique
can use his experience to restart the optimiza- presented in par. 5-7.2, an elementary opti-
tion algorithm at a meaningful distribution of mal design problem will be solved under a
design variables which may be quite different number of loading conditions and a variety of
from the design which resulted from previous constraints. The effect of designer-computer
calculations. His experience, thus, aids him in interaction on rate of convergence is ex-
starting with different trial designs. amined as well as the effect of changing
structural configuration.
Perhaps even more important than trying
various distributions of design variables, the Figure 5-37(A), shows the geometry and
designer can utilize the display of Fig. 5-35 to dimensions of the structure being considered.
change the configuration of the structure This structure has been studied by Schmit
based on information he accumulates during (Ref. 37), Sved and Ginos (Ref. 38), and
iterative design and based on his experience. Corcoran (Ref. 35). Three independent load-
For example, he might try taking member 2 ing conditions are applied to the structure.
out of the structure and optimize based on These are as follows: 40K at 45 deg; 30K at
the modified configuration. Very often, 90 deg; 20K at 135 deg. The allowable stress
significant gains are made in this manner. level for members 1 and 3 is 5 Ksi and for
Precisely this behavior occurs in the three member 2 it is 20 Ksi. The density of the
member truss being considered. material is taken as 0.10 lb/in.3 and Young's
modulus as 104 Ksi. Starting from the feasible
There are actually compelling mathematical solution, bx = 8.0, b2 = 2.4,b3 = 3.2, Schmit
reasons for allowing the designer to make (Ref. 37) arrived at the solution bt = 7.099,
5-54
AMPC 708-192
shown in this figure. The final solution

obtained by Corcoran was bl = 4.241,
b2 = 2.038 with J = 7.55 lb.
Considerable experimentation was done

with this problem. Starting from a feasible
point bt = 10, b 2~ 5, b, =5, the solution
obtained without interaction was bx = 7.064,
b2 = 1.971, b, = 2.835 and the minimum was
/= 15.97 lb. The variation of weight with
respect to iteration number is shown by Curve
(A) Three-bar Truss 1, Fig. 5-38. Next, by adjusting the step size
in interactive computing, the solution was
obtained in only five iterations. This is shown
by Curve 2, Fig. 5-38. It was observed that
member 2 never reached its allowable stress
level. As a second starting point, the area of
member 2 was initially chosen to bring its
stress to the allowable limit. The minimum
reached in this case was the same as before,
Curve 3, Fig. 5-38. Another solution was
obtained by starting from an infeasible point
(B) Corcoran Truss
&i = 5.0, b2 = 1.5, b3 = 0.10. The solution in
Figure 5-37. Trusses (Example 5-7) this case was bx = 6.98, b 2 = 2.30, b3 = 2.68
with./ = 15.971b, Curve 4, Fig. 5-38.
Next, member 3 was omitted from the

b2 = 1.849,^3 =2.897, for which J= 15.986 structure. Starting from a point bx = 10,
lb. Sved and Ginos (Ref. 38) have shown that b2 = 5, the solution obtained was , = 8.0,
this is only a local minima and by omitting b2= 1.5 with J= 12.812 lb, Curve 1, Fig.
member 3, they obtained the solution as 5-39, which is same as reported in Ref. 38. At
by = 8.5, b2 = 1.5 with /= 12.812 lb. They
have also shown that it is impossible to reach
this minimum by an iterative optimization
method unless member 3 is omitted from the
calculations by the designer. Corcoran (Ref. . Curve 2
35) has considered configurational optimiza- !urve 1
tion of this three-bar truss. By using horizon-
tal coordinates of nodes 1, 2, and 3 also as
design variables, he arrived at an optimum
structure shown in Fig. 5-37(B). As a result of
this configurational optimization procedure,
Iteration Number
members 1 and 3 were combined and their
Figure 538. Iteration vs Weight Curves for
orientation is shown by member 1 of Fig. Example 5-7, Three-bar Truss
5-37(B). Member 2 attained an orientation as With Stress Constraints Only
j-jj
AMCP 706-192
bi = 16.0,b2 = 11.31,and J= 33.94 1b. Thus,

the optimum weight obtained for this statical-
ly determinate case is approximately 70%
higher than the optimum weight obtained for
the statically indeterminate case.
It was found that interactive computing

yielded convergence more rapidly than was
Iteration Number the case in the batch mode. It is expected that
even more significant reduction in computing
Example 5-7, Three-bar Truss time will occur in large scale problems.
This problem was also solved by omitting

this point an interesting observation was member 2 from computation. The results
made. The maximum horizontal and the obtained in this case are given in Columns 3
vertical deflections of node 4 were as follows: and 7 of Table 5-15. The truss optimized by
with three bars, zv =0.689 x 10~2 in., Corcoran (Ref. 35) was also solved here by
z 2= 0.595 x 10~2 in.; with two bars, first imposing the stress constraints only and
zi = 0.239 xl0_1ia, and z2= 0.20 x lO'Mn. then by considering all the constraints. The
Thus, although the optimum weight obtained results of these cases are given in Columns 4
by omitting member 3 is approximately 24% and 8 of Table 5-15.
lower than the weight obtained by including
member 3, the deflections of node 4 in the The key point in the solution is that the
former case were approximately four times configuration of the optimum design is not
greater than in the latter case. obvious from analytical considerations. A
designer's experience and insight are required
One might be led to believe that if to select candidate configurations and then
deflection or frequency constraints were en- obtain the optimum design analytically. The
forced, then the optimum structure might not global solution in this case must be chosen by
be statically determinate. To investigate this comparing relative minima. It may be ex-
possibility, displacement as well as buckling pected, in structures with greater redundancy,
and natural frequency constraints were im- that certain members may be removed during
posed. The deflection limits were taken as interactive computation when they are ob-
zx =0.005 in. and z2 = 0.005, and the served to approach their allowable lower
lower limit on natural frequency was taken as limits.
3830 Hz. With the starting point bx = 10,
b2 = 5, b3 = 5, the solution obtained was An interesting point, illustrated by Table
*i = 9.18, b2=2A6, 03=3.85, and 5-15, is that a statically determinate truss is
J = 20.59 lb, Curves 2 and 3, Fig. 5-39. When optimum when only stress constraints are
member 3 was omitted, the starting point was imposed. Quite the contrary, when the full
taken ash] = 10, b2 = 10, Curve 4, Fig. 5-39, range of constraints are imposed, a statically
and as bx = 18, b2 = 10, Curve 5, Fig. 5-39. indeterminate truss is optimum (not consider-
The solution obtained in this case was ing the configurational optimization).
5-56
AMCP 706-192
TABLE 5-15
OPTIMUM THREE-MEMBER TRUSSES (EXAMPLE 5-7)
El With Only Stress Constraints With All Constraints

No. Final Area, in/ Final Area, in.'
1 2 3 4 5 6 7 8
1 7.064 8.500 7.991 Corcoran 9.180 16.00 8.485 Corcoran

Truss Truss
2 1.971 1.500 4.246 2.160 11.310 4.247
3 2.835 4.243 2.039 3.850 8.485 11.410
Wt, lb 15.970 12.812 17.300 7.555 20.59 33.94 24.000 20.115
Max.
Defl, in. 0.00689 0.02390 0.00766 0.02559 0.005 0.005 0.005 0.005
2. Example 5-8. Transmission Tower computation and the Curve 1 of Fig. 5-41
shows the variation of cost function with the
Fig. 5-40 shows the geometry and dimen- number of iterations. The computations of
sions of the transmission tower to be studied. this case were monitored to determine which
This problem has been considered by cross sections went to their lower bounds.
Venkayya and others (Ref. 39). The tower
has 25 members, 10 joints, 18 degrees of One set of members which attained their
freedom, and is designed for 6 loading condi- lower limits of cross-sectional area were num-
tions. The structure is indeterminate, with a bers 10, 11, 12, and 13. It was observed that
degree of indeterminacy of seven. these members carried small forces and could
be removed without causing collapse of the
The tower was designed by first imposing tower, so they were removed from the tower.
only stress constraints, and then by imposing The final values of areas of cross section of
stress, displacement, buckling, and natural the resulting structure are given in Column 2
frequency constraints. Design information is of Table 5-17. Curve 2 of Fig. 5-41 shows the
given in Table 5-16, and the final results variation of cost function with respect to the
obtained are shown in Tables 5-17 and 5-18. design cycle. The final weight in this case was
Table 5-17 shows the results when only stress slightly less than the previous case.
constraints are considered, and Table 5-18
gives the results for the corresponding cases The next member that reached its lower
when all the constraints are considered. For limit was number 1, so it was also removed
results given in Column 1 of Table 5-17, all from the structure. The results of this case are
the members of tower were included in the given in Column 3 of Table 5-17 and Curve 3
5-57
AMCP 706-192
75 in.
Figure 5-40. Transmission Tower (Example 5-8)
5-58
AMCP 708-192
TABLE 5-16
DESIGN INFORMATION FOR TRANSMISSION TOWER (EXAMPLE 5-8)
For each member of the structure, the modulus of elasticity E., the specific weight p/V the constant a. (moment
of inertia of /th member,/. = ai>.2), andthe stress limitsare 104 kips/in.2, 0.10 lb/in.3, 1.0, and 40.0 kips/in.2, re-
spectively. The lower limit on the area of cross section of each member is 0.10 in.2 for the case with stress
constraints only and 0.01 in.2 for other cases. There is no upper limit on the member sizes. The resonant frequency
for the structure is 173.92Hz andthe displacement limitsare 0.35 in. on all nodesand in all directions. There are six
loading conditions and they are as follows (all loads are in kips):
Load Direction of Load Load Directionof Load

Cond. Node x Y z Cond. Node X V z
1 1.0 10.0 -5.0 1 0 10.0 -5.0

2 0 10.0 - 5.0 2 - 1.0 10.0 -5.0
1 2
3 0.5 0 0 4 -0.5 0 0
6 0.5 0 0 5 -0.5 0 0
1 1.0 -10.0 -5.0 1 0 -10.0 -5.0

2 0 -10.0 -5.0 2 - 1.0 - 10.0 - 5.0
3 4
3 0.5 0 0 4 -0.5 0 0
6 0.5 0 0 5 -0.5 0 0
1 0 20.0 - 5.0 1 O -20.0 -5.0

5 6
2 0 - 20.0 - 5.0 2 0 20.0 - 5.0
of Fig. 5-41. The final weight in this case was 16, and 17 reached their lower bounds but
86.94 lb, which is given slightly less than the removal of all of these members rendered a
previous case. Finally, members 14, 15, 16, structure that was geometrically unstable.
and 17 were at their lower limits of cross- However, members 14 and 16 or 15 and 17
sectional area. Removal of any of these could be removed without causing the col-
members, however, would cause collapse of lapse of the structure. Results with members
the structure. Members 2 and 5 or 3 and 4 15 and 17 removed are given in Column 5 of
could be removed to make the structure Table 5-17 and similar results are obtained by
determinate. The results for a statically omitting members 14 and 16 from the com-
determinate structure, obtained by removing putation. The next set of members that were
members 2 and 5, are shown in Column 4 of at their lower bounds and could be removed
Table 5-17. The final weight in this case was without making the structure unstable were
106.97 lb. It may be noted that this statically numbers 1, 12, and 13. These were also
determinate structure yielded only a local removed from the structure and the results
optimum, Curve 4, Fig. 5-41. obtained in this case are given in Column 6 of
Table 5-17. Two other members could be
Another sequence of removing the mem- removed from the structure to make it
bers that reached their lower limits of area of statically determinate. Results obtained by
cross section was also tried. Members 14, 15, removing members 4 and 5, and then numbers
5-59
AMCP 706-192
TABLE 5-17
OPTIMUM TRANSMISSION TOWERS WITH STRESS CONSTRAINTS ONLY

(EXAMPLE 5-8)
Final Area, in.2
El.
No. 1 2 3 4 5 6 7 8
1 0.100 0.100 _ 0.100 _

2 0.376 0.377 0.346 - 0.384 0.364 0.272 -
3 0.376 0.377 0.346 0.100 0.384 0.366 0.272 0.272
4 0.376 0.377 0.346 0.100 0.387 0.363 - 0.272
5 0.376 0.377 0.346 - 0.385 0.365 - -
6 0.471 0.470 0.494 0.779 0.465 0.484 0.775 0.779
7 0.471 0.470 0.494 0.779 0.463 0.482 0.779 0.775
8 0.471 0.470 0.494 0.779 0.464 0.481 0.779 0.779
9 0.471 0.470 0.494 0.779 0.463 0.479 0.779 0.779
10 0.100 - - - 0.103 0.103 0.182 0.182
11 0.100 - - - 0.103 0.103 0.182 0.182
12 0.100 - - - 0.100 - - -
13 0.100 - - - 0.100 - - -
14 0.100 0.100 0.100 0.165 0.151 0.152 0.302 0.302
15 0.100 0.100 0.100 0.165 - - - -
16 0.100 0.100 0.100 0.165 0.151 0.152 0.302 0.302
17 0.100 0.100 0.100 0.165 - - -
18 0.277 0.279 0.292 0.413 0.278 0.288 0.413 0.413
19 0.277 0.279 0.292 0.413 0.277 0.288 0.413 0.413
20 0.277 0.279 0.292 0.413 0.274 0.287 0.413 0.413
21 0.277 0.279 0.292 0.413 0.273 0.287 0.413 0.412
22 0.380 0.374 0.363 0.547 0.445 0.436 0.669 0.669
23 0.380 0.374 0.363 0.547 0.334 0.370 0.447 0.447
24 0.380 0.374 0.363 0.547 0.442 0.436 0.669 0.669
25 0.380 0.374 0.363 0.547 0.336 0.370 0.447 0.447
Wt, lb 91.13 87.90 86.94 106.97 89.94 88.95 113.69 113.68
Max.
Defl. in. 2.288 2.305 2.311 3.489 2.486 2.453 3.614 3.615
2 and 5 are given, respectively, in Columns 7 tained by removing members 1,2, J, 15, 16,
and 8 of Table 5-17. Computations were also 19, and 20 was optimized. The cross-sectional
carried out by removing members 2 and 3, areas of various members at the optimum
and members 3 and 4 along with members 1, point were as follows: 3,4(0.100); 6 to
12, 13, 15, and 17. Results obtained in these 9(0.779); 10,11(0.182); 12,13(0.446);
cases were the same as those shown in 14,17(0.302); 18,21(0.775); 22,25(0.537);
Columns 7 and 8 of Table 5-17. For this and 23,24(0.751). The optimum weight in
reason, these results are not reproduced here. this case was 118.1 lb and the maximum
Finally, another determinate structure, ob- deflection at this point was 3.861 in.
5-60
AMCP 706-192
TABLE 5-18
OPTIMUM TRANSMISSION TOWERS WITH ALL CONSTRAINTS

(EXAMPLE 5-8)
Final Area, n.2

El.
No. 1 2 3 4 5 6 7 8
1
0.010 0.010 0.010
2 2.092 2.339 2.393 - 2.263 2.389 0.548 -
3 2.075 2.386 2.404 0.548 2.264 2.384 0.548 0.548
4 2.095 2.339 2.393 0.548 2.021 1.826 - 0.548
5 2.083 2.385 2.404 - 1.920 1.915 - -
6 2.357 2.085 2.076 7.132 2.389 2.452 6.596 6.699
7 2.354 2.084 2.076 6.857 2.186 2.042 6.483 6.296
8 2.350 2.113 2.083 6.895 2.411 2.430 6.596 6.686
9 2.335 2.1 12 2.082 7.101 2.095 2.123 6.476 6.47 1
10 0.035 - - - 0.666 0.621 2.102 2.054
11 0.035 - - - 0.658 0.630 2.102 2.047
12 0.087 - - - 0.090 - -
13 0.084 - - - 0.071 - - -
14 1.113 1.114 1.139 1.785 1.461 1.485 4.172 4.101
15 1.113 1.114 1.139 1.735 - - - -
16 1.112 1.117 1.146 1.727 1.438 1.498 4.170 4.167
17 1.112 1.117 1.146 1.798 - - - -
18 2.056 2.047 2.027 4.317 2.161 2.171 4.692 4.645
19 2.058 2.034 2.022 4.390 2.158 2.173 4.692 4.664
20 2.046 2.047 2.027 4.400 2.403 2.538 4.985 5.108
21 2.058 2.034 2.022 4.328 2.415 2.524 4.989 5.038
22 2.822 2.878 2.886 5.655 4.187 4.035 6.746 6.909
23 2.808 2.878 2.886 5.730 2.915 2.873 5.086 4.781
24 2.803 2.926 2.895 5.743 4.124 4.086 6.743 7.039
25 2.785 2.926 2.895 5.648 2.908 2.881 5.086 4.749
Wt, lb 590.32 596.64 597.82 060.6 625.37 626.70 142.7 1139.9
Max.
Defl, in, 0.350 0.350 0.350 0.350 0.350 0.350 0.350 0.350
All these tower configurations were also mum weight of the tower increased as more
optimized by imposing all constraints; i.e., redundant members were removed from the
stress, displacement, buckling, and natural structure.
frequency. The results of these cases are given
in Table 5-18. Curves 1 to 4 of Fig. 5-42 show
the variation of cost function with respect to 5-7.4 INTERACTIVE COMPUTING CON-
the iteration number for results of Columns 1 CLUSIONS
to 4 of Table 5-18. It can be observed from
the results of Table 5-18 that, for the casein Computing times for this interactive com-
which all constraints were imposed, the opti- puting approach are considerably shorter than
5-6"1
AMCP 706-192
140
870-1
Curve 4
* 100-
2 4 6
Figure 5-41. Iteration vs Weight Curves for Iteration Number
Example 5-8, Transmission
Tower With Stress Constraints Figure 5-42. Iteration vs Weight Curves for
Only Example 5-8, Transmission Tower
With Al I Constraints
had been experienced when the same prob-
lems were solved in the batch mode. Second,
and probably more significant, interactive
computing allows the designer to alter the
right. For the case when only stress con-
structural configuration in a systematic way
straints are imposed, results of Table 5-15
to seek the global optimum design. This is not
indicate that minimum weight designs for
to say that a mathematically precise method
trusses with multiple loading may be statically
of obtaining a global optimum has been
determinate. However, the results of the
found, for no such method is known. It
appears, however, that the technique pre- second example given in Table 5-17 indicate
that all statically determinate trusses may not
sented here makes strong use of the designer's
be lighter than the indeterminate trusses.
knowledge and intuition, and gives him a tool
with which to seek a global optimum in an
organized way. For the case when all constraints are
imposed, results of Tables 5-15 and J-18 show
The results presented for the two examples that statically indeterminate trusses are lighter
solved in par. 5-7 are of interest in their own than the determinate trusses.
REFERENCES
1. Z. Wasiutynski and A. Brandt, "The of Optimum Design of Structures",/!/;/;/.

Present State of Knowledge in the Field Mech. Rev., Vol. 16, No. 5, May 1963,
5-62
AMCP 706-192
pp. 341-350. Minimum Weight Design of the General

Truss, Division of Solid Mechanics, Struc-
2. C. Y. Sheu and W. Prager, "Recent tures, and Mechanical Design, Case
Developments in Optimal Structural De- Western Reserve University, Report No.
sign",^/?/?/. Mech. Rev., Vol. 21, No. 10, 24, June 1968.
October 1968,pp. 985-992.
10. L. P. Felton and M. F. Rubinstein,
3. R. A. Ridha, "Minimum Weight Design of Optimal Structural Design, preprint of
Aircraft Landing Gear Reinforcement paper presented at SAE, Aeronautic and
Rings", Proceedings AIAA/ASME 9th Space Engineering and Manufacturing
Structures, Structural Dynamics and Meeting, October 1968.
Materials Conference, April 1968.
11. G. Sved and Z. Ginos, "Structural Opti-
4. W. A. Thornton and L. A. Schmit, Jr., mization Under Multiple Loading", Int.
"Structural Synthesis of an Ablating J. Mech., Sei., Vol. 10, 1968, pp.
Thermostructural Panel", Proceedings 803-805.
AIAA/ASME 9th Structures, Structural
Dynamics and Materials Conference, 12. W. A. Thornton and L. A. Schmit, Jr.,
April 1968. The Structural Synthesis of an Ablating
Thermostructural Panel, NASA Report,
5. L. A. Schmit, Jr. and T. P. Richer, "A NASA CR-1215, December 1968.
Structural Synthesis Capability for Inte-
grally Stiffened Cylindrical Shells", Pro- 13. W. M Morrow, II and L. A. Schmit, Jr.,
ceedings AIAA/ASME 9th Structures, Structural Synthesis of a Stiffened Cylin-
Structural Dynamics and Materials Con- der, NASA Report, NASA CR-1217,
ference, April 1968. December 1968.
6. R. Luik and R. J. Melosh, "An Allocation 14. W. R. Spillers and J. Farrell, "On the
Procedure for Structural Design", Pro- Analysis of Structural Design", J. Math.
ceedings AIAA/ASME 9th Structures, Anal, and Appl., Vol. 25, 1969, pp.
Structural Dynamics and Materials Con- 285-295.
ference, April 1968.
15. W. J. Woods, "Substructure Optimization
7. W J. Woods and J. H. Sams, III, "Geo- in the Theory of Structural Synthesis",
metric Optimization in the Theory of Proceedings AI A A 7 th Aerospace
Structural Synthesis", Proceedings Sciences Meeting, January 1969.
AIAA/ASME 9th Structures, Structural
Dynamics and Materials Conference, 16. M. J. Turner, "Optimization of Struc-
April 1968. tures to Satisfy Flutter Requirements",
Proceedings AIAA Structural Dynamics
8. R. L. Barnett and P. C. Hermann, High and Aeroelasticity Specialist Conference,
Performance Structures, NASA Report, April 1969.
NASA CR-1038, May 1968.
17. C. P. Rubin, "Dynamic Optimization of
9. M. J. Schrader, An Algorithm for the Complex Structures", Proceedings AIAA
5-63
AMCP 706-192
Structural Dynamics and Aeroelasticity 27. Manual of Steel Construction, Sixth Edi-
Specialist Conference, April 1969. tion, American Institute of Steel Con-
struction, 1967.
18. R. L. Fox and M. P. Kapoor, "Structural
Optimization in the Dynamics Response 28. Y. Nakamura, Optimum Design of
Regime: A Computational Approach", Framed Structures Using Linear Pro-
Proceedings AIAA Structural Dynamics gramming, Master's thesis, Department of
and Aeroelasticity Specialist Conference, Civil Engineering, M.I.T., Cambridge,
April 1969. Massachusetts, 1966.
19. W. C. Hurty and M. F. Rubenstein, 29. D. Kavhe and J. Moe, "Application of

Dynamics of Structures, Prentice-Hall, Nonlinear Programming to Optimum
Englewood Cliffs, N. J., 1966. Grillage Design with Nonconvex Sets of
Variables", International Journal for
20. O. C. Zienkiewicz, The Finite Element Numerical Methods in Engineering, Vol.
Method in Structural and Continuum l,No. 4, 1969,pp. 351-378.
Mechanics, McGraw-Hill, New York,
1967. 30. D. M. Brown, and A. H. S. Ang, "Struc-
tural Optimization by Nonlinear Pro-
21. R. E. Beckett and J. Hurt, Numerical gramming", Journal of the Structural
Calculations and Algorithms, McGraw- Division, Proc. ASCE, Vol. 92, No. ST6,
Hill, New York, 1967. December, 1966, pp. 319-340 and Vol.
93, No. ST5, October, 1967, pp.
22. S. H. Crandall, Engineering Analysis, 618-619.
31. F. Moses, and S. Onoda, "Minimum
23. T. Kato, Perturbation Theory for Linear Weight Design of Structures With Appli-
cation to Elastic Grillages", International
Operations, Springer-Verlag, New York,
Journal for Numerical Methods in Engi-
1966.
neering, Vol. 1, No. 4, 1969, pp.
24. J. S. Przemieniecki, Theory of Matrix 311-331.
Structural Analysis, McGraw-Hill Book
Co., New York, 1968. 32. W. McGuire, Steel Structures, Prentice-
Hall Inc., Englewood Cliffs, N J., 1968.
25. V. B. Venkayya, N. S. Khot, and V. S.
Reddy, Energy Distribution in an Opti- 33. Proceedings, Army Symposium on Solid
mum Structural Design, Technical Report Mechanics, 1970 Lightweight Struc-
AFFDL-TR-68-158, Wright-Patterson Air tures, Army Materials and Mechanics
Force Base, Ohio 45433, March 1969. Research Center, Watertown, Mass.,
1970.
26. D. Johnson and D. M Brotton, "Opti-
mum Elastic Design of Redundant 34. J. Arora, Optimal Design of Elastic Struc-
Trusses", Journal of the Structural Divi- tures Under Multiple Constraint Condi-
sion, Proc. ASCE, Vol. 95, No. ST12, tions, Dissertation, University of Iowa,
December, 1969,pp. 2589-2610. 1971.
5-64
AMCP 706-192
35. P. Corcoran, "Configurational Optimiza- ference on Electronic Computation,

tion of Structures", International journal Structural Division of ASCE, September
of Mechanical Sciences, Vol. 12, 1970, 1960,pp. 105-132.
pp. 459-462.
38. G. Sved and Z. Ginos, "Structural Opti-
36. R. Douty and S. Shore, "Technique for mization Under Multiple Loading", Inter-
Interactive Computer Graphics in De- national Journal ofMechanical Sciences,
sign", Journal of the Structural Division, Vol. 10, 1968,pp. 803-805.
ASCE, Vol. 97, No. ST1, January 1971, 39. V B. Venkayya, N. S. Khot, and V. S.
pp. 273-288. Reddy, Energy Distribution in an Opti-
mum Structural Design, Technical Report
37. L. A. Schmit, "Structural Design by AFFDL-TR-68-158, Wright-Patterson Air
Systematic Synthesis", Second Con- Force Base, Ohio 45433, March 1969.
5-65
AMCP 706-192
CHAPTER 6
THE CALCULUS OF VARIATIONS AND OPTIMAL PROCESSTHEORY
6-1 INTRODUCTION Example 6-1: The shortest path between

two points, {t,x) and (t\xl), in the t-x
The problems of Chapters 2 through 5 are plane is to be found. As shown in Fig. 6-1, the
all optimal design problems in which the particular path chosen between the two
design variables were elements of R", i.e., a points has a length associated with it. The
vector of n real numbers uniquely specified problem is to choose the curve x(t), t a t Q
the design of the system being investigated. In tl which has the shortest length. For a
many important, real-world, optimal design smooth curve x(t) the length is given by
problems the design of a system cannot be
1/2
specified so easily. For example, the thrust dx
vector acting on a rocket during takeoff must J(x) = 1 + dt. (6-1)
dt
be continuously oriented in time so that the fl
rocket remains stable and follows a certain
Note that in this example the quantity J(x)
path. In this example, the angles the thrust
to be minimized is a real number once the
vector makes with the rocket must be speci-
function x(J), t Q t Q f1 is chosen. In this
fied at each instant of time during takeoff. It
sense J(x) is a real valued function of a
is clear that a function specifies the thrust
function or curve.
direction rather than a finite number of
parameters.
Example 6-2: (The Brachistochrone): Giv-
Examples of this kind of problem abound en two points (t,x) and (t1 ,x*) in a vertical
plane that do not lie on the same vertical line,
in the aircraft guidance literature and in the
find a curve x(t), t < t a t' Joining them so
optimal control literature. Typical design or
control variables in these problems are thrust, that a particle starting at rest will traverse the
motor torque, control surface settings, etc. curve without friction from one point to the
All these variables must be specified through-
out the entire interval of time an aircraft is in
the air. Similar problems arise in the presently
developing field of optimal structural design.
In this field the design variables are generally (/',!')
variables that describe the distribution of
material in structural elements.
In order to illustrate the kind of problem

to be treated in this chapter, two classic
examples will be given. Figure 6-1. Shortest Path
6-1
AMCP 706-192
other in the shortest possible time. Candidate

curves are shown in Fig. 6-2.
T = J(x)
[2g(x - x0)] ln
(6-4)
This notation makes it clear that T depends

on the entire curve transversed by the parti-
cle. The Brachistochrone problem, therefore,
is reduced to finding a curve x{t), t < t a t1 ,
that passes through the two given points and
Figure 6-2. Curve for Minimum Time makes T as small as possible.
In Examples 6-1 and 6-2 it is clear that a

curve, or equivalently a function character-
Let m be the mass of the particle andg be
izing the curve, is to be found as the solution
the acceleration due to gravity. Since the
of the optimization problem. Further, the real
particle starts at rest at (t,x) and there is no
valued quantities to be minimized are deter-
friction,
mined by curves or the functions character-
izing those curves. These real valued quanti-
mv2 = mg(x x) (6-2) ties, therefore, are functions of functions.
Such a real valued function is called a
functional. The functional notation J(x) in
where v is velocity, Eqs. 6-1 and 6-4 is then interpreted as a real
valued function of the function x(t),t Q(<
1/2
t1 . The most common kind of functional
dr/ +
\dT/ encountered in calculus of variations is the
integral.
1/2
1 + fdx -dt-
(6-3)
dt dr' The optimization problem considered here
might be stated as: find the function x(t), t
< t < t1 , that minimizes the functional/(x).
where T is time. Solving Eq. 6-2 for v,
A glance at the functional defined in Eqs. 6-1
substituting this into Eq. 6-3 and solving for
and 6-4 reveals a basic flaw in this statement
dr yields
of the optimization problem. In both cases,
the functional are defined only if the func-
1 + dt
tion x{t) has anintegrable derivative on( Q /
dr = s tl, i.e., it doesn't make sense to admit all
[2g(x -x0)}1'2 functions as candidates for an extremum. The
problem is more reasonably stated: find the
The total time T required for the particle to function x(t) t Q t < t', in a class of
remove from(f,x) to (t{,xl) is then functions D, that minimizes the functional
6-2
4MCP 706-192
J(x). The admissible class of functions here spaces may be described in a similar manner
plays a role similar to the constraint sets of as
Chapters 3, 4, and 5.
C'(a,b) = {x(t), a-i t< b \ x(t) has i
The idea of classes of functions required continuous derivatives ( . (6-6)
here is basic to the mathematical field called
Functional Analysis. Classes of functions in It should be understood here that x(t) may be
this field are called function spaces. Consider, a vector valued function and the differ-
for example, the collection of all continuous entiability requirement in Eq. 6-6 refers to
functions x{t) on0< (< 1. The graphs of each component.
several such functions are shown in Fig. 6-3.
Function spaces may be thought of as sets
of elements, where elements in the function
space are really curves or functions. In this
way the problem of minimizing J{x) may be
viewed as picking the element (curve) in the
appropriate function space that makes J(x) as
small as possible. This approach makes mini-
mization of a functional sound very similar to
the programming problems of Chapter 2. With
this mental analogy one may begin his study
Figure 6-3. Examples of Continuous Functions of the calculus of variations armed with a
powerful intuitive tool.
It is clear that there are infinitely many The basic ideas of function space theory
continuous functions but that not all func- are presented very clearly in Ref. 1, Chapter
tions are contained in this class. For example 2.
0 <s t t 1/2 In connection with vector spaces, it is often

x(t) = necessary to require that a function is small,
1, 1/2 < t < 1 or near the zero function. For this purpose it
is required that size of a function be defined.
is not continuous so it is not in the class. This is done by defining a norm as a function-
al IMI on the function space of interest with
To expedite the development that follows, the following properties:
some notation will be introduced. The collec-
tion of continuous functions on 0 <s t < 1 || x || > 0, || x || = 0 implies x is the zero
described previously is called a function space function (6-7)
and is denoted
|| JC || = || a || || x || for real a (6-8)
C(0,1)= \x(.t),0< t< 1 \x(t)
is continuous } . (6-5) \\x+y || < || * || + || j; || . (6-9)
A large number of important function Examples of norms include
6-3
AMCP 706-192
max x(t) (6-10) xft)
for xeC(?V; and i(t)+i -J

_ - - ~Z xft)
f"
1/2 X(t) - 5
x\\ = x2(f)df (6-11)
J ,u
for square integrable functions x(f)- For a

discussion of the basic ideas of functional
analysis as they apply to optimization theory
the reader is referred to Ref. 1. Figure 6-4. A Neighborhood of x(t)
With the idea of norm defined, one can

speak of relative minima of functionals. The
functional J(x) has a relative minimum at some modification of the basic optimization
xeD if there is a 6 > 0 such that problems. Without a thorough knowledge of
the theory, the designer will probably have no
/(x) < J(x) idea of how to modify the existing theory to
suit his purposes.
for allxe> with
6-2 THE FUNDAMENTAL PROBLEM OF
< 6 (6-12) THE CALCULUS OF VARIATIONS
This simply says that J(x) has a minimum in a Examples 6-1 and 6-2 have features in
sufficiently small neighborhood of 2. It is common that allow for the formulation of an
interesting to look at a neighborhood of a entire class of problems containing these two.
curve in C(t,tl) where norm is defined by For the sake of generality, let the variable x(t)
Eq. 6-10. In this case Eq. 6-12 simply be a vector valued function of the real
demands that x(t) be within 6 of x(t) for all t variable t, i.e.,
in t Q t ss t1. The neighborhood of x in this
case is simply the collection of all continuous x,(0
curves which can be drawn between x(t) + 6 x(t) = (6-13)
andx(t) 6, as shown in Fig. 6-4.
xn(t)_
The present chapter will be devoted almost
exclusively to the theory of the calculus of where x;(f) are real valued functions of t.
variations and optimal process theory. Con-
structive methods for these problems will be The problem considered here may be for-
treated in the chapters to follow. A knowl- mulated as Definition 6-1.
edge of this basic theory is essential for
successful application of the theory of opti- Definition 6-1 (Fundamental Problem of
mal design. It has been the experience of the the Calculus of Variations): Find a function
author that most real-world problems require x(0 inC2^0,/1) which satisfies
6-4
AMCP 706-192
x.(t) = x , for some indices k ia such candidate that must then be the solu-
tion. If there are several candidates, other
x- (tl) = x J, for some indices 1 <s ;' s n methods must be used to choose the solution.
This problem will be discussed later.
(6-14)
Graphically, the method of obtaining con-
and which minimizes ditions on the solution, x{t), of the funda-
mental problem will be to allow small changes
J(x) = \
f Fit, x, x')dt (6-15) inx(0 and examine the behavior of J(x). An
admissible, small perturbation is illustrated in
Fig. 6-5. The equation for this curve is x(t) +
where F is a real valued function of all its
arguments and
' dxx
dt
(6-16)
L- dt Figure 6-5. Perturbation from Optimum
If the reader wishes he may consider x(t) as

being a real valued function of /, the general- 67?(0 where is a small real number and TJ(0
ization to vector valued functions is simply a is any member of C2(t,tl) such that
matter of notation. The conditions, Eq. 6-14,
specify some or all of the components of x(t) Vj(.t) = 0, for i withx.(/) = xj0'
at the end points of the interval t Q / < t'. ) (6-17)
This corresponds to demanding that the rift1) = 0, for/' with ^.(f') = xj 1
curves in Examples 6-1 and 6-2 pass through
given points.
To examine the effect of this perturbation

6-2.1 NECESSARY CONDITIONS FOR of J{x), substitute x +er\ inEq. 6-15,
THE FUNDAMENTAL PROBLEM
t1
Only necessary conditions for solution of
the fundamental problem of Def. 6-1 will be
developed here, i.e., the existence of a solu-
J(x + er})= |
I t
F{t, x + er), x' + er\')dt
(6-18)
tion, x(t), inC2(?,f') first will be assumed. A
set of conditions that x(t) must satisfy then Recall that x(t) is a local minimum of J(x)
will be derived. These conditions then may be subject to Eq. 6-14, i.e., any small change in
employed in particular problems to find x{t) increases J{x). For any given function
functions x(t) that are candidate solutions of q(t) mC2{t,t') and satisfying Eq. 6-17, x{t)
the problem. Hopefully, there will be just one + er?(0 is in C2(t,tx) and satisfies Eq. 6-14
6-5
AMCP 706-192
for all E. Therefore, for this r\(t), J(2 + 617) is a be treated independently, i.e., each is required
real valued function of e. Further, for E - 0, to be zero. One of the major results which
J(x + erj) has a relative minimum and it is follows is a direct application of Lemma 6-1.
assumed that F(t,x,x') is twice continuously
differentiable in x and x' so that J(2 + erj) is a Lemma 6-1: If M(t) is a continuous func-
twice continuously differentiable function of tion on t <i t < t1 and if
E. Theorem 2-2 then applies, and it is required
that
n
J(x + eri)
ae k=o
0. (6-19)
f M(t) 77(f) dt = 0 (6-22)
for all 77(f) in C2(f, t1) withr?(f) = 77a1) = 0,

then M(t) = 0, f < t < tl.
The object now is to transform condition,
Eq. 6-19, into conditions onx(f). Performing
The ideas involved in the proof are easily
the differentiation indicated in Eq. 6-19,
seen graphically. In Fig. 6-6 a point t*, t <
ae ie=o
V+ r}'}dt = G, (6-20)
X X I
4 (
where the arguments in the partial derivatives Figure 6-6. Graphical Proof of Lemma 6-1
of / in Eq. 6-20 are x(t) and x'{t). It is
important to remember that Eq. 6-20 is
required to hold for any rj(t) in C2(t,tl) t* < t1 , is shown where M^t*) = 0. The curve
which satisfies Eq. 6-17. 77.(t) is then constructed so that neither
function is zero in the interval a < t < b.
Integrating the second term in the Their integral over the entire interval is then
integrand of Eq. 6-20 yields nonzero which is a contradiction of Eq. 6-22,
so M{(t*) = 0.
oF d OF
; Vdt Since the two terms in Eq. 6-21 must each
dx dt ox'
be zero,
i-^-T [t',x(tl),x'(ti)~]r,(t1) (6-21)

ox
dF
I! dx " dt \ox')
r}dt = 0 (6-23)
Since the behavior of 77 inside the interval t for all 77(f) in C2(t,tl). In any subinterval of
< t < tl and at its ends are independent, the t < t < tl where x(t) is continuously
integral and boundary terms in Eq. 6-21 may differentiable, the quantity
6-6
AMCP 706-192
[dF/dx d/dt (dF/dx')] is continuous. forallt?(?), r\{tl) satisfying Eq. 6-14, and
Therefore Lemma 6-1 implies
f-[f_0,Jc(/-0),Jc'(f-0)]
dF dF dx
=0 (6-24)
dx dt \dx
ML fff 0,x(f+0),x'(f+0)] (6.28)
in that subinterval. dx'

at each point t of discontinuity of x'(t).
If, however, x(t) has ajump discontinuity
at some point i then [dF/dx d/dt (dF/dx1)] Condition, Eq. 6-26, is a second-order
need not be continuous at i and Lemma 6-1 differential equation in x(t) and is called the
may not be applied over any subinterval Euler-Lagrange equation. Condition, Eq. 6-27,
containing t. Since Eq. 6-24 must hold in is called a transversality condition. For each i
subintervals on both sides of t, this equation or; such that i?-(/) or rjAt1) is not specified
may be integrated from t 6, 6 > 0, to t to by Eq. 6-14, Eq. 6-27 implies dF/dx] (t) = 0
obtain or dF/dx'. (r1) = 0. The condition, Eq. 6-28,
at discontinuities (called corners) in x(t) is
- f'
17'" J
dF
dx
dt + C. (6-25)
called the Weierstrass-Erdmann corner condi-
tion.
One further necessary condition will be

The vector dF/dx is piecewise continuous so
important for further development. Define
the right-hand side of Eq. 6-25 is continuous.
the Weierstrass E-function as
Therefore dF/dx' is continuous even at t.
E(t,x,x',w) = F(t,x,w) -F(t,x,x')

These results may be stated in the form of
(6-29)
a theorem.
dF
Jtx.x'Kw-x').
ax
Theorem 6-1: The following conditions
must be satisfied by the solution of the
The proof of the Weierstrass necessary condi-
problem of Def. 6-1, x(t), whose derivative is
tion may be found in Ref. 2, page 149. The
piecewise continuous;
result only will be given here as Theorem 6-2.
W,*(r),*'(r)] Theorem 6-2: If the function x{t) is the

dx
solution of the problem of Def. 6-1, then it is
d I dF necessary that
(6-26)
EU,x(t),i'(.t),w] > 0 (6-30)
at points of continuity of x(t)
for all t < t < t' and all finite w.
^7[r ,x(r ).i'(' )]r?(r )
, I , 1
dx The Weierstrass condition of Theorem 6-2
generally is not used to generate candidate
[t,x(t ),x'(t )]r)O) = 0
o o
(6-27)
dx solutions of the fundamental problem.
6-7
AMCP 706-192
Rather, when solutions of Eq. 6-26 are This is simply an algebraic equation be-
determined, Eq. 6-30 is used to eliminate tween t and x. Since there will be no
unsuitable functions, i.e., it may very well constants of integration, it will not generally
disqualify a function which satisfies Eq. 6-26. be possible to pass the resulting curve through
particular points. This means that a solution
The derivation of necessary conditions for to such a problem generally will not exist.
the fundamental problem is only very lightly
covered here. Further, the theory of sufficient Example 6-3: Minimize
conditions is completely neglected. For out-
standing and complete treatments of these
topics see Refs. 2, 3,4, and 5. 1 x2dt
J 0
6-2.2 SPECIAL CASES AND EXAMPLES
for
In many problems the form of the function
x(0) = 0, x(l) = 1.
Ftt,x,x') allows for simplification for the
Euler-Lagrange equation, Eq. 6-26. In any The condition, Eq. 6-33, is
case, Eq. 6-26 may be written, using the chain
rule of differentiation and the notation
2x = 0.
3F d2F
F Fx , and But it is, therefore, impossible to satisfyx(l)
* ~?)x' ' ~dx"F'x' dtdx
= 1 so the problem has no solution.
2
F
F To get an idea of what has gone wrong,
x'x a \ax
note that since x2(t) > 0 for each t,
to obtain
T TF
1x2 (t)dt > 0
Fx-Fx;-x' Fl'x-x" *'*
0
0. (6-31) for any curve on 0 *s t < 1. It is, therefore,

clear that if there were a curve which mini-
This is simply a second-order differential
mized / l x2dt, then the minimum value of
equation for x(t). / o
the integral would be non-negative.
Several special cases with examples will
now be considered. It was noted that no minimum exists.
However, consider the family of curves
Case 1. F does not depend on*';
x, (t)=t".
F = F{t,x). (6-32)
These curves all satisfy the end conditions and
Eq. 6-3 1 in this case is
J(x)= I t2ndt=-
F(t,x)=0. (6-33) 2n+ 1
6-8
AMCP 706-192
Therefore, it is possible to choose n large The function F from Eq. 6-1 is

enough so that / x*dt is as close as desired
to zero. However, the limit of *(0 as n F=[\+{x,)2]ll\
approaches infinity is the function
The form of the Euler-Lagrange equation in
Eq. 6-35 applies in this case to yield
x (0!
- [i + u')2]"3/2 x"=0.
and this is not even a continuous function. Since OO2 > 0, [ 1 + (x')2 ] * 0 and x'(t) is
The class of functions xn (t) are illustrated in required to be continuous so [ 1 + (x')2 ] =
Fig. 6-7. and it, therefore, is required that
x"(t) = 0
or
x{t) =at + b,
where a and b are constants. This implies that

o 1 t the shortest path between two points in a
Figure 6-7. Minimizing Sequence plane is a straight line. This shouldn't shake
anyone up.
In this illustration, a solution of the prob- The end conditions yield

lem exists in the class of piecewise continuous
functions but not in the class of twice x(0) = b = 0
continuously differentiable functions. This
problem, therefore, should serve as a warning and
that not all innocent looking calculus of
variations problems have solutions. x(l) =a= 1
Case 2. F depends only onx': Therefore the solution of the problem is
F=F(x'). (6-34) x(t) = t.
Eq. 6-3 1 is in this case Case 3. F depends only on t and x':
Fx,x.x" = 0. (6-35) F = F(t,x'). (6-36)
Example 6-4: Using the formulation of Eq. 6-26 is, in this case,
Example 6-1, find the shortest curve in the t-x
plane which passes through the points (0,0)
Fxx.(t,x') = 0
and (1,1). dt
6-9
AMCP 706-192
or This reduces to
Fx.(t.x') = C (6-37) 1={ 2gx[1+|VJ2]} 1/2

C,
where C is an arbitrary constant. or
Case 4. x is a real valued function, and F xi^+fx')1] = C,

depends only on* and*':
where Cj is a new constant.
F=F(x,x'). (6-38)
The solution of this differential equation is
Eq. 6-3 1 is, in this case, a family of cycloids in parametric form
Fx-Fx'xx'-Fx.x.xH = 0. t = C2 + r(s - sins)

2
Multiplying by x' yields
and
x'Fx-(x'PFx.x-x'x"Fx.x. = 0.
x = *- (1 coss).
2
This is just
The constants C\ and C2 are to be deter-
(F-x'F,)=0,
x
mined so that the cycloid which passes
dt through the given points is fixed.
so It should be noted that each of the

problems treated here reduced to the solution
F-x'Fx.=C (6-39) of a nonlinear differential equation. This is
characteristic of problems of the calculus of
where C is an arbitrary constant. variations. The reader is undoubtedly aware
that it is only in the simplest cases that closed
Example 6-5: Solve the Brachistochrone form solutions of these differential equations
problem of Example 6-2. may be obtained. Further, questions of exis-
tence and uniqueness of solutions are by no
means trivial.
The function F from Eq. 6-4 is
1/2 6-2.3 VARIATIONAL NOTATION AND
1 +(x >)2
SECOND-ORDER CONDITIONS
2gx
Eq. 6-39 applies in this case and yields ForJ(x)= F(t,x,x')dt,

1/2
~\+(x')2~ (x'P
l/2 1/2 define the first variation ofJ(x) as
_ 2gx {2gx) [\+(x'P]
8J(x) = -d-J(x+e8x) (6-40)

=C de e=o
6-10
AMCP 706-192
A"
de Jf
F(t,x + eSx,x'
T 2
d F
r ~^ {t-x'x')bxT
x
+ e8x')dt T
e=o b2F
n-
+ 28x , (t,x,x')bxf
dxdx
3F
(t,x,x')bx
dx
T 2
d F
+ 8x' (t,x,x')bx' dt
dx2
dF
+ ,(t,x,x)bx dt.
dx'
Define
Note that all this does not require that x(t) d2F
be the solution of the fundamental problem.
dx2
If, however, x(t) = x(t) is the solution of the
fundamental problem, then it is clear from
d2F
Eq. 6-19 that it is necessary that 5=2
dxdx'
8J(x) = 0, (6-41)
and
for all 8x(t) for which x(t) + ebx(t) satisfy the
d2F
end conditions in the fundamental problem. C=
dx'
In a way quite similar to the definition of
With this notation,
the first variation, the second variation may
be defined as
d2
b2 J(x) = J(x + eSx)
de E=0
b2J(x)=\
I"
If 0
(bxTASx + oJBbx'
+ 8x'TC8x')dt
Performing the differentiation, this is
If F(t,x,x') has three derivatives, then by
J(x)--
d2
T" I F(t,x +e8x,x' +i 8x') dt
e=o
Taylor's formula
J(x+ebx)=J(x) + \
de
le=o
Sx1
d2J d*J
e< +
"de2 "d?
bF1 e=o e=e
(t,x +ebx,x' +e5x')
dx
(6-42)
T
dFT
+ 8x' (t,x + ebx.x + ebx') dt
dx' e=o where 0 < e < E. If we computed d3J/de3, it
6-11
AMCP 706-192
would involve a sum of terms each containing which has large derivatives. One might, there-
third degree terms in 8x and 8x'. For 8x and fore, be led to believe that the derivative term
8x' sufficiently small, this term may be in the inequality of Eq. 6-43 is dominant.
neglected to obtain a second-order approxi- This would then require C to be positive
mation when E = 1 so that semi-definite.
J(x + 8x)~J(x) + 8J + 82J. To show that this is the case, assume that
there is a point t", t < t < t1 and a nonzero
It is clear then that 6J and 82J play the vector h such that hTC(t*)h = - 2 < 0. For
role of differentials in the theory of func- any continuous bx'(t) such that dx'ft*) = h,
tionals. there is an interval
Further, if x(t) yields a relative minimum t*-< +

t< t* a> 0, suchthat
for the fundamental problem, then/fx + e8x)
is a relative minimum at E = 0. It is, therefore, bx'(t)TC(t)bx'(t)< -< 0
necessary that
inf*-a< t < t* + OL-
d2J
>
de2
e=o Define
This is just / OL
h sin I \,t*-a< t< t*+a
7T
2
8 J( x)> 0 Sx(t) = '
0, elsewhere
or
so that
(8xTA8x+8xTB8x' I \ir{t - r*)l ^
h cos , t* <x< t < t + a
Sx'(t)
+ 8x'TC8x')dt> 0 (6-43)
0, elsewhere
for all 8x(t) such that x + 8x satisfy the end

conditions for the fundamental problem. In Now, Eq. 6-43 is
what follows it will be convenient to limit
8x(t) to those variations which satisfy 8x(t) 0< I (8xTA8x + 8xTB8x' + 8x'TC8x') dt
= 8x(tl) = 0.
If 8x'(t) is small for all /, then 8x(t) must

also be small since 8x(t) = 0. On the other
hand, it is possible to choose 8x(t) which is hTAh
zero at the endpoints and small for all t, but
6-12
AMCP 706-192
incorrect. This implies /; C/i s> 0 for all A and

Bh t . Therefore C(t) is positive semi-definite.
Since
+ COS2 [at]. \hTCh) dt C(t) =

d2F
dx'2
this result may be stated as Theorem 6-3.
sin2 6 dd Theorem 6-3: A necessary condition for

J IT the fundamental problem to have a relative
minimum at x(t) is that
b2F
Isin0 II cos 0 \d6 - [t,x(t),x'(t)]
dx
J it
be positive semi-definite for allf, t < t < t1.
Gelfand and Fomin (Ref. 2, p. 104) indi-

cos2 0tf0, cate that people are prone to argue that
/ 7T
positive definiteness of d2F/bx'2 at each
point of the solution is a sufficient condition
for an extremum. They point out, however,
where that this is not the case and, in fact, that no
local condition can provide sufficient condi-
M_ nyx \hTA(t-)h I tions. For a treatment of sufficient conditions
see Refs. 2, 3,4, and 5.
P= max \hTB{t)h I
f
6-2.4 DIRECT METHODS

Therefore, integration in the preceding in-
equality yields The direct methods of the calculus of
variations seek to generate a sequence of
functions [x>n> (t)] such that, if is the
1 \ infimum oiJ(x) over all admissible x, then
0< M+ P- -et-
\7r a I
lim
J[x<">\ = . (6-44)
However, since a may be chosen arbitrarily

small and > 0, the right side will be negative Direct methods are capable of showing
for sufficiently small a. But this is a con- existence of solution as well as construction
tradiction. Therefore, the assumption that of approximations of the solution. It is
there exist t* and h such that hTCh < 0 is generally very difficult to prove existence of a
6-13
AMCP 706-192
solution of the nonlinear boundary-value where a. are constants. Classical, trigo-

problem in the necessary conditions for the nometric Fourier series is an example of this
fundamental problem. It is often possible, kind of representation.
however, to show that the sequence [x(n) (t)]
converges to a function x(t) which is the In the Ritz Method, the nth function in the
solution of the fundamental problem, i.e., minimizing sequence is formed by
lim
J[x<n>] =J(x) = Z. (6-45) x<n>(t)= 2 a.4>t(t) (6-46)
It is clear, however, that a sequence which where the 4>/t) are chosen so that x^nHt)
satisfies Eq. 6-44 may very well fail to satisfies the end conditions associated with
converge to an admissible function (?/ This the fundamental problem. This expression k
must necessarily be the case if no solution of then substituted into J(x) to obtain
the fundamental problem exists. From an
engineering point of view, one may not be too J[x<n>] =
concerned with existence of a limit of the
sequence [x<nl (t)] .Provided it is possible to
successively reduce /, consistently better re- I'1
sults are being obtained and the process will FL.Z afaU), arfft^dt. (6-47)
be continued until no further meaningful
reduction in / may be achieved. For an The object now is to choose the coefficients
outstanding treatment of convergence of di- ap i = 1, ..., n, so that J[x^"^ ] is as small as
rect methods, see Ref. 2, page 192, and Ref. possible. For this purpose, it should be noted
3, page 127.
that the right side of Eq. 6-47 is simply a
function of parameters. The problem is now
The problem of primary interest to the
to minimize this function without any other
engineer is the construction of a minimizing
constraints. For this purpose, any of the
sequence. There are many ways of generating
methods of Chapter 2 may be used.
such a sequence, only two of which will be
treated here. These methods are known as the
Ritz Method and the Method of Finite Differ- The property
ences.
J[x(" + l>} < J[X("1]
6-2.4.1 THE RITZ METHOD
follows readily from the method of deter-
The Ritz Method is based on the idea of mining the af. It is clear that by choosing
representing functions by using linear com- an+l = 0, x(n+1) (0 = x'"'(t). However, by
binations of known functions; i.e., given 4>/t), allowing a+l to be nonzero, a larger number
i = 1,2, . . . which preferably form a complete of functions are available as candidates for
set, a function is represented by minimum of J[x<-n+1)] than J[x<") ]. The
minimum of /[x(" + 1)] will, therefore, cer-
tainly not be greater than that of J[x<n> ] and
y(t) = 2 a.<t>.(t)
/= l ' ' this is the desired result.
6-14
AMCP 706-192
In practice, the rate of convergence of This is a positive definite quadratic form, so it

[x<")(t)] depends strongly on the functions has a unique minimum which may be ob-
<t>/t) chosen. The number of terms required to tained by setting its first derivatives equal to
obtain a reasonable approximation of the zero. This yields
solution is greatly reduced if these functions
are chosen based on a reasonable engineering 38 20 1
estimate of the form of the solution. By a, + a2 =0
105 ' 45 2 - 3
making a judicious choice of the 4>/t), a good
approximation of the solution may be ob- 20 , 2488 2
tained with as few as two or three terms. a, + a2 =0.
45 3465
Example 6-6: In solving the boundary- The solution of these equations is

value problem
! = 0.9877
x" + (1 +t2)x +1=0
a2 = - 0.05433.
x(- 1) = x(\) = 0,
Substituting these coefficients into Eq.
it is necessary to minimize the functional 6^9,
J(x)-
i: [x'2 _(i + t2)x2 -2x]dt
(6-48)
x<2)(t) =75fe(3969 - 4200r2 + 23lr4).
In particular,
subject to the end conditions x(\) = x(l) =
0. x<2)(0) = 0.93344.
In order to minimize J(x) of Eq. 6-48, by If the three term approximation

the Ritz Method, choose
xO~> =aj(l _?2) +a2(i -t*) +a3(\ -t6)
<t>i(0 = (\-t2i).
is determined in the same manner,
If for a first approximation n = 2 is chosen,
Ô) (0) = 0.93207.
2 2 4
x<- H0=ai(l ~x ) + a2(l -x ). (6-49)
This might lead one to believe that both
Substituting x<2) into Eq. 6-48 and integrat- x(2V^ and x^Ht) are good approximations
ing yields of the solution.
6-2.4.2 METHOD OF FINITE DIFFER-

ENCES
.1244 , 1 2 The Method of Finite Differences, as its

+ a, a,
-I
~~ a
2
3465 2 3 5 name implies, is simply based on the replace-
6-15
AMCP 706-192
ment of derivatives and the integral by finite chapters that these side conditions generally
approximations of these continuous opera- include both equality and inequality con-
tions. A grid, t = t0, ti , ..., tn+1 = t' , is straints. An extension to inequality con-
placed on the interval t < t < tl and the straints will be given in par. 6-4.
value of xft) only at the grid points is sought,
i.e., only the parametersx( =x(tj), i = 1,..., n, The problem to be treated here is given in
and perhaps x(t0) or x(tn+l), are sought. Definition 6-2.
Replacing derivatives by finite differences Definition 6-2 (Problem of Bolza): The

and the integral by a finite sum, the problem problem of Bolza is a problem of finding u(t),
is to determine the x. which minimize b, x(t), t < t < tn, which minimizes
J(x,)= 2 t;X.,
J = g0(b,t',x>)
1=0 ' ' t.i+i-h
(6-50)
x Oi+l-ti)) + f0[t,x(t),u(t),b]dt
('t"
The problem is now simply an uncon- subject to the conditions
strained minimization problem in a finite
dimensional space and may be solved by the dx
methods of Chapter 2. f(t,x,u,b), t < t< t" (6-5 1)
lit
6-3 A PROBLEM OF BOLZA gjb,t',x')+ I La[t,x(t),u(t),b]dt

J t>
6-3.1 STATEMENT OF THE PROBLEM
= 0, a = 1,..., r (6-52)
Many real-world optimal design problems
cannot realistically be reduced to the finite <j>(t,x,u,b) = 0,
dimensional form of Chapter 5. In many
problems the system varies continuously in = ],..., q,t < tu t" (6-53)
time or space, so functions rather than just
parameters must be determined. Examples 6-1 where
and 6-2, par. 6-1, are extremely simple, yet
even they involve distribution of the con- xt(t) ut(t)
trolling factor over space and time.
x(t) = u(t)-
As has been seen in previous chapters, xjt) ujt)

optimal design problems involve ideas of
design variables and state variables. Further,
b,
since the system being designed must be
capable of performing certain functions, side
conditions on the state and design variables
occur. It has been observed in previous (6-54)
6-16
AMCP 706-192
f(t,x,u,b) is expected to be only piecewise continuous.

:
The resulting state x(t), therefore, will have
f(t, x, u, b) only a piecewise continuous derivative in
f(t,x,u,b) general.
and t < t' < tv, where (t',x') are inter- The allowed discontinuities of f0, f, La,
mediate points, / = 1, ...,77 - 1. and 4>a play an important role in many
real-world problems. This feature allows for
For the problem considered here it will be completely different forms of state equations,
required that the conditions, Eqs. 6-53, shall constraints, etc., for different ranges of state
not determine any component of x(t) ex- and time. It is, therefore, possible to routinely
plicitly. This is equivalent to requiring that account for sudden changes in system be-
the rank of the matrix havior such as reverse in direction of frictional
force, motion of objects in a space where
physical barriers or restraint surfaces exist,
30,
(t, x, u, b) (6-55) logic built into the system which changes
Lbuk configuration as in staging of rockets, etc. It
q X m
should be clear that these features are re-
shall be q for all admissible values of the quired in order to treat many realistic prob-
arguments. In case some constraint function lems.
should depend only on x(t) and t; this
constraint is called a state variable constraint. For a discussion of the effect of these
This kind of constraint will be discussed in a discontinuities on more detailed necessary
later paragraph. conditions and sufficient conditions, see Ref.
8.
The vector variable x(t) is called the state
variable, u(t) is called the design (or control) 6-3.2A MULTIPLIER RULE
variable, and b is called the design (or control)
parameter. Eqs. 6-52 contain the boundary As mentioned in par. 6-3.1, real-world
conditions on the state variable and functions optimal design problems require at least the
which determine the end points of the inter- complexity of the Bolza problem of Def. 6-2.
val, t and /". The independent variable t In fact, the system designer requires all the
may be time or a space-type variable, depend- tools the mathematical theory of optimal
ing on the problem being considered. processes can give him. This requirement
points out one of the obstacles to engineers in
The functions f0,f,La, and <j>a are assumed utilizing the modern theories of mathematics.
to be continuously differentiable at all points This text cannot possibly present the mathe-
except (t'.x1), j = 1,..., TJ 1. At these points matical theory required of the research math-
the functions may have jump discontinuities; ematician who is developing the theory of
i.e., the functions will have limits along any optimization. The approach taken here to
path, but limits along different paths may by-pass this obstacle is to accept a key
have different values. In general, even for theorem of Functional Analysis and then
problems with very regular functions, u(t) proceed to develop the tools required for
may havejump discontinuities. Therefore u(t) solving problems of optimal design. A very
6-17
AMCP 706-192
powerful theorem of Liusternik and Sobolev, 6-3.3 NECESSARY CONDITIONS FOR

Ref. 6, page 209, will be used to obtain THE BOLZA PROBLEM
necessary conditions for the problem of
Bolza. The simplest Bolza problem is the one
having all its functions three times continu-
Theorem 6-4: If dt), b, t', and x(t) pro- ously differentiable. Even in this case, how-
vide a solution to the Bolza problem of Def. ever, lit) may be only piecewise continuous.
6-2, then there exist multipliers X0 > 0, ya, ot To include this possibility, let /* be a point of
= 1,..., r, \/t), i = 1,..., n, and ft(t), = 1,.... discontinuity of any component of vit).
q, not all zero, such that
Before computing the variation called for
in Eq. 6-56, it should be noted that t, t'\ t*,
6.7=0, (6-56)
and tn are not fixed but must be determined.
where This means that these special points must be
treated as parameters that are to be deter-
mined, much as the design parameter b. At
J = X0g0 (b, t', x') + 2yagaa (b, t', x>)
a =I " first glance, this may seem to introduce no
essential complication into the problem. The
behavior of the allowed variations in xft),
\0f0(t,x,u,b) (6-57) however, must be treated very carefully.
Let i be a typical point t' or t * where x(t)

dxt
+ 2 Xft) f.(t,x,u,b) may very well be discontinuous. The function
i=\ ' ~dT x(t) will be changed at i by both the
independent variation in x(t), Sx(t), and the
shift in the point (t, Si). Denote the total
+ 2
_ 1 y'a L(t,x,u,b)
a
a =1 change in x(i) due to both of these sources by
Ax(i). It must be assumed that there are no
other points t' or t* arbitrarily near i, so
+ 2 ixjt)<t>(t,x,u,b) } dt.
0=1 e limits from the left and right exist. For t= i,
xft) is continuous so the total change Axft) in
Note that the symbol 6/ is the first xft) due to Sx(t) and Si is continuous and
variation of / as defined in par. 6-2. For
proofs of this multiplier rule, the reader is Ax(i) = Ax(i - 0) = Ax(i + 0).
referred to the literature (Refs. 2,5-9).
where
This theorem says nothing about the con-
tinuity and differentiability properties of the Ax(i -0) = 8x(i -0)+ i(i -0)Si
solution x(t), tit), and the multipliers \/t)
and pJt). In general, piecewise continuity is and
all that may be expected of tit). Eq. 6-51
then implies x(t) has a piecewise continuous Ax(i +0) = bx(i +0) + x(i+0)S i.
derivative. The properties of \/t) and (t)
will be determined when necessary conditions It should be noted that this condition
are derived. imposes restrictions on 8x(i 0) and bx(j +
6-18
AMCP 706-192
0). In particular, they are not necessarily the

same so Sx(t) on f <s t < tv need not - 2 p{i)<j>Jt,x,u,b)
necessarily be continuous.
r^-59)
Before enforcing Eq. 6-56, put J in the
so that Eq. 6-58 becomes
form
J=\0g0(b,t',x') + 2 laga{b,t',x')
I {\0fo(t,x,u, b) + 2 \,.(f)
J=G +
n dt
\T(t)-H]dt
dt
(6-60)
\dxt
, dx
+ 2, \La(t,x,u,b) (O-r -H] dt.
dt
+ 2iiiftW(t,x,u,b) \ dt Eq. 6-56 may now be applied to yield
+ r> 3C . . , 9G . 3G
i> * r n *
+ 5?" +8b
9r" 9
f5-J8)
where the argument of the pair of braces is

the same as that of the integrand of the first
integral. Note that t* and t' are simply typical
i>'f a* Sx
elements of their respective classes. 'dH 9//

5M 56 rff
9
For convenience in the development that
follows, define
G - Kgo(b,t>,x>) +2 -v (b t1 x1')
LI a? 9JC
-l 'assav ' ' '

a 9# 9//
5 66 dt
9 a
n
H,x,u,bXy,n) = \TU)fU,x,u,b)
. T, , dSx 9//
- ~k0fo(t,x,u,b)
9// ,. 9//
"^7 ('.*..*) dt
du ab
6-19
AMCP 706-192
f -0 + \T(tr> _0)Ax" -Hit" -0)6?"

XTit)*m_Hit) Of
dt t> + 0 -\T(t +0)Ax +H(t + 0)8t.
rfx(f + 0) Since the variables Ax', St', 8b, 8x(t), ana

\T(t +0)
It 8u{t) are arbitrary, (Ax1 are taken as arbitrary
along with 8t' so that 8x' = Ax' x'8t' is
-#(r +0 ,x, u, b)\8t fixed) Lemma 6-1 applies. Application of this
Lemma yields
dx(t) f*-0
\T(t) H(t) 8t* Theorem 6-5: If (x(t), u(t), b, t'\ x') is a
dt t* + 0 solution of the Bolza problem of Def. 6-2,
then there exist multipliers X0 > 0, X^f), i =
dxit" -0) 1,.... n, ya, a= \,...,r, n(t),= l,...,q, not
X(r" - 0)
* all zero, satisfying the conditions:
dHT
-Hit7' -0,x, u, b) lot".
J'
Integrating the first terms in each integrand
dh
dt x
Tfor t = t! (6-61)
by parts yields ML 0, for t # ft (6-62)

du
9G n dG A dG p . 1/"
0 = Ax + ... + Ax" +- 5f dH
dx dx* dt dt = 0 (6-63)
dG , dG
+ ...+T5f* +^-(5ft
I
3r> 36
l_r-\(r) = o
3x
1
t'
T
L dt dx dG
+ \(tr>) = 0 (6-64)
~dx^
dH dH
+ 5+ 8b dt
du db r + \(f> -0J-\(t' +0)=0
dx'
If r f"
dt L+H(t +0)=0
+ [\T(t' -0)-\T(t> +0))Ax' a_

-#(f -0)=0 > (6-65)
3f
[Hit' -0)-H(t> +0)]6t'
dG
-H{t> -0) + H(t> +0J=0
+ [\T (t*-O)-\T(t*+0)] Ax(t*) Mi
[H(t* + 0)-H(t*+Q)] 8t* H(t* -0)-H(t* + 0) = 0 (6-66)
6-20
AMCP 706-192
X(f* -0)-Mt* +0)=0 (6-67) equations to determine the unknowns. It does

not assert that a solution of Eqs. 6-61 through
Note that the necessary conditions, Eqs. 6-67 exists. Existence theory for these prob-
6-61 through 6-67, are linear and homoge- lems is a difficult question that is treated in
neous in the multipliers X0, \(0, 7a, P(t). It Refs. 10 and 11.
is, therefore, permissible to choose the magni-
tude of one multiplier so that the remaining The conditions of Theorem 6-5 are very
multipliers will be uniquely determined. It nearly the famous Pontryagin Maximum Prin-
seems reasonable that if the necessary condi- ciple (Ref. 12). The condition that completes
tions obtained by setting 6/ = 0 are to be the Maximum Principle is an inequality which
related to minimization of J, then X0 should follows from the Weierstrass condition of the
not be zero. This is indeed the case and if X0 calculus of variations. This condition is given
is required to be zero by the necessary as Theorem 6-6.
conditions, then the Bolza problem is "abnor-
mal" in a sense. Most meaningful problems Theorem 6-6: In addition to the condi-
are normal as defined in Refs. 7, 8, and 9 and tions of Theorem 6-5, the solution of the
require X 0 = 0. In solving problems using the Bolza problem must satisfy the condition:
necessary conditions of Theorem 6-5, one
should first verify that Eqs. 6-61 through 6-67 H[t,x(t),U,b,\(t),y,0]
have no solution if X0 =0. It is then
permissible to put X0 = 1 so that the < H[t,x(t),u(t),bMt),y,0] (6-68)
remaining multipliers are uniquely deter-
mined. for all admissible U and all t, t =s t < tv.
Even though Eqs. 6-61 through 6-67 are For proof of this theorem see Refs. 8 and
very complicated, it is interesting to note that 13.
they provide just the right number of equa-
tions to solve for all the unknowns. Eqs. 6-51 Another useful result is the following iden-
along with Eq. 6-61 form a system of 2n tity:
first-order differential equations for x{t) and
X(f). Further, the first and last members of dH__dH_
Eq. 6-64 may be considered as 2n equations ,fort*f'. (6-69)
dt at
in boundary conditions onX and x. This is the
proper number of boundary conditions. The
This condition is useful in case H does not
second equation of Eqs. 6-64 provides any
depend explicitly on/. Then H is constant
jump conditions in \{t) at the intermediate
between the points t>', and at these points it
points t', 0 < j < r). Eqs. 6-65 may be
may have discontinuities governed by the
interpreted as determining t>, j = 0, 1,..., 77,
third equation in Eq. 6-65.
and Eq. 6-66 determines t*. Eq. 6-67 simply
states that X is continuous even at jump
To prove this relation holds, compute
discontinuities in u. Finally, Eq. 6-63 deter-
formally
mines the design parameter b.
It should be clear that this argument only dH M m du dH dx dH dh
+ + +
shows that there are the proper number of dt at du dt ax dt ah dt
6-21
AMCP 706-192
dx M M d\> x- andy-axes are fixed in inertial space. In the

Since- = /, = 0, = - , problem discussed here, 6(t) is to be chosen
dt du dx dt
so as to direct the motion of the particle.
M _ Hence 6{t) is the control variable.
and =f , this is
dH_-_dH__ d>J_ fTdX_ bH Denoting horizontal and vertical compo-

dt dt dt dt at nents of velocity of the projectile by u and v,
respectively, the motion of the projectile is
as required.
governed by the equations
6-3.4 APPLICATION OF THE BOLZA

PROBLEM x =u
y=v
In order to obtain familiarity with the
Bolza problem, several examples will be con-
. F (6-70)
sidered. In order to illustrate the basic ideas u = cos 0
associated with the Bolza problems, these m
examples will be elementary. In real-world . F
problems the engineer should be prepared for v = sin 6 g,
complexity that will probably force him to
use a numerical method of solution. For
examples of the Bolza problem in the field of where
dt
aerodynamics, a field which contributed
greatly to optimal design theory, see Refs.
15,16, and 17. The projectile is fired from a gun at time t
= 0 with x(0) = y(0) = 0 and initial velocity
Example 6-6: Maximum Range Rocket- w(0) = V cos 0O, "(0) = V sin 0O, where V is
assisted Projectile the muzzle velocity of the projectile. The
problem at hand is to choose B0 and d(t) so
A projectile of mass m is acted on by a that at some future time T, the projectile will
fixed force F as shown in Fig. 6-8. The angle hit the earth as far as possible from the launch
of 0(t) is measured from the x-axis, where the point, i.e., y(T) = 0,x(T) = maximum.
In the notation of the Bolza problem, 0(f)

is a design or control variable u{t), d0 is a
design parameter b, T is terminal time f, and
(x,y,u,v) is the state. The quantity to be
minimized is
J = go(l>,t',x') = -x(T).
Figure 6-8. Particle in Motion Boundary conditions on the state variables are
6-22
gl =x(0) = 0 dG
= 7i = \(0)
g2 = y(0)=Q 3x(0)
g3 =u(0) - Vcose, =0 (6.71) dG

= T2 = \(0)
g4 = v(0) - Vsin 0 = 0 MO)
(6-77)
Ss =y(T) = 0. dG
= 73 = X(0)
du{0)
As defined in Eq. 6-59,
dG
G = _\ox(T)+7ix(0) +y2y(0) 74 = \(0)
3v(0) /
+ 73 ["(0) -Vcosd0] (6-72) 1
dG
dx(T)
Vo= \(T)
+ 74 [v(0)- Fsin0o] + fsy{T)
F dG
H = \u +X v+\ cose
dy(D
= x5 = \(T)
K
(isind^
(6-78)
(6-73) dG
= 0 = \(T)
where variable named subscripts are used for du(T)
the A's.
dG
0 = \(T)
Theorem 6-5 yields as necessary conditions dv{T)
dH dG
= H(T-0) (6-79)
dT~
dH and
= 0
\ = dy
(6-74) \ cos + Xv sin 0 < X cos 6
dH
X = = -X
du + \ sine (6-80)
dH for aU admissible .
$*-
Eqs. 6-74 yield

dH \F \F
- = 0 = - sin+ cos 6 (6-75)
do m m
dG | dH
dt = 0 73 Vsin0
ae
de0 - J o
. 74 K cos e, (6-76) Xv = 4 - 2 '
6-23
AMCP 706-192
The last two equations of Eq. 6-78 imply |3 = ;2F

%i T and 4 = 2 T. If we assume the problem x(0 =r v +cose,
2m '
is normal, X0 = 1 so X (D = 1 = t so
y(t)=tV sine, t2g (6-84)

X, = 1
t2F
+ sin0.
2m
\, =*2
(6-81) By use of these equations, the last equation in
\=r Eq. 6-71 yields
X =i2(.T-t) 1 TF
T(VSm8, -gT+smdo)=0.
2 2m
Substituting from Eqs. 6-77 and 6-81 into This implies

Eqs. 6-75 and 6-76
-8T
TsinS, -i7Tcos8, =0 sin 8, (6-85;
and 2m
_ (T _ t) sin 8 + 2 (T- t) cos 0=0. The one condition which has not been
used is Eq. 6-79. By substitution of Eqs. 6-81
and 6-83 into Eq. 6-73, Eq. 6-79 becomes
For all t = T, this is
TF
V+ m-J cos d0 + i-i I ysin 8,
- sin 8 + 2 cos d = 0.
h
These equations imply TF
+ sin0 ~gT
m
2 =tan8, By use of Eq. 6-82 this becomes
(6-82)
8(0=60
V+ ) cos2 60 + [V+ I sin2 0O
Integrating the last two equations in Eq.

-gTsm8, =0,
6-70,
or
tF TF
u{t)= V cos d0 + cos 0O V+ =^rsin0o. (6-86)
m m
(6-83)
tF Combining Eqs. 6-85 and 6-86,
v(t) = V sin 60 + sin 0O - St
m
TF\ / TF\ 1
Integrating again, V+ K+ = g2T2
m / \ 2m ) 2
6-24
AMCP 706-192
or illustrated in Fig. 6-9 and the time scale is

shown in Fig. 6-10.
F2 e2\ 2 3FV
_*_ T + T+ V2 = 0
2m2 1 } 2 m The times ?i and t2 may actually coincide,
depending on the problem parameters. These
so times play the role of t' in the Bolza problem.
The equations'of motion of the spacecraft are
1/2
9F2V2
T=-
3FV
2 m
-t
4m2
4P ( A
\2m2 2 /
taken as
2(
-^ (6-87)
. v2 M h{t)F sine
u = - *
Substituting T from Eq. 6-87 into Eq. 6-85
(6-88)
then gives an easy equation for 6 0.
uv , h(t)F cos 6
v = f
While the results of this problem are not
particularly useful, the solution does illustrate
the use of the various conditions in Theorem m =h(t) 4
6-5 in generating a candidate sclution of the
problem. The reader, however, should not be where
led to believe that all Bolza problems may be
solved in closed form as in this example.
In more general problems the adjoint equa-

tions, Eq. 6-61, cannot be solved so easily in
closed form. Further, the equation dll/du = 0
may not yield so simple a condition as Eq.
6-75 for the design variable. It is often of
value to keep a procedure in mind for
determining the various unknowns as in this
problem, even though more realistic problems
may require numerical methods at each step
in the procedure.
Example 6-7: Minimum Fuel Orbit Trans-

fer
Figure 6-9. Orbit Transfer
A rocket equipped with a constant thrust
engine is to transfer from a circular earth
orbit of radius r0 to one of radius R > r0
Thrust On Thrust Off Thrust On
using a minimum of fuel. The time allowed 1 1 1
for this transfer is T. Further, it is possible to
shut the rocket down during one time interval
of the transfer if desired. The orbits are Figure 6-70. Thrust Program
6-25
AMCP 706-192
and the boundary conditions are and
r(0) =r0. (0) = 0 H = \u + X.

r2 j 1
l/2
v(0)=(M//-o) ,m(0) = wo,
r(T)=R, u(T) = 0
(6-89)
+ X.
MV h (t)F sin = (M/^)1/2 + \nh(t)q.
where The necessary conditions of Theorem 6-5

arc
r =radius
w = radial velocity
i*f)-4f
7 = tangential velocity K = -X+x
m = mass of spacecraft
\.=-\A )+ X, (6-90)
fi = gravitation constant ft
MO-F sinift
F = thrust \=\, L W
4 = mass flow rate during thrust
ft(0-F COS0
+ X.
<t> - thrust orientation
h(t)F cos<p
Since m0 m(t) is the amount of fuel o = x
consumed up to time t, the object here is to m
minimize (6-91)
, h(t)F cos (p
J - m0 m(T).
XM(D = -1 (6-92)
For use in Theorem 6-5, define X(r, -o;=\(r, +0)

(6-93)
G = m0 - m(T) + 7i [r(0) - r0 ] + y2 a(0) \(t2 -0)=\(t2 +o)
+ 73[f(0)-(M//-0)1/2] Hit, -0)=H(t1 +0)

(6-94)
+ 74[w(0)-w0] +7s Wr)-] H{t2 -0)=H(t2 +0)
+ 76"(7") + 77rv(71-(M//?)1/2l the ti are those shown in Fig. 6-10.
6-26
AMCP 706-192
The prospect of solving this set of equa- Definition 6-3 (Problem cf Optimal De-
tions in closed form is dim indeed. A general sign): The optimal design problem is a prob-
procedure can be discussed, however, and the lem of finding u(t), b, x(t), t <s t < tn, which
actual solution can be obtained using numeri- minimize
cal methods discussed in a later paragraph.
J = g0(b,t'\x')
From Eq. 6-91, <p{t) may be determined as
4>(t) = Arctan ( JL (6-95)

+ I
1/' fn [t,x(t),u(t),b]dt
subject to the conditions

(6-96)
for t not in /x <: t *s t2, where 0 need not be

defined. The result of Eq. 6-95 may be
t# = f(t, x, u, b), t < t < t", t # t' (6-97)
substituted into Eqs. 6-88 and 6-90 so that
these equations become a set of eight first-
c
order differential equations for the state and
adjoint variables. Eqs. 6-89 and 6-92 form a
ga{b,t>,x>)+ I La [t,x{t),u{t)Mdt
set of eight boundary conditions for these
variables. Eqs. 6-93 show that the adjoint
variable is continuous and Eqs. 6-94 deter- = 0, a= 1, ...,r (6-98)
mine tx and t2. A numerical procedure may
be used to solve this problem. The resulting
adjoint variables may then be substituted into ga(b,t',x')+ I La [t,x(t),u(t),b]dt
Eq. 6-95 to obtain the explicit design (or
control) variable. A problem of this kind is
discussed in Ref. 18. The method used there < 0, a/ = r + 1,..., Y (6-99)
to construct a solution is completely different
(p(,t,x,u,b) = 0,= 1, ...,q'
from the one proposed here.
(6-100)
t a to t*
6-4 PROBLEMS OF OPTIMAL DESIGN
AND CONTROL and
The Bolza problem of par. 6-3 is of almost <p(t,x,u,b)< 0, = q' + 1, ...,4,

the generality required for optimal design. K (6-101)
The principal shortcoming of that problem is t < t< t*.
in the lack of generality in the constraints. It
has been noted in preceding chapters that The variables and functions appearing here are
meaningful optimal design problems generally identical to those in Def. 6-2.
involve inequality constraints. It is the pur-
pose of this paragraph to extend the Bolza The inequalities in this problem are treated
problem to account for inequality constraints. here in the manner presented in Ref. 13. The
inequality constraints are first reduced to
The problem treated here is given in Defini- equality constraints, and the results of par.
tion 6-3. 6-3 are applied. In order to perform this
6-27
AMCP 706-192
reduction, define the slack variables va, a = r' required to have rank q. For this purpose, the
+ 1,..., r and w At), = 4' + 1,..., q by design vector must be considered as (uT,
wT)T, where wT = (w -+1, W
J-
ga(b,t',x')+ |
I. I[/,x(0,K(f),*]df
(6-102)
The matrix, Eq. 6-55, becomes
90t 901
+ v2 =0,SL = /+ L...,i 0, ..., 0
3, 3
and
^(U,,i)) + ws2(O = 0, 9^
,0,...,0
(6-103) 3! 3
)3 = </'+!,...,? (6-104)
The constraints, Eqs. 6-102 and 6-103, are V+i

2w
3! ,'+l
3w
equivalent to Eqs. 6-99 and 6-101, respec-
tively, where va and w (t) are interpreted as
design parameters and design variables. With
3
these equality constraints replacing the in- *< , ..., ^ , 0,...,
n 9
2vv
equality constraints, the optimal design prob- 3"i 9"m *
lem becomes a Bolza problem. The necessary
conditions of par. 6-3, therefore, may be
applied to this modified problem. This matrix is required to have rank 4. In
order for this to be possible the number of
The form of the constraints, Eq. 6-101, has columns, m +4 q', must be greater than or
a great deal to do with the behavior of the equal to the number of rows, 4, or m 4' >
problem. If some function 0. depends only 0. Further, it is obvious that the firsts' rows
on t, x, and b then the problem is compli- must be linearly independent, or the entire
cated in intervals in which 0 = 0. This kind matrix could not possibly have rank 4. Next
of constraint will be referred to as a state note that if wa = 0, then the ath row must be
variable inequality constraint and will be linearly independent of all the other rows
treated separately. If <P does depend explicit- since it has the only nonzero element in the m
ly on u, then the constraint is referred to as a + a q'\ column. Therefore, linear inde-
design variable inequality constraint. This pendence of the rows from 4' + 1 to 4 need
problem will now be investigated. only be considered for those a with wa = 0.
By Eq. 6-103, this is the same as 0a = 0.
6-4.1 DESIGN VARIABLE INEQUALITY
CONSTRAINTS
The conclusion is, then, that the matrix,
In order to apply Theorem 6-5 to the Eq. 6-104, will have rank 4 if and only if the
problem of Eqs. 6-96, 6-97, 6-98, 6-100, matrix
6-102, and 6-103, the independence of condi-
tions expressed by Eqs. 6-100 and 6-103 must 30/
, 0,- = 0 (6-105)
be verified; i.e., the matrix, Eq. 6-55, is "iiT.
6-28
AMCP 706-192
is of full row rank. This simply says the and the conditions, Eqs. 6-1 lOthrough 6-113,
gradients of all constraint functions (which are unchanged. Further, Theorem 6-6 yields
are equalities) with respect to the design
variable must be linearly independent. Assum- H[t,x(t),U,bMt),yfl,v,w]
ing this is the case, Theorem 6-5 may be
applied. < H[t,x(t),uV),bMt),0,v,w] (6-1 15)
Define
for all admissible U.

G =\0io + 2 yaga (6-106)
The condition, Eq. 6-112, in scalar form is

G
'= 2 yaavl (6-107)
a=r+1 " -2pwp = 0, = q' + 1,..., q . (6-116)
//=X^/_X0/0_ 2 yaLa_ Su K
a~ 1 =1 If w = 0, then by Eq. 6-103 <f> = 0. If w
J=0,<t>< 0 and /* = 0. Therefore, Eq. 6-116
(6-108) is equivalent to
i,,^' HU)<t>(t,x,u,b) = 0, = q'+l,..., q .

2
#' = - 7^ -
(6-109) (6-117)
Condition, Eq. 6-114, is

The quantities tf = # + H' and G = G + G'
take the place of H and G in Theorem 6-5.
Necessary conditions for the optimal design 1y'a v a + f
I 'a a dt = 0,
problem are, therefore, J fO
d\ _ u=r' + 1,.... r
(6-110)
dt dx
M or
=0 (6-111)
du
2yava(] +(i - t) = 0, a = r'+ 1, ...,/-.
ffl'
=0 (6-112)
Since 1 + ?" _ f =0,
"J,f
dG d_H_ 7ava =0, a =r' + 1,. (6-118)
dt = 0 (6-113)
db db
If v =0, then by Eq. 6-102, the constraint,

dC
bv "
f dt = 0 (6-114)
Eq. 6-99, is an equality. If va = 0, then the
constraint, Eq. 6-99, is a strict inequality and
ya = 0. Therefore, Eq. 6-118 is equivalent to
6-29
AMCP 706-192
dGT
7a ga(b,t',x')
3GT
La[t,x(t),u(.0,b] dt = 0. + X(r*)=0 (6-124)
3x"
a?=r'+ 1,..., (6-119) j+\(t> -0)-Mti +

0)=o
ox'
The conditions, Eqs. 6-116 and 6-1 18,

ft3- + H{t +0) = 0
imply H' = 0 and G' = 0 so that H=H and G
= G. The necessary condition, Eq. 6-115, is,
therefore,just
9G
_H{tr> -0)=0 > (6-125)
H [t,x(t),U,bMt),7,0]
%6--H(t> - 0) + H(t> +0)=0

< H [t,x(OMt),bMt),7,0] (6-120)
it'
for all admissible U. It is further shown (Refs.
5,10,12) that X0 > 0, ya > 0, *> = r' + 1,..., H{t* -0J-H(t* + 0J=0 (6-126)
r, and M(0 > 0, |3 = 4' + 1, ..., 4, f <
t < f. X(f* -0)-Mt*+0)=0 (6-127)
The conditions obtained through applica- it)4>(t,x,u,b) = 0, =\,...,q (6-128)

tion of Theorem 6-5 to the optimal design
problem may now be stated as Theorem 6-7.
7a | ga{b.t',x')
Theorem 6-7: If [x(t), u(t), b, t>, x'] is a
solution of the optimal design problem of
tv \
Def. 6-3 and if the matrix, Eq. 6-105, has full
+ 1 La[r,jc(r)p(0,] dt \ =0
row rank, then there exist multipliers X0 > 0,
J t<
\(t), i = 1,..., m, ya, a = 1,..., r, ya > 0, a =
/' + l,..,r. n(t),p = l,.... 4,*i,(0 > 0,p = 4' +
a= l,...,r (6-129)
1, ..., 4, not all zero, and functions G and H
of Eqs. 6-106 and 6-108 such that
3HT ,for^?; (6-130)

dh
,fOTt=t t' (6-121)
dt
and
M
=0, fOTt^t' (6-122)
du tf [f(x(r),C/,&,X(0,7,0]
9G
db
6-30
r dH
db
dt = 0 (6-123)
//[r,x(f),u(f),,X(/),7,0]
for all admissible U.

(6-131)
AMCP 706-192
It should be noted that, just as in Theorem plicitly on u(t), then this equation is of the
6-5, the number of conditions here is just form required in the problem treated in par.
equal to the number of unknowns, so one 6-4.1. If not, then differentiating through this
might be led to believe that a solution may be equation with respect to t and using the chain
found. Existence of a solution is, however, a rule of differentiation
very difficult question that is treated in Refs.
10 and 11. d2K *K 2*fl
+ 2
""** bt2 Utax
6-4.2 STATE VARIABLE INEQUALITY
CONSTRAINTS
r
24> df
+/ " f + 3v
(6-133)
In many meaningful design problems con-
straints may involve restrictions only on the where all the arguments are omitted. If the
state variable. This is the case when some <p right side of Eq. 6-133 depends explicitly on
of Eq. 6-101 depends only on t, x, and b. To u{t) then this equation is of the form treated
study this problem, just one such constraint in par. 6-4.1.
needs to be considered, i.e.,
This process continues until
<j>M,x,b)< 0,t < t< r" . (6-132)
Let t~ < t< t+, t~ < t+, be an interval in O-/0 [t,x{t)M (6-134)
which <S> of Eq. 6-132 is an equality. It is dt"
clear that b(j>Jbu = 0, so the matrix, Eq.
6-105, has a zero row in this interval and involves u(t) explicitly in its right side and
hence cannot be of full row rank. Theorem u(t) can be determined as a function of x(t)
6-7 cannot be applied directly, so further
and b, as in par. 6-4.1. The integer i> A 1
analysis is required.
is defined to be the first integer for which this
is true. The constraint, Eq. 6-132, is then
In the interval t < t< r = 0 so it is
called a f^th order state variable inequality
necessary that
constraint.
b<t> b<j> dx
= 0 From the theory of ordinary differential
dt bt dx dt
equations (Ref. 14), Eq. 6-134 throughout
From Eq. 6-97, dx/dt may be replaced by / t~ < t < f+ and
and this relation becomes
At-,x(r),b] =0 (6-135)
b<t>[t,x{t)M
dt bt
d0
- [r,x(r),b] =0, i = \,...,p._.
d<t>M,xU),b] dt1
+ f[t,x(t)Mt),b]
ox (6-136)
If the right side of this equation depends ex- are equivalent to 0 = 0 throughout t~ <
6-31
AMCP 706-192
t < t+. This, of course, requires that <j> Theorem 6-8: If [x{t), u{t), b, t!, x'] is a
have v piecewise continuous derivatives, solution of the optimal design problem with
and / have v&l piecewise continuous state variable inequality constraints, then
derivatives in t~ < t < t+. The point there exist multipliers X0 > 0, \.(/), i = 1, ,
t~ plays the role of a t' in the problem n, ya, a = 1, ..., r, ya > 0, a = r' + I, ..., r,
stated earlier in this paragraph. Hv(t),= l,...,qi,M/J(r)>0,j3 = q'+l,...,4>and
TT i = 1,..., v and associated with a state
variable constraint and 4> = 0 in?" < t< t+,
It will be assumed that when the right side
of Eq. 6-134 is used in place of </> in d\ 3HT
computing the matrix, Eq. 6-105, this matrix r= -- ,foi t=t', r, t+ (6-141)
at ux
has full row rank. In this case Theorem 6-7
may be employed. To utilize this theorem, ML-
define 0, for t t1, r, t+ (6-142)
3u
,.*<o
V/=l d
(6-137)
3G
3b
+
LJ:
3G
3b
3H_
3b
dt = 0 (6-143)
<t>, = 0
3GT
dt -\(t) =0
3x
where v& = 0 if 0 involves u explicitly, 3GT
dx*
G = XQg0+ 2 \ga (6-138) 3GT
et=l
+ \(t> -0)-\(t> +0)=0 <6-144)
~3xi~
V
- *~l \{ dl
% [t-,x(t),b]
-, )\ (6-139)
G=2 2 T- +\(r -0)-\(r +0)=0
1=0
dt1 ax-
where this sum on/3 is extended only over the \{t -0)-\(t+ +0)=0
indices associated with state variable in-
equality constraints, T~ are multipliers, and +H(t+0)=Q
d- -Hit" -0)=0
ff = X'/-X0/0- 2 yaLa
a=l dt*
(6-140)
fl?, "uV* -H(t' - 0) + H(t' + 0) = 0 (6-145)
With G = G + G and H replacing G and H in 3G

Theoiem 6-7, a set of necessary conditions for -H(r -0)+H{t- +0)=0
at-
this problem are obtained. They are easily
computed and are given here as Theorem 6-8. H{t+ -0)+H{t+ +0)=0
6-32
AMCP 706-192
H(t* -0)-HU* + 0)=0 (6-146) these subintervals is not known before the
solution is computed. The generality of the
X(t* -0)-X(t* +
0J=0 (6-147) problem makes it difficult to discuss all its
intracies without resorting to special cases and
lt(t)4>V,x,u,b) = 0, 0=1, ., q (6-148) examples.
Ja 8a{b,tl,xl) 6-4.3 APPLICATION OF THE THEORY OF

OPTIMAL DESIGN
i: L, [t,x(.t),uU)M dt\ =0, In order to develop some familiarity with

the methods of the preceding subparagraphs,
several examples will be treated here. These
a= 1,..., r (6-149) problems will be idealizations of real-world
problems but will illustrate the basic ideas
dH 3H ^ ^ which carry over into more complicated
(6-150) problems.
Example 6-8: Time-optimal Steering of a

and
Ground Vehicle (Ref. 19)
To illustrate the concepts presented in par.

< #U,x(0,"(OAX(0,7,0] (6-151) 6-4.2, an optimal vehicle steering problem will
be solved. This problem is chosen because of
for all admissible /. its clarity of formulation and solution. A
ground vehicle (a tractor in this case) is to be
The full set of necessary conditions em- steered so that it begins at a given point and is
bodied in this theorem is awesome from a steered so that it reaches a given straight line
computational point of view. The differential path in the shortest possible time. The vehicle
equations for x and \ are subject to multi- and the line it is to reach are shown in Fig.
6-11.
point boundary conditions that involve a set
of undetermined multipliers. In a gross sense,
Point A, midway between the rear wheels,
Eqs. 6-147 may be viewed as determining
is located by the coordinatesxÛ) andx2(f)-
intermediate points in f < t < tn and the
associated boundary conditions on x{t) and
MO-
Use of the theorem is further complicated

by the fact that the design variable may be
determined as the solution of Eq. 6-142
which satisfies Eq. 6-15 1. This means that u
will be determined as a function of x, b and
/v
all the multipliers. The expression for u will
generally take different forms in different
subintervals of t < t < t" and the spacing of Figure 6-11. Ground Vehicle
6-33
AMPC 706-192
Orientation of the vehicle is specified by a + 74 [x2(T)-x\] +-ysx3{T)

third variable x3(t). Steering of the vehicle is
accomplished by choosing the angle 0(f). H = Xj V cos x3 + X2 V sin x3 + \3a tan8
From physical grounds it is clear that the
state of the vehicle is described by x(t) =
-Mi (0 - 0O) - 02 (0o -0)-
[x1(t),x2(t),x3(t)]T and the vehicle is con-
trolled through choice of d(t). The conditions of Theorem 6-7 are
It is assumed that the rear axle of the bH

X, =__ =0
vehicle moves with a constant velocity V. In
this case, motion of the vehicle is governed by
the differential equation M
x2 = - =0
OX 2
i, = Fcosx, (6-154)
i2 = V sin x3 (6-152) A, = =XiKsin^3

9^3
a tan e
X2 V cos x3
where a = V/L. At the initial time t = 0,*J(0)

bH
= x\, JC2(0) = x2, and x3(0) = x\. The 77-= 0 =\3a sec2 e-Mi +M2 (6-155)
be
terminal time T is not determined but it is
required that x2(T) = x\ and x3(T) = 0 since
\!(T)=0 (6-156)
the vehicle must be tangent to the target line
at time T.
X0 = Xi (T) V cos x3 (T) + \2 (T) Ksin x3(T)
The steering angle is limited by

+ \3(T)a tan 6(T) (6-157)
e0 < e < e, (6-153) Pl(6-80) = 0

(6-158)
and as an idealization it is assumed that any M2O0 -fl) = o
steering angle in 80 < 9 < 60 may be and
chosen instantaneously. For a reasonable
problem it is clear that 60 < ir/2. Further, for dH bH
definiteness, assume lx3\ < TT/2 and x^ > x. 0. (6-159)
dt at
AU other initial conditions can be obtained
from these by reflection in Fig. 6-11. The first two equations in Eq. 6-154 yield
The problem is now in the form described

in par. 6-3. For use in Theorem 6-7,
A2 (0 = f2
G=\0T+7l [Xl(0)-x\
and Eq. 6-156 implies J, =0. The last
+ 72 lx2(0)-x2] +y3[x3(0)-x] equation in Eq. 6-154 is then
6-34
AMCP 706-192
X3 = -S2 ^cosx3. Since x < 7r/2, forf small, either 0(f) =0O
or 0(f) = - 0O. From Fig. 6-11, it is
Using the first equation in Eq. 6-152 to reasonably clear that 0(f) = 0O and Eq. 6-152
replace V cosx3, this is can be integrated to obtain
X3 =-2*l- x1(t)=xl +R [sinQc +bt) *
Therefore, sin x ]
MO = -2*i(0 +3- (6-160) x2(t) = x2 -R [cos (x3+bt) (6-162)
The behavior of 0(f) may be isolated to COS X? ]

two different cases. The first is |0(f)l = 8,.
The second is 10(01 < 00, in which case Eq. x3(t) =x + at tan 0O,
6-158 implies Mi(0 = iMO = 0. Eq. 6-155
then shows that X3(f) = 0. By Eq. 6-160 then where
Xi (f) is either a constant or 2 = 3 = 0. This
and Eq. 6-157 then implies X0 = 0 so all X. are b =<ztan 0O
zero. This is forbidden by Theorem 6-7, so
*i(0 is a constant when \9(t) | < 0O. But if R = V/b.
*i(0 is constant *i(0 = 0 and the first
equation in Eq. 6-152 implies x3(t) = 0. The This path is just a circular arc with center at
last equation in Eq. 6-152 implies 0(f) = 0.
(Xj - R sin x", Xj + R cos x3) and
counterclockwise motion.
It is clear then that if 10(01 < 0O for some
interval of time, the path of the vehicle must Similarly, if 0(f) should become 8, at
be a straight line parallel to the x2-axis in Fig. sometime t* where x^f*) =x*, x2(t*) =x*,
6-11. and x3(f *) = x| then the path is described by
Since the last two terms in H are zero, the x1(t)=x*-R [sin (x*-bt) '
only explicit dependence of H on 0 is through
the term \3{t)a tan 0(f). The inequality, Eq. sinx.
6-131, states that 0(f) must maximize//. It is
clear then that if X3(f) = 0, then x2(f) = x| +R [cos (X? -of) V (6-163)
0(f) = 0 sgn [X3(01 (6-161) cosxJ]
where x3(f) = x* -af tan0o.
This path is a circular arc with center at (x* +

sgn^r =,
q R sin x*, x* R cos X3) and clockwise
motion. Since this circular arc must be tan-
Further, it is clear that 0(f) = 0 is possible gent to the line x2 =x\, thex2-coordinate of
only when A3(f) = 0. the center must be x2 R = x* R cos x*.
6-35
AMCP 706-192
Note that by Eq. 6-152, x3(t) must be

x = +R
continuous, so xt and x2 are continuous. * ~r(*2 cosx" +x* -R cosxj).
Therefore, the tangent to the path of points
in the (xltx2)-plane is Further, the relation noted just below Eq.
6-163 is
dx2 x2
=-r- = tan x3
(IX x Xj x\ -R =x* -R COSX3
and this slope is continuous. This means that These equations yield
segments of the optimal path where 6 = d0,
0, or 60 must be tangent where they intersect. x* = x\-R sinx +\R
2
-{xl2 -R
With this information, the solution of the
problem may be constructed geometrically.
-x2 -R cosx)2 1/2
In Fig. 6-12 the initial arc, which is

described by Eq. 6-162, is shown leaving X*=^(*2+*2+COSX3-.R).
(x\,x\). A whole family of second arcs is
shown corresponding to different values of
It may be noted by examining the family
of paths in Fig. 6-12 that if x\ > s=R + x +
R cos X3, then the first arc has been followed
beyond a time t where x3(f) = ir/2. At the
point xt(f) = x" -R sinx" +R, x2(i) = x7 +
R cos X3 it would have been possible to
construct a vertical portion of the optimal
path. This construction is shown in Fig. 6-13.
The extremal paths constructed for x2 > s

satisfy all the conditions of the theorem so
<*?.*%>
Figure 6-12. Extremal Arcs
From the construction of Fig. 6-12 it is

clear that the point of tangency of the two lx1('),x2("))
circles (x*, x*) is at the middle of the line
joining their centers, i.e., (44)
**=T(*i ~
R sin
*3 +xf + R sinxj) Figure 6-13. Extremal Arcs With Straight
Section
6-36
AMCP 708-192
that they may be optimum. It is clear that for this case, Eq. 6-122 provides no information.
x\ < s there is only one possible solution of It is then required that the inequality, Eq.
the problem. For x\ > s this is not the case as 6-131, must be used to determine the design
shown for x\ = x\. Both the extremals variable. For a complete treatment of this
leading to the path x2 - x2 satisfy the subject, see Ref. 20.
necessary conditions of the theorem. It is,
geometrically, relatively clear that these are The problem treated in this paragraph is
the only two possibilities so the one with the not as complicated as most optimal design
shortest time required to get to x2 = x\ is to problems occuring in the real-world. It does,
be chosen. The test, Eq. 6-151, may eliminate however, illustrate some of the features and
one candidate. It seems clear that when the difficulty encountered in most realistic
extremal with straight line exists, it is best. optimal design problems. This problem should
convince the reader that the solution of
It should be noted that if x\ > s + 2R, it is optimal design problems is not simply a
impossible to intersect x2 = xl with only two matter of plugging numbers into formulas.
circular arcs so the extremal with a straight Even though analytical methods will be
section is required. stressed in subsequent work, the effective
solution of this class of problems requires a
This problem illustrates many of the basic sound understanding of the theory of optimal
ideas and complexities involved in optimal design.
design and optimal control theory. Some of
the features are worth noting because they Example 6-9: A Constrained Brachisto-
will arise later: chrone Problem.
1. Pieced extremals. The conditions of The problem considered here is similar to

Theorem 6-7 give a set of curves or solutions Example 6-4 but with a constraint added. It is
that must be pieced together to form the required to find the path through (0,0) which
optimal path in state space. In the vehicle lies above the line x2 - h + x x tan a in the
steering problem, these curves or arcs are put (x1,x2)-plane and that carries a particle,
together geometrically. In more complex without friction, to the vertical line xl = x\
problems, this will have to be done analyti- in the shortest possible time. The problem is
cally using the conditions of Theorem 6-7. shown in Fig. 6-14.
2. Multiple solutions. As seen in the fore- This problem will be treated as an optimal
going problem, more than one candidate design problem. On the assumption that there
solution may be constructed. Condition, Eq. are no discontunities in the velocity vector,
6-13 1, must then be used to choose the best conservation of energy yields
of these candidates.
1
3. Singular arcs. It occasionally happens, as -mv = mgx2
in the vehicle steering problem, that there will
exist a set of values of the state variables and or
multipliers such that the function // does not
V = (2gX2)l/2.
depend explicitly on the design variable. In
6-37
AMCP 706-192
1/2 1/2
0 = (2g*2) sin H - (2gx2) tan a cos ^
0 (6-167)
which does contain u explicitly. The con-

straint, Eq. 6-166, is, therefore, a first-order
state variable inequality constraint.
x z =h +x. tana
1
In order to employ Theorem 6-8, define

multipliers yf, T~, X. such that
Figure 6-/4. Bounded Brachistochrone
The equations of motion of the particle are G = r+7,x1(0) + 72x2(0)

then
+ 73 1*1(7')-*}]'
Xi = {2gx2) cosu
(6-164) G ~T~ (*j x\ tan a h)
x2 = (2gx2) sin u
1/2
H=Xi(2gx2) cos > (6-168)
where u is the angle between the x ,-axis and
the tangent to the path on which the particle 1/2
+ X2(2gx2) sin
is to travel. This angle u specifies the curve, so
it is the design variable. The location of the -"(2gx2)
particle is specified by the point (xl,x2) so
this is the state variable. The boundary x(cos u tan a sin u).
conditions are
Xl{0) = x2(0) = 0
Necessary conditions from Theorem 6-8 are
(6-165)
Xi{T)=x\.
X, =0
The object is then to find u(t), xt(t), and
1/2
x2(t) such that a particle starting at rest at X2 = -g(2gx2) [Xj cos
(0,0) reaches x(T) = x\ in minimum time T. (6-169)
The path is required to satisfy the constraint + X2 sin /x(cos u
4> = x2 xl tana h<0. (6-166) - tan a sinu)\
1/2
Since the constraint of Eq. 6-166 does not (2gx2) [-X, sinw+X2 COSJ;
involve the design variable u explicitly, the
problem contains a state variable inequality + ^(sin u +tan a cosu)] = 0 (6-170)
constraint. Computing <j> and substituting
from Eq. 6-164 yields \2(T)=0 (6-171)
6-38
AMCP 706-192
-T~ tana + Xj (r -0)
-X, (t- + 0)=0

i (6-172)
the optimum curve is given by
/IT
2
Wi(, 0< t < t
T~ +\2{r -o) u(t) = < a, t-< t< t+
-\2(t- + 0)=0 K u2(T-t), t+ < t< T
l-H(T-0) = 0 (6-173) where
r(a -IT/2 +cot a;1

-H{r -0) + H((- +0)=0, (6-174)
2h cot a 1
and
"g (a + cot SL) j1/2
dH w2
- 0. (6-175) ,2(x\ +hcota) [
dt
Ideally, the solution for u(t) might proceed TT/2 a

t~ =
by solving Eq. 6-170 for u as a function of A
and n- This result could then be substituted
into Eqs. 6-164, 6-167, and 6-169. The t+ = T-a/(2u>2)
variables A, x, and n could then be determined
and the results substituted back into the and
previously derived equation forî. This would
be the desired solution. It is clear that these 1 1/2
(x\ + h cot a)(a + cot a.)
steps would be extremely messy so a huristic g
argument will be used here to suggest a
solution. This solution can then be checked in 2h ir 112
the conditions Eqs. 6-169 through 6-175. cot (a.-+cota

g 2
It might be expected that when 0 = 0, then Fig. 6-15 shows solutions for tan a = 1/2
the curve is a cycloid as in Example 6-5. and several values of h
Whenever <p = 0 it is clear that u = a. This is,
in fact, the case and as presented in Ref. 21 The reader may very well get the impres-
the solution is a cycloid for sion from these examples that analytical
solutions of general optimal design problems
h 2
> -IT - a tana are extremely difficult to obtain. This is
\2 indeed the case. Therefore, either numerical
methods must be used to solve the equations
i.e., the optimum path does not touch the given as necessary conditions in the theorems,
constraint surface. or some direct computational method must
be used to solve the optimal design problem.
For Some numerical methods of solving the neces-
sary conditions are presented in the next
paragraph. Several optimal structural design
1- a ) tan a
problems are solved in Chapter 7 to illustrate
6-39
AMCP 706-192
boundary-value problem to a sequence of

linear boundary-value problems whose solu-
tions converge to the solution of the non-
linear problem.
6-5.1 INITIAL VALUE METHODS (OR

SHOOTING TECHNIQUES)
In order to develop the main ideas without

getting bogged down in notation, consider the
problem of finding y(t) = {y^t),..., yn(t)]T,
that satisfies
Figure 6-15. Bounded Brachistochrone

Solution ^=f(t,y)J < t* t1 (6-176)
these methods. Direct methods of solving and

optimal design problems are presented in later
chapters. ^(f0) = yj ,for some!
(6-177)
6-5 METHODS OF SATISFYING NECES- yftl)=yli , for some/'
SARY CONDITIONS
where the total number of conditions in Eq.
The previous three paragraphs of this chap 6-177 is n. In order to further simplify
ter have been devoted to obtaining necessary notation, assume the components of y{t) have
conditions for optimization problems of vary- been numbered so that the first equation in
ing degrees of difficulty. It has been observed Eq. 6-177 holds fori = 1,..., k < n.
that these necessary conditions generally
reduce to some sort of boundary-value prob- Since initial-value problems are so efficient-
lem, usually nonlinear. The object of this ly integrated forward in time, the missing
paragraph is to explore ways in which the conditions onj' at t may be estimated as
boundary-value problem may be solved. This
topic has received considerable treatment in yAt) = t.,i = k + l,...,n (6-178)
the recent literature, so it will be treated only
briefly here. and Eq. 6-176 integrated from t to tx using
the full set of initial conditions from the first
Two different methods will be discussed equation at Eq. 6-177 and Eq. 6-178. The
here and will be applied to optimal structural value of yXt1) obtained from this integration
design problems in the next chapter. The first will probably not satisfy the second equation
method is based on a reduction of the inEq. 6-177, i.e.,
boundary-value problem to a sequence of
initial-value problems whose solutions con- y,-(t1;!:)*y1ri = k+ 1,..., (6-179)
verge to the solution of the original problem.
The second method reduces a nonlinear where = (k + 1, .... )r and the notation of
6-40
AMCP 706-192
Eq. 6-179 is introduced to illustrate the The initial value problems, Eqs. 6-182 and
dependence of the final values of y on|. 6-183, for i = k + 1,..., n may be integrated
from t to tl to obtain the derivatives
It is clear that a solution of the problem required inEq. 6-181.
can be obtained if % can be found so that Eqs.
6-179 are equalities. To simplify notation, Once these derivatives have been deter-
define the column vectors mined, the new estimate |_ in Newton's
Method is given by
Ht1 ;D = iyjU1 ;)] for those; inEq. 6-177
and 1=?0 W1;*0)
[y(t1;H0)-y1
a?
y1 =
lyj ] for the same/.
(6-184)
The conditions which are to determine are
The process is repeated with ' playing the
yV1;Z)=y1. (6-180)
role previously occupied by .
Any numerical method of solving algebraic
equations may be used to solve Eq. 6-189. If a This method of finding the partial deriva-
scheme like Newton's Method or a Gradient tives is direct in nature but requires the
Method is to be used, it must be possible to solution of Eq. 6-182 n k times. Further,
compute both the differential equations, Eq. 6-176 and
6-182, must be programmed.
dy
as (t'\?)
(6-181) A second method of constructing the par-
tial derivatives of Eq. 6-181 (or approxima-
where is an estimate of the solution of Eq. tions of them) is to use a difference quotient,
6-180. These partial derivatives may be ob- i.e., Eq. 6-176 is solved for and + 5 where
tained or approximated in a number of ways. 5 = (0, ..., e, ..., 0)r where i indicates the ith
position and E is small. Therefore
The first method of determining the deriva-
tives in Eq. 6-181 is to observe that y(t) = <W ;6) ^ y1 ;$ + S) - yjt1 ;Q
y(t-) and further, that the dependence on
is very regular (Ref. 14) so that by{.t&)$ (6-185)
exists. Differentiating formally with respect
to inEq. 6-176,
Once these approximate derivatives are
determined, the algorithm, Eq. 6-184, may be
zy\ a/ ay (6-182) used.
dt
This approximate method of constructing

and
the partial derivatives requires that the differ-
ential equation, Eq. 6-176, be solved n k
(0)= [0,..., 0,..., 1,0 . .01
additional times. It, therefore, requires
approximately the same amount of computa-
/ = *+!,. (6-183) tion as the previous scheme, but all the
6-41
AMCP 706-192
computation is performed with the same set The Generalized Newton Method for
of differential equations. This method is solving Eqs. 6-186 and 6-187 is similar in
illustrated in the problems of pars. 7-2 and philosophy to the Newton method of solving
7-3. algebraic equations. An estimate of the solu-
tion, y<0Ht), is made and the right side of Eq.
A third scheme which makes use of differ- 6-186 is expanded about ;v(0)(f) using Tay-
entiation formulas for definite integrals is lor's formula to obtain
developed in par. 7-4.
^- =^-lt,y^\t)]y^+f[t,/Ht)]
6-5.2 A GENERALIZED NEWTON METH- dt dy
OD
(0)
Ityv'VHy^'V) (6-188)
A second method which is used to solve the dy
necessary conditions for optimization prob-
lems is a Generalized Newton Method of where y^Ht) is required to satisfy
solving boundary-value problems. It has been
pointed out in the foregoing that Bolza >'J1)(r)=>',0,forthoseiinEq. 6-187
problems and optimal design problems may
be reduced to nonlinear boundary-value prob- y)l)(tl )=y}, for thosej in Eq. 6-187
lems. The method employed here was devel-
oped forjust such problems (Refs. 22, 23). (6-189)
In order to introduce the Generalized The boundary-value problem for >>(1)(0 is

Newton Method for boundary-value prob- linear so that if it has a solution, that solution
lems, consider the system of first-order equa- may be obtained by superposition techniques,
tions or any other technique for solving linear
boundary-value problems, for that matter
dy (Ref. 24).
- =g(y,t) (6-186)
dt
The function y^Ht) is taken as an im-
where proved estimate for the solution of Eqs. 6-186
and 6-187. This estimate then replacesy(0){t)
y(0 = LM')> - yn(t)]T and in the preceding analysis. If k is the iteration
number for this process, then yik)(t) is
g(y,t) = Igify.t), .... g(y,t)]T determined by
In addition to satisfying Eq. 6-186, y{t) is dyW df

required to satisfy ^ \t,y (*- i) (t) ,,(*)
dt dy I
yf{t0) = yf, for some i

+f[t.y(k-l\t)} (6-190)
(6-187)
yft1) =
y)> fr some;', 9/
[t,y<-k- D {t)\y(k-l\t)
dy
where the total number of conditions in Eq.
6-187 is n.
6-42
AMCP 706-192
and the boundary conditions and
y\k\t) = yf ,for those i inEq. 6-187 x((t) = xf for some i

(6-194)
yfht1) = yj, for those/' in Eq. 6-187 xJt1) = x. for some /
(6-191) where the total number of boundary condi-

tions in Eq. 6-194 may be less than, equal to,
The sequence of approximations to the or greater than n, x(t) = [x,(0 -, xn(t)]T,
solution \y^k\t)] is considered to have con- u(t) = [u, {t), .... mU)]T.
verged when the difference between succes-
sive iterates is sufficiently small. Theorems Defining
given in Ref. 23 show that if the initial
estimate of the solution >,(0)(?) is sufficiently
// = X0/ +;2 X./,,
accurate, then under rather restrictive condi-
tions, the sequence [y^Ht)] converges to the
solution of Eqs. 6-186 and 6-187. Further, the necessary conditions of Theorem 6-5 are
the convergence is quadratic in the sense that
the error at the k + 1st iteration is propor-
dh _ -_ M
tional to the error squared in the kth itera- (6-195)
tion. This kind of convergence is extremely dt ax
nice.
= 0 (6-196)
Even though it is difficult or impossible to du
verify the hypotheses of the convergence
and
theorems in Ref. 23, it has been observed in
practice (Ref. 23) that good convergence is \V) = 0, rîinEq. 6-194
nevertheless obtained in many real-world (6-197)
problems. X a1) =0,s=tj inEq. 6-194.
Since the discussion in this paragraph is on The argument used in applying the General-
ways of solving optimization problems, the ized Newton Method to the problem of
Generalized Newton Method will be applied determining x{t), u(t), and $t) from Eqs.
more directly to this class of problems. For 6-193, 6-194, and 6-195 through 6-197 as
the present, consider only the following prob- follows:
lem:
1. Solve Eq. 6-187 for
minimize J
I' f(t,x,u)dt (6-192) u =u(t,x,$ (6-198)
and substitute this expression into Eqs. 6-193

subject to and 6-195.
dx 2. These differential equations then form

= fU.x.u) (6-193)
dt 2n first-order, nonlinear differential equations
6-43
AMCP 706-192
in 2n variables. Further, there are exactly 2n an auxiliary parameter e; is introduced

boundary conditions inEqs. 6-194and 6-197. through
This nonlinear boundary-value problem is
now solved by the Generalized Newton Meth-
od. ^f{t)H[^Ji(t)]dt = ei (6-202)
3. The solution x(t), \(t) is then sub-

stituted into Eq. 6-198 to obtain the optimal where
design function.
0, s< 0
Since the Generalized Newton Method, as H(s)
presented here, is only capable of solving \,s> 0.
two-point boundary-value problems, inequa-
In a sense, e. is a measure of violation of Eq.
lity constraints may not be treated explicitly.
6-201. The procedure in solving an optimal
Rather, the general optimal design problem
design problem with a constraint of this kind
with inequality constraints must be reduced
is to solve a sequence of problems with Eq.
to a problem with only equality constraints.
6-202 replacing Eq. 6-201, and ej(k) ap-
For example, for problems with constraints of
proaching zero as k becomes infinite; i.e., a
the form
modified design problem is solved imposing
Eq. 6-202 in place of Eq. 6-201 withe/0) > 0
<j>,(t,x,) < 0, (6-199)
chosen. This solution is carried out through
use of the Generalized Newton Method de-
where 4>t depends explicitly on u a trans- scribed. The problem is then solved again with
formation may be performed by introducing K
0 < ec(i)' < e( ' beginning the iteration with
an auxiliary design variable (slack variable) the solution of the preceding problem. The
aft) through the relation process is repeated with 0 < e(fc) < e^-1'
until changes in successive solutions are suf-
4>i(t,x,u)+<xf{t)=0. (6-200) ficiently small.
It is clear that with the new variable, Eq. The Generalized Newton Method presented
6-200 is equivalent to Eq. 6-199. The neces- here has been discussed by many authors and
sary conditions of Theorem 6-5 may now be generally has received favorable comments.
applied and the Generalized Newton Method For a more detailed discussion and examples,
utilized just as in the preceding case. see Refs. 23, 25 through 28. An outstanding
treatment of the Generalized Newton Method
In case the optimal design problems with also appears in book form (Ref. 29). A very
state variable inequality constraints, a differ- rigorous treatment of existence and con-
ent technique for elimination of inequalities vergence properties of the method is given
has proved effective. For constraints of the which applies to the control problems dis-
form cussed. The reader should note that some
writers follow Bellman in calling the method
4>,(t,x) < 0 (6-201) described here, "Quasilinearization".
6-44
AMCP 706-192
REFERENCES
1. D.G. Luenberger, Optimization by Vec- vances in Control Systems, Vol. 3, C.T.

tor Space Methods, John Wiley, & Sons, Leondes Ed., Academic Press, New York,
New York, 1969. 1966.
2- I.M. Ge If and and S.V.Fomin, Calculus of 12. L.S. Pontryagin, V.G. Boltyanskii, R.V.
Variations, Prentice-Hall, Englewood Gamkrelidze, and E.F. Mishchenko, The
Cliffs, New Jersey, 1963. Mathematical Theory of Optimal Pro-
cesses, John Wiley & Sons, New York,
3. N.I. Akhiezer, The Calculus of Variations,
1962.
Blaisdell, New York, 1962.
13. L.D. Berkovitz, "Variational Methods in

4. L.E. Elsgolc, Calculus of Variations, Per-
Problems of Control and Programming",
gamon, London, 1962.
/. Math. Anal. Appl., Vol. 3, pp.
145-169, 1961.
5. M.R. Hesteness, Calculus of Variations
and Optimal Control Theory, John Wiley
14. EA. Coddington and N Levinson, Theory
& Sons, New York, 1966.
of Ordinary Differential Equations,
6. LA. Liusternik and V.J. Sobolev,Ele-
ments cf Functional Analysis, Ungar,
15. G. Leitmann, Ed., Optimization Tech-
New York, 1961.
niques, Academic Press, New York, 1962.
7. E.K. Blum, "The Calculus of Variations,

16. W.F. Denham, Steepest-Ascent Solution
Functional Analysis, and Optimal Con-
of Optimal Programming Problems, Ray-
trol Problems", in Topics in Optimiza-
theon Report No. BR-2393, Raytheon
tion, G. Leitmann Ed., Academic Press,
Company, Bedford, Mass., 1963.
New York, 1967.
17. A.V. Balakrishnan and L.W. Neustadt,

8. M.R. Hesteness, "On Variational Theory
Ed., Computational Methods in Optimi-
and Optimal Control Theory", J. SIAM,
zation Problems, Academic Press, New
Series A: Control, Vol. 3, No. 1, pp.
York, 1964.
23-48, 1965.
9. GA. Bhss, Lectures on the Calculus cf 18. D.S. Hauge, "Solution of Multiple-arc
Variations, University of Chicago Press, Problems by the Steepest-Descent Meth-
Chicago, 1961. od", in Recent Advances in Optimization
Techniques, A. Lavi and T.P. Vogl, Eds.,
10. E.B. Lee and L. Markus, Foundations of John Wiley & Sons, New York, 1966.
Optimal Control Theory, John Wiley &
Sons, New York, 1967. 19. H.M. Hung, Time-Optimal Steering of a
Ground Vehicle, presented at the 1968
11. W.W. Schmaedke, "The Existence Theory Spring Meeting of SIAM, Toronto,
of Optimal Control Systems", inAd- Canada.
6-45
AMCP 706-192
20. H.J. Kelly, R.E. Kopp, and H.G. Moyer, Blaisdell, Waltham, Mass., 1968.
"Singular Extremals", in Topics in Opti-
mization, G. Leitmann, Ed., Academic 25. G. Paine, The Application of the Method
Press, New York, 1967. of Quasilinearization to the Computation
of Optimal Control, Report No. 67-49,
21. A.E. Bryson Jr., W.F. Denham, and S.E. Department of Engineering, UCLA,
Dreyfus, "Optimal Programming Prob- August 1967.
lems with Inequality Constraints I: Nec-
26. R.E. Kopp and H.G. Moyer, "Trajectory
essary Conditions for Extremal Solu-
Optimization Techniques", in Advances
tions", AIAA J., Vol. 11, November
in Control Systems, Vol. 4, C.T. Leondes,
1963, pp. 2544-2550.
Ed., Academic Press, New York, 1966.
22. L.V. Kantorovich and G.P. Akilov, Func-

27. A.P. Sage, Optimum Systems Control,
tional Analysis in Normed Spaces,
Prentice-Hall, Englewood Cliffs, N.J.,
Pergamon Press, London, 1965.
1968.
23. P. Kenneth and R. McGill, "Two-Point 28. B.D. Tapley and J.M. Lewallen, "Com-
Boundary-Value-Problems Techniques", parison of Several Numerical Optimiza-
in Advances in Control Systems, Vol. 3, tion Methods", J. Opt. Theory and Appl. ,
C.T. Leondes, Ed., Academic Press, New- Vol. l,No. 1, 1967, pp. 1-32.
York, 1966.
29. PL. Falb, Some Successive Approxima-
24. H.B. Keller, Numerical Methods for tion Methods in Control and Oscillation
Two-Point Boundary- Value Problems, Theory, McGraw-Hill, New York, 1971.
6-46
AMCP 708-192
CHAPTER 7
OPTIMAL STRUCTURAL DESIGN BY THE INDIRECT METHOD
7-1 INTRODUCTION As systems become more complex and

more emphasis is placed on minimum cost,
7-1.1 THE CLASS OF PROBLEMS CON- the designer is unable to make all the trade-
SIDERED off analyses mentally. A method of design
synthesis, therefore, is necessary which is able
Since the beginning of engineering disci- to include all requirements on the system and
plines, the engineer has attempted to develop the requirement of minimum cost in a unified
structures and machines that perform some design procedure. One such method for opti-
specified task. In the case of structures, a mal structural design is illustrated in this
frame or truss is required to support a given chapter.
system of loads. Likewise, machines and
machine elements are required to support 7-1.2 HISTORICAL DEVELOPMENT
loads while they perform some function.
Very early in the development of me-
The objective of the examples treated here chanics of materials, methods of determining
is to illustrate organized methods that the stress and displacement for given bodies under
engineer may use to obtain a load-carrying the action of given forces were emphasized.
system which is best in some sense that is As these methods became better developed,
associated with the particular application. In the question arose as to how a structure might
design of commercial goods, the dollar cost of be proportioned to satisfy certain require-
an element is probably the index that is to be ments and be best in some sense. Problems of
minimized (Ref. 1). In military and aerospace this kind were considered by Lagrange (Ref.
applications, while dollar cost is important, 2) in 1771 and by Clausen (Ref. 3) in 1851.
frequently weight cost is even more essential.
In the example problems presented here, the Until very recent years, methods of the
criterion of minimum weight will be chosen. calculus of variations were not sufficient for
treating realistic design problems. Probably
Until very recently, most design procedures for this reason, design problems were stated in
depended on the engineer's intuition and terms of a few parameters that specified the
experience in proportioning a load-carrying structure. For example, uniform beams of
system. An analysis of the proposed con- undetermined depth are placed in a given
figuration was then made to determine wheth- configuration. The depths are then deter-
er the system met all requirements placed on mined so that the structure supports the given
it. If not or if the preliminary design was loads and is as light as possible. For a detailed
obviously excessively strong, the procedure bibliography of this development through
was repea ed until a satisfactory solution was 1963, see Ref. 4. For a more current bibliog-
obtained. raphy, see Ref. 5.
7-1
AMCP 706-192
Another important method of design devel- conditions that the solution of the optimal
oped principally by Prager and Drucker (Refs. design problems must satisfy. Once this task is
6,7,8) is limit analysis. In this method of complete, the design problem is reduced to
design, the structure is allowed to reach a the determination of solutions of the neces-
state of collapse due to plastic action of the sary conditions that are candidate solutions of
material. The resulting design is, therefore, the optimal design problem. The term "in-
safe for application of the given loads even direct" seems to describe this process quite
though permanent deformation of the struc- well.
ture results. If the loads must be applied
many times in the life of the structure, it will As discussed in par. 6-5, any method of
generally be required that all material in the solving the nonlinear boundary-value problem
structure must remain in the elastic range at contained within the necessary conditions is
all times. For this reason, methods had to be admissible. In this chapter, two problems will
developed for elastic design. be solved by shooting techniques. The prob-
lems of par. 7-2 are treated by the shooting
In 1960, Joseph B. Keller published an technique of par. 6-5. The problems of par.
article on column design (Ref. 9) which 7-3, however, are treated by a modified
renewed interest in elastic, minimum weight shooting technique.
design. Several papers have subsequently been
published by Keller and his associates in 7-2 A MINIMUM WEIGHT COLUMN
which a class of eigenvalue problems is treated
(Refs. 10,11,12). The methods employed in A lightweight column of length T is to be
these papers are elegant but are not easily designed to support a given load P. The
adapted to realistic engineering problems. material is specified and has yield strength
a
max- The particular support considered is
A new method of optimal design has been shown in Fig. 7-1. In problems considered
developed by J.E. Taylor and W. Prager since
1967 (Refs. 13,14,15). This method is based
on an energy representation of the structural
element under consideration. A particularly
nice feature of the method is the ability to
obtain sufficient conditions for certain classes
of design problems. However, no unified
method of constructing solutions has been
presented.
7-7777-7- / / ff f
7-1.3 METHODS EMPLOYED
(A) Undeflected (B) Deflected
The theorems of Chapter 6 will be em- Figure 7-1. Column Under Consideration
ployed here for the solution of optimal design
problems, Use of the results of Chapter 6 to here, the cross section is assumed to depend
construct solutions of optimal design prob- on only one design variable, u(t), 0 < to T.
lems is called an indirect method of solution. The problem is to determine u(t) that mini-
This is so, because one first obtains a set of mizes the weight or, equivalently, the volume
7-2
AMCP 706-192
the optimal control problem considered in

J= 1 A[u{t)\dt (7-1) par. 6-4.
Jo
For use in Theorem 6-7, construct
and satisfies the conditions,
' Px,
H- \0A(u) + XjX2 X2
. Eliu)
EI(u)^ +Px =0,
at
(7-2)
M max
_A(u) _
*(0) = 0, -CD = 0
at
G = X1x,(0) + X2x2(7T).
and
Conditions, Eqs. 6-121 and 6-124, yield
a m ax < 0 (7-3) d\i _ M_ _ P\2

A{u)
dt ~~ dxi ~ EI(u)
where
d\2 _ dH
(7-5)
x(?) = lateral deflection of the column dt dx2
t = distance measured along the column Xj(0) = o,x,(r) = o.
A(u) = area of the cross section The system, Eq. 7-5, reduces to
/() = smallest moment of inertia of the

d2\2 P\2
area of the cross section about a centroidal
dt2 EI(u)
axis. All cross sections are assumed to have
(7-6)
two orthogonal axes of symmetry with P
d\2
acting through their intersection. X2(0) = 0, =0
By defining x, = x and x2 = dxt/dt, Eq.

Eq. 7-6 for X2(f) is identical to Eq. 6-14 for
7-2 reduces to the system
x(t). Both problems are homogeneous, how-
ever, so X2(r) and x(t) may differ by an
dxi
= *2 =/i arbitrary constant multiplier, say X0, i.e., put
~dt
X2(0 = \0x(t). This problem is normal (Ref.
17), so X0 7^ 0 may be chosen as one.
dx Px,
2 (7-4)
dt EI(u) Condition, Eq. 6-128, of Theorem 6-7 is, in
this case, (P/A(u) _ amax) = 0. Two
xl(0) = 0,x2(.T) = 0 possibilities now exist; eitherM = 0, oiP/A(u)
m ax = 0. In the second case, u is just the
The problem is thus reduced to the form of algebraic solution of
7-3
AMCP 708-192
P/A(u) = a (7-7) dx
x(0) = 0, (T)= 0
at
In the remaining case, n = 0 and condition,
Eq. 6-122, of Theorem 6-7 is where the choice on the right side of Eq. 7-12
must correspond to the selection inEq. 7-11.
M___M _tfh [1//(u)1 0. This boundary-value problem is solved by

9M 9M I 3
an iterative method based on Newton's
Algorithm. The missing initial condition is
(7-8)
taken as dx/dt(0) = C. Integration of the
The design variable u(t) is thus determined resulting initial value problem from 0 to T
in subintervals of [0,T] by either Eq. 7-7 or yields an error dx/dt(T;C) in the final value.
7-8. So that the results of the present method
may be compared with those obtained by This notation is chosen to emphasize the
Keller (Ref. 9), choose A = u and / = au2. dependence of x on the estimate C of the
This corresponds to having the geometric missing initial condition. The objective is to
shape of the cross section fixed and allowing find C so that dx/dt(T;Q = 0. Once C is
all dimensions to vary as u1/2. found, the initial value problem for x(t) may
be solved and u(t) determined from Eq. 7-11.
With this form of A{u) and I(u), Eqs. 7-7
and 7-8 become In order to employ Newton's Algorithm,
d/dC [dx/dt(T;Q] is needed. It is obtained
P
l = "max (7-9)
by formally differentiating Eq. 7-12 with
and respect to C to obtain
2Px2
Eau*
(7-10) 1.
dt2
EaP
or
Condition, Eq. 6-131, of Theorem 6-7 1/3

requires that the expression for u be chosen x-w r (7-13)
which satisfies the constraint, Eq. 7-3, and
3 \m)
makes H as large as possible. This criterion
yields the choice between f(0) = 0,(0)= 1
dt
1/3
UP where the choice on the right side of Eq. 7-13
a(0! {Pia ) or (
""' ' Ea
x2/3|(7-ll)
must correspond to the selection in Eq. 7-11.
The order of taking derivatives has been
When Eq. 7-11 is substituted into Eq. 7-2, changed and the notation f = dx/dC intro-
duced in obtaining Eq. 7-13.
d2x Px_
dt2 ' Eoi (%ax/^)2 or
The iterative method for determining C is
then:
2/3
Ea
v- "4/3 (7-12;
IP Step 1. Make estimate C= CQ
7-4
AMCP 708-192
Step 2. Integrate the differential equation A FORTRAN program was written to

inEq. 7-12 with perform the iterative procedure. The program
was run on an IBM 360-65 Computer and
dx required approximately 0.1 sec per iteration
(0)= C0, x(0) = 0, and Eq. 7-13.
at and only four to six iterations to converge.
Step 3. Make the adjustment in C The results for this design problem are
given in Table 7-1 and Fig. 7-2. For loads
dx _ /<# above 6794 lb, the cross-sectional area is
Cl = C0 (P / (71
dt / dr ' determined by ^4 = P/omiX and the resulting
column is stable. A meaningful optimal design
Step 4. Return to Step 2 with new esti- problem then exists only for P < 6794.
mate C= Ct , and repeat.
7-3 A MINIMUM WEIGHT STRUCTURE
The equations derived here must be WITH ANGULAR DEFLECTION RE-
changed only slightly to solve column prob- QUIREMENTS
lems with other end conditions and other
forms of cross section. 7-3.1 STATEMENT OF THE PROBLEM
For a numerical example of this problem, The problem considered in this paragraph is
let the cross section be circular with variable the design of a portable communication tower
radius. In this case a = l/(47r). For the exam- of height L which will support a line-of-site
ple, let omax = 20,000 psi, = 3x10' psi, and transmission unit, a laser transmitter, for
7= 10 in. example. In order for the transmission beam
TABLE 7-1
RESULTS FOR COLUMN PROBLEM
Volume of
P, lb Volume, in.3 Uniform Column', in. Saving,'.
50 0.260 0.291 10.6

100 0.361 0.412 12.4
200 0.507 0.595 14.7
500 0.806 0.923 12.7
1000 1.140 1.300 12.3
1500 1.408 1.600 11.9
2000 1.640 1.840 10.9
3000 2.048 2.260 9.3
4000 2.412 2.600 7.3
5000 2.765 2.910 5.2
6794 3.397 3.397 0.0
1
Minimum Weight Uniform Column
7-5
AMCP 706-192
A
r The general configuration of the tower is
shown in Fig. 7-3. Three vertical members
with cross-sectional area A{t) are arranged on
the vertices of an equilateral triangle of
altitude hit). Here t is a coordinate measured
along the length of the tower. In order to
maintain the spacing of the vertical elements,
small cross members are inserted.
P=S0 P=100 P=500 P=1000 P= 1500
A(t)
Figure 7-3. Tower Considered
It is assumed that the tower is constructed

of a given material with density p. Further, it
P - 5000 is assumed that h cubic units of material are
P = 2000 P = 3000 p = ei94
required per unit height of tower in order to
maintain the spacing of the vertical elements.
Figure 7-2. Profiles of Optimal Columns
The coefficient is to be determined from
design experience. For this configuration the
to hit the receiving unit, the top of the tower total weight of the tower is
on which the transmitter unit is mounted
must undergo only a certain allowable rota-
W= \ 3p lA(t)+h(t)) dt.
tion 6 when the tower is exposed to a given
extreme uniform wind load Q pounds per unit Jo
length of tower. Since 3p is a constant, W is minimized if and
only if
It is required that the tower be as light-
weight as possible so that it may be trans-
ported and erected without the aid of heavy
machinery. For this reason, the design crite-
rion is minimum weight. However, one addi-
V=
i: [A(t) + h)] dt (7-14)
tional requirement must be placed on the is minimum. The objective in the design
tower. In transportation and election it must problem is to choose/4(f) and h(t) for 0 < t <
be strong enough so that it is not damaged by L, so that V is as small as possible and the
rough treatment. Therefore, it is required that given conditions are met.
the moment of inertia of the cross-sectional
area of the tower be greater than a prede- Lateral deflection of the tower due to the
termined limit I0 everywhere. lateral wind load is determined by elementary
7-6
AMCP 706-192
beam theory. The differential equation for considered in par. 6-4, define
displacement is (Ref. 14)
XI ^
EJ(t)x" = -M(t) (7-15)

x^ =34
where
with this notation and that of par. 6-4, the
E = Young's modulus of the material second-order equation 7-2 is equivalent to
I(t) = minimum moment of inertia of Xj - x2 ft

the cross-sectional area of the (7-17)
tower M_
El
x = lateral displacement
d2x The design problem will now be solved for
two admissible configurations of the tower.
dt2
Mit) = bending moment of tower. In order to compare results obtained for
the various configurations considered, a tower
The moment of inertia of the cross section is with properties of Table 7-2 will be treated.
I)=-A(t){h(t)]2. (7-16)
TABLE 7-2
CONSTANTS
In order to prevent damage in handling, it is
L = 360 in. 6 = O.OOOlrad
required that
Q = 8.35lb/in. = 0.25in.
I(t) > I0 > 0, 0 < t < L.
E = 3x 107 lb/in. / = 172.8in?
or in the notation of par. 6-4,
<j>=I0 -I(t)< 0,0< t-i L.

7-3.2 TOWER WITH ONE DESIGN VARI-
If I(t) = /0 gives a tower with angular ABLE
deflection less than or equal to 6 at the top,
then this is the optimal tower and no further For the problems considered in this para-
work is required. On the other hand, if this graph, A(t) will be held constant with the
tower has angular deflection greater than 8, value of A. Two ways of mounting the tower
then the tower is not admissible and it is on the earth will be chosen. The first method
required that the angular displacement is will be to fix the base of the tower rigidly to
equal to 0. This is the only situation con- the earth and leave the top unsupported. The
sidered here. second method will be to pin the lower end to
the earth and support the top with guy lines.
In order to fit this problem into the form Since^4 is constant
7-7
AMCP 706-192
V=
r
I
J 0
[A+h(t)]dt
par 7-3.1. Eqs. 7-7 7 become
x, X-i = fi
=AL+
r
Jo
h(t)dt.
Minimization of V in this case is equivalent

, _ 3Q(L - t)2
AEAh2
with boundary conditions

h
(7-20)
to minimization of
*,(0) = 0
J=
r
Jo
h{t)dt.
In the notation of par 6-4,

(7-18) x2(0)=0
x2(D = e
(7-21)
The problem stated here will now be solved

/o =h(t). using Theorem 6-7. Define
7-3.2.1 METHOD 1. TOWER WITH BASE 3g(l - t)2

H = Xi x-i + Xj Xnh
RIGIDLY FASTENED TO THE L 2EAh2
EARTH
The tower considered here is shown in Fig. "('-M

7-4. The bending moment M(t) due to the
and
G = yiXl(0)+y2x2(0) + y3[x2(L)-d].
The conditions of Theorem 6-7 yield
A', = 0
(7-22)
X2 = Xi
Figure 7-4. Loading of Tower

~6Q(L-t)2~
o=- = -x0-x2
2EAhl
wind load Q is
4
+ 11Ah (7-23)
M(t)=-^(L-t)2 (7-19)
X1(I)=0. (7-24)
where t is measured upward from the bottom
of the tower; the other symbols are defined in The general solution of Eq. 7-22 is
7-8
AMCP 706-192
X,(0 = Si Since Cases 1 and 2 cover all possibilities,
It
1/3
Xa(0 = -a -fi*- 3gagtf-Q2
h or
24
Condition, Eq. 7-24, implies 1 = 0, so (7-27;
1/2
34
Ai (0 = 0 24
f7-2j;
MO = -$a where by Eq. 6-131, the choice in Eq. 7-27
must be made which makes
For the determination of A<7), two cases
must be considered:
3tag(I-Q2
H=-h
4&4A2
Case 1: 4> = 0. In this case
a maximum.
2
-Ah -I0 =0
Let /* be a point of transition from one
expression in Eq. 7-27 to the other. Since
so Theorem 6-7 requires H to be continuous,
1/2 when the two expressions of Eq. 7-27 evalu-
3ô ated at t* are substituted into//, a common
/! = = Ai
2A value must occur.
Case 2: <t> < 0. In this case Eq. 6-128 is Evaluate H as a function of both h2 and
hi-
M0 = O
and since <t> = 0,ju = 0. Substituting this result //(/i,) = î =-Ai

into Eq. 7-25,
2
1 "36a(-0 "
X 6ag(J,-Q'_n r7-26; 2 2EA J A2
h
If X0 = 0, then, 2 = 0 and from Eq. 7-26, A, --
l
Xi = X2 = 0. This, however, violates the
condition stated in the first sentence of
Theorem 6-7. Therefore, X0 = 0 and it is H(h2) = H2 =-h2
permissible to choose X0 = 2.
Further, since X0 > 0, 2 > 0 in order that 3{a(-02

Eq. 7-26 hold. Eq. 7-26 then yields 2,4
3Q(-Q2 = -A2
A = = A, 2 /*
2,4
7-9
AMCP 706-192
Then form for 0 < t < t*. The problem is now to

determine %2.
3 1 h2
H2 -H,l = h22 + A, +
2 2 h2 The condition which has not yet been
satisfied is x2(L) = 0. Substituting Eq. 7-27
into Eq. 7-20 yields
(2Ai +Aj)(A, -A2)2
2h]
1/3
1 30 2/3 -2/3
If h2 < hi at any point i, thenA2 violates (L - t) ' %2 ,
2 V 2EA
the constraint and h=hi.lfh2=h1 at any t,
there are no alternative choices forA. If h2 > xi for 0 < t < t*
hi at any point t, the choice h = hi must
maximize H{t); this implies H2 //[ < 0. But Q{L - t)2
this is impossible from the above because h i , for t* < t hi , then it is required that h=h2. From r7-29;
this, it is concluded that
A(0 = maxtAjCr), A2(f)] withx2(0) = 0.
Since h(t) is defined as the maximum of two Integrating Eq. 7-29 first from 0 to f* (as
continuous functions, h is continuous. It given by Eq. 7-28) and then from t* to L
follows that all points t* of transition from yields
one value of Eq. 7-27 to another can be found
by equating the two expressions of h(t).
3L5/3 / 3Q\1/3 -2/3
(L)= b
The point t is then determined by ** 15" I IS)
1/2
3/p 1/
V0/12(23/4x3+10/0/6)_r3/2
hi(t*) =
2A
1/3
(24)1/4101/2//4
_ 3{aQ(j,_/)i (7-30)
2fe = ,(f*).
Eq. 7-50 is solved numerically for 2. Once
This solution yields two values of f*. The 2 is determined, then Eqs. 7-27 and 7-28
requirement 0 s ;* < L results in a unique completely specify the tower. Results are
value of/* shown in Table 7-3(A) and Fig. 7-5(A).
3' /a ii/4. (7-28)

This design problem has been solved analyt-
2 Ai\ Q2 ically. As will become apparent as more
realistic problems are treated, one should not
For 0 < t < L, the first form of A in Eq. expect to obtain solutions in this way. In
7-27 is monotone decreasing and is zero at f = most problems, numerical methods must be
L. It is, therefore, clear that the second form applied to solve the differential equations
of A must hold forf* < t <z L and the first arising in the Theorem 6-7.
7-10
TABLE 73(A) TABLE 73(B)
WEIGHTS OF SIMPLY SUPPORTED WEIGHTS OF GUY-LINE SUPPORTED
TOWERS, ONE DESIGN VARIABLE TOWERS, ONE DESIGN VARIABLE
A = 6.0 in. W = 2121.5 1b A = 3.9 in/ W = 1362.11b
A = 6.6 in.2 W = 2112.4 lb A = 4.2 in.2 W = 1357.6 1b
A = 6.9 in.2 W = 2111.41b A = 4.434 in.' W = 1356.61b
A = 7.2 in.2 W = 2112.31b A = 4.5 in.2 W = 1356.7 1b
A = 7.9 in.2 W = 2119.5 lb A = 4.8 in.2 W = 1358.81b
TABLE 73(C)
WEIGHTS OF TOWERS
Guy-line Guy-line Guy-line

Cantilevered Cantilevered Cantilevered Supported Supported Supported
Number of
Design
Variables 0 12 0 1
Best Weight W = 2440.6 1b W= 2111.4 1b W= 1827.91b W = 1563.991b W= 1356.61b W - 1265.71 lb
h = 63.7 in. h
max = 914in hmax = 80 2 in h = 46 in. hmax = 46 5 in h
max = 36 55 in
A = 7.96 in.2 6.97 in.2 10.03 in.2 A = 3.84 in.2 4.434 in.2 4.95 in.2 >
3
o
o
1
o
O)
(O
M
AMCP 706-192
x x
i ~ i /i
, _ Qt{f - L) _ (7-31)
=/2
2EI
The boundary conditions in this case are
*!(())= 0
jca(0) =e } (7-32)
(A) One Control Variable (B) Two Control Variables
x1(L)=0
Figure 7-5. Tower With Base Rigidly Fastened
to the Earth
The quantity to be minimized is still given
by Eq. 6-17. In the problem considered here,
7-3.2.2 METHOD 2. TOWER WITH BASE

3Qt(t-L)
PINNED TO EARTH AND WITH H = \iX2 +X2
4EAh2
TOP SUPPORTED BY GUY LINES
The tower considered here is shown in Fig. _X0A-M(/o ~4^2)

7-6. It is convenient here to locate the
G = ylXl(0)+y2[x2(0)-d] +y3Xl{L)
coordinate system at the top of the tower.
The bending moment generated by the uni-
Conditions of Theorem 6-7 are
form wind load Q is M = - g/2 t(t - L) so
the differential equation for bending is
dXj _ 3/7
=0
EI(h)x" = t(t-L) . dt dx
d\2 _ dH
= -X,
Define xt = x and x2 = x[; this is equiva- dt ~ dx-.
lent to
so
Xj(0 = -b -iit.
Also,
\2(L) = 0
This equation implies %2 = i I, so
Figure 7-6. Tower With Guy Lines
7-12
AMCP 708-192
X2 =f,( -t). of Eq. 7-31 and solution of the boundary

conditions, Eq. 7-32, for x are not feasible.
Two cases must now be considered: Therefore, the shooting technique of par. 6-5
is employed. Numerical results for this prob-
Case 1. 0 = 0. In this case lem are given in Fig. 7-7(A) and Table 7-3(B).
Âh2 -/0=0
so
1/2
37o
1 (7-33)
2A
Case 2. <j> < 0. In this case M

Substituting into dH/dh = 0,
Xo =
2^
As in the previous case X0 = 0 so it is

permissible to put X0 = 2 and obtain
(A) One Control (B) 'Two Control
1/3
Variable Variables
3$iQf(-02
(7-34)
2 4
Figure 7-7. Tower With Base Simply Supported
and Top Supported With Guy Lines
Eqs. 7-33 and 7-34 along with Eq. 6-131
yield
The modification of Newton's Algorithm

h(t) or consists of using a correction of at most 10%
of the values of the unknown iteration param-
1/3
3^Qt(L-t)2 eters. It has been found in particular problems
(7-35) that where the Newton Method fails to
2EA
converge, this method will converge. The rate
of convergence, particularly for the first few-
the choice in Eq. 7-35 being made which
iterations, is slowed by the modification,
makes
however.
3^Qt(L-t)2
H=- h 7-3.3 TOWER WITH TWO DESIGN VARI-
4EAh2
ABLES
largest.
The same two methods of supporting the
The problem is thus solved when , is towers will again be considered. Here, how-
determined. In this case, analytical integration ever, both A and h will be allowed to vary
7-13
AMCP 706-192
along the tower and play the role of design is permissible to take X0 = 1- The system is
variables. then two equations for h and A with solution
1/4
Eqs. 7-16 and 7-17 remain the same in the 2
h(t) - (3iQ a -t)V
present problem. However, fromEq. 7-14 2E
1/3
) (7-37)
I/2
f0=A(t) + h{t). 3t2Q'< (1-0
A{t)
7-3.3.1 METHOD 1. TOWER WITH BASE

RIGIDLY FASTENED TO THE
EARTH
Case 2. 0=0. This is
Fig. 7-4 applies and Eqs. 7-19, 7-20, and

jAh2=I0 (7-38)
7-21 hold in this problem. Substitution into
Eq. 6-108 yields
Eq. 7-38 along with Eq. 7-36 is a system of
three equations in h, A, and n and solving for
3Q(L - t)2
H = \\X2 + X2 -\0(A+h) h and A yields
4EAh2
1/3
H[I0 - \ Ah2
"-(- . (7-39)
The equations for Xt and X2 are just as in the
3/0|32
preceding work, so again, A{t)
Xi (t) = 0 The design variables are chosen by Eqs.

7-37 and 7-39, depending on which makes H
X2(0 = -b largest. The problem with differential equa-
tions, Eq. 7-20, and boundary conditions, Eq.
Since h plays the role of wj and^l the role of 7-21, is now treated by the shooting tech-
u2, Eq. 6-122 is:
nique of par. 6-5. Numerical results are given
in Table 7-3(C).
dH
3/i
.W-*^-,f,*-o 7-3.3.2 METHOD 2. TOWER WITH BASE
PINNED TO EARTH AND WITH
= 2 2 TOP SUPPORTED BY GUY LINES
7 X0 + + M -, h =0.
5A 4EAh 3
Fig. 7-6 applies and Eqs. 7-31 and 7-32
(7-36) hold for this problem. Substituting into Eq.
6-108 yeilds
As before, two cases must be considered:
3Qt(t - L)
H = X1x2 +X2
AEAh2
Case 1. 4 < 0. This implies p = 0. From
Eq. 7-36, it is clear that X0 = 0 implies 2 = 0 i
which contradicts Theorem 6-7. Therefore, it

X0(A + h) - Ji0 jAh
7-14
AMCP 708-192
The equations for A, and X2 are just as in

the preceding case, so again
Xi =n
hit)
Ait) =
3/oY/3
40
3/o0
I (7-43)
Xj ={,(-/)
The design variables are chosen as in Eqs.

Eqs. 6-122 forthe design variables are
7-41 or 7-43, depending on which makes H
largest. The problem with differential equa-
tions and boundary conditions, Eq. 7-20, is
<sh 2EAh2
now solved by the shooting technique. Nu-
merical results are given in Table 7-3(C).
+ M ? Ah = 0
(7 40)
-- A,*i3Qftf_Q' 1 " 7-3.4 DISCUSSION OF RESULTS
dA ' 4EA2h2
For both types of tower considered (simply
+ p^ h2 =0 supported and towers with top supported by
guy lines), the variables A and h could be
As before, two cases must be considered: fixed at constant values, large enough that
deflection requirements are met. For a given
Case 1. <j>< 0. This implies M = 0. configuration of the tower, there is one pair
of constant values A and h which yield a
From Eq. 7-40, it is clear that if X0 =0, then tower at least as light as any other combina-
i = 0, and Xi = X2 = 0. This contradicts the tion of constant A and h. For both types of
theorem, so X0 =0 and it is permissible to tower, finding these values is a matter of
take X0 = 1. The system is then a set of two simple algebra. Results for both towers are
equations for A and A which yields given in Table 7-3(C), referred to as towers
1/4
with no design variables. With/4 held constant
3ig ,1/4 ( _ 0l/2
and h{t) treated as a design variable, the
hit)
2E problem can be solved for several different
(7-41) values of A for both types of tower. Sum-
1/4
A,t)-\(BfL) â-ry/j maries of these solutions are given in Table

7-3(A) and Table 7-3(B). If the tower weights
are then plotted as a function of fixed values
of A, a minimum, or best, weight canbe found
(Table 7-3(C)). These optimum towers show a
Case 2. 0 = 0. This is reduction in weight over the no design vari-
able case of 13.5%and 13%for the simply
(7-42) supported and guy-line supported towers,
respectively. Results for solutions of the
Eq. 7-42 along with Eq. 7-40 is a system of problem when both^4(/) and hit) are treated
three equations for h, A , and M- Eliminating ju as design variables are also given in Table
and solving for A and/4 yields 7-3(C). These represent reductions in weight,
7-15
AMCP 706-192
over the case where only the spacing h(t) is maximum deflection at a given point could be
allowed to vary, of 6.1% and 6.7% for the among these.
simply supported and guy-line supported
towers, respectively. Similarly, these towers Note that Figs. 7-5 and 7-6 are not scale
represent respective reductions in weight over drawings of the towers, but are representative
the completely uniform (no design variables) of the general shape of the respective towers,
tower of 18.8% and 19%. as viewed on one face. Sample profiles of the
four possible structures are presented in Table
7-3(A), (B), and (C).
All four configurations of the structure
described in the preceding subparagraphs have 7-4 MINIMUM WEIGHT DESIGN OF
been successfully prescribed by a digital com- BEAMS WITH INEQUALITY CON-
puter approach. An IBM 360-65 Computer STRAINTS ON STRESS AND DEFLEC-
was used and the programs employed Runge- TION
Kutta integration with the Newton's Method
described in the text. Convergence depended The problems treated thus far in this
on getting a good starting value for the chapter have only design variable inequality
multiplier . This was made more difficult by constraints. In engineering design one often
the fact that has no physical significance so encounters problems in which it is required
that, consequently, engineering intuition was that the state of the system satisfies inequal-
no help. To find a sufficiently close starting ity constraints. As seen in Chapter 6, state
value for , (i.e., a value for which the error is variable constraints are more tedious to treat
of order two or less) several different values and have features not encountered in prob-
of were investigated, each increasing from
lems without state constraints. A class of
the previous value by a factor of 10, and the
beam design problems including state variable
first value very close to zero. Once a starting
constraints is presented in this paragraph to
value that would allow Newton's Method
illustrate some of the features and difficulties
iterate effectively was found, convergence
that can arise in this difficult class of prob-
occurred in ten or less iterations, taking less
lems. While the problems solved are of limited
than two minutes of computer time. This
practical value, they do illustrate typical
time could be reduced with increased sophisti-
features that can arise in state variable con-
cation of the computer program.
strained problems.
Other optimal design problems can be 7-4.1 STATEMENT OF THE PROBLEM

approached with the method of this para-
graph. Also, different parameters could be Beams which are loaded in a general way
treated as variables. For instance, several (such as in Fig. 7-8) are considered in this
materials of different densities and stiffness paragraph. The cross sections of the beams are
characteristics could be used in the same assumed to depend on a vector parameter u(t)
structure, and the choice could be left as a = [ux{t), u2{t), ..., um(t)} and to be sym-
design parameter. The loading could be a metric with respect to vertical and horizontal
function of the height above the ground axes. The vertical axis of symmetry is as-
rather than be constant. Other restrictions sumed to lie in the plane of loading. The
might also be imposed; maximum width or beams are made of a homogeneous, isotropic,
7-16
AMCP 706-192
where d is the distance above the neutral axis

of the cross section.
M~ rmTKl\4 fl^rrjmh) V
M
T
M, M R
For each t, let dj [t,u(t),M(t),V(t)) , i = 1
r 7 T
and 2, be the distances from the neutral axis
where
Figure 7-8. Beam Loaded in a General Way ap(0 = 2
linearly elastic material with Young's modulus

E. Small deflection, elementary beam theory
is used throughout this paragraph. Also, the 11/2
effect of the weight of the beam on deflection
is neglected. \l2(t)b2lt,dl(t)])(
Since, for a particular beam, the cross and

section is determined by u(t),
t. l | M\t)d\(.t)
A(t) =A[u(t)] (cross-sectional area), l2U)
1/2
(7-44) V2(t)Q2U,d2(t)f
I2{t)b2[t,di(t)],
7(0 =I{u(t)\ (moment of inertia),
respectively, are the maximum principal
(7-45) stresses that occur in the cross section at t.
The values dt and d2 may be determined by
b(t,d) = b[u(t),d] (width of cross section the methods of ordinary calculus, for each t.
atd),
The problem is to determine u(t) so that
(7-46) the beam, subjected to a given ioading,
contains as little material as possible and still
Q{t,d) = Q[u(t),d] (first moment of area satisfiesthe following conditions:
above d, about the
neutral axis), 1. Principal normal stress is less than or
equal to some allowable normal stress
(7-47)
2. Principal shear stress is less than or

and equal to some allowable shear stress
C(0 = C[u(t)] (half-depth of beam),

3. Stiffness is bounded away from zero
(7-48) (otherwise an infinitesimal change in
7-17
AMCP 706-192
load can cause the deflection to be where q(t) is distributed load;

discontinuous).
S,[x<'>(0), xa\T}] = 0, s = 1, .... B, i < 3;
4. Beam deflection at each point is
bounded by two given functions Xx (t) (7-51)
and X2(t), i.e., X^t) < x(t) <s X2(t)
withX,(f) < X2(t). apy(t)' < a max (7-52)
In structural design problems, it is fre- for all f in (0,T), where M(t) is bending
quently sufficient to require only that the moment;
maximum flexural stress be less than <Jmax
T
and the maximum direct shear stress be less pW-< Tmax (7-53)
than 7-max. These conditions are considerably
easier to enforce than the conditions on for all t in (0,70 where V(t) is shear;
maximum principal stress.
/[(/)] >/o. (7-54)
If the beam is subjected to several loadings,
then the problem is more tedious but is no where I0 is a constant greater than zero; and
more difficult mathematically. Corresponding for all/in (0,71,
to each loading there is a a deflection curve,
bending stress, and shear stress that must Xi(t) < x(t) < X2(t) (7-55)
satisfy the stated conditions.
It is assumed that the functions appearing
Further, since the beams are made of above have the following properties:
homogeneous material, the weight of a beam
will be minimum if and only if its volume is 1. q has a piecewise continuous derivative
minimum. Therefore, in the following the in (0,70.
quantity to be minimized will be volume.
2. A, I, C, and Q are piecewise twice
The given problem is now stated mathe- continuously differentiable.
matically: A vector function w(f) is sought
which causes the functional 3. X, and X2 have continuous second
derivatives in (0,70.
J A [u{t)\ dt (7-49) A solution is sought with the following

properties:
to be a minimum subject to the following

1. uXt), j = 1, ...,m, are piecewise continu-
conditions:
ous in (0,71.
d2 d2x\ 2. x(t) is piecewise four times continuous-

q(t) (7-50)
ly differentiable in (0,70-
at all but a finite number of points in (0, T), In case only maximum bending stress and
7-18
AMCP 706-192
maximum direct shear stress are to be In the statically determinate case, the
bounded by amax and rmax , respectively, differential equation, Eq. 7-50, and the
constraints, Eqs. 7-52 and 7-53, are replaced boundary conditions, Eq. 7-5 1, reduce to
by
d2x Mjt)
(7-58)
\M{t)\ C[u(t)] dt2 EI{u)
|o(r)| = (7-56)
I[u(t)]
and
and
g ,[x(0),x'(0),jc(r),*'(O] =0,
W(t)\Q[u(t),d3(t)]
\T(t)\ = (7-57)
I[u(.t)]b[u(t),d3(t)] 5= 1,2. (7-59)
for all t in (0,T) where d3(t) = d3 [t,u{t), The boundary-value problem, Eqs. 7-58
M{t), V(t)\ is the distance from the neutral and 7-59, is equivalent to a boundary-value
axis where the absolute value of the direct problem with a system of first-order equa-
shear stress is largest. The distance d3{t) may tions. The new problem may be written as
be determined by the methods of ordinary
calculus. dxx
(7-60)
~
In the case of beam design with multiple
loading requirements, there is still just one dx2 Mjt)
(7-61)
design variable u{t). However, c rresponding ~~dT EI{u)
to each loading there is an additional state
variable (deflection) that must satisfy condi- and
tions identical in form to Eqs. 7-50 through
7-55. The problem is to determine u(t) so that gs[x1(.0)>x2(0),xl(D,x2(T)] =0,
the functional, Eq. 7-49, is minimum subject
to the condition that Eqs. 7-50 through 7-55 s = 1, 2 (7-62)
are satisfied for each loading.
w here x t is defined to be x. In terms of this
notation, Eq. 7-57 becomes
7-4.2 NECESSARY CONDITIONS FOR THE
BEAM DESIGN PROBLEM jr,(r)< JC,(O< Jr2(0.
The treatment which follows applies only It is now clear that the beam design
to statically determinate beams, i.e., beams problem is contained in the class of problems
loaded in such a way that reactions at all to which par. 6-4 applies. The quantities
supports (and hence also shear and bending appearing in par. 6-4 will now be identified
moment) are determined completely by the with the physical quantities associated with
conditions for equilibrium of the beam. beam so that necessary conditions for the
Changes in formulation of the problem which beam design problem may be stated.
are necessary to consider statically indeter-
minate problems will be indicated below. Conditions, Eqs. 7-52, 7-53, and 7-54,
7-19
AMCP 706-192
correspond to Eqs. 6-101; Eq. 7-55 corre-

sponds to two restrictions of the type ex-
pressed by Eq. 6-132. The differential equa-
tions, Eqs. 7-60 and 7-61, correspond to Eqs.
6-97, where
/l =*2 (7-63)
(7-66)
M(t) (7-64)
/=
EI(u) *2 (0,04=0
EI(u)
and in Eq. 6-96 and
fo=A(u). (7-65) 05,05 < 0
The ends of the beam are located at the (7-67)

known points / = 0 and t = T. Therefore, t + X"1(t),4>5 =0
EI{u)
and t' in the general variational problem are
known. Also, boundary conditions will Since the only explicit dependence of Eqs.
generally be separated; i.e., some conditions 7-52, 7-53, 7-54, 7-63, 7-64, and 7-65 on t is
will be given at 0 and others at T. through M{t) and V(t), the points of discon-
tinuity of functions (f in par. 6-4) are
The state variable constraints, Eq. 7-55, are denoted wa , including points of discontinuity
of second-order since of M{t) or V(t); i.e., the o)a correspond to
points of application of concentrated loads or
moments. Therefore, the w are known.
j* v*2 xO Xj x\ - X2 x2
At each point f~, the deflection curve is

is not an explicit function of u, but tangent to one of the curves
xi =X1{t) (7-68)
mr(*a -x1)=X"2(t)-x'i(t)
or
X'i(t) +
EI(u) X2{t) (7-69)
is an explicit function of u. The same argu- and there is a neighborhood of t~ in which

ment holds for xx Xl > 0. this is the only point of tangency.
In terms of the notation of Eq. 6-101, At each t~, the deflection curve becomes
tangent to one of the curves, Eqs. 7-68 or
*'=Vf)-%x* 7-69. Further, there is a neighborhood of t~
in which Eq. 7-55 is a strict inequality to the
* P max left of t~, and Eq. 7-68 or Eq. 7-69 holds to
7-20
AMCP 706-192
the right of t~. At the corresponding point t*, dxt

= x2 (7-71)
the deflection curve leaves Eq. 7-68 or Eq.
7-69. Eq. 7-68 or Eq. 7-69 then holds in {t~,
/*), and Eq. 7-55 is again a strict inequality and
immediately to the right oft:.
dx2 M{t)
(7-72
If x,(0) or *i(r) is not fixed by the It Eliu)
boundary conditions of Eq. 7-62, then one
part of Eq. 7-55 may be an equality at 0 or 7. at all but a finite number of points ir1(0,7)
In this case 0'5' of Eqs. 7-66 and 7-67 need not
be zero. X0 > 0
The t* are points where one or more of *i=*i

Eqs. 7-52, 7-53, and 7-54 changes from strict
inequality to equality. Condition, Eq. 6-105, x2 =h -Hit
is assumed to hold at these points.
and
According to Eqs. 6-106, 6-108

Xo + X? + X^ > 0
M{t) for all t in (0,7), where j and 2 may have

H = - X0,4 () + Xt x2 X2
Eliu) different values in subintervals of (0,7) which
are bounded by the tt and t~;
Hi 01 11242 11343
Mit)
M4 04,2 ~Ms 05,2 A.Q A2
du. du. Eliu)
G= 2 7sgs + ZyB+a(t-o>a)
s - 1 a
- [Ml 01 +M2 02 +M3 03 (7-73)
au.
M7-70)
G = T0|4>6 04 (t ) + T7]4)6 04 Op
+ M4 04.2 +M5 05,2l =0,/= 1, ..., m
+ 7",5,5 05 (^)+ '"1,5,6 05 OJ) M,-0, = O,i= 1,2,3, (7-74)
+ '"0,4,f 04(^) + '"l,4,/- 04 0r") and
+ T
+ T.S.r 05 0,") l,5,r 05 (0 , M4 04.2 = MS 05,2 =0 (7-75)
Theorem 6-9 and Eq. 7-70 yield Theorem at all but a finite number of points in (0,7);
7-1.
X.(a+0)-Va -0) = 0,
Theorem 7-1: Necessary conditions for the
minimum weight beam problem are: i= 1,2 (7-76)
7-21
AMCP 708-192
M^ + 0)_A,0;_0)-r4r the boundary conditions gs = 0, s = 1, 2, must

be satisfied along with the conditions
+ rsr=0 (7-77)
fc.
(7-85)
\2(t~r + o)-\2o;-o)-T~l4l. V) = I?1^v- i= 1,2
+ r~lSr =0 (7-78) and
Xi(^+0)-X,(f- -0)-r4S 3*,

\.{T)= i 7 ,i= 1 2- (7-86)
+ To 56=0 (7-79)
and the Weierstrass condition
X2(f-+0)-X2(r--0)-r746
H(xe U, X., t) < #(*,, u, x,, /;

+ TTs6=0 (7-80)
must be satisfied for each t in (0,T), where /

H(&a+0)-H(ua-0)+yB + a=0
is any function which along with*! and x2
satisfies Eqs. 7-52, 7-53, 7-54, 7-55, 7-60,
H(t~r + 0) _//(?r _ 0) - T~04 X'2 (g
7-61, and 7-62 with u replaced by U. The
statement of Theorem 7-1 is now complete.
-^X'Ut^ + T^rXW;)
If there is only a scalar control variable

+ SrX'l(t'r)=0 (7-81)
u(t), then the condition of Eq. 6-105 will be
violated at points t* which are intersections
^(/- +o) _H{t- -O)-746 âp
of intervals in which Eq. 7-52, 7-53, or 7-54 is
-TT46 ^'Op+rss X'i(/-) an equality. With an additional hypothesis,
however, the conclusions of Theorem 7-1 are
+ T75 6 î'(C) = 0 (7-82) still valid.
//a;+o)-//(r*-o)=o At a point t* = coa, it is assumed that 1^1 =

i>2 = 0, where 4/t > 0 and i//2 > 0 are any
and two of the constraints of Eqs. 7-52, 7-53, and
7-54. If ^ is defined as 4/ = min (\pi ,02), then
H(t+B +0)-H{t* _0)=0 i > 0 replaces the conditions 4*i > 0 and 4>2
> 0. It is assumed that d^t/du and 9i//2/3
for all <x, r, 6, and ??; at each of the points S = are not zero at t*. The new constraint now
tr and t~ , either satisfies the conditions of Eq. 6-105.
0, (S) = <j>\ (S) = 0 (7-83) If Theorem 6-8 is applied to the new

formulation of the problem, the result is
or identical to Theorem 7-1 with the exception
of Eqs. 7-73, 7-74, and 7-76. However, the
02 (5) = 02 (S) = 0; (7-84) new conditions on u are identical to those
7-22
AMCP 706-192
implied by Eqs. 7-73, 7-74, and 7-76. The finite number of unknown constants, or
roles played by Mi and 2 in Theorem 7-1
would simply be combined in a new variable 2. The beam is supported in such a way
M. This result may be stated as Corollary 7-1. that an infinite number of constants are
required to specify the reactions (e.g., abeam
Corollary 7-1: Let there be a scalar design on an elastic foundation).
variable u(t) and assume that any two of the
inequality constraints, Eqs. 7-52, 7-53, and In the first case, the unknown constants
7-54 are equalities at t*. If the first partial appear in the expressions for M and V. By
derivatives of these two constraint functions defining new state variables, x( with i > 3, to
with respect to u are not zero at (*, then be these parameters, the following differential
Theorem 7-1 holds. equations must be satisfied:
One further result may be easily obtained. dxi

If b\j/1/du and d\p2/du are nonzero at t* and 0,i> 3.
are of the same sign, then u is continuous at
In this way, statically indeterminate prob-
lems of the first type are reduced to varia-
To prove this, it is supposed first that tional problems to which Theorem 6-8 ap-
d\pjdu > 0 and u(t* + 0) = u(t* -0)-E,E plies.
> 0. Taylor's theorem (Ref. 16, p. 56) implies
For statically indeterminate problems of
\p1[t*,u(t*+0)] =\p! [t*,u(t*-0)] the second type, however, more basic changes
in formulation must be made. The fourth-
d$1[t*,u(t*-0)-0e] order differential equation, Eq. 7-50 must be
du treated since q(t) may be q[t,xl{t),x2(t)}.
The fourth-order equation is equivalent to the
where 0 < 8 < 1. But, dîu > 0 and e > 0, first-order system
so
0 = $i [t*,u+Q)] < ^, [t*,u(t* -0)] =0 = x2
which contradicts the assumption E > 0. An dx2 x3

identical argument holds in the remaining =
dt EI(Uj)
cases, so"[t*,u(t* -0)] =0.
dx3
7-4.3 STATICALLY INDETERMINATE *4
dt
PROBLEMS
and
A statically indeterminate beam may be
classified as one of two types: dx*
-q{t,xx,x2)
dt
1. The beam is supported in such a way
that all reactions are determined to within a where xx = x, x3 = M, and x4 = V. Theorem
7-23
AMCP 706-192
6-8 now applies to statically indeterminate T1 .+ 3 rcan be eliminated and the tr deter-
problems of the second type. mined as functions of the parameters t and
2. Note that 1 and 2 to the left of t~ need
7-4.4 SOLUTION OF THE EQUATIONS OF not equal 1 and 2 to the right. In exactly
THEOREM 7-1 the same way, the t~ are determined by Eqs.
7-79, 7-80, and 7-82. Continuity of H at t*
The Lagrange multipliers, X0, Xu X2, are and t* determines the ft and t* as functions
not uniquely determined by Theorem 7-1. of 1 and2-
However, if the solution, x(t) and u{t), is
normal (Ref. 17, p. 214) for the problem, it is The equations previously enumerated
permissible to put X0 = 1. Theorem 7-1 then determine u-, t* tr, f^, and t* as functions of
determines the remaining \i uniquely. Abnor- t, 1; and 2. The problem would be solved
mal solutions are peculiar in that there may by direct integration of Eqs. 7-71 and 7-72
be no other functions, x(t) and u(t), near and application of g{ = g2 = 0 if |t and 2
them which satisfy the conditions, Eqs. 7-50 were known.
through 7-55. The procedure adopted in this
subparagraph is to assume there is a normal The conditions which determine ^ and 2
solution and then attempt to solve for it. If are Eqs. 7-83, 7-84, 7-85, and 7-86. Eqs. 7-85
this fails, there is either no solution or the and 7-86 after elimination of7i and y2> yield
solution is abnormal, in which case a special two equations in j and 2. If there are w of
analysis is required. In the following, X0 will the points tr and t~ , they subdivide (0,T) into
be taken as 1. w + 1 subintervals. Since these are the only
possible points of discontinuity of Xj and X2 ,
In order to determine u{t), consider any there are just w + 1 pairs, 1 and 2, which
interval in which z 0 Q Z < 4, of the represent 2w + 2 unknown. Thus, there are
inequalities, Eqs. 7-52 through 7-55, are 2w + 2 equations from which the 2w + 2
equalities and the remaining 4 z are strict values of ] and 2 may be determined. If this
inequalities. Eqs. 7-74 and 7-75 show that the is not the case, the problem has no solution or
4 z multipliers corresponding to the 4 z the solution is abnormal.
strict inequalities are zero. Then, Eq. 7-73 is a
system of m equations in the m functions It is assumed now that u{f) and points t'r
uAt) and the z nonzero multipliers. Further, and t~ are known functions of t and 2. A
the z equalities of Eqs. 7-52 through 7-55 numerical method is developed which can be
yield z equations in the ut). Thus, there are used to solve the equations given above for 1
m + z equations which are to determine the m and 2. A numerical solution is required since,
+ z unknowns. The nonzero Xt) are first even for very simple problems, the function
eliminated and the u,{t) are then found as fi(t&\ ,2) is far too complicated to integrate
functions of t and the parameters t and 2. in closed form.
At points tr one part of Eq. 7-55 is an Expressions, Eqs. 7-85 and 7-86 generally
equality, say the jth part (j ~ 1 or 2). In this yield two easy relations between j and 2.
case, only T0i/ + 3|,. and Tij; + 3,r can possibly Eqs. 7-83 and 7-84, however, require succes-
be nonzero. Eqs. 7-77, 7-78, and 7-81 are sive integration of /20,i ,2). To complicate
three equations from which i"0/ + 3r and matters, some limits of integration are the
7-24
AMCP 706-192
points t~ and t~, which are themselves func- First, the equations of Theorem 7-1 are
tions of 1 and 2. Therefore, Eqs. 7-83 and written out in detail and simplified for the
7-84 form a system of nonlinear, finite case of beam with rectangular cross section of
(nondifferential) equations ini and 2- variable depth. In pars. 7-4.5.1, 7-4.5.2, and
7-4.5.3, three specific examples are con-
A Generalized Newton Method (Ref. 18, p. sidered. These examples range from an easy
220) is used to solve this set of equations. A problem in par. 7-4.5.1 to a rather complex
generalization of Leibniz' Rule is used to one in par. 7-4.5.3.
compute the required derivatives of integrals
with variable limits of integration. This rule is The cross section considered here is shown
(Ref. 16, p. 80). in Fig. 7-9. For this particular cross section, if
d_ (sAr.
h(t,r)dt
dT
)gl(T.
g
^T) bf2(t,T)
I
dt
dr
(7-87)
dg2 (T)
+ f2 [g2(T).T]
dr
Figure 7-9. Rectangular Cross Section

dgt (r)
~h [gl(T),T]
dr
r
max * 1/2amax> ^ 7"52 and 7-53 are
where 9/2(f,T)/dr is piecewise continuous, and
satisfied if and only if Eqs. 7-56 and 7-57 are
g, and g2 are differentiable. It is assumed,
satisfied. This result may be proved by ex-
also, that f2 and 9/2 /dr are continuous at t =
pressing o and T as functions of d and
gi(r) and g2(r).
applying methods of ordinary calculus. The
restriction rmax > 1/(2 omilx) is necessary,
since at the extreme fiber of the beam, the
7-4.5 BEAMSWITH RECTANGULAR
principal shear stress is half the principal
CROSS SECTION OF VARIABLE
normal stress. This relation between rmax and
DEPTH a
max 's no restriction for design of metallic
beams. Yield stresses for steel and other
Three examples are considered in this
common metals satisfy this condition.
subparagraph. In each example, the beam
cross section is rectangular with fixed width In this case, there is only one design
and variable depth. Also, the constraint, Eq. variable h(t); Eqs. 7-44 through 7-48 are
7-55, is taken as
- A < x(t) < A

Ht)=h3(t) (7-89)
(7-88)
where A > 0 is a constant. A{t)=bh(t) (7-90)
7-25
AMCP 706-192
bit) (7-91) In any interval where Eq. 7-54 is an

equality,
Qit)=~h2(t) (7-92)
bh3 (t)
O
12 -Io
and
so
C(t)=-h(t). (7-93) 1/3

12/p
hit)- (7-95)
b
M(t) and V{t) are assumed to be known,
piecewise twice continuously differentiable In any interval, its, f J), where Eq. 7-88 is
functions of t whose discontinuities occur at an equality, direct differentiation and use of
points t = wa . Eqs. 7-71 and 7-72 yield
Eqs. 7-89 through 7-93 and Eq. 7-70 along

withXo = 1, yield 0 = LiMî;
(7-96)
Eh2 (t)'
12M(Q
H= -bh + \xx2 - A2 A"3.
Eb
(7-94)
In order for Eq. 7-96 to be satisfied, it is
necessary that Mit) = 0 (hence also V{t) = 0 =
The procedure outlined is now used to gO)) must be identically satisfied in (t~, /*).
determine A(f)- In any interval where Eq. 7-56 ff this is the case, hit) is given by Eq. 7-95.
is an equality,
61AfC0l In any interval where Eqs. 7-56, 7-57, 7-54,

bh2it) and 7-88 are all strict inequalities, Eqs. 7-84
and 7-75 show that M,(f) = 0, i = 1,..., 5. Eq.
so 7-73 then is
1/2
6\M(t)\ 36M(r)
hit)- 6+ X2 (0 h-*it) = 0 (7-97)
bo 6
In any interval where Eq. 7-57 is an so

equality,
1/4
36X20)M(f)
A(0 =
3Wit)\ Z>2
Ibhit)
It is worthwhile to note that in order for
so Eq. 7-97 to hold the product \2it)M(t) must
be positive. That is, X2(0 andA/(/) must have
3|K(/)| the same algebraic sign throughout any inter-
*(') =
26T val in which Eq. 7-97 holds.
7-26
AMCP 706-192
In a more compact notation, 2b a3max x\1"

c2 = ( J
\ 32 /
3|K(f)l
if|T|=T
2b T
'4ft*V'4
1/2 -=(\
>9E)
6|M(Q|
if I a I = a
and the function sgn ( ) is defined by the
h(t)=4 (7-98) relation
1/3
12/p
if/ = /,
q sgn(<?) = q
1/4
36X2 (f)Jtf(?) for real g
if IT I < T
Eb2
|a| < a max' Equations which determine the special
V
and I > I0. points t;, t*, t~, and t* may now be found. In
the problem at hand, Xx (t)a.n& X2 (t)are
The Weierstrass condition shows that the constant, so their derivatives are zero. Eq.
largest of the expressions in Eq. 7-98 is the 7-81 is then
proper value of h{t).
H(t;-0)=H(t; + 0) (7-100)
Eq. 7-98 in Eq. 7-64 yields
Experience has shown that on both sides of
t-, Eqs. 7-56, 7-57, and 7-54 are strict
-CtMiowior3, if r =T
l l max inequalities. Assuming this is this is the case,
Eqs. 7-94 and 7-98 together with Eq. 7-83 or
-C2|M(r)r]/2 7-84 may be used to simplify Eq. 7-100. The
if a = a
' l l max result is
xsgn [M(t)],
[x2;-owa;-o)]1/4
/(f) = if/= 4,
1 1 (7-99) = [x2(/; + 0W(/; + 0)]"4 . (7-101)
3/4
-c3ix2wr ' if T
l l< rmax
Eq. 7-98 and / > 0 imply M(t;) # 0, so if
x \M(t)\ lh a M(t) is continuous at t;, then Eq. 7-101
' l l< max>
reduces to
k x sgn [M{t)}, and I > I0,
X2(f;_0)=X2(f;+0) (7-102)
Points of discontinuity of M{t) must be

where
checked in Eq. 7-101 as possible t-.
32b2T3m ax
Eq. 7-96 shows that points t* and t~ can
Ci =-
9E occur only in intervals where M(t) (hence also
7-27
AMCP 706-192
V{t) and q(t)) is identically zero. Since this

situation is not common in practical prob- -b V{Q)
2br
lems, such intervals will not be discussed here.
C,X2()M()| V(Q)
According to Theorem 7-1, the points t* 1/4
are determined by the condition 1/4
()
[X2()M()]
mrv -o) = HI + o) (7-103) 1/4

-C3[X2(M)1
Using the definition of C3 in this equation

By direct computation it is seen that the
and manipulating the result yields
partial derivatives with respect to h of the left
sides of
1/4
3|K(g)r ^64ftT,.x
1
max X2()M() 9E
\a | -a max < 0
1
3/4
V{Q) I4 64b2 T*
and 0.
M)M() 9E
/ - KO (7-105)
By putting
are all negative at points where M(t) = 0 = l'/4
V(t). The result stated just below Corollary K()
7-1 then shows that points of intersection of LM)M)J
intervals in which one or more of the above
and
inequalities is an equality may be determined 1/4
by the condition C = (64b2 max/
-9E-
W -o; = w; + o) (7-10-/;
Eq. 7-105 becomes
If Eq. 7-104 is used to determine t*, then 3P4 -4CP3+(A = 0.

points wa must be checked in Eq. 7-103 as
possible t*. The roots of this equation in P are C, C, C( 1 +
v/2/), and C(l - >/20, where /2 = - 1. The
Let Q = f * = wa be defined to be a point fourth powers of the last two roots are not
real, so the only real solution of Eq. 7-103 is
of intersection of two intervals such that |r| =
T
max on one sê of t = Q and Eqs. 7-56,
2
7-57, 7-54, and 7-88 are strict inequalitieson _ /( 64b rmax \\
I V(Q) X2()JM().
the other. Point Q is to be determined by Eq. A 9E I
7-103. Due to continuity at t = Q, Eq. 7-103
may be written as (7-106)
7-28
AMCP 706192
Similarly, S = /* = co is defined to be a where A < B < C, and

point of intersection of two intervals such
that lcr| = amax on one side of S and Eqs.
7-56, 7-57, 7-54, and 7-88 are strict inequal- da fcn,a)d'f)dv
ities on the other. Just as above, Eq. 7-103 U*.<) J*,0)
reduces to
drjdv
3P-4CP +C4 =0
where
-1 1/4
~M{S)~
p= dg i(a)
U2(S)J x /tei(a), a]
Ja
and
1/4
+ rp.oo d*2 (a)
/(Tj,a)dT)
da
(7-109)
Therefore, the only real solution of Eq.
Leibniz' Rule, Eq. 7-87, is used repeatedly to
7-103 at t = Sis
obtain Eq. 7-109.
JH(S)=l-=^-IX2(S). (7-107)
7-4.5.1 A PROBLEM WHICH CAN BE
SOLVED ANALYTICAL LY
In deriving Eqs. 7-106 and 7-107, it was
assumed that Q and S were not equal to any As a first example, the cantilever beam of
<ja. The coa are, therefore, possible choices
Fig. 7-10 is considered. This problem is simple
for Q and S and must be checked in Eq.
enough that a solution can be obtained
7-103.
analytically.
In particular problems, the following two
Boundary conditions for this beam are
identities are used:
*i(0)=*2(0) = 0. (7-110)
f(v)du = {C-B) \ f{v)dv
+ I 1 f{y\)dr\dv
)
n:
+ I I f(n)dt\dv, (7-108)
Figure 7- 70. Simple Cantilever Beam
7-29
AMCP 706-192
The bending moment and shear are and
M(t) = M > 0 (M constant) dx2
and
and application of Eq. 7-1 lOyields
V(t) = 0.
For simplicity, let /0 > 0 be small enough

so that
1/2
*2(0=|
l:
o
Mn)dv (7-113)
(l2/0\1/3 < (jM ) and
If this is the case then I(t) > I0 for all t.

Further, since r(t) = 0, Eq. 7-98 may be
simplified to
*i(0 =
l:|: f2(r))dr)dv (7-114)
The inequality /2(f) < 0 and Eq. 7-113

imply x2 < 0 in (0,7), so there can be no
6M ' points t~. The only possible point at which
h(t) =<
([" bo max _
if|a| = <7
Eq. 7-88 is an equality is t = T. Since f2{t) <
0, Eq. 7-114 implies x1(t) < 0. Therefore, if
1/4
36X2(.t)lM Eq. 7-88 is an equality, it must be Xi(T) =
if la I < a_ A. Further, since there are no t", t\, t~, or
Eb2
wa, \i and \2 are continuous and there is
(7-111) only one pair of constants, 1 and %2, to be
determined.
Also, Eq. 7-99 becomes It is assumed first that xx(T) > A. In this
case, Eq. 7-86 implies
2ftc \1/2
^H , if | a | = a Xi(D = f1=0
3E2M ) ' ' max
/2(0=< and
4i2M\1/4 - 3/4
|Xj(OI
9E *i(T) = t2 -7-^=0
which in turn implies (^ = 2 = 0. Since A2 =

0 throughout (0,7), the second part of Eq.
(7-112) 7-111 cannot occur. Therefore, the beam of
minimum weight is uniform. Eqs. 7-112 and
Integration of the differential equations 7-114 then yield
dXf
= x-. Ci(r)=
IF -(-5^rJ T
7-30
AMCP 706-192
Therefore, Eq. 7-111 with |o| = crmax is the >*A/V/4

solution of the problem provided the deflec- xdT) = (T-t*)A\^-j
tion requirement A is such that
bo3max TA
(7-115) Xf1-^[(7'-r*)1/4 -r1/4]
2
6E M
1/4
A (4b2M\
If the deflection requirement A does not 9E j
satisfy Eq. 7-115, then it is necessary that
xt(T) = A. Otherwise, this argument would x?r3/4[-|(r-f*)5/4 +T1/4-t*
hold, and the deflection at T would violate
Eq. 7-88. Therefore, the additional boundary 2bo3max \\1/2 (j _ t*\2
condition, 3E2M
g3=xl(D + A = 0, The right-hand side of this equation may be

simplified by eliminating either 1 or t*
must be satisfied. The two constants i and through use of Eq. 7-116. Since ^ does not
2 must now be found.
have as much physical significance as t*, it is
eliminated. The conditions x^(T) = A
The only useful relation given by Eq. 7-86
becomes
is
1/2
A = l'2fcgmax
2
3E M
This implies 2 = 1 T, so that
3/4
(T-t*)2 --f-TW (T-t*)
10
X2(D = ,(r-0,
and only 1 remains to be found. (7-117)
On physical grounds, it is expected that the The derivative of the right side of Eq.
beam should be stiffest near / = 0 in order to 7-117 with respect to t* is zero at t* = T and
reduce the deflection at T efficiently. Also, positive everywhere else. This means that Eq.
7-117 has at most one solution. Eq. 7-116
since X2 is largest at t = 0, the second part of
Eq. 7-111 would tend to stiffen the beam then determines 1 and the problem is solved.
there. It is assumed, therefore, that there is
just one point t* having I a I < amax on its As a numerical example, the beam of Fig.
7-10, having the following properties, is con-
left and I a on its right. Eq. 7-107 for
sidered :
r* is
2 T= lOin.
M I(a max \|t (T-t*). (7-116)
b= 1 in.
Eq. 7-114 with t = T may be integrated
ff
using Eqs. 7-108and 7-112 and becomes max = 30,000 lb/in.2
7-31
AMCP 706-192
E = 107 lb/in.2 q(t) > 0 for all t in (0,7). The load q(t) of this
form implies V(t) is non-negative at zero and
and decreases monotonically in (0,T). M(t) is zero
at both t = 0 and t = T. Further, on either side
M = 450 in.-lb of the point where M(t) has its (non-negative)
maximum it is monotone.
If A > 1 in., then the beam of minimum
weight is uniform with h = 0.30 in.
For a more meaningful problem, A = 0.5
in. is considered. Eqs. 7-117 and 7-116 yield
fTTTTTTr^
rFT
t* = 7.7 in. and |i = 2.17. The precise shape
of the optimal beam is given by Eq. 7-111. By
putting Eqs. 7-111 and 7-90 in Eq. 7-49 and Figure 7-12. Simply Supported Beam With
Positive Distributed Load
performing the indicated integration, the vol-
ume of the optimal beam is found to be 3.59
in? A plot of the profile of the optimal
beam may be made by direct substitution into
q(t)dt # 0, thenA/(0 and V(t) cannot
Eq. 7-111. This profile is shown in Fig. 7-11. Jo the same point. There is, therefore,
be zeroat
no danger that the optimal beam will have
h(t) = 0 at any point. For this reason, the
constraint, Eq. 7-54, is not imposed here.
0.431 in. |.30 in.
Li Since M(0) = M{T) = 0 and V(0) # 0 *

7.7 in.
V(t), \T\ = rmax is expected near the ends of
Figure 7-11. Cantilever Beam of Minimum Weight the beam. Toward the center of the beam,
M(t) becomes large, so |CT| = omax is expected
there. If the beam requires stiffening, the
additional material can best be used near the
By elementary computation, it is seen that point of maximum deflection. This argument
the uniform beam which has *i (10) = 0.5 indicates that (0,T) should be broken up into
in. and satisfies Eqs. 7-56 and 7-57 is 0.378 subintervals as shown in Fig. 7-13.
in. deep; so its volume is 3.78 in? The
designed beam, therefore, has 5.3% less vol- In terms of previous notation, tu t2, t$,
ume than a uniform beam which will satisfy and f5 correspond to the notation t*\ and f3
the same stress and deflection requirements. to /;. For certain ranges of A some of the
subintervals shown in Fig. 7-13 will not
7-4.5.2 SIMPLY SUPPORTED BEAM appear.
WITH POSITIVE DISTRIBUTED
LOAD Provided tt and t$ separate intervals in
which \T\ = Tmax and \o\ = amax, they are
The beam considered here is simply sup- determined by Eq. 7-104. The points t2 and
ported (see Fig. 7-12) with piecewise con- /4, when they exist, are determined by Eq.
tinuously differentiable distributed load q(t), 7-107. Finally, f3 is determined by Eq. 7-102.
7-32
AMCP 706-192
II 11 V V ii ti
1 __ .1- 1 -
o
-^^
~
t- c
I
1
re
1 1
Figure 7-13. Subdivision of the Beam With Distributed Load
The boundary conditions for this problem Eq. 7-102 for f3 yields
arc
t3 =
n2
xt(0) = 0 TU + fj
and Assuming t2 and f4 exist, the equations that

determine them are
xl(T)=0.
m ax ,
Eqs. 7-75 and 7-76, therefore, yield M{ti)=\z)Slt% (7-120)
X2(0) = \2OT = 0.
and
The constants i and 2 in A2 may have

different values on opposite sides of t3. To
the left of t3,
M(t4)- Ml f) (7-121)
In this case, /t and fs are determined by Eq.

X2(0)=2 -ZlxO = 0 7-105. If t2 orf4 does not exist, then tx orf5
is determined by
so
64b2T*
(7-1 18) V{t,) ~W
(7-122)
To the right of t3,
or
Xa(D = -iT=0
, 64ft2r4
F(fs)|4 =
so 9
M = b(l-y)=?2(l-^). (7-11 9) xfi(l-y)Jf(r5). (7-123)
The constants ?i and f2, which are intro- It is noted that Eqs. 7-120 through 7-123
duced in Eqs. 7-118 and 7-119, are now to be can be solved easily for f1 and f2, but, in
determined. general, not so easily for the t(. In the
7-33
AMCP 706-192
development which follows, it will be con-

venient to use Eqs. 7-120 through 7-123 to
solve for ?! andf2 as functions of the tr
R2=A + (T-t4)\
i; Mv,$2)dri
I
The conditions that are to determine the
unknown t{ are Xi(t3) = A and x2(t3) = 0. + (7"-/s)| /i(ij)di?
However, for computational reasons, a more
convenient set of equivalent conditions is
*i(0) = 0
and
(7-124)
If fi(V,$*)dvdv
*,(D = 0
where Xi (/3) = A and x2(ti) = 0 are used as

(7-125)
rr
it, it,
f2{r\)dt}dv
initial conditions for integration.
Conditions, Eqs. 7-124 and 7-125 may be

rr f2{n,)dr)dv = 0 , (7-127)
written explicitly as
where .Ri and i?2 are introduced for nota-
tional purposes.
i?,= A + r, P f (v)dr) + t P
fi I| 2 | f (v)dr)
2 2
It is assumed now that q{t), b, E, T, <7max,
Joo j, /.
t, and Tmix are given. The equations that
determine the t( are different in four distinct
ranges of the deflection requirement.' These
ranges of A are described in the following:
{
+ t3\ f2{r\,U)dn
1. A0 denotes the largest deflection that

occurs when the subinterval (t2, f4) does not
appear, i.e., when the beam is specified by
f2(rj)dr)dv only the first two parts of Eq. 7-98.
ff A > A0) then the beam specified by the

first two parts of Eq. 7-98 is the one of
f2(y\)dt\dv minimum weight.
2. For A slightly less than A0) there exist

points t2 and t^ that are determined by Eqs.
f2(r,,$i)dndS = Q (7-126) 7-126 and 7-127.
j t2 J t,
As A decreases, points t2 and f4 move
and toward zero and T, respectively. There is a
7-34
AMCP 706-192
value of A, say A = A,, for which either t2 or method of solution will be described for the
ti, first coincides with t\ or ts, respectively. Case 2.
For definiteness, assume t2 = t, when A =
A,. An iterative method called Generalized
Newton Method (Ref. 18) is used to solve for
3. For A slightly less than Aj, points /s t2 and t4. The procedure begins by estimating
and ?4 are determined by Eqs. 7-126 and values i2 and ?4; and then making a correc-
7-127. tion according to the formula
As A decreases, points tx and t4 move l)Ri dRf -1
toward 0 and T, respectively. There is a value H h a?7 QrT

Ri
of A, say A = A2, for which tA first coincides
with ts, (7-128)
4. For A < A2, points t1 and ts are bR2 bR2

determined by Eqs 7-126 and "M?.7. L
U
-
u h 3r4_
Ri
This explanation of the behavior of the ^ is where [ ]"' denotes matrix inverse, and t2
not the result of a mathematical analysis. It is and f4 are improvements on the estimate.
expected on physical grounds and has been
vahd in each case treated. Eqs. 7-102 and 7-107 determine t3 =
f3(f i ,?2), i = f i U2), and f2 = f2(r4). By use
The values of A { and A2 could be obtained of this information, the derivatives in Eq.
analytically. However, their determination 7-128 are computed by the chain rule of
would be of the same order of difficulty as differentiation. For example,
the optimization problem considered in this
paragraph.
dR i
= *i.r,
9r,
A, and A2 can be determined by a trial (7-129)
and error scheme. For example, to determine + (bRt tbRt bt3\ adf,
A,, t2 is put equal to tx and Eq. 7-107 3, dr3 3f, ) dt2
determines fi. Then, t3 is guessed and Eq.
7-102 solved for f2. Numerical integration of where
Eqs. 7-60 and 7-61 indicates the correction
that is to be made in t3. When t3 is located Rl , =t2[f2{t2 -0)-/2(/2+0)]
accurately, the resulting deflecting at 13 is A{.
A similar procedure is used to determine A2 is the partial derivative of R t with respect to
(t4 is put equal to /5). t2 with all variables in R , taken as indepen-
dent,
The solution of Eqs. 7-126 and 7-127 for
the t( differs only in certain details depending
on whether the given value of A is in the dt)
3fi
range described by Cases 2, 3, or 4. The
7-35
AMCP 706-192
40 in.
G?7?dV
b 0.25 in.
and E 107 lb/in.2
15,000 lb/in.2
= t3f2(t3-0).
and
Similar expressions for the remaining deriva-
a
tives in Eq. 7-128 are derived with the aid of max = 30,000 lb/in.2
Eq. 7-109.
For this problem, it was found that
The iterative procedure for determining t2
and f4 is: A0 = 0.774 in.
Step 1. Make an estimate, t2 and f4 , At = 0.728 in. (t2 =?,)
Step 2. Solve Eqs. 7-120 and 7-121 for?, and

and 52,
Step 3. Compute, numerically, all the A2 = 0.470 in. (74 = ts).

integrals in Eqs. 7-127, 7-128, and
the remaining derivatives corre- As has been noted, the magnitude of the
sponding to Eq. 7-129, deflection requirement A plays a major role in
the outcome of a particular problem. In order
Step 4. Compute the right side of Eq. to emphasize the effect of A on the properties
7-128, and of the optimal beam, the numerical example
given was solved for eight different values of
Step 5. With this improved estimate return A. The results are presented in Table 7-4.
to Step 1.
In Table 7-4, the first column contains the
This procedure has been programmed for a values of A considered. The following seven
digital computer. The program was arranged columns give information which, when sub-
in such a way that only A0) At , A,, M{t), stituted into Eq. 7-98, completely specifies
V(t), and the physical properties of the beam the optimal beam. The next column gives the
need to be specified. Many different loading volume of this optimal beam. The final two
situations may thus be considered without columns give the volume of the lightest beam
altering the program appreciably. of constant depth which satisfies the condi-
tions of the problem and the percent saving
As a numerical example, a beam with the realized when the optimal beam is used
following properties is considered: instead of this uniform beam. Dashes have
been inserted in the table when the quantity
t lb/in. to be tabulated does not exist.
7-36
o
9
TABLE 74
RESULTS FOR SIMPLY SUPPORTED BEAM WITH q (t) = t
Vol. of
Unif.
A, fi. h. t3. t. ts. 5i U Vol., Beam', Saving, %
in. in. in. in. in. in. in.3 in.3
>0.775 0.053 39.89 14.00 18.15 29.6
0.75 0.053 10.21 19.42 22.54 39.89 2.77 104.7 14.12 18.15 28.6
0.65 0.048 0.048 19.99 28.41 39.89 3.60 144.0 14.75 18.15 23.0
0.50 0.040 0.040 20.32 37.79 39.89 5.27 217.7 16.08 18.58 15.5
0.40 0.034 0.034 20.32 39.90 39.90 7.12 294.1 17.33 20.01 15.4
0.30 0.028 0.028 20.32 39.92 39.92 10.45 431.7 19.07 22.03 15.5
0.20 0.022 0.022 20.32 39.94 39.94 17.91 740.0 31.82 25.21 15.5
0.10 0.014 0.014 20.32 39.96 39.96 45.17 1866.0 27.50 31.78 15.6
1
Volume of lightest beam of constant depth which satisfies all the requirements of the problem.
AMCP 706-192
For each value of A, the iterative procedure and

used to solve the problem required approxi- V(t)=0.
mately 40 sec per iteration on an IBM 1410
Computer. Further, three to six iterations In this problem A > T/2 is assumed.
were sufficient to obtain convergence of the
residue to seven decimal places, so computing Since Mit) is never zero, the requirement
time was not excessive. |or| < %ax implies h ^ 0 for all t. I = 0 is
impossible, so the requirement I > / is not
It is shown (Ref. 18, p. 222) that if the enforced.
sequence of approximations constructed by
the Generalized Newton Algorithm converges, Eqs. 7-98 and 7-99 in this case are
then it must converge quadratic ally, i.e., the
1/2
error at the n + 1st step is proportional to the // 6M \ ifl l = a
square of the error at the nth step. This rapid ({j^>
I \ max / ' -
convergence was observed in the numerical
h(t) = (7-130)
calculations and explains why only three to 1/4
six iterations were required. 36X2(_t)M(t)
([ Eb2
. if Iff I < ff
7-4.5.3 A PROBLEM OF A MORE and

GENERAL TYPE 1/2
C,M
xsgn[Af(f)], if|ff|=cr
The beam considered here is loaded as
fa(0=<
shown in Fig. 7-14. 3/4 1/4
C3|X,(OI M
x sgn [M(t)], if \a\ < a
For a given deflection requirement A there

M
(f are three possibilities concerning attainment
of the maxium deflection. Either
Figure 7-14. Beam With an Inflection Point
1. \x(t)\< A for all t,
2. \x(t)\ = A forjust one t, or
Boundary conditions are, as in the simply 3. \x(t)\ = A for two distinct t.

supported case,
The third possibility occurs here because M{t)
*i (0) = x1(T) = 0. changes sign. In Case 1, |ff| = 0max through-
out the beam determines h(t). The Case 2
For the given loading, may be treated in exactly the same way as the
problem in the preceding paragraph. Case 3 is
{ M, if 0< / < A considered in detail here.
M(t) =
-M, liA < t <, T Assume there are two points at which
7-38
AMCP 70G-192
\x(t)\ = A; the beam will not probably be EM _1_

(7-134)
subdivided as shown in Fig. 7-15. r2
t7 (7-135)
I - I
f,
f
f t,
f If I-
and
EM _1
Figure 7-15. Subdivision of the Beam T 1 (7-136)
r.2
Points f2 and tn, where Case 3 occurs, are In obtaining the expressions for t3 and f6>
tr, so they are determined by Eq. 7-102. use was made of the fact that t3 < A < t6.
Further, xl(t2) = A and *i(?7) = A. Points The result, \2{t)M{t) > 0, from Eqs. 7-97
fi> ts, t6, and t% are <* and are determined by and 7-98 shows that this is true. To prove
Eq. 7-107. this, assume for definiteness t3 > A. Since
M(t) changes sign at A and X2 (t) is con-
In this problem, t2 and t-, may be points of tinuous there, X2W) = 0. But, since X2(f) is
discontinuity of X4 and \2. The conditions, continuous near A, it is arbitrarily near zero
Eqs. 7-85 and 7-86 on K2 are X2(0) = X2(D = in a neighborhood of A. Eq. 7-130 then shows
0. Therefore, X2 may be written as that hit) is arbitrarily near zero in a neighbor-
hood of A and this violates the condition \o\
< cm ,v. Likewise,' A < t6.
max
, if 0 < f < ?2
X2(0 = < , if f 2 < ^ < 17 Conditions that determine the f;. are
, if f 7 < f f. x1(.t2) = A,x1(t2) = 0
and
Solving Eqs. 7-102 and 7-107 for the t.
yields the following: x1(r7)=-A,A;2a7) = 0.
_/ EM \ _1_ For computational reasons, it is more con-

(7-131) venient to use the following equivalent set of
conditions:
ti =- (7-132) Xi(0)= 0
Si +!
x2(f7) = 0
1 EM \ J_
ti =-. (7-133)
xt(t7) + A = 0
7-39
AMCP 706-192
and
x,(D = 0
+ (?7-^6)\
i; /2(7)rfT?
where Xi(?2) = A and x2(t2) = 0 are used as

initial conditions for integration.
More explicity, these equations are

n: h(ri2,U)dr)dv
Rt =xl(0) = A + t1 \
J no
f2(n)dn
n;
+i l f2(n)dndp
+ t2\
r fMi)dti
+ 1 I h (?,f 2 ,f 3)dvdv = 0
(7-139)
+1
n
'o
1 hW)dndv
J
o
and
4 = *(7} = -A
rf i
i *i
fMi)dqdv = 0
(7-137) + (7"-f8) f /2(7?,?4)df?
s:
2=X2(r7)=\ /2(7?,?2,f3)^
n: f2(V,U)dvdv
rr
+ \ /2<u)ftj
f2(xi)dr\dv = 0 (7-140)
/2(r?,f2,f3)dT? = 0 (7-138)
where the i?; are introduced for notational
purposes.
A3 = x,(f7) + A = 2A
Generalized Newton Method is used to
solve Eqs. 7-137 through 7-140 for the fr An
initial estimate (, / = 1, 2,5, 4, is made; and a
+ (t1-ti)\
C /2(T),?2)f3)dij
correction is computed according to the
formula
7-40
AMCP70G-192
W= 10,000 lb/in2
Ti~ }f R'I
and
?2 fa -1 R2
[a/?,] m,x = 20,000 lb/in?
F3 fa L *> J R3
It was noted that thre distinct situations
may occur depending on the value of A. In
. /- RA
this example, the problem breaks down as
(7-141) follows:
1. If A > 0.509, then*! (t) < A for all?,
where terms on the right side of Eq. 7-141 are
computed in terms of the r The corrected 2. If 0.156 < A< 0.509, then there isjust
guess ff then takes the place of t and the one point / for which \xx (t) = A, and
process is repeated. 3. If A < 0.156, then there are two values
of t for which |xx (?) I = A.
The derivatives of the Rt with respect to
the f are computed by the chain rule of The numerical example given was solved
differentiation, Eq. 7-109, and Eqs. 7-137 for eleven different values of A. The results of
irough 7-140. Just as inEq. 7-129 several of these calculations are presented in Table 7-5.
these derivatives must be determined by The first column of this table consists of the
successive numerical integration. values of A considered. The following ten
columns give information that, when
The matrix of derivatives which appears in substituted into Eq. 7-130, completely speci-
Eq. 7-141 has sixteen elements. Twelve suc- fies the optimal beam. The next column gives
cessive definite integrals appear in one or the volume of this optimal beam. The final
more elements of this matrix. Therefore, two columns give the volume of the lightest
considerable computation is involved in each beam constant depth that satisfies the condi-
iteration. All this computation was incor- tions of the problem and the percent saving
porated in a single computer program. realized when the optimal beam is used
instead of this uniform beam. Dashes have
As a numerical example, the beam of Fig. been inserted in the table when the quantity
7-14 having the following properties is con- to be tabulated does not exist.
sidered :
For each value of A, the iterative procedure
used to solve the problem required approxi-
M = 1,100 n.-lb mately two minutes per iteration on an IBM
1410 Computer. However, three to five itera-
T = 40 in. tions were sufficient to obtain convergence of
the residue to seven decimal places, so com-
b = 0.5 in puting time was not excessive. This rapid
convergence is, again, characteristic of the
E = 107 lb/in? Generalized Newton Method.
A = 25 in. An interesting sidelight of this particular
7-41
>
I
4^ 2
o
-o
u
o
e
TABLE 7-5
RESULTS FOR BEAM WITH S-SHAPED DEFLECTION CURVE
Vol. of
Unif.
A, fa- 7, Si f2 t3 U Vol., Beam',
in. in. in. in. in. in. in. in.3 in.3 Savinqs, %
XD.509 16.3 16.3 0
0 40 8.27 13.59 23.93 3.35 69.00 16.7 17.6 5
0 30 5.56 12.75 25.00 4.94 92.91 17.4 19.4 12
0 23 4.18 11.94 25.00 6.57 111.96 17.9 21.1 18
0 16 2.84 10.96 25.00 9.69 146.37 18.8 23.9 28
0.15 2.32 11.56 24.51 31.02 32.02 33.90 11.84 8.45 234.5 1 180.41 19.0 24.3 29
0.14 2.11 11.54 24.56 29.66 31.93 35.09 13.10 9.75 263.99 225.17 19.3 25.0 30
0.10 1.31 11.53 24.58 27.93 31.77 37.50 21.02 16.45 431.98 440.45 21.3 28.0 31
0.075 0.88 11.49 24.74 26.95 31.67 38.42 31.10 24.89 643.33 695.92 23.3 30.7 32
0.05 0.51 11.45 24.89 26.15 31.58 39.12 53.56 43.62 1113.10 1256.79 26.5 35.2 33
0.02 0.15 11.40 25.00 25.37 31.49 39.75 182.14 150.63 3794.73 4457.51 35.9 47.7 33
'Volume of lightest beam of constant depth which satisfies all the requirements of the problem.
AMCP 706-192
example concludes the present subparagraph. optimal beams with deflection requirements
A plot of volume (of optimal beam) versus greater and less than 0.156 have considerably
deflection requirement for the problem con- different form. The beam profiles of Figs.
sidered here is given in Fig. 7-16. From this 7-17 and 7-18 illustrate this difference graph-
ically. As A decreases toward 0.156, the jump
in h(t) at / = 25 (see Fig. 7-17) becomes more
40n pronounced. However, for A very slightly less
than 0.156, the profile is continuous, much as
30-
in Fig. 7-18.
7-4.5.4 CONCLUSIONS
20-
The examples considered in pars. 7-4.5.2

10-
and 7-4.5.3 are of the order of complexity
that might be found in actual practice. In
these examples, a saving of material up to
0 01 02 03 0.4 0.5
33% is realized when nonuniform, optimal
A, in. beams are used instead of uniform beams. For
Figure 7-16. Volume vs Deflection more complex loading situations, the saving
Requirement may be even greater. From an engineering
viewpoint, such savings are significant.
graph, it appears that the volume of the
optimal beam is a continuous function of
deflection requirement. This is a rather re- In structural applications, this saving may
markable result in view of the fact that be offset by additional cost of fabrication.
n
1.14
All dimensions in inches
T
_L
0.81
I I I
0 3.2 25 40
Figure 7-17. Profile of Optimal Beam for A = 0.16
All dimensions in inches
1.21
IT 0.81
0 2.6 24.5 31.0 33.7 40
Figure 7-18. Profile of Optimal Beam for A = 0.15
7-43
AMCP 706-192
However, for applications in which weight is a forming nonuniform beams is not prohibitive,
premium, such as in aerospace work, fabrica- such as in the manufacture of reinforced
tion of minimum weight structural members concrete beams, then nonuniform optimal
may be quite feasible. Further, if the cost of beams may be used to advantage.
REFERENCES
1. R. A. Ridha and R. N. Wright, "Minimum Plastic Bodies", Proc. Symposium

Cost Design of Frames", /. of the Struc- IUTAM,yy. 139-146, Warsaw, 1958.
tural Division, ASCE, Vol. 93, No. ST4,
pp. 165-183, August 1967. 9. J.B.Keller, "The Shape of the Strongest
Column", Archive Ratl. Mech. Anal.,
2. J.L. Lagrange, "Sur la forces des Ressorts Vol. 5, pp. 275-285, 1960.
plies", Mem Acad., Berlin, 1771.
10. I. Tadjbakhsh and J.B. Keller, "Strongest
3. T. Clausen, "ber die For archi- Columns and Isoperametric Inequalities
tektonischer Sulen", Bulletin physico- for Eigenvalues", J.Appl. Mech., Vol. 29,
methematique de l'Academie, T. IX., pp. No. l,pp. 159-164,March 1962.
386-379, St. Petersburg, 1851.
11. J.B. Keller and F.I. Niordson, "The
4. Z. Wasiutynski, and A. Brandt, "The Tallest Column", J. of Math, and Mech.,
Present State of Knowledge in the Field Vol. 16, No. 5, pp. 433-446, 1966.
of Optimum Design of Structures", App.
Mech. Reviews, Vol. 16, No. 5, pp. 12. F.I. Niordson, "The Optimal Design of a
341-350, May 1963. Vibration Beam", Quarterly of Appl.
Math., Vol. 23, No. 1, pp. 47-53, April
5. C.Y. Sheu and W. Prager, "Recent Devel- 1965.
opments in Optimal Structural Design",
App. Mech. Reviews, Vol. 21, No. 10, pp.
13. W. Prager and J.E. Taylor, "Problems of
985-992, October 1968.
Optimal Structural Design", J. Appl.
Mech., Vol. 35, No. 1, pp. 102-106,
6. D.C. Drucker and R.T. Shield, "Design
March 1966.
for Minimum Weight", Proc. 9th Inter-
national Congr. Appl. Mech., Brussels,
1956. 14. J.E. Taylor, "Minimum Mass Bar for
Axial Vibration at Specified Natural Fre-
quency", AIAA J., Vol. 5, No. 10, pp.
7. W. Prager, "Minimum-Weight Design of a
1911-1913,October 1967.
Portal Frame", Proc. ASCE Vol. 82,
1956.
15. J.E. Taylor and C.Y. Liu, "On the Opti-
8. D.C. Drucker, "On Minimum Weight De- mal Design of Columns", AIAA J.,Vol.
sign and Strength of Nonhomogeneous 6, No. 8, August 1968.
7-44
AMCP 706-192
16. C. Goffman, Calculus of Several Vari- Chicago, 1946.

ables, Harper and Row, New York, 1965.
18. J.F. Traub, Iterative Methods for the
17. G.A. Bliss, Lectures on the Calculus of Solution of Equations, Prentice-Hall,
Variations, University of Chicago Press, Englewood Cliffs, New Jersey, 1964.
7-45
AMCP 706-192
CHAPTER 8
METHODS OF STEEPEST DESCENT FOR OPTIMAL DESIGN PROBLEMS
8-1 INTRODUCTION A second desirable property of a general

method of optimal design is that it apply
As seen by the examples of Chapter 7, routinely to a large class of real-world optimal
solution of the necessary conditions for the design problems. To be useful to the working
general problem of optimal design is difficult. design engineer, the method should apply
Even in idealized design problems numerical whenever the designer has developed the
methods must normally be employed to capability to analyze the system to be de-
construct a solution. signed. Further, the method should be ex-
plicit enough so that a senior engineer can set
The numerical techniques for the indirect the problem up for computation and a less
method presented in par. 6-5 and in Chapter 7 experienced junior engineer can program the
are iterative in nature. Each of the techniques algorithm for use on a digital computer.
requires that an estimate of the solution be
made before the iterative process may be The methods to be developed in this
initiated. In many cases, particularly in new chapter and applied in the next have manv of
problem areas, the designer may have only a these nice-to-have properties. The basic idea
gross notion of what to expect of the solution of these direct methods is to simplifv the
so his initial estimate may be poor. basic design problem so that it will readily
yi'iid information which allows the designer
Convergence of the techniques ot Chapters to make a small improvement in an estimated
6 and 7 are reported to be very poor unless optimum design. After the improvement is
good estimates of the solution are available. made, a new and better estimate of the
In fact, these iterative techniques often di- solution of the optimal design problem is
verge for poor estimates of the solution. On obtained. The process is repeated successively
the other hand, if a good initial estimate is to obtain small improvements in the best
available, these methods converge very rapid- available estimate of the solution until the
design obtained is sufficiently near the opti-
mum.
This discussion illustrates the need for a
workhorse technique that may be used even The basic method of simplification of the
when only poor estimates of the solution of design problem is to expand functions in-
the optimal design problem are available. The volved in the problems through use of Tay-
method should be capable of making steady lor's Formula. In this way, a simplified
improvement in an estimated solution and, in problem is obtained which serves as a good
fact, converge to the solution. Rate of con- approximation of the original problem pro-
vergence could be sacrificed for dependability vided only small changes are allowed in
if required. certain variables.
-1
AMCP 706-192
8-2 A STEEPEST DESCENT METHOD FOR and

THE BASIC OPTIMAL DESIGN PROB-
LEM 4>(t,u) = 0, = 1, .... q', t < t < t',\
8-2.1 THE PROBLEM CONSIDERED (p(t,u)<z 0, = q'+ l,...,q, t < t< t1.
In order to present the basic ideas of the (8-4)

Method of Steepest Descent, consideration
Just as in Chapter 6, the variables x(t), (/),
here will be limited to optimal design prob-
and b are vectors, x(t) = [xjW, ..., xn(t)]T,
lems with fixed endpoints, no discontinuities
u{t) *= [Ml(0, ..-, um)]T, and b =
in the basic problems, and no intermediate
lbu...,bk]T.
conditions on the state variable. As seen in
par. 6-4, this eliminates state variable in-
Any inequality constraints of the form
equality constraints from direct treatment.
All these features of more general optimal
G)(t,x,u,b) < 0 (8-5)
design problems will be treated in par. 8-3.
can be transformed into a constraint of the

Specifically, the problem treated here is to
form
find u(t), t < t < tl, and b which minimize
J = g0(b,x,x1)+ I f0[t,x(t),u(t),b] dt \ j\co[t,x(t),u(.t),b]
(8-1)
subject to the conditions + \o>[t,x(t),u(t),b] | *=0 (8-6)
^-=f(t,x,u,b), tu < t< t1

at which is then a constraint of the kind of Eq.
(8-2) 8-3.
Ax,x1) = 0 s = 1,..., n
The class of problems considered is, there-
K =ga(b,x,x1) fore, fairly general. The essential features that
are not included are variable limits of inte-
i
+1 LAt,x(t)Mt),b]dt = 0,
a= \,...,r'
gration, discontinuities in functions of the
problem (u(t) may still be discontinuous),
intermediate conditions on x(t), and state
variable inequality constraints.
>(8-3)
= g(b,x,xl)
8-2.2 EFFECTS OF SMALL CHANGES IN
DESIGN VARIABLES AND PARAM-
f
+ 1 L [t, x(t),u(t),b] dt< 0, ETERS
The basic idea of the direct method of

a = r' + 1, ..., r solving optimal design problems is to first
8-2
AMCP706192
construct an estimate w(0)(0, (0) of the ot= 1,..., r, and

solution and then find small changes S(f), 6b
such that u<-\t) + 8u(t), M> + 8b is an 30
improved estimate in some sense. Before the 50, =~Su,=l,...,q. (8-10)
improvements can be determined, analysis of
their effect on the problem must be per-
In all the formulas, Eqs. 8-7 through 8-10,
formed.
the functions are evaluated at [f,x(0)(r),M(0>
(t), fc(0)] where x(0\t) is the solution of the
In this analysis, Su{t) and 8b are required
boundary-value problem Eq. 8-2 for x(t) with
to be small so that a first order Taylor
wa)=M(0>(Oandb = (>.
expansion of the functions of the problem is a
good approximation. Since x(t) is the solution
To simplify the work which follows, it will
of a boundary-value problem involving u(t)
be convenient to eliminate explicit depen-
and b, it is clear that 8u(t), 8b will cause a
dence of Eqs. 8-7, 8-9, and 8-10 on 8x(t).
change Sx(f) in x(t). It is assumed here that
This elimination is performed through use of
the boundary-value problem for x(t) is well
the differential equation adjoint to the linear
posed, see Ref. 1, page 227, that 8u{t), 6b
equation for 5x(f) in Eq. 8-8 (Ref. 2). This
small implies 8x(t) small. Using this fact,
equation is
dX df
db dx dx1 X + h(t) (8-11)
dt dx
where the function hit) will be chosen to

8x+ 6u (8-7)
)tAjx- ^r obtain results needed later in the develop-
ment.
dt Note that for any solution \(t) of Eq. 8-11

db / and any solution 8x(t) of Eq. 8-8,
dhx df
=
~T~ T~ x + 8u + 8b I
dt dx du db I -d-(Ar 5x) = 6x + AT = - 47 6x
(8-8) dt dt dt dx
ae 1
i&x + s-bx =0, s= I, ...,n I
dx 0 dx1 / + hTSx + XTKSx + XT^8u
dx du
d
8
ft. 8a dga
*=-^
az>
6b +
T7o
dx *+-
dx1
*'
+ XT~-8b = hT8x + XT^-8u
db du
ox + 8u + X^6b.
J
\ \ dx du db
t"
Integrating this equation from t to t1 and

+ 6b) dt (8-9)
using the fundamental theorem of calculus,
8-3
AMCP 708-192
KT (t^xit1) -\T (t0)8x(t) = \bb db} .

dt.
(8-12) (8-14)
9/ a/
J: fcrx + X7"-^-5w + X7'-^-5Z>
9u 9
dt. Likewise, put h(t)
and define
- bLr[t,uw,b(0)]
dx
\Va(t) as the solution of Eq. 8-11 with
By choosing the function ft(/> and the boundary conditions onX a(t) and X "(t1)
boundary conditions on \(t) apprqpriately, determined by
the identity Eq. 8-12 will yield the desired
relationships. First, put hit) = df0T/dx
U(0\b(0)] and define \J(t) as the solution
of Eq. 8-11 with boundary conditions on
\J(t) and XJ(t1) determined by Sx + - Sx1 (8-15)
9x 9x'
\jT(t*)5x1 -\jT(t)bx for all 5x and Sx1 satisfying the second

equation of Eq. 8-8. The identity Eq. 8-15
^ 5x l Sx' (8-13) determined boundary conditions just as Eq.
9x dx
8-13 did. Substituting tfrom Eq. 8-12 into Eq.
8-9 yields
for all 5x and Sx1 satisfying the second
equation of Eq. 8-8. To see that the second
equation, Eq. 8-8, and Eq. 8-13 determine
conditions on \J(t) and X/(r1), consider the
following procedure. Determine n of the 2 Jt
variables Sxf., Sx1., i = 1,..., n in terms of the
remaining n of these variables. Now substitute 3/
-+\v So dt .
the variables Sx0., Sx1j just found into Eq. db
8-13. Eq. 8-13 may now be written as a linear
(8-16)
combination of n independent Sx0,-, Sx1 r
Since Eq. 8-13 must hold for all n indepen-
In terms of the adjoint variables \J(t) and
dent variables Sx0,., Sx1. previously identi-
X a(t), the quantities 5/ and 5â are now
fied, the coefficients of all these variables
given explicitly as functions of 8u(t) and 6b.
must be zero. This is then a system of n
The problem is now reduced to determining
equations involving only \JU), Xy(f') and
8u(t) and 8b, which yield the greatest re-
known quantities. This procedure will be
duction in / subject to the linearized con-
carried out in detail in particular problems.
straints of the problem.
Substituting from Eq. 8-12 into Eq. 8-7
It should be noted that the boundary-value
yields
problems for XJ(t) and X a(t), a = 1, ..., r,
have solutions if the boundary-value problem
3go of Eq. 8-8 is well posed. This is a basic
67 = '
db
8b +
r [ au,
property of adjoint boundary-value problems
which is proved in Ref. 2.
8-4
AMCP 708192
8-2.3 A STEEPEST DESCENT APPROACH Note that the column vector 0(f) may have
different components at different points in t
The problem of determining 8u(t) and 8b < t < t'. In order to assure that constraints
to minimize (maximum negative of) 6J must are satisfied, it will be required that
deal with the inequality constraints, Eqs. 8-3
and 84. The argument to be used here is: 84/a =-a4/a, a= 1,..., r'
simply ignore a constraint function that is
negative before an iteration begins. If, on the 8ipa < -a\j/a, a = r'+ l,...,r (8-19)
other hand, a constraint function is positive,
and a&A
it is required to be reduced. For example, if
4>a > 0 or 0 for some t, then it is
required that 8i>a = a\jja , a = 1,..., r' and and
S'Pa < ~axlJ<X' a = r' + 1>- rand 4>a > Oand
50^(0 = - c^W, 0=1,.... q and S<^(0 < 6fy(/) = - c0(f), =h-,q' '
- c4>(t), = 4' + 1,..., 4 and <j>g(t) > 0 where
0 < a 1 and 0 < c < 1. The magnitude of a 60,(0 < -c4>(t), = q'+l,...,q (8-20)
and c are chosen so that the required changes
and GB(t).
5\jja and 84 At) are not excessively large. If
$a and 4>e(t) are not so large that the linear
approximation is violated with a = l,orc = 1, Finally, define
then a or c are chosen as one.
For convenience, define two sets of indices (8-21)
A= {a|*a[x<0>,u<0>,<>] >o}
dt
and a* 1 L a*
Bit) = { \<p0[t,u(o\t)] > 0} .
(8-22)
It should be noted that the collection(0 of

indices may change with the variable /. A* (0 = +-J x a
au du
Define the column vector of elements 0

r
a foralla^4 (8-23)
with 0 s> 0
and
r
a
(8-17)
e* =
**'*
aGA
db
and a similar column vector of functions 0(f)
with <t>M) > 0 for all a^ . (8-24)
Note that A* (%)is a matrix of functions with

(0 (8-18) m rows and the same number of columns as
-^B(t) there are indices in A. The matrix $* of
.8-5
AMCP 706-192
constants has k rows and the same number of G = \X0AJ T+yTy>T)8b

columns'as A* (t).
+ y08bTWb5b + yTcijJ (8-29)

Using the matrix notation of Eqs. 8-21
through 8-24 in Eqs. 8-14,8-16, and 8-17.
such that
oH jT T ,
8J = S.jT8b+ \ Aji (t)8u dt (8-25) -=0 = _\0A' -TrA*
.fjiT(trfi>--2y08uTW, (8-30)
du
8^ = ^ A* (t)8u dt . (8-26)
Before 8u(t) and <56 are determined, some

3G (' M
mechanism must be set up for requiring that dt = 0
b8b
these variations are actually small. For con-
venience, put
= X0^r + yT^ T
+ 2y08bTWb
2 T
dP = 8b Wb8b +
I T
I 8u Wu(t)8u dt
(8-27)
(8-31)
where Wb and Wu{t) are chosen as positive In the following development, it will be
assumed that the problem is normal so that it
definite weighting matrices and dP is to be
is permissible to put X0 = 1. Solving Eqs. 8-30
chosen small enough that 8u{t) and 6b are
and 8-31 for5(f) and 8b, respectively, yields
sufficiently small.
The problem is now reduced to finding

Suit) = - *, (t)-1
Su(f) and 8b which minimize 6/ subject to 27o
Eqs. 8-19, 8-20, and 8-27. This problem is
now a special case of the Bolza problem of
90T '
par. 6-4. According to Theorem 6-7, there AJ{t) + AHt)y+ -T- M(0
ou
exist multipliers X0 s> 71 n > 0
(8-32)
t\~l L J
for a > r\ p(t) = V , At) > 0ior> and
q , and 7o with
8b = - Wb ' (F + fi*17) (8-33)
27o
H=[-\0A'T-yTA*T(t)-nT(t)daj
* In order to complete the determination of
8u(t) and 8b, p(t) and y must be eliminated
-y08uTWu8u + firc<j> from Eqs. 8-32 and 8-33. A direct analytical
(8-28) elimination of MC) and y is not feasible at this
8-6
AMCP 706-192
point since Theorem 6-7 requires or
2y0 du "
HT(t)[b4>(t) + ci>{t)) =0
and some components of 5^ +a$ and 50(0 + AJ(t) + A* 0)7 + -v- MO) + c0 = O
9u
c$0) will be zero. A certain amount of logic is
required to find 7 and p(t). Since only small It is assumed that at points where B(t) is not
Su(t) and 8b are admitted, if \p and 00) are
empty, 90/9 has full row rank, i.e., all
zero for (0)O). *(0) then very likely they constraint functions which are zero or posi-
wlalsobe zero for(0)0) + 8u(t), *(0) +5*. tive are independent. Since Wu(t) is non-
Following this line of reasoning, it will be singular, the matrix
assumed first that
90 90r
8\p +a\p = 0 u (8-34)
au du
and
is nonsingular. Therefore,
500) + c0O) = 0 . 0 _1
MO) = -A O)
Then 7 and n(t) are determined by substi-

tuting S0) and 8b from Eqs. 8-32 and 8-33 30 Wul (AJ + A^ y) - 27oc0 ,
into these equations. The multipliers 7 and du
p(t) are then determined and checked for the (8-35)
proper sign. If ya > Ofora> r' andpJt) > 0
for > q', then this assumption is admissible. At points where B(t) is empty, put 90/9u =
If, on the other hand, ya < 0 for some a > r', 00) = 0 and A*0) = 1. In this way, MO) is
then 8\pa +a\pa = 0 is incorrect and it must consistently defined by Eq. 8-35 for all t.
be that S0Q +a\pa < 0 should occur. This is
equivalent to simply removing \\>a from \p and Substituting Eq. 8-35 into Eq. 8-32
recalculating. Likewise, if Jt) < 0 for some
> q\ then 50^0) + a$(t) < 0 should occur
and 0O) should be removed from 0(f) and 8u(t) = - W~l (A'+A*?)
27o
the multipliers recalculated. So much for the
semi-mathematics, now to the calculations
based on this argument. + W~l J&I A*"1
27o " 9"
If B{t) is empty for all t, then 0 is not 9

0" 1
defined and it) need not be determined. In
case B(t) is not empty, is to be required that
90
50 + c0 = 0 = 8u + c<t>
du cw-1 4r
ou
A0_1
*
AMCP70G-192
or and
i
bu{t)=_w-,
, T , dd>J _i
M 4><t>
x(/ A (8-40)
-a * Mw~\
Eq. 8-37 becomes
y
x(A +A*y)
TT (^/+Af^7) + cM|t =ai. (8-41)
-cW-t-T- A* * (-36) Z7

3
where/ is the identity matrix.
Since W (r) is positive definite so is W"1,
Substituting 6u(t) and bb from Eqs. 8-36 and there is a nonsingular matrix 5(f) such
and 8-33 into Eq. 8-26 and then enforcing d\p that W~x(t) = sT(t)s(t). By direct multiplica-
= a\j/, tion, it may be verified that
1 T f''
-^T* W-i (y + *7)-cl A*r
yTM^y=yTl+TW-1 Z*y
f1
i
x"'--1 Â#_1**-2^-^ A*'
3 + \ yTA* Wj1
/ 30r -I 30 . \
xr" 1 \ / A* - w" 1 /)
Zu
x(AJ + X'Jy)d; = -ai

du
(8-37*
"-- *-'-),*
Defining
-yTZ* W'1 9*y
T
A*TW~l
J fo
/ 30r -i 30 , \ , $>-*-
(8-38)
x
-farsT jsA^y
M
i,i, =n*Tw-1 if +[
i A* Vr1
3 du
x sA*y dt> 0.
(8-39)
(8-42)
8-8
AMCP 706-192
Therefore, M, , is at least positive semi- Defining

definite. In the development that follows, it
will be assumed that M, is positive definite 30 7 -l _30
and, hence, nonsingular. 6ul(t)=W- 3w A* du
w;
Eq. 84 1 is now solved for y to obtain

x(Ar-A*M-]lM^J) (8-44)
(8-43)
6uH = -Wj>[l-&A*-lwu-
It should be noted that if the set of indices A
is empty, $ does not exist soM, . is not even
.A*M;l<fi4,-cM)
defined. If, in this case, M, , is defined as
one and 0 zero, then y = 0 in Eqs. 8-41 and (8_45)
" a A* 0
8-36 reduces, appropriately. In this way, a
single mathematical analysis holds in all cases.
Substituting Eq. 8-43 into Eqs. 8-33 and

fi1=W*-l<* -**^J**/> (8-46)
8-36 yields
and
8u(t) =-W~1fl - 30r-A@-i 90 , t>b* = -W-i&*M;l(a4,-cM^) (8-47)

9 du "
the expressions for 5(f) and 6b are simply
-A-' + A^JWr' (aii-cM, J

1
8w(?): SMHO + S" ^)
2
(8-48)
1 27o
AÂf **M*J
2>o
and
8b = - riSd1 +562 (8-49)

wo
and The variations Su(t) and 6b from Eqs. 8-48

and 8-49 could now be substituted into Eq.
8b = W.- 1
^ 8-27 to determine y0 However, since dP has
b
27o no real physical significance, one mightjust as
well choose 70- It is interesting to note that
-W^^M-\{a^-cM^)
the terms 8u2{t) and 8b2 are not multiplied
by an undetermined parameter. Further, note
that each term in the definitions, Eqs. 8-45
2-y * ip <l) \I/J
and 847, of these quantities involves0 and 0.
8-9
AMCP 706-192
In fact, if 0 and if are zero or empty, 8u2 {t) = of Steepest Descent. First, Relation 1 states
8b2 = 0. It appears that 8u2(t) and 8b2 may that the changes Su'Cr), 8bl and 8u2(t), 8b2
be interpreted as making corrections in con- are orthogonal. Relations 2 and 5 show that
straint errors, or keeping constraint functions 8u2(t),8b2 provides the requested reduction
from being violated. Actually, this and more in the constraint functions. Relations 3 and 4
is true. show, as might be expected due to the
orthogonality of Relation 1, that 51 (f), 8bl
Theorem 8-1: The following identities has no effect on the constraint functions in 0
among 6u'(0, 8u2{t), 8b1, and 5b2 of Eqs. and \p. Finally, Relation 6, along with
8-44 through 8-47 hold: Eqs. 8-48 and 8-49, simply states that if 4> = 0
and 00) = 0 then 8u(t), 6b provides a
reduction inJ.
1. 8b1 Wb8b2 + \
C u1
T W 8u2dt = 0
u
Before stating a computational algorithm,

it is important to develop a test for con-
vergence to the solution of the original
problem. The procedure here will be to show,
2. 2* 8b2 +
I
" A* 8u2dt = -a\jj
through use of the necessary conditions of
Chapter 6, that as the solution of the original
problem is approached, Su1 (/) and 8bl must
approach zero.
By Theorem 6-7, at the solution of the

3, V*18b>'1 +
+1\ A* T8u1dt = Q
A* problem, Eqs. 8-1 through 8-4, there are
multipliers w((0. i = 1, - n, va, a
- h > r,
and L(0, (3=1, , q suchthat for
30
4. du1 =0 (8-50)
3
and
30
5. 8u2 =-c4
au (8-51)
it is required that
6. -HJ 8b1 AJ 8u dt s 0 . du>

(8-52)
dt "3JT
By considering the case when 4> and ^ are
empty, it is clear that in order for 8u, 6b to dH
be in the negative gradient direction of J (i.e., = 0 (8-53)
du
5u = A7 and 66 = Q.J), 70 > 0 is required.
3 f'' 3/7
The six relationships of Theorem 8-1 give dt = 0 (8-54)
db ~ \ 3b
the designer an intuitive feel for the Method
8-10
AMCP 706-192
p
aâ =0. <*=r'+ l,...,r (8-55) Further conditions from Theorem 6-7 are,
fromEq. 6-124,
$(0*fl(f) = 0, = q'+\,...,q. (8-56)
dG1
1 - _w(/) = 0 (8-62)
Corresponding to the definition of 4 and 4> dx
in Eqs. 8-17 and 8-18, define 3, L, i, and fas
and
containing only components of v, L, g, and ,
corresponding to elements of ty and 0. In this 9Gr
+^) = 0 (8-63)
notation - and due to Eqs. 8-55 and 8-56
Eqs. 8-52, 8-53, and 8-54 become
Multiplying Eqs. 8-62 and 8-63 by fix0 and
d fix1 yields
-^-Zir!L+zh.+yT ^- (8.57)
dt dx dx ax
|l SXO+PTJ8 8xo_uT(to)5xo=0
9i dx" dx
T of d/o -7- ;r 4>
du du
"' du
"" ou (8-64)
(8-58) and
T1 Sx1 + i>' _ 8xl

+ v' dx dx1
db db
+ uT01)oxl =0. (8-65)
These equations hold for all fix0 and Sx1.

) V 8* to db)dt
Adding Eqs. 8-64 and 8-65,
(8-59) ^L8x+^t>x>
dx dx1
Substituting from Eq. 8-11 into Eq. 8-57
yields + i>T(j*8xO+j%8x\
\dx dx1 I
cfw d\J dX*
+ + P - wr(f)5x0 + u>T(.tl )Sx1 = 0 . (8-66)
dt dt dt
Again, Eq. 8-66 holds for all fix0 and fix1.

(8-60)
ox ox ax Substituting terms from Eqs. 8-13 and 8-15
into Eq. 8-66,
or
\jT(t1)Sx1 -AJT(t)8x
(w + X7 + X* v) =
+ PT\\* T(t1 )5x' - X* r(r)5xI
df
(w+X-7 +X*). (8-6 1)
ax - uTU)ox + cjr(?' )fixJ = 0 (8-67)
8-11
AMCP 706-192
for all 6x and dx1 satisfying the second Premultiplying the transpose of Eq. 8-70 by
equation of Eq. 8-8. By collecting terms, Eq. (30/9") W"1 yields
8-67 becomes
30 , 30r -
/7 7 T l
o/tf') + A V ) + V {t )JSx' 3 " 3 c
- Lr(/) + XjT(t) + v X1* r(f)lx0 =0 30 bfT

du u
\ du
x^'Jl
du
(8-68)
30 . (df , bLT\
- W-1 -i X^ + )i>.
for all 5x and Sx1 satisfying the second du " \ du du
equation of Eq. 8-8.
(8-72)
Eqs. 8-61 and 8-68 constitute the bound-

The coefficient of in Eq. 8-72 is just A*(0
ary-value problem adjoint to Eq. 8-8, where
of Eq. 8-34 which is nonsingular. Therefore,
the dependent variable is (o> + ~KJ + X* 3). Due
to the assumed well posed nature of Eq. 8-2,
30
the boundary-value problem of Eq. 8-8 has a = -A* ^-(A^*).
unique solution for all 6w(f) and 5ft. It is
shown in Ref. 2, Chapter 4, that in this case (8-73)
the adjoint boundary-value problem, Eqs.
8-61 and 8-68, has a unique null solution, i.e., Substituting Eq. 8-73 into Eq. 8-71
w(f) + XJ(t) + X* (f) = 0, t < t < tl . 9gor

V
+ X* 1
3ft 3& J o L 9*
(8-69)
Substituting for co(t) from Eq. 8-69 into m+

3ft
dLT
3ft
dt = 0
Eqs. 8-58 and 8-59 yields
3 or, in the notation of Eqs. 8-22 and 8-24,

T
\ du du du
HJ + 9* v = 0 . (8-74)
T 30
(8-70)
Substituting for % in Eq. 8-73 into Eq.
8-70,
and
3ft AJ-A*v + -^ A* ' W~lAJ

du du "
30r A-i 30 ,
+ - A* W"1 A* 3 = 0
du du "
(8-71) (8-75)
8-12
AMCP 706-192
or,
Premultiplying Eq. 8-75 by A* W~l and
integrating yields
V bu bu " I
x\AJ -A^T/^J-0. (8-79)

J fo
Eq. 8-79 is the desired result, 5ul (t) = 0.

l
x A* W~l AJ J dt
bu bu " Substituting 3 from Eq. 8-78 into Eq. 8-74
yields
c< A* WÂ* Z'-9*M;IM =o

and this implies, by Eq. 8-46, that Sb1 = 0.
V ,-i
A* W~l .0_ A*~ It is now possible to state a computational
bu algorithm employing the results of the pre-
ceding analysis and discussion.
00
x y- /-A* I d/ i? = 0 (8-76)
A Co mpu ta do nal A Igorith m:

Premultiplying Eq. 8-74 by ^ W~' yields,
Step 1. Make an engineering estimate
j* w/'-'+* W-1fl*3=0. (8-77) w(0)(O, &(0), of the optimum de-
sign function and parameter.
Adding Eqs. 8-76 and 8-77 finally yields
Step 2. Solve Eq. 8-2 for *(0) correspond-
ing to w(0)(0, b(0K
M.. -M, v=0
Step 3. Check constraints and form \jj and

so
0(/) of Eqs. 8-17 and 8-18.
Mr] M. , (8-78)
Step 4. Solve the differential equation
8-11 with h and the boundary
conditions of Eqs. 8-13 and 8-15
Substituting!/from Eq. 8-78 into Eq. 8-75, to obtain XJ(t) and X a(t), re-
spectively.
A7 -h*M~\ M, ,
Step 5. Compute A/(/)) 7, A* (t),and *
in Eqs. 8-21 through 8-24 and
u
bu bu A*(OinEq. 8-34.
1
+ ML A*- te-W7iA+M7iM Step 6. Choose the correction factors a and
l
**m*j=0
du bu ein Eqs. 8-19 and 8-20.
8-13
AMCP 708-192
Step 7. Compute M^j, M, ,. and M^ ^ in Substituting these expressions into Eq. 8-25
Eqs. 8-38, 8-39, and 8-40. and using Eqs. 8-44 and 8-46,
5/=
Step 8. Choose y0 > 0 and compute y - Wo [/rpv(/-*^;*v)
and MO) of Eqs. 8-43 and 8-35. If
any components of 7 with a > r',
or /J(0 with 0 > q', are negative,
redefine $ and <ji(f) by deleting J ,0
corresponding terms and return to

Step 5. M/-Â*-^r'
3w du "
Step 9. Compute bul(t), bu2 (t),8bl, and
bb2 of Eqs. 8-44 through 8-47. x [A> -A*M^M^)dt
Step 10. Compute -- (M^-M^JM-IM^
where
u(i>(0=u<>(f)- = bul{t)
'To
2
+ Su (0 MJJ=2,1W^Z'+ [' AjT W^
ft(D = ft(0)__L Sb'+Sb2 .

27o
ou du "
Step 11. If the constraints are satisfied and
61 (0 and Si1 are sufficiently With Eq. 8-80 it is possible to request a
small, terminate. Otherwise, return reasonable magnitude for 6/and compute the
to Step 2 with M(0), bw being 7o which should give this reduction inJ. In
replaced by w(1\ >(1). this way, it is possible to choose a reasonable
7o. Experience with this method on structural
design problems, of the kind discussed in the
An algorithm of this kind invariably in- following chapter, has indicated that a request
volves a certain amount of computational art. of 2% to 10%reduction in the cost function
The critical element of this algorithm is the on the first iteration gives a value of 70 that
choice of the parameter 70 in Step 8. Once yields convergence. Often, this value of To
the constraints are satisfied to acceptable must be adjusted during the iterative process
accuracy, bu2(t) and bb2 will be approxi- to prevent divergence or to speed con-
mately zero and l/(270)can be viewed as a vergence.
step size in the direction 5M1 (0> bb1. In
this case the change in u(t) and b is This matter of choosing step size in Step 8
requires a great deal more attention. With a
bu(t) = - bul(t) little experience one can develop a "feel" for
2TO
how to adjust 70 to get good convergence,
bb 5ft1 even in complex problems. A feasible auto-
27o matic method of choosing 70 is desirable for
8-14
AMCP 706192
use on high-speed computers. No reliable

method is known to the writer at this time.
8-3 A STEEPEST DESCENTMETHOD FOR

A GENERAL OPTIMAL DESIGN PROB-
"['L- [t,x(t),u(t),b]dt = 0,
a= !,...,/'
LEM
> (8-84)
8-3.1 THE PROBLEM CONSIDERED
*=*(*.''.*'>
The basic optimal design problem with
fixed endpoints, no discontinuities, and no 1 La[t,x(t)MO,b]dt<0,
intermediate constraints was treated in the
preceding paragraph. The problem considered a = r' + 1,..., r
here will be a generalization of that problem
to include features such as variable endpoints, and
discontinuities, and intermediate constraints.
The basic idea of the method of solution will *,(*.) = 0, t Qt< f\
be the same as in the preceding paragraph.
Accounting for the additional features of this
problem, however, introduces some com- (8-85)
plexity into the derivation of equations. <t>(t,u)< 0,t<t< /",
The problem to be treated here is to

= q'+\,...,q.
determine u(t), t < t o tv, b, and t, tl,...,
tn which minimize Note that this is just a special case of the
problem of Def. 6-3. Equality constraints will
be included in Eqs. 8-84 and 8-85 in a natural
J=g0(b,t>,x>)
way during the development. It is assumed
that for given u(t),b,t', andx; the boundary-
,[t,x(t),uV),b]dt (8-81) value problem, Eq. 8-82, has a continuous
solution x(t). If constraints of the form
u(t,x,u,b) < 0 occur, they may be replaced
subject to the conditions by a constraint of the form
dx
f(t,x,u,b), t < t< t*,tt>\ {oj[t,x(t)MO,b]
dt
H8-82 (8-86)
+ \w[t.x{t)Mt),b]\) dt = 0.
d (t.x.t7'.xr>) = 0.s= 1 n 1
Constraints of the form of Eq. 8-85 are easily
treated in a direct manner so they need not be
ni(ti,xi) = 0,i=\,...,v (8-83) reduced to the form of Eq. 8-86.
8-15
AMCP 708192
Just as in the method of par. 8-2, the where Ax1 is the total change in x' at the
idea here will be to estimate u(0)(Y),&(0) point t', Since x{t) is to be continuous before
and t, .... tn, and then to allow small and after the variation, the total change in
changes 8u(t) and bb. The object is to x(t) must be continuous at each point t. This
determine 8u(t) and 6b which yield the requires that
greatest reduction in J and which satisfy
the constraints. 8xi(ti -0)+f(f -0)8f = AxJ
8-3.2 THE EFFECT OF SMALL CHANGES = 5x''(f' + 0)+f(f + 0)8f (8-89)

IN DESIGN VARIABLES AND
PARAMETERS i = 0,1,. . ., q, where 5x(f) and dt' are
indpendent changes in x(t) and f.
Before the optimum changes in u(t) and b
may be determined, the effect of these The independent variation in x(t), Sx(f), is
changes on J and tya must be assessed. Since J related to 8u(t) and 6ft through the varia-
and *pa have the same form, the expressions tional equation
for change of a general functional
d bf bf 3/
(ox) = 5x + 8u+8b,
dt ox du db
Q=g(b,t> ,x>) +1 F[t,x(t)MO,b]dt
t < t < tn,t= t'. (8-90)
(8-87)
The boundary conditions, Eq. 8-82, require

will be determined and the result will be
that
applied to J and 4*a
bds bds b6s

Expanding Q to first-order terms in the ~n Ax +~8t +- Ax"
variables u(t),b,f, and x7, yields 3x 3/ 3*"
8Q= de8b+ de de 3e
Tb -ir^0+-+^ (> +
,
5/" = 0 (8-91)
3r"
+ ...+|[^4!_6ro+...+|r5f/ s = 1,..., n. Finally, the relations of Eq. 8-83

dx* bt dt'
require that
+ ...+-^-9f -F(t +0)5r + ... 3S2' . da' .

Ax' + of = 0,
!
3x ' bf
+ [F(t> -0)- F(t> + 0)] Of'' +
:
0, 1 77. (8-92)
+ F(tr> -0)5f
It is clear that the variations 5x(f), 8u(t), 8b,
Bf, and Ax' are not all indpendent.
f'" / dF dF bF \
8x+ 6u+ 5b dt,
)lAl7 ^ lF ) In order to express S in terms of only
8-16
AMCP 706-192
Su(t) and 8b which are to be determined, Or, using the definition of Ax', this is
introduce the adjoint variable X(t) just as in
Eq. 8-91, ('* i
1 dx 8xdt = -XT(t + 0)Ax
J t
dh _ lf_ dFT
(8-93)
dt dx ax
+ Xr(f +0)f(t +0)5/
Integrating the identity below Eq. 8-91 from
t' to t'+1, 0 < /,/ + 1 < qr,one obtains,just as + ... + [\T(t> - 0)
in Eq. 8-92,
-XT(t'+0)]Ax'
Xr(^+1 -0)8x(t'+1 -0)
- [XV - o)/o> - o)
-XT(t' + 0)8x(t> +0)
-XT(t'+0)f(t'+0)]8t'
>' ' dF ^.Tdf

8x + XJ 8u + ... + Xr(r" -OJAx"
dx du
- XT(tn - 0)/(f - 0)6f
+ XT-^8b) dt. (8-94)

db
+1 (V-^5
J
t
Note that the boundary conditions on \{t)
and the properties of X(t) at the points t', j = + XT -di 8b \dt
I, ..., TJ 1, have not yet been specified.
These boundary and intermediate conditions Substituting from Eq. 8-95 into Eq. 8-88,
on X(t) will be the major output of this yields
subparagraph.
9
8Q = -r- 5b + \+XT(fi) Ax + ...
Summing all the formulas, Eq. 8-94, overj db dx
= 0,1,..., 7] 1, one obtains
[j*, -XT(t'-0) + Xr(t' + 0) Ax*
\_dx'
8xdt=- XT(t +0)SxU + 0)
\ ,o dx
+ ... + \T(t> -0)6x(t' -0)

-*[ 9x"
1
-\T(tj + 0)8x(t> +0)+... (8-95) + 0)
+ XT(tn -0)8x(t" -0)

-XT(t +0)/(f +0) 6f
8u + XT~8b) dt
) \ du db +... + + F(t' - 0) - F(f' + 0)
dt>
8-17
AMCP706192
+ \T(t> -0)f(t> -0) + ...+

-x'V+oW+o) />
+ Xr(fT> -0)/(f -0) of =0
+ ...+ -2- + F (tv _ 0) (8-97)
+xra* -o)/(/" -o)jsf for all Ax' and St' satisfying Eqs. 8-90 and
8-92.
("[(ifrfk In order to determine conditions on X(f;)

based on Eq. 8-97, a Lemma is required.
Lemma 8-1: For A, Bj: i = 1, ..., m < n in

\9ft 9ft/ J
R" ifATx = OforallxinR" suchthat
(8-96)
BTx=0, i= l,...,w
;
The quantities Ax' and S/ appearing in Eq.
8-96 are required to satisfy the conditions of then there exist constants u>, suchthat
Eqs. 8-91 and 8-92. The objective now is to
choose the boundary and intermediate condi-
tions on X(t) so that Eq. 8-96 is independent A' x + 1=1
X cj.Bj
i '
x=0 (8-98)
of Ax' and St'; i.e., so that
for allx in/?" .
9x
+xr(f) W + For proof of this Lemma see Ref. 3, page
12.
9? Consider the expression of Eq. 8-97 as

J* -\T(ti-0)+\T(ti+0)\Ax>
-dx> 1 ATx, where the components of .4 depend on
the values of \(t') and x = (Ax07",..., Ax'r,
+ ... + ^--Xr(f) Ax" of0, ..., St")7. The equalities Bfx = Oarejust
9x" Eqs. 8-91 and 8-92. Denoting the multipliers
T.,S = 1,..., n, and 7,,;' = 0, 1,..., 77, Eq.
ceomes /
|L_F(r+0)
- Xr(f + 0)/(r + 0) St
ô-^l,^
9g
+ +
912
Ax
+ 7o
Ix5"
T
- F(?/ + 0) + \ (t> - 0)f(t' - 0)
\T(t> +0)f(t'+0) St' +... + -|- _ xr(r' - 0) + xr(r> + 0)

9x'
AMCP706192
an'
+ 7-
7 ~ Ax'
> bx>.
^_\T{tn _0)+ 2T "
Zg r 3x" 3x"
+ ... + X (f - 0)
3x"
an"
+7 =C (8-102)
n '" 3x"
be s 9^"
+ 2 r s ~ +y Ax"
i=i 3*" " dx"
-^ - F(t + 0) - Xr(f + 0)/(f + 0)
^ _F(r +0) -\T(t +0)fu +0) 36 3ft

+ +
s?/S-# ô 1^T=0 (8-103)
"3?
" ae 3J2"
ST, * + 5f
= 1' 3? y 9ro
%- + F(t> -0)-Fit1 - 0)
/ at'
+ + j-+/7(f/_o;-Fa -0)
3f'
+ \T(t> -0)f(t> -0)
+ x7V-o)/u>-o) - \T(t> + 0)f(t> + 0)
bQ,f
r
r , , 0^ + y. -= n (8-104)
- X (f' + 0)/(/'+0)+7,-^7" bt' "' bt'
+ +
+ F(f> - 0) + Xr(f - 0)/(f - 0)
3fi
+ Xr(f -0)/(/" -0) + 2 rs **j- " 39 an"

1=1 bt* + 2r0^+7 -- = 0. (8-105)
3ft""
rn 9fu 5r" =0 The object is now to eliminate the 7. and TS
in order to obtain explicit conditions onX(f).
(8-99) Postmultiplying Eq. 8-100 by f(t + 0) and
adding Eq. 8-103 yields
for all Ax/, or', / = 0,1,..., )? Therefore, n
g(t + 0) - F(t + 0) + 2 T(t + 0)
3# " 30
+ 7o"(f +0)=0 (8-106)
3xu s=i 3x
where
+ 7o (8-100)
3x . 3g
be
08 rr. . bl>
_XT(^-0) + XT(r'+0)+ 7,^ = 0 + -*-/[?' 0, x(f'' 0), u(f'' + 0), */
3x' 9x' 9x!
(8-101) (8-107)
8-19
AMCP 706192
9(t'0)-- Eq. 8-102 by f(P - 0) and adding to Eq.

dt< 8-105 yields
ae
e flf o,x(t' 0), u(t' 0),b]
+ -LL. g(P - 0) + F(P - 0) + 2 T
A('" - )
dx' s=l
(8-108) + 7 &(P -0) = 0.
and Solving for 7 and substituting into Eq. 8-102

yields
n'(r' 0) =
at'
dw f[t> 0,x(t' 0),w(f' 0),b] .

+ 21L
X(P -0)=-
y
+ 2 7, 2i_
9*i s=i a**1
dx'
(8-109)
1
0 0
UP -o) + F(r'
Since n (t,x ) = 0 is to determine t, the n"(fi -CM
total derivative with respect to t, fl(t0 + 0), aW
+ 2 rs6(P - C)
should not be zero. Therefore, 70 may be dx"
determined as
(8-112)
1
7o = - where g {P 0)andoj(f - 0) are defined in
n(f+0)
Eqs. 8-107 and 8-108. Eq. 8-112 then serves
g(t +0)- F(t + 0) as a boundary condition on \(t) at P.
+ 2 Td(t +0) (8-1 10) In order to use Eqs. 8-111 and 8-112 as
explicit conditions on \(t) at t and P, the
parameters rs must be eliminated. This may
Substituting Eq. 8-1 lOinto Eq. 8-100 yields be accomplished by alegbraic manipulation in
9^ " deT particular cases. To illustrate this idea on a
problem which has been treated extensively in
the literature (Refs. 5,7,8,9), consider the
case in which a full set of initial conditions
n(t0 +o) Ug(t + 0) is given, i.e.,
ei(t.x)=0, 5=1,...,. (8-113)

F(t +0)
In this case, Eq. 8-112 yields
3S2 or
+ 2 re. (t + 0) (8-111)
dx dgT __^ 1
UP - 0) =
a** fti^-o)
This is then a boundary condition on \(t) at
t.
g{P ~0) + F(P -0) an"
3x" (8-114)
In exactly the same way, postmultiplying
8-20
AMCP 706-192
Eq. 8-111, on the other hand, is just a vector

equation with n components which deter- \(t'+0)-\(tf -0)=-ad (8-116)
ad
mines Tj ,,Tn. It gives no explicit informa-
tion on X(t +0).
git' 0) + Fit' -0)
Now the boundary conditions on\(t) have
been determined. It remains.however, to de-
termine jump conditions on X(?) at the
intermediate points t'. Postmultiplying Eq.
8-101 by fit1 0) and adding to Eq. 8-104
yields
git' 0) + F{t' - 0) - Fit' + 0) + \Tit'+ 0)[fit' OJ-fit' + 0)]

dx'
T j
-X it' -0)\fit iO)-fit' -0)}
This equation, the boundary conditions at
Eqs. 8-111 and 8-114, and the differential
+ \Tit' + 0) [fit' 0)- fit' + 0)]
equation, Eq. 8-93, are to determine the
adjoint variable, X(f), t Q t < t7*. The
yfW(t' 0) = 0 (8-115)
boundary and intermediate conditions on X(f)
were constructed so that Eq. 8-97 holds and
where the notation of Eqs. 8-107 through
in turn, Eq. 8-96 becomes
8-109 has been used. The choice of limit from
the right or left (the plus or minus sign,
respectively) in Eq. 8-115 is left open for
now. One or the other alternative will be
chosen for computational convenience.
. BF T df ,
dt
It is assumed that the condition fi/it',x') = \bb db '
0 determines t' as a function of x', so it is
required that the total derivative of SV with or
respect to t', 0,'it' 0), not be zero.
Therefore, from Eq. 8-115,
dt bb
Idb J, \db bb
(^T^Y>>- <8-"7
\git'0) + Fit'-0)
n'it'0)
-Fit' +0)
- \T(t' - 0) \fit> 0)-fit' - 0)] This equation meets the objective of this
subparagraph, namely, determination of the
dependence of SQ on 8b and Suit) explicitly.
+ \Tit' + 0) [fit' 0)- fit' + 0)] [ .
Since Q was any functional, this result can be
applied to the particular functionals of the
Substituting this expression into Eq. 8-101, present problem, / and <i . To obtain 6J and
8-21
AMCP 706-192
Si//a, define \J(t) and ^"{t) as the solution

fa =_3,
of Eqs. 8-93, 8-111, 8-114, and 8-116with
db
g = g0andF=/0forX/(0 (8-118)
J? (8-122)
and
and
a
g=aandF = Ifor X* (/). (8-119)
A*(0= ^4-^ X*(f). (8-123)

In this notation, Eq. 8-117 yields du du
In this notation
5/ =
db ' 3 f0 La* jT
5/ = % 6b+\ AjT (t)8u(t)dt (8-124)
J ,0
+.X7 (t)\dt \ 6b
and
In 9
8\jj = %'"<* 8b + (t)8u(t)dt .
,7 3/
6u{t)dt (8-125)
du
and The problem of this paragraph is now in

approximately the same state as the problem
of par. 8-2 was in Eqs. 8-14 and 8-16. Before
6
*M a+ ),IV proceeding to derive a steepest descent al-
gorithm, however, several comments are in
order.
dr \ 8b
First, the choice of limit from the right or

+ left was not made in Eq. 8-116. This choice is
J L"3lT
generally made depending on the distribution
3/ of boundary conditions on the state variable.
8u{t)dt
If most of the boundary conditions on x(t)
are given at t, for example, then most of the
For a more compact notation define boundary conditions on \(t) will be given at
f. Since the adjoint equations are linear,
superposition techniques may be used to solve
db } 0 L 3* 36 J the boundary-value problem. These tech-
niques involve several integrations of Eq. 8-93
(8-120)
from tv to t with different starting con-
ditions at tn. These integrations must account
(8-121)
au du for the jump condition, Eq. 8-116. The
8-22
AMCP 706192
integration then proceeds from the right and Sl'lt'.xit')] #0.

Eq. 8-116 should then be used to determine
\(t> - 0) in terms of \(t> + 0) so that the One might argue that the function x{t) is
integration may continue. For this reason, the close to the actual state and examine the
minus sign is chosen in Eq. 8-116 so that effect onlJ'[t',x(t')] of altering?', i.e.,
dJV_
8tof[tf,x(t0] 6t>

\(t> -0) = \(t> + 0) + - bt>
912' dx .. .
+ 5t> =n'(t' 0)5t>
ox dt
git' - 0)
n'it' - o) (8-127)
+ F(ti -0)-F(t' +0) where the plus or minus sign is chosen

depending on whether St> should be positive
+ \(t> +0)[f(t> -0) or negative to make 2/'(f/ + t'),x(t' +?') = 0.
The change 5?y is then chosen and if it is not
too large, the state equations need not be
-fit'' +0)] (8-126) re-integrated. This argument corresponds to a
Newton-type algorithm for the determination
Since the state equations have previously been of t!. This procedure should be used after
integrated, g(t' - 0) and &'U> _ 0) can be every variation in u(t) and b and subsequent
computed inEqs. 8-107 and 8-109. integration of the state equations, since x(t)
will be altered with an accompanying altera-
The second matter that requires discussion tion in t'.
is the determination of t1 and its variation,; =
0,1, ..., Tj. If the state equations form an 8-3.3 A STEEPEST DESCENT COMPUTA-
initial-value problem (all initial conditions TIONAL ALGORITHM
given) then one can make an estimate for u(t)
and b and integrate Eq. 8-93 from t toward The problem of determining 8u(t) and 6b
?" (or tn toward t if all boundary conditions which reduce J and satisfy other constraints
are given at ?"). As the integration progresses, will now be solved just as in par. 8-2. As in
Sl'it.x) can be monitored and the value of t the preceding paragraph, if some <j> [?,(?)] or
for which it is zero is called t'. The situation is <pa is less than zero, it will be ignored. If, on
not so easy in case the state equations form a the other hand, t/>a > 0 or 4>[t,u(t)] > 0,
boundary-value problem. then it will be required that
One method of determining t' requires that Wa=~aâ

a reasonable estimate of t' be available,
perhaps from engineering intuition or prelim- and
inary analysis. The state equations are then
integrated using the engineering estimates for 50p = - c<p
u{t) and b. It is likely that for the solution
where 0 < a < 1 and 0 < c < 1.
8-23
AMCP 706-192
Just as in par. 8-2, define two sets of and

indices
'A*(0
A*(t)' (8-133)
(0) (0) (0) J
,4 = {a|*a[x , ,* ]>o} ' aEA
That is, the columns of * and A* (%)are S.*"

and
and A a (t)for those CY with \pa > 0. Now,
B(t)={[<t>[tMHt)\ > 0}
i;
AjT\t)bu{t)dt (8-134)
where w(0)(/) and ft(0) are the beginning and

estimate of the design variable and design
T C'v T
parameter, respectively, and x(0)(0 is the 0*
5i// =r 5ft
ZU + V A*
+ \ K* (t)8u(t)dt .
associated solution of the state equations. J?
Further, define the column vectors of con- (8-135)
straint functions
The problem of this paragraph is now to
find bu{t) and 5ft to minimize 8/, subject to
<^ = Eqs. 8-130 and 8-131. Although the symbols
OLEA (8-128) have a slightly different origin, this problem is
precisely the same as that given by Eqs. 8-19,
and 8-20, and 8-25 of par. 8-2. All the analysis
required to determine 8u(t) and 5ft follows
00[t,u(oHt)f and Theorem 8-1 holds. The only difference is
4>(0 = (8-129) that tl in Theorem 8-1 must be interpreted as
L j3G5 . tn in the present problem.
By the argument of par. 8-2, it will be The algorithm of par. 8-2 may now be
required that given with references to equations of this
paragraph.
5ij/ = -a\J/ (8-130)

A Igor iihm:
and Step 1. Make an engineering estimate

u<-0)(t), ft(0) of the optimum de-
50(O = -c0(O- (8-131) sign function and parameter.
Using the notation of Eqs. 8-122 and Step2. Estimate t, t', .... ?", and solve
8-123, define the matrices Eq. 8-82 forx(t).
râ -1 Step 3. Adjust t' as required by the discus-

* = sion below Eq. 8-127 and recom-
oc&A J (8-132) pute x (t)'ti required.
8-24
AMCP 706-192
Step 4. Check constraints and form ^ small, terminate. Otherwise, pro-

and 0(f) of Eqs. 8-128 and 8-129. ceed to Step 13.
Step 5. Solve the differential equation, Eq. Step 13. Adjust f, t1 ,..., tn as required by
8-93, with boundary and inter- the discussion below Eq. 8-127.
mediate conditions of Eqs. 8-111, Return to Step 2 with uw, M>
8-114, and 8-116. The solutions being replaced by (1) and (1).
corresponding to the functions,
Eqs. 8-118 and 8-119, yield \J(t) For an alternate development of the al-
and X a (t) .respectively. gorithm in the special case of a full set of
initial conditions, see Refs. 5 and 7. Several
Step 6. Compute ZJ, AJ(t), * , A* (?), and example problems are solved in Ref. 5.
A*(f), in Eqs. 8-120, 8-121, 8-132,
8-133, and 8-34, respectively.
8-4 STEEPEST DESCENT PROGRAMMING
Step 7. Choose the correction factors a FOR A CLASS OF SYSTEMS DE-
and ein Eqs. 8-130 and 8-131. SCRIBED BY PARTIAL DIFFEREN-
TIAL EQUATIONS
Step 8. Compute M^j, M, ,, and M. in
Eqs. 8-38, 8-39, and 8-40. 8-4.1 THE CLASS OF PROBLEMS CON-
SIDERED
Step 9. Choose 7o > 0 and compute y and
(t) of Eqs. 8-43 and 8-35. If any Thus far, all problems considered have had
components of y, with a > r', or their state variable specified by algebraic
(t), with > q' are negative, equations or boundary-value problems in
redefine \p and <j>(t) by deleting ordinary differential equations. It is possible,
corresponding terms and return to however, that the state of the system being
Step 6. considered is governed by a boundary-value
problem with partial differential equations. In
Step. 10. Compute 5u'(0, 8u2(t), 8bl, such cases, the state and design variables are
and 8b2 of Eqs. 8-44 through functions of more than one independent
8-47. variable. One may then think of the design
variable as being distributed over an area,
Step. 11. Compute volume, or higher dimensional space. For this
reason, such problems have been described as
u^Kt)=u^Ht)--^ bul(t) distributed parameter systems.
27o
A great deal of work has been done on
+ a2o) distributed parameter systems which have a
time-like variable (Refs. 12,13); i.e., a variable
M >=*> - 8b1 + 8b2 . which makes the governing differential equa-
27o
tion hyperbolic or parabolic. In this para-
Step 12. If the constraints are satisfied and graph, consideration will be limited to static
S1^) and 8b1 are sufficiently problems such as equilibrium of plates, shells,
8-25
AMCP 706-192
etc. These problems are described by linear, |a|=a, +...+ak,\\=i + ...+k.

elliptic, partial differential equations (Ref. 1).
The object of the problem is to determine
The boundary-value problem treated will u{x), x&Sl, v{x), xEF, and b such that
be denoted
J- S g0(x,z,v,b)dr
T
L(u,b)[z] =Q(x,u,b),xen (8-136)
B{v,b)[z] = q(x,v,b),x^T (8-137) + SS Mx,z,u,b)dn (8-140)
where x = (x1,x2,-,xk)T is the independent is a minimum subject to the constraints of

variable which ranges over the domain SI in Eqs. 8-136 and 8-137,
Rk with boundary P. The vector u(x) =
[](%), ..., um(x)]T is the design variable 4>a= S ga(x,z,v,b)dr
r
over 2, v(x) = [vj(x),.... VT(X)]T is the design
variable over T (boundary design), and b =
+ SS La(x,z,u,b)dSl = 0,
(bi,..., b,) T is the design parameter. The state
variable z(x) = [zx(x), ..., zn(x)]T is to be
determined by the boundary-value problem, =!,...,/
Eqs. 8-136 and 8-137, which is linear once u, (8-141)
v, and b are specified. It is important to note, â= S ga(x,z,v,b)dT
r
however, that the problem depends in a
nonlinear way on u, v, and b.
+ 1/ I (x,z,w,)rf2< 0,
An example of the form of the differential
operatorsL(u,b)[z] andB(v,b)[z] is = /'+ 1, ...,r
(j>j(x,u) = 0, xGfi, i= 1,..., J'

L(u,b)[z] = 2 a<x,u,b)
Ii < T) (8-142)
^.(x,u)< 0, xen, i= + i;...,
aialz (8-138)
, *en and
bx*1 ...foejj*
and W/(JT,V)=0, xer, / = 1,...,?'

(8-143)
B(v,b)[z] = 2 bAx.v.b)
p
u.(x,v)< o, *er, 7=f' + i,-.,f.
IPKti
The method of solving this problem will be

3|/3|z (8-139)
x , *er similar to the methods of pars. 8-2 and 8-3.
An estimate w(0)(*), v(0)(x), and 6(0) will be
made and changes sought which reduce J,
where subject to the constraints of the problem.
Before desirable changes in u<-\ v(0>, and
a = (!,..., at)r, = 03i,...,jJt)r 6(0) may be determined, of course, their
8-26
AMCP 706-192
effect on function in the problem must be forx6I\ where

examined.
A,L(x,u,b)[z] =L(x,u,b)[z]
" U
8-4.2 EFFECT OF SMALL CHANGES IN

DESIGN VARIABLES AND
a
AbL(x,u,b)[z] = L(x,u,b [z]
PARAMETERS
(8-146)
It will be assumed in the following that the AB(x,v,b)[z] = B{x,v,b) z]
boundary-value problem, Eqs. 8-136 and av
8-137, is well-behaved in the sense that small
changes bu in (0), 5v in v(0), and 6b in b*0' A B(x,v,b)[z] =B(x,v,b)[z]
yield a new solution z(0) + Sz (where z(0) is ob
the solution corresponding to the estimated
design functions and parameters), where Sz is For convenience in the following develop-
small. ment, the arguments of L and B will always
be taken as w(0), v, and b(0).
To first order, 8z must satisfy the lin-
earized boundary-value problem. The functional J and \p are of the same
general form, so, for their analysis define
L[<>,a<>] (Sz) + AuZ,[U(),<0>] [z<>] bu

P= J g(x,z,v,b)dV + tt F(x,z,u,b)dn.
+ AhL[u(\bW][z<)]Sb r n
(8-147)
92,[x,u(0),b<0)]Su
Once the dependence of P on changes in u, v
Bu
and b is determined, the result may be applied
directly to J and \j/a.
3b (8-144)
To first order terms,
forjcfi and
z + 5v + 6b I dr
oz dv db
0 ( 0 m w r V
B[v'- \b^](bz)+At,B[v - \b ][z ]bv '/(
bF
+ A&5[v<0>,b<0)][z(0)]Sb XT 02
bz + ^ bu + *P 6b)dJ2.
3 3b 7
a? (8-148)
U,v<>,b(0>]5v
dv
In order to make use of Eq. 8-148 in the
determination of bu, Si>, and 6b, it is desirable
+ /"x/),b]5b
db to eliminate explicit dependence on bz. This
is done through use of the adjoint operator
(8-145) L * defined by
8-27
AMCP 706-192
CC{XTL(u,b)[8z} - 8zTL*(u,b)[\}} dl The objective now is to eliminate explicit

si dependence of 8P on 8z. This may be done by
requiring that
= \A[\]T C[8z]dT (8-149)
- 8z - A\\]TC[8z] (8-152;
where A [A] and C[8z] are differential opera-

be explicitly independent of 8z for all 8z
tors. The form of the operators^, C, and L*
satisfying Eq. 8-145. This may be interpreted
is determined by integrating\T L(u,b)[8z] by
as requiring that on T certain components of
parts.
Sz be determined from Eq. 8-145 in terms of
Putting
8v, 8b, and the remaining Sz. The coefficients
bFT of all components of 8z remaining in Eq.
L*(u,b)[\] (8-150)
dz 8-151 must then be set equal to zero. These
Eq. 8-148 becomes equations will then yield boundary conditions
8P
bg
8z~A[\T}C[8z] +-8v
bv
+ -& 8b\ dV
for X(x).
Assuming all this calculation has been

completed and AQc) determined, Eq. 8-151
may be written as
bF SP= f$AT(x)8udtt + f IlT(x)8v dV

+ |U \' L(u,b)[5z] + dtt-Su
&n
+ ZTSb
+ 8b \ d,
bb where
Substituting from Eq. 8-144 for L{u,b)[8z] ,
bFT bQT
this is A(x) = - -AuL(.u.b)[z]\ + -2-\
bu " by
bF
8P- \T&L(u,b)[z\
bu Tl(x) = coefficient of 8v in Eq. 8-15 1 after
n
substitution
+ \T -$-\ 8u
l"
bFT
bF -AfcL(u,ft)[z]X
+ { -\TAbL(u,b)[z] bb
+ A' dtt bb
ty
+ / [coefficient of Si1 in Eq. 8-151
A
dg
+ \l -r- 5z~A[\)TC[8z] + -&5v after substitution] dV
(8-151) (8-153)
+ 8b}dr.
bb
8-28
AMCP708192
By putting Y
a
(8-158)
F= (8-154)
g=go> fo LaGA
and where
g=ga, F = La (8-155) A = (a | i,a > 0) (8-159)
one obtains It will be required that
5\jj= -Ci4/ (8-160)

8J = SSA
JT
(X)8U dl + SnjT{x)5vdSl
n r
where Cj is a constant between zero and one.

+ / 6b (8-156)
The idea here is to drive ty toward zero if a
constraint is violated or will be violated by a
and
change in the design variables or parameters.
6â = JjAâ (x)Sw dl For convenience in later development,

a.
define
T T
+ SnK ix)Svdr+2K db
A* (x)= ( A* ; CCGA)
(8-157)
respectively
n*(x)=( n*;ae^) > (8-161)
* = ( je* ; aEA ) .
The expressions, Eqs. 8-156 and 8-157, give
the desired explicit dependence of 6J and In this notation,
8â on 5u,Sv, and 8b. The problem is now
reduced to determination of 8u, 8v, and 6b
which give the greatest reduction in J subject 8\p= ff A* (x)T8u dl
to the constraints of the problem.
8-4.3 A STEEPEST DESCENT COMPUTA- + JTT*' (x)SvdT +* 8b. (8-162)

r
TIONAL ALGORITHM
Likewise, define
The procedure will now be to choose 8u,
8v, and 8b so as to minimize 5/ subject to the
constraints Eqs. 8-141 through 8-143, just as
*(*) = (8-163)
in pars. 8-1 through 8-3. In order to insure
iED(x)_
Eq. 8-141, define
8-29
AMCP 706-192
and dP2 = f!$8uTWu8udn.
<Z>(x) = (8-164) + 5 8vTWv8vdT + 8bTWb8b

JE(x)_ r
where (8-171)
where dP is "small". The choice of dP will be
D(x)=[i\ <j>j{x)> 0] (8-165) discussed later.
and The design variables and parameter, 8u, 6v

and 8b, are now to be chosen to minimize 57
E(x) = [; | co;.(x) > 0] . (8-166) of Eq. 8-156 subject to Eqs. 8-160, 8-167,
8-168, and 8-171.
It will be required that
A multiplier rule of Liusternik and
84>(x)< - C24>(x), jcen (8-167) Sobolev, Ref. 14, page 209, will now be
applied to the present problem. It guarantees
and the existence of multipliers, n(x), xE.l, ^x)
> o, i > ', v(x), xsr, VjM >oj> f, 7, ya
8(x) < - C3(x), JcGr (8-168) > 0, a > r', X0 > 0, and 7o suchthat
where 5(5/) = 0 (8-172)
0 < C2 < 1 and 0 < C, < 1 for all bu, 8v and 8b, where
30
6$(x) = (x)u(x) (8-169)
du 57 = Jj r-X0A/r(jf)-7rA*7'(x)
and
9j
Sw (x) = - (x)8v(x) (8-170) 7o 8uT W..
ov
Before determining 8u, 8v, and 8b, a device \8udn

should be introduced to insure that these
quantities are small as is required in order that
f[-\njT{x)-yTn*T(x)
the preceding first order analysis is a good
approximation to reality. The engineer should
choose positive definite weighting matrices
yQ8vTWv
Wu(x), Wv(x), and Wb so as to associate a
relative importance to all the variables. It is
del;
then required that -vT{x) 8vdT
[-Xo #' YT^ - To 8b W,b 8b.
(8-173)
8-30
AMCP706192
Using 5/ of Eq. 8-173 inEq. 8-172 At this point it is assumed that the problem
is normal so that X0 = 1 may be chosen. Eqs.
5(8/) = 0= Cf\~\0AjT(x) 8-175 through 8-177 yield
L
si
-yTA*7(x)
-2yQ8uTWu
8u{x)= 2^-W-Hx)
T,-^ d<t> 2
- ' (x) 8 udSl
du
-AJ(x)-A+(x)y- ^~ /*(*)
du
J - \0lljT(x)
xen (8-178)
-yTn*T(x)-27o5vTWv 8v(x)
2To
3d)
82vdV
.v'ixy
dv x ["ny(x)-n^7
A,7'
"$>4> 3d)-,T
v{x) ,xer (8-179)
2y08bWh 82b ,
and
(8-174)
6b
=^yTWb1 <-fi/-*7)- (8-180)
for all 82uQc), xefi, 82v(x), xsr, and 82b.
This implies
Assume for the present that Eqs. 8-167 and
J
- \0 (x) - A* (x) -. 27o Wu 8u{x) 8-168 are equalities. Substituting Eqs. 8-178
and 8-179 into Eqs. 8-167 and 8-168 yields
30 T
(x) = 0 (8-175)
du
forxEQ, C2</>, A'Efi
- \0UJ(x) - IT* (JC)T - 2y0Wv8v(x)
3wr and
v(x) = 0 (8-176)
dv
1 3c , , r
W~l - llJ - IT* 7
forx^V, and 27o ?>V
3>r
7
-X0e -8*7-2T0W6 = 0. (8-177) c =-c,w, xer.
3v
8-31
AMCP 706-192
Since Wu 1 and Wy i are positive definite, the bcbT

matrices (90/3) W'1 O0r/a) and (3w/9v) v(x) = w: I- A^
27o 9v
W~* (dT/dv) are positive semi-definite. It
will be assumed that they are positive definite 9)
w-i
and hence nonsingular. In case 0 or t is 9v
empty, then the terms multiplying and Y do
not exist. In this case simply define /i = v - 0 x (-ny - n* 7)
and 00/d") W"1 O0r/a) = (9/9v) W~l
(9wr/3v) = 1. In any case,
A"
ov
27oC3w ,xer
M(x) =A* 27oC20+iiw;1
(8-184)
x(-A7 -A1* 7)|, xS2
rS-181
where
In order to determine y, these expressions are
A* = - W~' - 5 substituted into Eq. 8-160. Using Eq. 8-182,
9 " 9 the resulting equation is
and
9v 1 i
~TZ~ M, j - -* M, , y - C2M, -
-"-7) , *er (8-182)
where C3JW, . =-Ci^ (8-185)
r
9cZ) , 9w
A" = 3v ^ 3v
Substituting from Eqs. 8-181 and 8-182 into where

Eqs. 8-178and 8-179yields
7 A
^-Wo^1 LI " 3u ' ^=ffA 3
30 -r
30 A
W-1 }
du " /
x (-AJ -A*?)
>Vv-ir A*
30 r 6 ..
A@-'27oC20j', xES~ x
9u 37TI)n'dr
(8-183) + fi* W^'fi7 (8-186)
8-32
AMCP 706-192
*.-/>,V(/-f A" 8v(x) = -TH-6V' (X) + 8v2 (x)

^7o
(8-192)
30 _,, and
X
-^K )A*da
8b = 5ft1 +8b2 (8-193)
dT 27o
^ p
\ 3v
where
!".-'!'
" \ du du " /
+ 8* H/;1^ (8-187)
x (AJ - A*M~l Alexen (8-194)
*.- ^K'^^'i*
(8-188)
and "V du du " )
M, . = I TVi>Tw-, dT W
A *dr.
3v
(8-189)
+ C,M ^ U>
>] (8-195)
It was shown in par. 8-2 that the matrices /-i if

Q^-^-A* 0,ên
in the integrands forM, , are positive semi-
definite. Therefore, M, , is at least positive
semi-definite. It will be assumed in what
follows that M, j is positive definite and,
v
therefore, nonsingular. Solving Eq. 8-185 for \ 3v dv " )
y then yields
x (n'-n^-^êr (8-196)
L~,uv~, r - C2M^- - CjM^^)
(8-190)
dv 3 "
Substituting y from Eq. 8-190 into Eq.

8-180, Eqs. 8-183, and 8-184 yield
+ C
3^)] (8-197)
8u(x) = - - 8ul (x) + 8u2(x) (8-191) 3)r _,

^7o Cs^-'-rAw to, xer
3v
8-33
AMCP708192
J
Sbl = Wri (i -ft*M~\Mn) (8-198)
4. -du2 =~C2<t>, in
du
and
90 ,
5. m-Su1 = 0, in $2
&b>=W-b1S.*M-\(-C1+ + C2Mii$
+ C
3^> (8-199) , 9w ,
6. -6v2 =-C3w, onT
3v
It should be noted that if there were no

constraints then 5, 8v, and 8b would reduce 7. 8vl =0,onr
dv
to - A7, - IT/ , and - Sy,
2To 2>O 27O 8. -ijTSbl - SnjT8vldv
respectively. In order that the change in
design variables and parameters should be in
the negative gradient direction, it is clear that = SS AJ 8u1dSl< 0
7o > 0 is required. The magnitude of 70
could be determined by substituting Eqs.
A computational algorithm may now be
8-191 through 8-199 into Eq. 8-171. How-
stated based on this development and the
ever, dP must be chosen so it may be just as
arguments presented in par. 8-2.
well to simply choose 7o in Eqs. 8-191,8-192
and 8-193.
Algorithm
Just as in the problems of pars. 8-2 and 8-3,

Step 1. Make an engineering estimate
the variations Su'(x), 8u2(x), 8vl(x), 8v2(x),
u<>(x), v<>(x), M> of the opti-
5bl , and 5b2 satisfy Theorem, 8-2.
mum design functions and param-
eter.
Theorem 8-2: The above variations satisfy
the identities
Step 2. Solve Eqs. 8-136 and 8-137 for
z<0)(x) corresponding to u(0)(*),
1. 8blTW.5b 2
+ S 5vlW8v 2
dT v<>(x)(andM0).
D p V
Step 3. Check constraints and form \p, <j>,

+ SS 8u,Wl8u2dSl = 0 and w of Eqs. 8-158, 8-163, and
n
8-164.
2. 2*T8b2 + S IT^Sv^r
Step 4. Solve the differential equation, Eq.
8-150, subject to the boundary
+ SS A* 8u2dSl = -Clip conditions generated by Eq. 8-152
a
with g and F given by Eqs. 8-154
and 8-155, to obtain \J and A*01,
3. Q"/T5b1 + S n*T8v1dr
r respectively.
+ SS A* 8u1d. = 0
n Step 5. Compute A/(x), nJ(x), ZJ, A* (x),
8-34
AMCP708192
11* (x), and * in Eqs. 8-153 and As in par. 8-2, if after several iterations the
8-161. constraints are all satisfied, 8u2(x), 8v2(x),
and 8b2 will all be zero. In this case,
Step 6. Choose the correction factors Ct,
C2, and C3 in Eqs. 8-160, 8-167,
5/= \Mlr-MTl M-\M, ) (8-200)
and 8-168.
Step 7. Compute Mij/j, ,.^

^**.^ .^
M
*4>> and where
M, in Eqs. 8-186 through 8-189.
M//=yr^&'/+ / n/7 W"
Step 8. Choose y0 > 0 and compute y,
//(*), and >(*) in Eqs. 8-190,
8-181, and 8-182. If any compo- x
1/ av
A" w.Jn'dr
av
nents of y with a > r', //(x) with i
> I', or v(x) with./ j> ?' are
negative, re-define I/*, $(*), and .11 A'VT
1
c(x) by deleting corresponding

terms and return to Step 5.
au ou
Step 9. Compute 8u1(x), 5u2(x), 8vl(x),
8v2(x), 8bl ,and So2 in Eqs. 8-194
through 8-199. Just as in par. 8-2, one can now specify a
reasonable, desired reduction 6J in J. The
Step 10. Compute formulation, Eq. 8-200, provides a means of
finding 70 that, based on the preceding linear
approximations, will yield the desired reduc-
M<)(x) = K<0)(X) SH'OO
27o tion in the cost function.
2
+ 5M (JC)
8-5 OPTIMAL DESIGN OF AN ARTIL-
LERY RECOIL MECHANISM"
v<l>(x) =v<>(x) 8v*(x)
2To
An artillery weapon mounted on tires or
2
+ 8v (x) tracks has some undesirable features. Unlike
the hard mount (weapon rests on a base
plate), the flexible mount will have a pitch
fe(D = bW 8bl +8b2. motion. During the recoil stroke, when the
2To
weapon is fired at 75-deg elevation, the tires
load up or compress; and when counterrecoil
Step 11. If the constraints are satisfied and
begins, the tires act like a spring and unload
u'fx), 8v1(x), and 8b1 are suf-
sending the tires off the ground. It is quite
ficiently small, terminate. Other-
obvious that, when the weapon comes to rest,
wise, return to Step 2 with
u<>00, V<>(JC), and Z>(0) re- the likelihood of it being zeroed in for the
placed by u(1)(*), v(')(~)%nd *The results of this paragraph represent the work of Mr.
6(1) .respectively. T. D. Streeter, Ref. 15.
8-35
AMCP 706-192
Figure 8-1. Howitzer, Towed, 105mm, XM 164
next round has been reduced considerably, modeling for the high-speed digital computer.
especially for high rate of fire weapons. This To do this, the steepest descent, numerical
phenomenon is known as a secondary recoil technique is used to minimize the hop or
effect, because an additional acceleration pitch motion of the weapon and, at the same
term enters into the recoil equations. Because time, to determine the necessary control
of this secondary recoil effect, the recoil rod design that will minimize hop.
mechanism design becomes much more dif-
ficult. For short recoil, the orifice areas in the The recoil equation for a rigid mount is of
recoil mechanism are designed at maximum the form
elevation (75 deg). Therefore, when elevation
is mentioned throughout the remainder of x+f(x)x2 + g(x)=h(t) (8-201)
this report, it refers to maximum elevation.
The weapon positioned for high-angle fire is where x is the displacement of the recoiling
shown in Fig. 8-1. parts, g(x) is a restoring force, and h(t) is the
The purpose of this paragraph will be to *The control rod in a hydraulic recoil mechanism is a rod of
develop a systematic recoil mechanism design variable cross section which moves through a larger orifice
during recoil and varies the area of the orifice to control
procedure characterized by mathematical recoil force level.
-36
AMCP 706-192
breech force due to firing. In the second term

of this equation, the expression for the effect
of the control rod orifice areas can also be R(t)
obtained from a predetermined recoil force*,
R(t). For the flexible mount, Eq. 8-201 is
coupled with the equation describing the Time
pitch motion of the weapon and thus yielding
two second-order nonlinear ordinary differen-
Figure 8-2. Recoil Force for a
tial equations with prescribed initial condi- Rigid Mount
tions. R(t) will be taken as the control
variable which is to be determined to mini-
mize hop (the pitch motion of the weapon)
subject to other design constraints. The 8-5.1 FORMULATION OF THE PROBLEM
orifice area is then determined to provide this
recoil force. During the recoil, counterrecoil cycle, there
are four different times which are of concern.
These are shown in Fig. 8-3.
This study was performed on a develop-
mental weapon, namely, the XM164. The
XM164 is a lightweight, split-trailed towed Time
105 mm howitzer with the XM44 hydropneu-

matic recoil mechanism. Unlike a rigid mount,
the XM164 is flexible and is fired while Figure 8-3. Time Intervals
resting on rubber tires.
In Fig. 8-3, the special times noted are:

For a rigid mount weapon, the resisting
force R(t) on the recoiling parts is designed
with a trapezoidal shape as shown in Fig. 8-2.
With the proper design of the control rod firing of round
orifice area, the flow of oil in the recoil
mechanism is controlled and such a force, as t1 = end of the recoil stroke
shown in Fig. 8-2, can be obtained. However,
when a force (shaped as in Fig. 8-2) is t2 = time at which maximum hop occurs
designed for the flexible mount, the question
is asked, "Can this force be applied with some t3 = end of counterrecoil
other 'best' shape, such that it will reduce the
pitch of the weapon?" This is the basic At these four times certain conditions must
question with which this design problem is be satisfied from the design requirements. At
concerned. time t the initial conditions for the state of
the system are given. At time t1 the displace-
ment of the recoiling parts is required to be
*The recoil force is the retarding force on the rearward equal to some specified value and the velocity
traveling barrel during recoil, due to throttling of cil
through the variable area orifice. of the recoiling parts must be equal to zero.
8-37
AMCP 706-192
At time t2 the velocity of the pitch motion intermediate and terminal constraint func-
must be zero (necessary condition for maxi- tions occur; xs and x3 are the velocities of
mum pitch), and the displacement of the the recoiling parts and pitch motion, respec-
pitch motion is to be a minimum. Note that it tively; x2 and xA are the displacements of the
will be possible for t2 to vary between tl and recoiling parts and pitch motion, respectively;
t3. Therefore, the hop or pitch motion will be iî = 0 is the constraint on the displacement
minimized for the entire counterrecoil stroke. of the recoiling parts such that at the end of
At the final time t3, which is the end of the recoil stroke the displacement will be
counterrecoil, the recoiling parts must return exactly equal to ??max inches. \fr2 = 0 is the
to their original position, and the velocity of constraint demanding that the recoiling parts
the recoiling parts will be some specified value return to the latch position at the end of
V3. This is to insure that the recoiling parts counterrecoil. ^ 3 = 0 is the constraint which
return to the latch position. It will also be requires that the velocity of the recoiling
demanded that the total cycle time be equal parts come into the latch position at a
to cT seconds. velocity V3 inchespersecond.fi1 =0 defines
the time at which the end of the recoil occurs;
Formulating this problem into the mathe- Q,2 = 0 defines the times at which the pitch
matical notation of par. 8-3 yields velocity is zero and the one with the largest
displacement is selected, thus defining the
Minimize/ = x^(t2) =g0 (8-202) time at which maximum hop occurs; and fi3
= 0 defines the total cycle time to be exactly
subject to the equality constraints: equal to cT seconds.
It was previously mentioned that the rod

force was taken as the design (control) vari-
^2 = x2(t3) -no - g,=0 able, instead of the orifice areas. Using the
rod force as the design variable simplifies the
$3 = x1(t3)- V3 =g3 =0 problem, and it also gives the engineer more
(8-203) insight into the design process since he has an
nl =xi(ti) = o intuitive feel for the force levels the weapon
system he is designing can tolerate. Thus,
n2=x3(t2)=0 immediately the engineer can specify an
admissible upper limit for the recoil force, say
n3 = t3 -cr = o fimax, for his design, and this value may be
varied by the engineer for any redesign. The
with the full set of initial conditions following inequality constraint, therefore,
must hold for all time /.
Jfi(0)=jf3(0) = x4(0) = 0, 4>=R(t)-Rmn< 0 Q<t<t3 (8-205)
(8-204)
X2(0)=770 The optimization problem has now been
formulated. All that must be done now is to
where $.. i = 1,2,3 are intermediate and put the problem into the steepest descent
terminal constraint functions to be satisfied; formulation. Par. 8-5.2 simplifies the equa-
ft1, 22, and J23 define the times at which the tions of motion for the XM164 Howitzer.
8-38
AMCP 706-192
8-5.2 EQUATIONS OF MOTION FOR THE V = recoil displacement

XM164 HOWITZER
f = distance from center line of trun-
A schematic diagram showing the moving nion to center line of recoiling
parts and variables defining the dynamic parts (vertical)
model of the XM164 Howitzer is shown in
Fig. 8-4. In explanation of this physical
idealization, the following variables are de- yt = distance from center line of
fined : spade to center line of trunnion
(horizontal)
yj = weight of recoiling parts
a
7 = angle of elevation of gun tube
Wb = weight of elevating parts less
recoiling parts zt = distance from center line of
spade to center line of trunnion
yj = weight of nonelevating parts (vertical)
d
M = mass of recoiling parts </> = pitch angle of weapon
Figure 8-4. Schematic of XM 164 105mm Towed Howitzer - Dynamic Model
8-39
AMCP 706-192
R(t) = rod force application
B(t) = breech force 0 = static value of <j>
8 = acceleration due to gravity c = damping coefficient
M = coefficient of friction * = spring constant of tire
Si = friction force (guide) qi = distance from center line of trun-

nion to rear of cradle (horizon-
S2 = friction force (guide) tal)
M = mass of elevating parts less recoil- q2 = distance from center line of trun-
b
ing parts nion to front of cradle (horizon-
tal)
i)b = distance from center line of trun- =
nion to mass center of Wb (hori- f2 distance from center line of trun-
zontal) nion to R(t) application (verti-
cal)
f6 = distance from center line of trun-
nion to mass center of Wb (verti- ct = distance from center line of trun-
cal) nion to bottom of rail (vertical)
= distance from center line of trun-

Md = mass of nonelevating parts
nion to top of rail (vertical)
yd - distance from center line of
L = distance from tires to trail spades
spade to mass center of Wd
(horizontal)
Y,Z = axes fixed in trunnion, parallel to
zd = distance from center line of x-, y-axes
spade to mass center of W,
(vertical) y,z = axes fixed in carriage
/a =
traverse moment of inertia of Wa H,Z = axes fixed in cradle
about its own CG
= traverse moment of inertia of Wb The differential equations to be solved are,

about its own CG Ref. 16:
= traverse moment of inertia of Wd M

a ^ -(? ytsin,y +Z
t cos
7)^
about its own CG
(.W +y, cosy +zt sin7)02 ]
= distance from center line of trun-
nion to center line of Bit) = R(t)-B(t)-Mesm(y+<j>)
8-40
AMCP 706192
_/i(|S, \+\S2 |) sgn(77) (8-206) tively, for the XM164 Howitzer. Eqs. 8-208
and 8-209 determine the guide friction.
JAfa(77 +yt COS7 + zf siny)2 + Affi [(rjft COS7
For small 0 the following approximations
+ 2
yt-$b sm7) + 0?6 snry + z, are made:
sin0 = 0
02
+ Id ! 0 + 2A/STJ0(T? + jr COS7 + zf sin7) COS0 = 1 s-
- Ma 02 (T? + yt COS7 + z sin7)

The cos (y + 0) and the sin (y + 0) then
become
x($-yt sin7 + z, COS7) = B{t) -, - ?)
, N . cosy
cos (y + 0) = cosy 0 0 siny
+ [/?(*) - M(I ^ I + I S2 |) sgn(rj)]
x (f yt siny + z, cosy) sin (y + 0) = sin7 - 02 siri7/2 + 0 cosy .
-g\Ma(T)+y, cosy +zr siny) From these approximations and the follow-
ing definitions, Eqs. 8-206 and 8-207 can be
xcos(7 + 0) +Md(yd cos0 zrf sin@) simplified. For simplification, define:
+ CON1 = Ma
Mb [yt cos0 zf sin0 + r\b cos(7 +0)
CON2= -Ma{t; -yt siny +zf cosy)

_ ib sin(7 + 0)] ) _fc(0 + 0jf) _ c0
CON3 = -n(\Si \ + \S3 I) sgn (rj)

(8-207)
CON4 = yt cosy + z siny
+
Ma[2r/4> +(ri yt cosy +z, sin7)0
CON5 =Mb[(nb cosy+.y, - f b siny)'
- (f + zt cosy - y, sin7)02 ] + (Vb siny +z; +fs cosy)']
+ Md{yd+zd)+Ia +/,+/,
= S, +S2 -Mag cosiy+t) (8-208)
CON6 = $ yt sir>7 + z, cosy
/>' = Si(<7i -i7)+52(<?2 -t?)
CON7 = f! -?
_(r) (? -^)+R(t) (? _?2)
CON8 = siny
-/*[|S, Itf-a)
CON9 = cosy
+ I 52 | (f _ /3)] sgn(TJ) . (8-209J
CON 10= -Afa -g- CON8
Eqs. 8-206 and 8-207 are the translational
and rotational equations of motion, respec- COM 1 = Ma g CON8/2
8 41
AMCP 706-192
CON12 = Ma - CON4 CON35=H -g-f6-CON9
CON13 = -2M. CON36 = CON20 + CON22 + CON25

+ CON28 + CON30
CON14= 2Ma -CON4
CON37 = CON21 + CON23 + CON26
CON15 = -k'd> , + CON29 + CON3 1
CON16 = Mfl-CON6 CON38 = CON24 + CON27 + CON33

+ CON34 + CON35-/c
COM 7 = M -CON6-CON4
CON39 = CON15 +CON36
CON18 = -g-Ma-com
CON40 = -Ma-g-CON9
CON19 = g-Ma'CON9/2
With these definitions, Eqs. 8-206 and
CON20 = -g-M -CON4-CON9 -207 may now be written as
CON21 = g-Mfl-CON4-CON9/2 CONl-Tj + CON2-0 = R(t) -B{t)
CON22 ; '-g-Md-yd + CON3+CON10
CON23 : g-Md-ydl2 + CON11-02
CON24 g-Md-zd + Ma'r\'4>2
CON25 = -g-Mb-yt + CON12-02
CON26 = g'Mb-y,n + 0-CON4O

(8-210)
CON27 = g'Mb-zt 2
[yWa(T?+CON4) +CONS]0 =
CON28 = -g'Mb-rib-COW
CON13"fj-0'Tj + CON14'7J0
CON29 = g-Mb rlb -CON9/2
+ CON38-^+CON32-r?0
CON30 = g'Mb-$b-CON8
+ CON39-C0 +CON16-02T?
CON31 = -g'Mb -f4-CON8/2
+ CON17-02 +S(f)-CON7
CON32 = ^\Mfl-CON8
+ [R{t) +CON3] -CON6 +CON18-7?
CON33 = g-M -CON4-CON8
+ CON19-7T02 +CON37-02
CON34 = M,-g-nb-com (8-211)
8-42
AMCP 706-192
Eqs. 8-210 and 8-211 can be put into the x, =17

following form
XJ =V
"nJ? +v12<j) = v13 (8-214)
(8-212) x3==V23 )
x4 = 0
where
When this is accomplished, the following
first-order equations yield the proper formula-
v,! =C0N1 tion that will be used in the steepest-descent
scheme:
v, 2 = CON2
*i =(v13'v22 -v12-v23)
vl3 =R(.t) -B{t) +CON3+CON10
+ CON11-02 +Man4>2 x2=xt=f2 (8-215)

+ CON12-02 +CON4O-0
v2l =0 x3 =v23/v22 =f3
"2 i = Ma(r\ + CON4)2 + CON5 *4 ~ x3 =

fl
8-5.3 STEEPEST DESCENT FORMULA-

v23 =CON13-TJ0Tj + CON14-TJ0
TION
+ CON38-0+CON32-5?0
The optimal design problem can be stated
as follows: Determine the design variable
+ CON39 -ci=x2(t1)-rio+vntx=0 ^
+ CON37-02 4/2 = x2(t3)-Vo =0
iP3=xl{ti)-V} =0
Eq. 8-212 can now be written as
(8-217)
1 1
ft =xl(t ) = 0
V = (v13-v22^v12v23)/(v11-v22))
(8-213)
n2 =x3o2)=o
<t>=v23/v22 .
a3 =?3 -C-r =o
By making the following definitions, Eqs.
8-213 can be put into first order form:
8-43
AMCP708192
4>=R(t)-R
r
* ' max
< 0 (8-2 18) 3v,
'22 \ / 3v. 1
M
"iil^-^l +"22'
dx. 3*.
and satisfying
l(viivl2) i= 1,2,3,4
x=f (8-219)
ML -v12x3(CON13-x2 +CON14)
where the components of the vector f are dx 1
given in Eq. 2-215 and initial conditions are
/("nv22)
*i(0) = X3(0) = x4(0)=0, x2(0) = r)o
ML VuV22[2Mav13(x2 +CON4)
dx2
The minimization problem stated here
starts with an estimated design for R(t),
+ Mav22xl - "12 (COm3>Xlx3
analyzes it, and then improves on the design.
The steepest descent technique of par. 8-3 is
+ CON32-x4+CON16-xi
used here to solve the design problem stated.
The first step in implementing the computa- + CON18 + CON19MC4)]
tional algorithm of par. 8-3.3 is computation
of auxiliary variables required for the al- -(V13V22 ~vl2v23)
gorithm.
^[2Mavll{x2+COm)]\l{v\lv\2)
8-5.3.1 DETERMINATION OF THE AD-
ML = iv22{2Max2x3 + 2-CON12jc3)
JOINT EQUATIONS bx3
-v12 (CON13-XJX2 +CON14-X!

The adjoint equations are, from Eq. 8-93
-C+2'CON16-x3x2
"a/" 3
x=- X, 0 < t * t
dx
+ 2-CON17-*3)]/(v11vaa)
where the vectors / and x are defined in Eqs. d

8-2l5andX = (X! ,X2,X3,X4)r. lL [v22(2-CONll-x4 +CON40)
ax4
- v12 (CON38 + CON32;t2

ML /3v22\
9*! =
hUrv
("11V22)
+ 2-CON19-x2x4
v
"(Sr)-vi;
+ 2-CON37-*4)]/(yll()
9/2
dvu dxx ax2 dx3 3x4
8/3 3v2;
- (vi3y22 - v12v23) dx: bx)~V23\bx '?)>'
8-44
AMCP 706-192
i= 1,2,3,4 8-5.3.2 DETERMINATION OF THE

BOUNDARY CONDITIONS FOR
3/s THE ADJOINT EQUATIONS
-^ = x3 (CON13-x2 + CON14)/>22
dxt
Because of the intermediate constraint
functions, we must evaluate X at?2-andf1_to
v22 (CON13-X!X3 4 CON32-x4
allow for any discontinuities which may occur
9x2
across?2 and?1. Since the initial conditions
tCONl6-xl +CON 8
for the adjoint equations are given at tz, these
equations are integrated backwards on the
+ CON19-xi)
time interval shown in Fig. 8-5. Integration is
carried out by integrating from t3 to t2 + .
-2Mav23(x2 +CON4)]/vl2
Application of Eq. 8-116 provides new initial
a/a (CON13-JC!X2 +CON14-*! -C

dx3 It
i" 1 1 1 11 H
1 ,2 ,3 Time
+ 2-CON16-x3x2 + 2CON17'*3)/v22
Figure 8-5. Recoil Time Interval
= (CON38 +CON32-A:2 + 2-CON19 conditions at t2'. Integration is then per-

3JC4
formed from /*-to t". Likewise, using new
initial conditions at tl~, integration is then
x2x4 +2'CON31-x4)/v22
performed to t.
3/4
3A "0, ^0, 3/4 3/4
7T--U ^~0. It is the object of this paragraph to
ax, OJt 0JC3
2 0X4
determine the initial conditions at t3, t2~ ,and
tl~ for the four different integrations per-
formed on the adjoint equations, i.e., for i>x,
The adjoint equations now become
i//2, ^3, and/.
3^_ 3/a
Since fi3 of Eq. 8-203 does not depend
3*! 3xx
explicitly on x, dl3/dx = 0 and Eq. 8-114
reduces to
3x2 3*2 ,3_r

X (8-220) ^3-)=^3--
3x3
3/i 3/3 Thus,
3x3 3x3
\J(t3) = (0,0,0,0)r
3/3
3x4 3jcd
xâ3) = (o,o,o,o)r
(8-221)
a3) = (o,i,o,o)r
where the partial derivatives are as computed
in this paragraph. x*ja3) = (i,o,o,o)r
8-45
AMCP706192
Integration of Eq. 8-220 backward from t3 Thus, Eq. 8-222 yields

to (t2 + 0) can now be effected, using the
initial conditions of Eq. 8-221. Since Eq.
8-220 is homogeneous in X and the initial 0
conditions on \J and X* > are zero, it is clear 0 \h(t2 -0)1
iUt2 - 0) =
0
that \J(t2 + 0) = xNf2 + 0) = 0. While J_ 0
X*2 (t2 + 0) and X*3 {t2 + 0) will not be zero,
o"
they may be treated as known. There is now 0
xN'2 -0) =
adequate information to use Eq. 8-116 to 0
_0_
determine Mt2 - 0) for all four adjoint
variables. ^(/2 0) = J\*Mr2+0)
11
| V* (r2 +0)|/(/2 -0)-/2 +0)1 0 (8-223)
2
In order to find X(r - 0) in Eq. 8-116, 1 i\U2 - 0) 1
_0
choose the minus sign alternative throughout
and obtain
and
9l! \*'U2 - 0) = A*'('2 +0)

X(?2 _0)=X02 +0) +
bx2
X*- (t2 +0)[fU2 -0) IV2 +0)]\
hit2 - 0)
- \n2(t2 - 0)1 \g{t2 -o)
+ XT(t2 + 0)[/(f2 -0) It may be noted that X*' (T + 0) = 0 will

result from integration of the homogeneous
,7"
dft2 Eq. 8-220.
f(t2+0)] 2
)( Bx
(8-222) Finally, at /' , Eqs. 8-107 and 8-109 yield
Using Eqs. 8-107 and 8-109 to determine gJ(tl -0) = 0

g(t2-0) and J2O2-0), one obtains at t2 - 0
g*'1 -0)=/2(r1 -0)
gJ(t2 -0)=h(t2 -0) g*(r> _0)=0
2
gl(t -0)=0 g*(fi -0) = 0
g2(t2 -0) = 0 and
g3(f2 -0)=0 fa1 -0) =/,(* -0)
and Replacing?2 with?1 ,andx2 withx1 in Eq.

8-222 provides the proper jump conditions at
2(t2 -0) = f3(t2 -0). t1 .They are
8-46
AMCP 706-192
x-V -o) = xJ(t, + 0) where WR(t) is a weighting factor, set equal

l"
( \jTU> +0)[/(/' -0)-/(r' + 0)1 0 to one in this example.

0
[ A(f'-O)
p_
With these factors defined, one must
~r choose the magnitude of constraint error
riI
X*'(f' 0) =
0
rLAC2 -o)ji 0
0
correction to be used, in Step 7. In the
0 _0 current problem, reasonable design estimates
led to small errors, so a = c = 1 was chosen.
X*"((' -0) = X*'(f' +0)
>(8-224)
rn
( X*'V +0)[/(f' -0)-/U' 1-0) 0 For Step 8, only the following routine
AC -0) I 0
_0_
numerical integrations are required:
\"'(_t' -0) = XV,(/' +0)
x*'7"' +Q)[/g' -0)-/(f' +0)]

AC -0)
M
<fj f A*Td(t)AJ dt
where
Eq. 8-220 may now be integrated from t3
to t, withjumps at t2 and t1 defined by Eqs. 1, if 0 < 0
8-223 and 8-224. This completes computation d(t)
required by Step 5 of the algorithm of par. 0, ifd>> 0
8-3.3.
8-5.3.3 COMPUTATION OF DESIGN IM- M,, = [ A*r d(t) A* dt

PROVEMENTS Jt"
The remaining steps of the computational

algorithm of par. 8-3.3 require only routine
calculation. Some of the key formulas are
M V0
fJ *o
A* i dt.
Eq. 8-80 was used to find 70 so that a ten

highlighted here to illustrate use of the
algorithm. In Step 6, the following calcula- percent reduction in cost function is sought.
tions are effected: From Eq. 8-43 (since a = c = 1)
A'W-Â'W
and from Eq. 8-35
M(0 = -A*(0_1 Id -d(0)(A'+A*y)

A'(%)==r'(%> i= 1,2,3,
_27o0].
W-'it), if 0 > 0 At any points where p(t) < 0, delete this

A*(0 = point from the domain of 0(0 and return to
if0< 0. Step 6.
8-47
AMCP 708-192
Finally, the improved design is provided by important factor in reducing the hop; i.e., the
Step 11 of the algorithm, where entries are faster the recoiling parts accelerate during this
computed directly from Eqs. 8-44 and 8-45. period, the greater the reduction in hop. As
While these formulas are a bit messy, they are one would expect, an increase in allowable
routinely programmed matrix computations recoil length also reduces hop significantly.
which are no real challenge to the computer. An increase in the maximum rod force will
also reduce hop, for example, if the recoil
8-5.4 RESULTS AND CONCLUSIONS force is allowed to obtain the value 24160 lb
in constraint set of Eq. 8-225, the hop can be
Fig. 8-6 represents the optimal recoil force reduced an additional 0.32 in. Fig. 8-7 shows
to minimize hop at 75-deg quadrant elevation
XM164 Control Rod Design Comparison Between

25 Resent Grooves and Optimum Grooves for Minimum
HOP at 75-deg Quadrant Elevation and 115%Maximum
20 J V /
f V Pressure Breech Force
V
/ \ ^
15 ' ^ 0.14
s. U- Present Groove Design
10 \ 0.12
V
\ /
~ 0.10
< 0.08
t
=S
4V %;
^
0 10 20 30 40 50 60 70 8090 100110120130140 o 0.06 V ^
o
Time, msec (3ro ove s foi Minimum N \
0.04 iop foi R(f) < 22,000 It)
Figure 8-6. Optimal Rod Force
0.02
'V
s\
0
j ^
0 2 4 6 8 10 12 14 16 1820 22 24 26 28
with the following constraints: Recoil Length, i
Figure 8-7. Optimal Control Rod Design

R(t)< 22000 1b
(8-225)
recoil length = 28 in.
a possible variable orifice area design for short
The resulting hop for this case is 1.53 in., i.e., recoil. The orifice areas were obtained from
the tires leave the ground 1.53 in. for a 115% the recoil force in Fig. 8-6. The resulting force
maximum rated pressure breech force. If the levels from the new groove design are indicated
constraints were relaxed, such that by the dotted lines from 0.110 sec to 0.13
sec, Fig. 8-6. The recoil force is the same as
R(OQ 235001b j the optimal shaped force curve from 0 to
(8-226) 0.110 sec. The increase in hop is approxi-
recoil length = 29 in. / mately 0.1 in. The recoil length changed a
very small amount.
the resulting minimal hop is 0.88 in.
An interesting side point is that of the
speed of convergence. The nominal design
The acceleration of the recoiling parts variable R(t) used for the first iteration was
during the first portion of counterrecoil is an such that at the end of counterrecoil the
8-48
AMCP 706-192
recoiling parts were 250 in. away from the to optimize tire characteristics for the final
latch position and the required final velocity configuration in the tire itself. In order to
of 6 in./sec, was 96 in./sec. In approximately obtain optimum weapon performance for
14 iterations, convergence was obtained flexible mount systems, such information as
which seems to be very fast if one considers tire performance could be incorporated into
the complexity of the equations involved. the mathematical model and perhaps tire
characteristics could also be optimized in the
Results from firing tests show a significant environment for which they are being used.
reduction (50% or more) in hop can be
achieved simply by increasing the tire pres- The technique used here has the capability
sure. Because tire performance information is to optimize many design parameters simulta-
not presently available, it was assumed neously. If there exist other sensitive param-
throughout this analysis that the spring rate eters, consideration should be given to opti-
of the tires was constant (linear spring). mize them along with the design variable R(t).
Therefore, it is not known what results would This study clearly indicates that weapon
be obtained under a nonlinear spring model. performance can be improved by using
Tire manufacturers are investigating methods methods of optimal design.
REFERENCES
1. R. Courant and D. Hubert, Methods of Leondes, Ed., Advances in Control Sys-

Mathematical Physics Vol. 2, Inter- tems, Vol. 4, Academic Press, New York,
science, New York, 1962. 1966.
2. C. Lanczos, Linear Differential Opera- 7. D. S. Hague, "Solution of Multiple-Arc

tors, Van Nostrand, London, 1961. Problems by the Steepest-Descent
Method", in A. Laviand T. P. Vogl, Eds.,
3. M R. Hestenes, Calculus of Variations Recent Advances in Optimization Tech-
and Optimal Control Theory, John Wiley niques, Wiley, New York, 1966.
& Sons, New York, 1966.
8. B. D. Tapley, and J. M. Lewallen, "Com-
4. H. J. Kelley, "Methods of Gradients", in parison of Several Numerical Optimiza-
G. Leitmann, Ed., Optimization Tech- tion Methods",/. Opt. Theory and Appl.,
niques, Academic Press, New York, 1962. Vol. 1, 1967,pp. 1-32.
5. W. F. Denham, Steepest-Ascent Solution 9. A. P. Sage, Optimum Systems Control,

of Optimal Programming Problems, Re- Prentice-Hall, Englewood Cliffs, New-
port No. BR-2393, Raytheon Co., Bed- Jersey, 1968.
ford, Mass., April, 1963.
10. E. J. Haug Jr., R. S. Newell, and T. D.
6. R. E. Kopp, and H. G. Moyer, "Trajec- Streeter, Optimal Design of Elastic Struc-
tory Optimization Techniques", in C. T. tural Elements, Tech. Rept. SY-R1-69,
8-49
AMCP708192
Systems Analysis Directorate, Head- 14. L. A. Liusternik and V. J. Sobolev,

quarters, US Army Weapons Command, Elements of Functional Analysis, Ungar
Rock Island, Illinois, 1969. New York, 1961.
11. E. B. Lee and L. Markus, Foundations of 15. T. D. Streeter, Optimal Weapon Stability
Optimal Control Theory, Wiley, New By a Steepest-Descent Method, Tech.
York, 1967. Rept. SY-R2-69, Systems Analysis
Directorate, U.S. Army Weapons Com-
12. P. K. Wang "Control of Distributed mand, Rock Island, 111, Aug 1969..
Parameter Systems", in C. T. Leondes,
Ed., Advances in Control Systems, Vol.
16. J. W Frantz and R. H. Coberly, Design of
1, Academic Press, New York, 1964.
Control Rod for XM44 Recoil Mecha-
13. W. L. Brogan, Optimal Control Theory nism (XM1 64 Howitzer) Using Coupled
Applied to Systems Described by Partial Rotational and Translational Equations
Differential Equations, Report No. of Motion, Technical Notes, U.S. Army
65-29, Department of Engineering, Weapons Command, Rock Island, Illinois,
UCLA, Los Angeles, 1965. August 1968.
8-50
AMCP 706-192
CHAPTER 9
APPLICATION OF STEEPEST DESCENT METHODS TO

9-1 INTRODUCTION particular problem. For this reason, it is more

difficult to give a general formulation of the
The same general class of optimal structural problem into which every structural design
design problems considered in Chapter 5 is problem will fit.
treated in this chapter by using the theoretical
results and computational algorithms de- Examples of difficulties that may occur in
veloped in Chapter 8. For a discussion of the particular problems include the dependence
essential elements of the design problem tobe of boundary conditions on design parameters
considered, the reader is referred to par. 5-2. and design variables, inequality constraints for
stress which involve state and design variables
The basic difference between problems that must be transformed to functional con-
treated in this chapter and those treated in straints, and interrelationships between eigen-
Chapter 5 is in the nature of the design values and state variables at particular points
variables and the associated state variables in the structure. Each of these peculiar
that describe the response of the structure. In features will be treated as it arises in a
this chapter, distribution of the material along particular problem. This requires that the
members of the structure will be continuous designer who is using continuous optimization
as opposed to discrete, as treated in Chapter methods for structural design must under-
5. Consequently, continuous variation of stand the origin of the methods well enough
stress, displacement, and eigenfunctions along so that he can alter the computational al-
members of the structure will need to be gorithms as required to fit his particular
determined. In this sense, the present problem problems.
is infinite dimensional whereas the problem
treated in Chapter 5 was finite dimensional.
While the previous discussion might indi-
The optimal structural design problem, there-
cate that one encounters only difficulties
fore, will involve boundary value problems as
when using continuous methods as opposed
opposed to algebra problems. to discrete methods, there appears to be a
potential for more efficient computational
The continuous design problem treated in methods in the infinite dimensional problem
this chapter will have more features to ac- than in the finite dimensional formulation.
count for than did the discrete problem. For Further, it is clear that the infinite dimen-
example, there will be differential equations, sional formulation can yield a true optimum
boundary conditions, pointwise inequality while the discrete formulation of the problem
constraints, and functional constraints. will generally yield only an approximate
Further, nonclassical analysis problems may optimum. In the following paragraph, a rela-
arise which require special techniques for each tively general formulation of the infinite
9-1
AMCP70G192
dimensional optimal structural design prob- T is the boundary of that region. L and B are
lem is given, and the computational algorithm differential operators on Q, and T, respective-
based on theory of Chapter 8 is developed. In ly. The functions Q and q are generally
following paragraphs, this theory and com- related to loads.
putational algorithm are applied to example
problems. Alterations in the general theory To better fix the idea of operators, con-
are made as they are required in the solution sider the simply supported beam of Fig. 9-1.
of individual problems.
9-2 STEEPEST DESCENT METHOD FOR p(x><
In the structural design problems treated \

here, the design is to be specified by a vector
design function u{x) = [w,(x), -, um{x)]T 2
and a vector design parameter b = [&i, ...,

b ]T, where the independent variable x may Figure 9-1. Simply Supported Beam
be a real variable, a two dimensional variable
x = [x,,x2 ]T, or a three dimensional variable
x = [xi, x2, x3]T, depending on whether The boundary value problem in this case is
material is to be distributed over a line, a simply
surface, or a volume. In addition to the design
variables u(x) and b, there will be state d2z
L (u)z = EI(U)-TY= -M(x) = Q (9-3)
variables z(x) = lzi(x),..., zn(x)f represent- dx
ing stress and displacement under load and
y(x) = [y\ (x),-> yq(x)]T representing mode and
shapes for vibration or buckling.
~z(.oy
For the purposes of convenience in nota- Bz (9-4)
tion and generality, the system equations will z(l)
be written in operator notation similar to that
used in par. 8-4. Only linear behavior will be Here, u(x) is a variable which uniquely speci-
considered so that stress and deflection are fied the beam cross section and determines
determined by the linear boundary-value the design, E is Young's modulus, z(x) is
problem deflection, and M(x) is the bending moment
that is computed from the distributed load
L(u,b)z = Q(u,b), xEtt (9-1) P(x).
and In this example, 2 is just the interval <0,1>

and T consists of the two endpoints x = 0
Bz=q, xer (9-2) and x = 1. The advantage in the notation of
Eqs. 9-1 and 9-2 is that it is convenient and at
In this notation, fi is the region over which the same time applied to a large class of
the material of the structure is distributed and problems. For an example of a problem in
9-2
AMCP 708-192
which 2 is a subset of two dimensional space, Constraints of the form i}(x,z,u,b)< 0 forxS2
see par. 9-7. will have to be reduced to functional con-
straints as in Eq. 8-6.
In addition to response due to static load,
it is necessary to treat the response of a Beginning with an engineering estimate of
structure to dynamic loads. One important the design variables u(x) and b, Eqs. 9-3
characteristic of a structure, which is classi- through 9-6 may be solved for z(x),y{x), and
fied as dynamic response, is natural fre- f. Aperturbation, {bu, 5b), in the design leads
quency. Another response, which must be to perturbations in the cost and constraint
treated, is buckling. Both buckling loads and functions
natural frequencies are determined, in the
present class of problems, by linear eigenvalue
problems. Again using operator notation,
5/ =
db aj- f& bz
these problems may be written in the form

9/i 9/i
+ 5u + 6b) dQ, (9-10)
K(u,b)y = $M(u,b)y, xSSl (9-5) du db
and 90.
60; = 1#- 5u, xGfi, i= !,...,/ (9-11)
Cy = 0, xer (9-6)
de, de..
de,. r nog,
where f is the eigenvalue and y(x) the 5.//,
v =_L hb+L 5f + U 8z
associated eigenfunction or mode shape. The > db 3f J J\fa
operators K andM generally relate to stiffness
and mass, respectively. In conservative prob-
lems they will be symmetric (Ref. ^(formal- + Êt 6U +^L 5b) dl,
bu db J
ly self-adjoint) but in nonconservative sys-
tems, they will not be symmetric. The more /= l,...,s (9-12)
general unsynimetric case is treated here.
The object, as in preceding work, is to
The optimal design problem is that of eliminate explicit dependencies on 5z and 5f.
minimizing First, the perturbation equation for 8z is
J = fo(b.n+ Sff1(x,z,u,b)dto (9-7)

n L(u,b)5z + [L(u,b)z]bu
du
subject to the pointwise constraints a dQ
+ [L(u,b)z]6b= -2-
db du
(9-13)
0,.(x,w) < 0, xGft, i= 1, ...,/ (9-8)
+ ab, xE S ~
and the functional constraints db
Bbz = 0, x on T .
fy = /(*.f) + / / gj(.x,z,u,b)dQ. < 0,
n In certain problems, the boundary con-
/=!,. (9-9) ditions may depend on the design parameter
9-3
AMCP 706-192
b. In that case, the perturbed boundary and

condition is
de,.
B8z + [B(b)z] 8b = 0
ob
r^-'t+ff \+j
instead of Eq. 9-13. In this case, methods of

par. 8-4 must be applied to particular prob- a<2 a
lems.
To eliminate 8z, integration of

8udQ.
/ / \TL5z dSl by parts yields operators L* in du _
ft"and*onr suchthat
( 3c,
\ db db
/ / [\TL8z -8zTL*\]dl = 0 (9-14) n
n
+ \*l
for all 8z and X satisfying B8z = 0 and B*\ = db
OonT.
Solving - -g- [L(u,b)z] dn\6b (9-18)
L*\ j_V>'' , xen

dz It remains only to eliminate explicit depen-
(9-15)
J dence on 5f. Under very restrictive hypoth-
B*X = o, xer eses, Kato (Ref. 2) has obtained a relationship
and among 5f, 8b, and 5. This relationship is
derived here formally. It is assumed that 5f
bz and 8y depend continuously on 8b and 5M
(9-16) and further that the following perturbation
formula holds:
and substituting into Eqs. 9-10 and 9-12,

a
using Eq. 9-13, one obtains K{u,b)8y + [K(u,b)y] 5"
ou
6/= % B ff\
of i+ n \ o + lK(u,b)y]8b
do
- - [L(u,b)z] +-ii- 8udl = 8$M(u,b)y + $ [M(u,b)y]8u

o / du _ du
[9/, + f-rr [M(u,b)y]8b + M(u.b)8y

ab
fr//
(9-19)
r/3 Just as in Eq. 9-14, integration by parts
+ X"
3ft may be used to obtain the operators K*
fW* and C* adjoint to K - $M and C. These
[L(u,b)z] dl 8b (9-1 V
ab operators are defined by the relation
9-4
AMCP 706192
// wT(Ky - $My)da = / / yT(k*w be emphasized that Eq. 9-21 is obtained by

n n strictly formal calculations. There are deep
mathematical questions concerning the gen-
- $Mw)dSl eral validity of this result. For a treatment of
the subject of perturbation theory in linear
for all w and y satisfying C*w = 0 and Cy operators, the reader is referred to Ref. 2.
= 0 onf. Let j satisfy
An expression for 6f in terms of 6 and 8b
K*y = fM*7, x Eft may not be substituted into Eqs. 9-17 and
(9-20) 9-18. Defining
m
C*jF = 0, JCGT
ae7
Premultiplying Eq. 9-19 by TlT ,
y integrating, db db
and rearranging terms yields
[L{ujb)z\ \J
T
ffy My<m\b$ =
a '
ff[yTK8y - $yTM8y] dQ, I/// y1 My da
+ II I > >yT^ [K(u,b)y]

ou
* i'-^ [K(u,b)y\\ y
?( [M{u,b)y]) y \\da
-$yT [M(u,b)y] \ bu da
(9-22)
W yT-^ [K(u,b)y]
A'/ = Ml
du
+
dQ1
du
tV1 [M(u,b)y]) da\ 8b lL(u.b)z) \J

db du
(9-21)
yTMyda
The first term on the right is
// [yTK8y - $yTM8v] da
x [i {Kiu> b)y] y
= ff [8yT(K*y - $M*y)] da = Q
n
S [ [M(u,b)y] ) y
by definition of the adjoint operator. It must (9-23)
9-5
AMCP706192
'do? Define vector constraint functions <j>(x)

M J
{\ to _ to containing those constraint functions 0.(x) >
0 and \jj containing constraint functional 41/
> 0. Define
-([.*]) _
A0(x) 5|iZ. | for all 2 with 0.(x) > oj
yTMydSl (9-28)
S*' | for all] with \pj > 0 (9-29)

b)y] y
and
-r( [M(u,b)y]) y f dl A*W! A *,' | for all/'with i//;. > 0
(9-24) (9-30)
and Defining A<j>(x) and Ai// to be desired
7- reduction in constraint error, the linearized
ae problem is to choose 6u(x) and Sb to mini-
3w 9
mize
[/.(a.*)r] x*/ SJ = ZjTdb+ Jf AjT8u dSl (9-31)
ij fJyTMydSl subject to
Z.T
A* 6 - A(j>< 0, forxJ2 (9-32)
[[K(u,b)y]) y
fi*r5fc+ //A*"r6u c/n-A^<0
LL [M(u,b)y]J y (9-33)
(9-25) and
Eqs. 9-17 and 9-18 become 8bTWbSb + // 8uTWu(x)5u dSl

n
5/ = ijT5b + f f AjTu dn (9-26)
-? s 0 (9-34)
and The weighting matrices Wb and Wu(x) are

chosen positive definite and is small. This
5 = Z4>iT8b + // A*'5u dft (9-27) optimization problem for 6w and 6b coincides
n with the problems of pars. 8-2 and 8-3 for
9-6
AMCP 706192
ordinary differential equations and with the

M^^ = 9* TW-1Z* + // A* TW-
problem of par. 8-4 for partial differential
equations.
Determination of a solution proceeds du t" du "

exactly as in Chapter 8 so only the resulting
computational algorithm is given here. and
A;r dV M_,
Steepest Descent Algorithm for Optimal M i><t>
-I! " du
MT} *4>dto
"">
Structural Design
Step 1. Make an engineering estimate of

Step 7. Choose stepsize 70 > 0 and evalu-
the solutionu(0)(x) and b(0).
ate
Step 2. For; = 0,1, ..., solve Eqs. 9-1, 9-2,
9-5, and 9-6 forz(", y(i), and ?(/). T = -JlC*[27o(A*+^)
</"/-
Step 3. Check the constraints of Eqs. 9-8 + M 0/J

and 9-9 and form <$> and \p contain-
ing the constraints not strictly
satisfied. and
Step 4. Solve the boundary value prob-

H(x) = -M7*[2y 0A
lems, Eqs. 9-15 and 9-16, for \J 44
and X1*/ corresponding to ty, > 0.
Solve the eigenvalue problem, Eq. + -^W~1(AJ + A*y)] .
ou "
9-20, for y.
If any components of y or (x) are
Step 5. Choose the corrections in con-
negative, delete the corresponding
straint errors A0 and A^.
components of \p and <$>, respective-
ly, and return to Step 5. Other-
Step 6. Evaluate HJ, AJ, $*>, and A*> in
wise, continue.
Eqs. 9-22 through 9-25 and com-
pute
Step 8. Compute
30 . 30r
Sul(x)=Wl7l [I ^~ M-}
du <*"
T x T 1
M.>jij
r = ^ W- & + jJA'" W-
-W
du "
u
\ du ">> du )
9-7
706-192
303 the method of steepest descent, to illustrate

5u2ix"i -m fr- du
M:
*4> the direct application of this technique to
optimal structural design. The mathematical
**- formulation of the problem is given in par.
du " , 7-2 and will be used here with a change in
notation to be consistent with par. 9-2.
*lA*At;l(A4,+M^)]
Fixing cross-sectional geometry and allow-
Sb^W-'^-^Mr'M.j) ing cross-sectional area at to vary, as in par.
7-2, yields
and
= au2 (x) (9-35)
b^W^WM^M+M,,)
The optimization problem is to choose up^; 0
and form <a x < I to minimize
w+1)(x) = w(;c) 5M '(*) J=f (9-36)

27o J o WflS
+ 5w2(x)
^=P0 -P< 0 (9-37)

and
<p=P/u(x) -^_ a 0 (9-38)
2To
Step 9. If all constraints are satisfied and and

1 1
5M (X) and 8b are sufficiently
small, terminate. Otherwise, return d2y 1
to Step 2. K(u)y = 4
2 = - P ^
2 y s PM(u)y
dx Eau
The results of Theorem 8-2 hold for the (9-39)

optimal perturbations, and it may be shown
that a necessary condition for convergence to >(0)
a local optimum is &ul(x) and Sb' approach Cz =
dy
zero. Discussions of Chapter 8 on use of the
dx (9-40)
algorithm apply. They will not be repeated
here.
where the coordinate system is as shown in
9-3 A MINIMUM WEIGHT COLUMN Fig. 7-1 withy replacing x and x replacing /.
A minimum weight column problem has In its present form, the boundary-value
been solved in pars. 5-4 and 7-2 to illustrate problem is just as in Eqs. 9-5 and 9-6 and is
the use of two optimization techniques. The self-adjoint, so y = y in par. 9-2. For use in
same problem is solved in this paragraph, by the steepest-descent algorithm, Eq. 9-23 is
9-8
AMCP706192
AJ = 1 (9-4D Specifically, the problem considered here is

the determination of the distribution of ma-
and Eq. 9-25 is terial along the centerline of a simply sup-
ported beam (see Fig. 9-2) so that the beam
y2(s) will be as light as possible and still have its
A* =-
3yHx)
3
EOLU (X)
/f aw2(x)
dx
(9-42)
fundamental frequency at least as large as a
predetermined frequency o>0. Further, so that
the beam can support a minimum level of
bending moment, it is required that the
The computational algorithm of par. 9-2 second moment of its cross-sectional area
now applies directly to the present problem. shall always be at least as large as a positive
The solution of the eigenvalue problem, Eqs. constant I0.
9-39 and 9-40, was obtained using the finite
element analysis technique outlined in par.
/77T7
5-4. The solution of this continuous problem ~&%h
required approximately the same time per T w
iteration as the discrete technique but fewer
iterations were generally required for con-
Figure 9-2. Simply Supported Vibrating Beam
vergence. Exactly the same results as given in
par. 7-2 were obtained. The reader is referred
As in the column problem of the preceding
to that paragraph for a tabulation of results.
paragraph and par. 7-2, the geometry of the
cross section is fixed and all dimensions are
9-4 A MINIMUM WEIGHT VIBRATING
allowed to vary proportionally. If the area is
BEAM denoted u(t), then
As was pointed out in par. 7-2, the paper I(t)=o.u2(t) (9-43)

by Keller (Ref 11, Chapter 7) in 1960
presented a mathematically elegant method of where a is the minimum second moment of a
designing the minimum weight column. The cross section with the given geometry and
same method was applied by Niordson (Ref. unit area.
12, Chapter 7) in 1965 to find the simply
supported beam of maximum natural fre- Since the material is to be specified with
quency for a given volume of material in the constant density, minimum weight is equiva-
beam. This method of solution resulted in a lent to minimum volume. The quantity to be
horribly nonlinear differential equation with minimized is, therefore,
serious singularities. While a solution was
obtained for the vibrating beam problem, it is =jL u(t)dt (9-44)
doubtful that the method could be extended
for the solution of multimember structural The constraint on I(t) discussed previously
design problems. The methods of Chapter 8, can now be written as
on the other hand, are quite general and will
be used in this paragraph to routinely solve a 0 = /o -<xu2(t) : 0 (9-45)
minimum weight beam design problem with
constraints on natural frequency. where I0 > 0 is given.
9-9
AMCP 706192
The remaining feature of the problem to be with boundary conditions

accounted for is the constraint on natural
frequency. If the beam withau2(t) = I0 has a yt(0) = y,(L) = 0
fundamental natural frequency of co0 or (9-49)
higher, then this is clearly the optimum beam. y2(0)=y2a) = o
On the other hand, if this beam has a natural
frequency below o>0, then there must be The boundary-value problem, Eqs. 9-48
points along the beam for which au2(t) > I0 and 9-49, is self-adjoint so y = y in par. 9-2.
and a meaningful design problem exists. The The optimal design problem is well-defined
inequality on natural frequency is and the notation of par. 9-2 applies directly.
From Eq. 9-23,
w< 0 (9-46)
AJ = 1
There are several ways in which the natural

and from Eq. 9-25,
frequency of vibration of a beam may be
related to the design of the beam u{t). The
iy\
relationship chosen here is the boundary-value A* =- uy\dx
Eau3 &
problem describing lateral displacement
during oscillation. It is given in Ref. 3 as
The computational steepest-descent algorithm
may now be implemented in a direct manner.
d2w
Eau2 :
poj uw
dx2 \ dx2 As a numerical example, the given problem
was solved with the data E = 3 x 107 psi, L =
(9-47)
w(0) = w(L) = 0 10 in., a = 1.0, and p = 0.00208 slug/in.3 The
eigenvalue problem was solved through use of
w"(0)=w"(I)=0 a finite element structural analysis program.
Even though there was no attempt at making
the computational routines efficient, only 7
where prime denotes differentiation with sec per iteration on an IBM 360-65 Computer
respect to x. were required. For most natural frequencies,
10 to 15 iterations were sufficient for con-
In order to put the boundary-value prob- vergence to within numerical accuracy of the
lem, Eq. 9-47, into the form Eqs. 9-1 and 9-2, computations. Results for a range of natural
define jvj = w, y2 = Eau2(d2yl/dx2), and frequencies are given in Table 9-1. The general
pu)2 = f. The problem, Eq. 9-47, is then shapes of profiles of several of the optimum
beams are shown in Fig. 9-3 to illustrate the
~d2Y2 optimum distribution of material.
dx2
Ly = = {My 9-5 A MINIMUM WEIGHT VIBRATING
d2yr FRAME
uy.
dx2 Eau2
The distribution of material along members
(9-48) of the frame shown in Fig. 9-4 is to be
9-10
AMCP 706-192
TABLE 9-1
L,
COMPARISON OF OPTIMAL BEAMS
Volume of Weight
2
r
Uniform Reduction
Beam of Optimal or
Frequency, Length Volume, Material 1 3
rad/sec 10in? in3 Savings,
%
X
3600 0.9353 0.8967 4.13 t/////

4000 1.1546 1.0583 8.34
4400 1.3971 1.2536 10.27
Figure 9-4. Portal Frame
4800 1.6627 1.4740 11.35
5200 1.9514 1.7189 11.92
5600 2.2631 1.9847 12.30
6000 2.5980 2.2705 12.61 frequency greater than or equal to a given w0.
10000 7.2165 6.3172 12.46 Further, as a form of strength requirement
I(x) > J0 > 0 is required.
*Uniform beam of lowest volume having required natural
frequency
For convenience, all members have the
same length and all cross sections have the
determined so that the frame is as lightweight
same geometry but may be scaled by a factor
as possible and has a fundamental natural
that varies with x. In this case, the area of
cross sections u.(x), i = 1,2,3, uniquely
determine the design of the beams when the
3,600 rad/sec
beam material is chosen. Further, the second
moment of the cross-sectional areas are
Ifo) =a. u2t(.x)
where a- is a constant depending on the

cross-sectional geometry chosen.
Defining
y2=EIiw'l
j>4 =EI2w'2'
ys =w3
Figure 9-3. Profile of Optimal Beam y6 = EI3W3
9-11
AMCP 706-192
the differential equations for vibration of the The perturbation boundary conditions
frame are from Eq. 9-51 are
y'i "1^1 5^,(0) = 0 *,() = -6j>s(0)
/.'-
lky2 0 8y\(0) = 0 8y'1(L)=8y'3(0)
y'i u2y3 8y3(.Q) = 0 5/3(I) = S/(0)

Ky = = pw2 ^My
y'i- 0
TT2n 8y3(L) = 0 8y2(L) = 8y4(0)
y'l "3^5 5y5(.L) = 0 8y,(L) = 8y6(.L)

l > (9-52)
y's-
EI3H]
L oj 8/ (2) = 0
8y'2(L) + 8y'6(0)
(9-50)
where f = pco2. Boundary conditions are
>i(0) = 0 7i(i) = -3'5(0) I

H
'0
8u2(x)dx yi(L)
/,(0) = 0 y[(L) = y'3(Q) (x)dx 8yi(L)
y3(0) = 0 y'3(L)=y'5(0) Two integrations by parts and elimination of

boundary terms through use of Eqs. 9-51 and
y3(L) = 0 y2(L)=y4(0) 9-52 yield
>(9-51)
ys(L) = 0 y*(L)=y6(L)
I 8yTKy dx = I yTK8y dx
Jo Jo
y'2(L)+y'6(0)>
H. 8u2{x)dx yx(L)
-
M ?| u2(x)dx ViiL)
Since the boundary value problem, Eqs.

(9-53)
The boundary-value problem, Eqs. 9-50

and 9-51, is written in self-adjoint form. The 9-50 and 9-51, is self-adjoint, y = y in the
last boundary condition in Eq. 9-51 is just general formulation. The derivation of Eq.
Newton's second law applied to horizontal 9-21 holds and Eq. 9-53 may be substituted
motion of Member 2. This boundary con- along with
dition does not fit Eq. 9-6 exactly due to
dependence on the design variable u2(x). It fL L
T
will have to be treated as a special case I y M8y dx = I 8yTMy dx
according to the comment following Eq. 9-13. Jo Jo
9-12
AMCP 706-192
to obtain This is precisely the form of Eq. 9-27, and the

remainder of the general derivation of the
steepest descent algorithm is valid. It should
I yTMy dx 5f =
be noted that this derivation is formal, and
rigorous verification of Eq. 9-54 is expected
to be extremely difficult.
[Ky-SMy]dx
Jo
The algorithm of par. 9-2 was used to solve
this problem in a direct manner. The eigen-
value problem was solved approximately by a
5u2(.x)dx ViiL) finite element method. Data for this problem
are a = 0.07958, p = 2.616xl04 lb-sec2/in.4,
E = 10.3xl06 psi,/0 = 0.009825 in.4, andi =
10.0 in. Weights of optimum frames are given
i 2y22
Eoi1uj
5! in Table 9-2 for several frequency require-
ments, and the profile of an optimum frame is
shown in Fig. 9-5.
6wo
:;J
TABLE 9-2
5, I t/jc
WEIGHT OF OPTIMUM FRAMES
Solving for Sf, C00, rad/sec 2000 3000 4000 5000
Optimum
sr: fc !
Weight, lb 1.73 2.56 3.59 4.69
3^5 )rf*.
2yl
-'O
8u7
Lâj! + ryi(i-)
2^1 Su3 \ dx
a*ui
(9-54)
By making the obvious choice for Af, Eq.

9-54 can be written
> >v ;;v
J:
1 T
5? = l' Af 6W dx (9-55) Figure 9-5. Profile of Optimum Frame
9-13
AMCP 706-192
9-6 A MINIMUM WEIGHT FRAME WITH TV -+T+\

MULTIPLE FAILURE CRITERIA
In most structural design problems, con-

straints of several kinds must be treated
simultaneously. In the present problem, con-
straints on member size, deflection, and
Figure 9-7. Free Bodies V-
buckling are enforced in the minimum weight
design of the frame shown in Fig. 9-6. Area is
the third member moves w (L) units to the

right.. Thus, by elementary beam theory (Ref.
3)
3niL)ET3
T=- (9-56)
The differential equation of deformation of

the first member is
[Ea,u21(x)w"]" = -q{x)
Figure 9-6. Laterally Loaded Frame
and the boundary conditions are
w(0) = 0
allowed to vary along the length of the first
and second members but the geometrical w'(0) = 0
shape of the cross section is fixed. Thus, ' (9-58)
w"(L)=0
Ii(x) = a.uf(x)
3EI3w(L)
- (Ea,uiw")' (L)= ;
i>3
where a; depends on the cross-sectional geom-
etry and Ufo) is the cross-sectional area of the
ith member at the point x. The size of the To get the boundary-value problem, Eqs.
third member is fixed and all members are 9-57 and 9-58, into the form of Eq. 9-1,
taken the same length. define
Free body diagrams of the members are z, =w

shown in Fig. 9-7. (9-59)
Z2 = <*! Hi w" = EOLy U\ z'[
The third member is uniform with constant
modulus EI3. Further, axial deformation of The boundary-value problem, Eqs. 9-57 and
the second member is neglected so the top of 9-58, is then
9-14
AMCP 706-192
z2 and
~
Lz- aiUf(x)+I0 < 0 /= 1,2 (9-67)
j" - [l/Ealu21]z2_
The constraint, Eq. 9-65, requires that the
(9-60)
'(*) axial load 7' in the second member be less
=e than or equal the buckling load. A limit S is
0 placed on horizontal deflection of the top of
the frame in Eq. 9-66. The constraints, Eq.
9-67, are included to insure that member
and cross section does not go to zero anywhere. A
more realistic constraint would be on bending
MO) "0~ stress but this would require a constraint of
the form of Eq. 8-6. This constraint will be
zi(0) 0 included in subsequent work but will not be
Bz = treated here.
22 (L) 0
3/3z,(Z,) The constraints, Eqs. 9-65 and 9-66, do not

z'2{L) 3 J3_ fit directly into the basic formulation of this
J
(9-61) text and require special treatment. The lin-
The equations which determine buckling earized forms of these constraints are
load P of the second member are 3 Eh
Szl(L)-SP< A$ (9-68)
Ay =/' = /> \y=PMy (9-62) and

Eqc2u2
and - 621 (I) S. A\p2 (9-69)
>(0)"
It remains to obtain expressions for5zi (L)
Q- (9-63)
and 6P explicitly in terms of du.(x). From Eq.
9-21, SP may be expressed in terms of 5w2 (x)-
The objective in the design problem is to

In order to obtain an expression for
choose i(*) and W2C*) to minimize the
Szi(L), a Green's identity similar to Eq. 9-14
weight of the first two members,
is needed. Integration twice by parts of
J = y\
i
'0
[ul{x) + u2(x)]dx (9-64)
/
fL
X7L8z dx yields
T
-L
The constraints to be enforced are j \ L8z dx= j( SzTLX dx
\T
-'0 Jn
*i =- (Mb. z1(L)-P< 0 (9-65) = (X] 8z'2 - Xi5z2

\ L3
+ \28z[ -X252,)
<p2=-Zi(L)-S< 0 (9-66)
9-15
AMCP706192
Choosing X such that IX = 0, substituting for Eqs. 9-1 and 9-5, are formally self-adjoint so
LSz from Eq. 9-13, and using the linearized L* = L, K* = K, and M* - M in the general
boundary conditions of Eq. 9-61, this be- theory. Since the boundary-value problem,
comes Eqs. 9-62 and 9-63, is self-adjoint, y = y and
r
the computation required in Step 2 of the
2Zn steepest descent algorithm is considerably
buidx =
ECLyU\ reduced.
This problem is now easily put in the form

-X.CD^*. (I) of the problem of par. 9-2. It was solved by
direct application of the algorithm of that
-X,(0)5z2(0) paragraph with the data S= 4 in., L = 100 in.,
E = 3.0xlO7 psi, a = 0,07958,10 = 0.0147
+ X',(0)Sz2(0) in?, and area of member 3 is 4.0 in? The
volume of the optimum frame for several
+ X2(Z)5z,(Z) values of q is given in Table 9-3. The profile
of an optimum frame is shown in Fig. 9-8.
-\2(L)SzdL)
(9-70)
TABLE 9-3
The adjoint variable is chosen to satisfy L\ VOLUME OF OPTIMUM FRAME

= 0 but no boundary conditions have yet been
specified. Choosing q, lb/in. 10 15 20 25
Optimum
Volume, 502.1 531.1 556.0 653.S
2El \l(L)-\2(L) = 1 in3
(9-71)
\1(0)=-K\(0)=\2(L)=0
9-7 A MINIMUM WEIGHT VIBRATING
the identity, Eq. 9-70, becomes
PLATE
f f 2z2(jc)X2(x)
5i(x)rfx In order to illustrate the use of the
steepest descent method in higher dimen-
(9-72) sional problems, a minimum weight vibrating
plate problem will be solved. A rectangular
Thus an explicit relationship between 5u and plate, Fig. 9-9, is specified by its thickness
Sz1 (Z) has been found and may be substi- function h(xx, x2) over the plate. The object
tuted directly into the linear constraints, Eqs. here is to choose h{xx, x2) such that the
9-68 and 9-69. With proper choice of nota- weight of the plate is as small as possible
tion, these inequalities fit into the form of subject to the constraint that the natural
Eq. 9-33. frequency of lateral vibration is greater than
or equal to a given frequency to. Further, due
In this problem the differential equations, to applied loads, a constraint of the form is
9-16
AMCP 706-192
3T57 w ,,
-*-*2
>>>>>/ / // / f
Figure 9-8. Profile of Minimum Weight Frame
Figure 9-9. Simply Supported Plate

h0(xu x2) -h{xu x2) < 0 (9-73)
enforced. In the present problem, h0(xlt x2) yi(a, x2) = 0,^,(0, x2) = 0
is taken as a constant h,.
y*(a, *2) = 0,y4(0, x2) = 0
Denoting bending moments (Ref. 4) by yx (9-75)
=
MX, y2 ~ M and y3 = Af and the lateral y2(x1,0) = 0,y2(x1,b) = 0
displacement by y4 = w, the equations
governing lateral vibration may be written y4(x1,0) = 0,y4(x1,b) = 0
(Ref. 4) in self-adjoint form as
3^4 12 The coordinate system and simply supported

-Oi - w )
dx\ Eh boundary of the plate are shown in Fig. 9-9.
<>2yt 12
-0,2 -î) To complete the formulation of this $rob-
ax: Eh3
K(h)y -;
lem in terms of the preceding theory, a cost
=f = JAW0.V
V. "6(1 + M) function is defined by
y
Sx1dX1 Eh3
32y, d2y2 d2y,

hy.
dX* ix\ bxl bx 2 J= I I jh(x1? x^dx^dxi (9-76)
ro So
(9-74)
where y is weight density of the plate
where ? = puy2 and w is the natural frequency material. The strength constraint in this prob-
of vibration of the plate. The boundary-value lem is taken as Eq. 9-73 and the eigenvalue
problem for the simply supported plate with constraint is
differential equations, Eqs. 9-74, is self-
adjoint with the boundary conditions pwl - f < 0 (9-77)
9-17
AMCP 708192
This problem is now in the form of the (Ref. 5). Data for this problem are a = b = 5.0
general problem of par. 9-2. The domain 1 in in., E = 3.0xl07 psi, p =7.43xl(T4 lb-sec2/
this case is simply the rectangular region of in.4, v = 0.30, and co0 = 1375 rad/sec. A
the plate that is of dimension two. Further, uniform plate withf = pco2 = 1400 was taken
since the boundary-value problem, Eqs. 9-74 as the initial estimate. The volume of the
and 9-75, is self-adjoint,^ = y in the theory of optimum plate is 10.71 in? Double sym-
par. 9-2. metry of the optimum plate was observed
about the axes through (a/2,6/2). One quarter
Using the definition of K and M, <p h0 of the optimum plate with contour lines is
h{xux2) in Eq. 9-73, and e;.(f) = poo2, -f in
i shown in Fig. 9-10.
Eq. 9-77, from Eq. 9-23,
A/=7 (9-78)
and from Eq. 9-25
A '
p6(yj 2w1y2+y21)+im+IJ)yl
Eh"
-H/n
a
hy\dxxdx2
(9-79)
This problem is now in the form of the

general problem of par. 9-2. It was solved by
direct application of the algorithm of that
paragraph. The eigenvalue and eigenfunction
for the variable thickness plate were deter-
mined approximately by the Ritz technique Figure 9-10. Contours o f Optimum Plate
REFERENCES
1. C. Lanczos, Linear Differential Operators, Mechanics of Materials.

Van No strand, London, 1961.
4. C. T. Wang, Applied Elasticity, McGraw-
2. T. Kato, Perturbation Theory for Linear Hill, New York, 1953.
Operators, Springer-Verlay, New York,
New York, 1966. 5. S. G. Mikhlin, Variational Methods in
Mathematical Physics, MacMillan, New
3. A. Higdon, E. H. Ohlsen, and W. B. Stiles, York, 1964.
9-18
AMCP 706-192
APPENDIX A
CONVEXITY
Convex functions and sets as defined in

Chapter 2 play an important role in optimiza-
tion theory. It is generally possible to obtain
much more comprehensive results in non-
linear programming problems that are convex
than in the nonconvex case. Some of the
more important results due to convexity are
given in Chapter 4, par. 4-2. (B)
In order that this appendix is self-

Figure A-1. Examples; Convex Case and
contained, definitions of general convex sets Nonconvex Case
and functions will be repeated here. For a
complete treatment of convexity, the reader
is referred to Ref. 1. and y. This is precisely the property which
characterizes convex functions. Analytically,
Definition A-I: LetD be a subset ofR". D this property is expressed by the inequality
is called a convex set if for any points x andy
in D, x + d(y x) is also inZ) for all 0 such f[z + 6(y-z)} <m + 6[f(y)-f(z)]
that0< 0 < 1.
for all 0 with0< 0 <; 1.
The collection of points x + 6{y x), 0 <
0 < 1, is just a straight line fromx toy. Def The same idea holds in R" where convex
A-l just says, then, that a set inRn is convex functions are characterized by
if the straight line joining any pair of points in
the set lies entirely in the set. For example, in Definition A-2: Let the real valued func-
R2 (the plane) the set of points inside the tion f(x) be defined on the convex subset D
unit circle is convex (see Fig. A-l (A)) whereas
the star-shaped region in Fig. A-1(B) is not
convex.
f(x)=x'
Convex functions have as their prototype

f(x) = x1 in R' . The graph of this function is
shown in Fig. A-2. Note in this figure that if a
straight line is constructed between any two
points [z, f(z)] and fy.fiy)], then this line is
above the graph of f(x) at all points between z Figure A-2. Graph o fx) =x2 in Rl
A-l
AMCP 706-192
of R" . Then/"(x) is called a convex function if All these desirable properties of convex
for any points z andj inZ), functions will go to waste unless one is able to
test a given function for convexity. The
f[z+0(y-z)] </(z) + fl[/(y)-/(z)] following three theorems provide useful tests:
(A-l) 1. Theorem A-3: If f(x) is twice contin-

uously differentiable in a convex subset D of
for all 8 with 0 < 8*1. R", it is convex in D if an only if the
quadratic form
Convex functions and convex sets are
related as is shown in 32/
ST S (or Sr-V2f-S) (A-3)
dXjdx-j
Theorem A-l: The set of points D inRn
which satisfy gf(x) < 0, i= 1, ..., m, is convex is positive-semidefinite at each point inZ).
if each of the functions g((x) is convex inZ?" .
2. TheoremA-4: If the functions qfx), i =
One further property of convex functions 1,..., r, are convex in the convex subset D of
is extremely important for applications. It is R" and a- > 0, z = 1,..., r, then
established by
2 a,o.(*)
Theorem A-2: If f(x) is differentiable and
convex on the convex subset D of R", then is convex in D.
f(x)>f(y)+Vf(y)'U-y) (A-2) 3. Theorem A-5: If g(x) is twice contin-

uously differentiable, g{x) < 0, and g{_x) is
for allz andy in>. convex; then, \j[g{x)} is convex.
REFERENCE
1. R. T. Rockafeller, Convex Analysis;

Princeton University Press, Princeton, N.J.,
1970.
A-2
AMCP 706-192
APPENDIX B
ANALYSIS OF BEAM-TYPE STRUCTURES
Finite and discrete element methods of

structural analysis (Refs. 1,3), require a
knowledge of the behavior of each element in
the structure. Once each element is described,
-4 Figure B-1. Basic Beam Element
$- *.
then the governing equations of the entire

structure may be derived. Energy methods are
generally used to obtain the governing equa- w(x)=-|(2x3 -39x2 +3) Ulx3
S.
tions.
B-1 ELEMENT ANALYSIS .3Zx2)+-(x3 -29x2

(B-2)
2
In order to apply energy theorems for the + 9. x)+-^-(x3 -9x2).
fi2
analysis of a structure, the potential energy
due to strain, kinetic energy, and change in
external dimensions due to bending must be It should be stressed that the longitudinal
described. The basic idea is to assign general- displacement s(x) is due only to longitudinal
ized displacement functions, which are of the strain in the beam and not due to the change
form expected in structural deformation and in length caused by the lateral displacement
that are uniquely specified when the displace- w(x).
. merit of both ends of the beam is known. A The potential energy PE due to deforma-
typical beam with its deformation sign con- tion of the beam is (Ref. 2):
vention is shown in Fig. B-1. The displace-
ment i, u2, u3, and 4 are components of
PE =
endpoint displacement and us and u6 are
endpoint rotations.
Jo \d*/
The longitudinal displacement of a point x,
0 < x < , on the beam due to longitudinal
1
*\" "{&)'<
V
(B-3)
strain is approximated by r
1 * , dx H
2 / 2
'o
s(x)- + u-> (B-1)
j El -^(12x_6)-|i(12x-6B)
Lateral displacement of the beam at a point
x is approximated by + -*1 (6x - 4J2) +-P- (6x - 29.)
B-\
AMCP 706-192
which is or
AH3 -Alt' 0 0 0 0 140 70 0 0 0 0

AH* AS. 2
0 0 0 0 70 140 0 0 0 0
0 0 156 54 -22n 13E
r 0 0 12/ -12/ -6/8 -6/R _ 1 -TpAi
*-i- f. 0 0-12/ 12/ 6/S 6/6
~l" 420 0 0 54 156 -13Q 222
0 0 -6/C 6/S 4/S 2

2/I 2 0 0 - 22Q --13Q 4S2 -3S2
0 0-6/6 6/e 2/s2 4/e2 0 0 13Q 222 -322 42
(B-6)
(B-4) The shortening of the beam ALL due to the

lateral displacement is
where u = [u^, u2, u3, w4, us, u6]T.

dx
Jo
Similarly, the kinetic energy KE of the
beam is 1 (dw\
f
Jo
*2 dw2\
KE= + |e?x
ft- (6x2 _6x)
(B-7)
2
i- (6x - 6fcc)
x- \ . x 3
-"l I ~ )+2
+ 5-(3x2 -411X+K.2)
^|(2x3 -3x2 + J23)

2
+^T(3x
2 2x) rfx
e
(B-5)
or
(2x3 -3fcc2)
o o
o o
3 3 1 1
+ (x3-2k' +fcc) oo -h
58 --h
5S ^r
20 ^
20
u. (B-8)
3 3 1
o o 5S 20 "20
1 1 e 12
0 0-^- -^r
20 20 15 "60
1 1 e
+ ^f(x3-x2) rfx 0 0
20 20 ~60 75
B-2
AMCP 706-192
B-2 VARIATIONAL PRINCIPLES V(u) = U(u) + J2() (B-9)
In most structural analysis work it is more Equilibrium states of a structure can now
convenient to use variational principles to be characterized in terms of the total poten-
describe the state of the structure than it is to tial energy. For unconstrained conservative
use Newton's laws directly. For use in analysis systems, a necessary and sufficient condition
of the structures considered here, the basic that u be an equilibrium state is that
variational principles are stated. These prin-
ciples apply to systems that are unconstrained
n W{u)
in the particular coordinate system chosen; bV(u) = 2 bu. = 0 (B-10)
1 = 1 out '
i.e., once the coordinates, up i = 1,..., n,that
describe the state of the system are chosen,
for all bu.. This is equivalent to
there are no algebraic relationships between
these variables. Further, it is required that all
dV(u)
force-state relationships for external or in- = 0, i= l,...,n. (B-ll)
du.
ternal forces are continuous, i.e., small
changes in state yield only small changes in
forces.
This result is proved in Ref. 2, page 23.
Let W denote the work done by all forces
on a structure due to an admissible displace- A second result is that a conservative,
ment. Then, the system is called conservative unconstrained system is in stable equilibrium
if the work W done in any displacement that at a state u if and only if the total potential
returns to the original state is zero; i.e., the energy is a relative minimum at u. This result
system is conservative if no energy is required is proved in Ref. 2, page 30. For the kind of
to change the state of the structure and bring structural systems considered here, it is
further shown, Ref. 2, page 211, that at a
it back to the original state along any path. A
buckling load P , the second variation must
structure would be nonconservative, for
be positive semi-definite, i.e.,
example, if sliding friction or viscous damping
were present.
b2V{u) = buT-&-f(u)bu> 0 (B-12)
Starting from some reference state of the
structure u0, define the stored energy in the
structure at state u to be U(u). Note that and further that there is a bu such that
since the system is conservative, U(u) depends
only on u and not on the path the state
_~ 92V
variable followed in getting from u0 to u. bzV{u)=buTT(u)bu =0. (B-13)
bur
Likewise, the negative of work done by the
external forces acting on the structure due to
the change in state variables will be denoted The inequality, Eq. B-12, shows that zero is
Sl(u). Again, 2(u) depends only on the final a relative minimum and Eq. B-13 shows that
state and not on the path from u0 to u. With relative minimum is attained for bu = bu. For
this notation, the total potential energy V(u) 52 V(u) treated as a function of bu, then, it is
is defined as necessary that
B-3
AMCP 706-192
1 p ' /
l62V(u)] =0, i= l,...,n m '
3/
35 i 2 /=i dZ 3Z
at 5 = Sr7. Thus where m. are element masses. In the notation

of Eq. B-15,
32 V
2 (u)Su = 0 (B-14)
3
M= S
p
'd^V1 /3/V1
i=l
The condition, Eq. B-14, determines buckling Putting

loads for the structural system. All this
analysis requires, of course, that V(u) is at
71- F (B-16)
least twice continuously differentiable.
Lagrange's equations of motion are simply
These laws are for static behavior of the (Ref. 2, page 239)
structural system. Dynamically, the structure
is governed by Lagrange's equations of
d (bL\ bL
motion. First, define the kinetic energy T as
the quadratic form
las-J-is;-0'^1'" (B 17)
-
B-3 EQUATIONS OF STRUCTURAL
T = iiTM, (B-15) ANALYSIS
2
where The given variational principles may be
applied to obtain the governing equations of
M = [my] structural analysis. The element properties
described by Eqs. B-4, B-6, and B-8 may be
and the m.. are generalized masses. The used to generate the potential and kinetic
generalized mass matrix M is defined by the energy of the entire structure. From the
transformation definitions, Eqs. B-9 and B-l 5,
z =f(Z) K= 2 PE> -2 (uk +uk) Fk (B-18)

i Jt
where
Z = [Z1,...,Zp] and
and Zt are components of physical displace- 2 KE> (B-19)

i
ment of the masses of the structure. There-
fore, where superscript /' denotes thejth element of
the structure, k denotes the components of
1 1 3/
P
T= Z Zf m. = 2
P
displacement at joints, Fk is the component
2 >=i 2 ;=i az of external force corresponding to uk, andw*
is the displacement in the same direction as
x m.
3/ uk but due to the A2 components of bar
3Z deformation. The displacements ~k will be
B-4
AMCP 706-192
determined from structure geometry and the 1 .7" . .

A; of individual members given in Eq. B-8. -uTKu -E' K'u>
2 2 i
For determining the equilibrium equations
B-ll, the displacements Tik are generally and
neglected since they are at least quadratic in u 1 1 T
iiTM =- S y M'ii'
so they will be small if no bucklingoccurs. It 2 2 /
is just these quadratic terms, however, that
predict buckling behavior of structures. where summation is taken over all elements of
the structure and K' and M' are defined in
ff a composite displacement vector u is Eqs. B-4 and B-5. The equations for displace-
formed from the components of all the ment, buckling, and dynamic motion may
member displacements, then matrices K and now be determined directly from Eqs. B-l 1,
M may be defined by B-14, and B-17.
REFERENCES
1. O. C. Zienkiewicz, The Finite Element 3. J. S. Przemieniecki, Theory of Matrix

Method, McGraw-Hill, London, 1967. Structural Analysis, McGraw-Hill, New
York, 1968.
2. H. L. Langhaar, Energy Methods in Ap-
plied Mechanics, Wiley, New York, 1962.
B-5
AMCP 706-192
INDEX
Absolute maximum, 23 Design variable space, 111

Absolute minimum, 23 Differentiability, 63
Adjoint equation, 8-3, 8-17, 8-28, 8-44 Direct methods, 2-2, 6-13
Admissible set, 21 Direction of steepest descent, 111, 29,
4-20
Basic feasible point, 36 Displacement constraints, 5 5, 526
Basic point, 36 Distributed parameter system, 825, 92
Basic variables, 36 Dual linear program, 34
Bolza Problem, 6 17
Boundary design variable, 826
Eigenfunction, 93
Bounded linear program, 33
Eigenvector, 57
Buckling constraints, 56, 525
Elastic structural design, 54
Elliptic partial differential equations, 826,
Calculus of variations, 61
9-3
Closed set, 16
Euler-Lagrange equation, 67
Collocation technique, 521 Exterior SUMT method, 4-14
Column optimization, 511, 72, 98
Computer aided design, 1 4
Conceptual design, 12 Feasible linear program, 33
Configuration optimization, 554 Fibonacci search, 27
Conjugate directions, 2 13 Finite dimensional optimal design, 48
Conjugate direction methods, 212 First order constraint qualification, 45
Conjugate gradient method, 2 14, 220 Fletcher-Powell Method, 2-16,2-2 1
Constraint, 110 Frame optimization, 536
Constraint set, 3-3, 4-2 Frequency constraint, 56, 524
Continuity, 16 Function analysis, 12
Control variables, 6-1,6-17,6-22 Function spaces, 63
Convergence, 16 Functional, 62
Convex function, 25 Functional analysis, 63
Convex programming problem, 45 Fundamental problem of the calculus
Convex set, 25 of variations, 64
Corners, 67
Cost function, 1-10, 1-15 Generalized Newton Method for boundary
value problems, 641
Degenerate linear program, 310 Generalized Newton Method, 2-11,2-19
Design sensitivity factors, coefficients, 114, Global properties, 25
5-52 Golden section search, 27
Design variable constraints, 526, 617, Gradient, 17
6-28 Gradient method, 28
Design variable, parameter, 19,21, 31, Gun barrel design, 53
4-8,5-5,6-1,6-17 Gun hop, 1-15,5-4
1-1
AMCP 706-192
INDEX (Con't.)
Helicopter armament, 52 Optimal process theory, 6 1

Optimality criteria (structures), 55
Indirect method, 22, 71 Orthogonal vectors, 16
Inequality constraint, 3 1
Infeasible linear program, 3 7 Perturbation equations, 58
Initial value methods, 640 Pieced extremals, 637
Inner product, 16 Pivot, 3-7
Interactive computation, 551 Pivot step, 37
Interactive computer aided design, 113 Plate optimization, 5-29, 9-16
Interactive graphics, 113 Pontryagin Maximum Principle, 621
Interior SUMT method, 4- 10 Portal frame optimization, 516, 910
Positive definite matrix, 24
Kuhn-Tucker necessary conditions, 46 Power method, 5-7, 5-24
Leibniz' Rule, 7-25, 7-29 Quadratic interpolation, 26

Light weight weapon structures, 52 Quasilinearization, 644
Linear independence (of vectors), 213
Linear programming, 3 1 Rayleigh quotient, 57
Linearization, 58 Recoil mechanism design, 115, 835
Relative maximum, 23
Mathematical programming, 4 1 Relative minimum, 23, 64
Minimizing sequence, 614 Ritz Method, 6- 14
Minimum weight structures, 51
Mixed SUMT method, 4-16 Second-order necessary conditions, 47,
Multiple failure criteria, 518 6-10
Multiple loading, 5-6, 5-28 Second-order sufficient conditions, 47
Self-adjoint operator, 9-3, 9-10,9-12,
Natural frequency constraints, 56, 524 9-17
Necessary condition, 2-3, 4-6, 4-7, 411, Sensitivity analysis, factor, 114, 552
6-5,6-18,6-30 Sensitivity function, 117
Newton's Method (equation solving), 74, Sequentially unconstrained minimization
7-25,7-35,7-40 techniques (SUMT), 4-10
NLP, 4-1 Set, 1-6
Nonbasicvariables, 36 Shooting technique, 6-40, 7-2
Nonlinear programming, 41 Singular arcs, 637
Norm, 16, 63 Simplex algorithm, 38
Normal Bolza Problem, 6-21, 8-6, 8-3 1 Simplex tableau, 36
State variable, 1-10,4-8, 5-5,6-17
One-dimensional minimization, 26 State variable inequality constraint, 628,
Operator equation, 826, 92 6-31,6-37,7-20
Optimal control, 6-1, 6-27, 6-33 Stationary point, 24
1-2
AMCP 708-192
INDEX (Con't.)
Steepest Descent Method, 2-8, 2-18 Truss optimization, 5 22

4-19,5-7,8-1
Step size, 1-11, 1-12,4-22,4-23, Unbounded linear program, 33
4-27,4-34,5-11 Unconstrained problem, 2 1
Stress constraints, 5-5, 525, 539
Strict absolute minimum, 23 Variational notation, 6 10,6 11
Structural topology, 314 Vector, 16
Subset, 1-6 Vector derivative, 1 7
Sufficient condition, 24, 68 Vector function, 16
Synthesis, 1- 1 Vector inequality, 16
System engineering, 1 2 Vibrating beam optimization, 5-14
Vibration constraints, 5 6
Taylor's Theorem, 17
Time-like variable, 825 Well posed problem, 83
Topological design, 554 Weierstrass-Erdmann corner condition,
Transversality condition, 67 6-7
Weierstrass necessary condition, 67
1-3
AMCP 706-192
(AMCRD-TV)
FOR THE COMMANDER:
OFFICIAL: JOSEPH W. PEZDIRTZ

Major General, USA
Chief of Staff
JOHN LYCAS
Colonel, GS
Chief , HQ Admin Mgt Ofc
DISTRIBUTION:
Special
ft U. S. GOVERNMENT PRINTING OFFICE : 1973 O - 540-B02 (5010A)

ENGINEERING DESIGN HANDBOOKS
Available to AMC activfties, 00D agencies, and Government agencies from Letterkenny Any Depot, Chambers burg, FA 17201.
Available to contractors and universities from National Technical Information Service (NTIS), Department of Commerce,
Springfield, VA 22151 EXCEPT WERE NOTED.
No. Title No. Title
AMCP 706- CP 706-
100 Design Guidance for Producibility "Helicopter Engineering Part One, Preliminary
104 Value Engineering Design
106# Elements of Armament Engineering, Part One, 202 "Helicopter Engineering Part Two, Detail Design
Sources of Energy 203 Helicopter Engineerin Part Three, Qualifica-
107# Elements of Armament Engineering, Part Two, tion Assurance
Ball istics 204 *He1 icopter Performance Testin<
108# Elements of Armament Engineering, Part Three, 205 "Timing Systems and Components
Weapon Systems and Conponents 210 Fuzes
109 Tables of the Cumulative Binomial Probabilities 211(C)* Fuzes., Proximity Electrical, Part One (U)
110 Experimental Statistics, Section 1, Basic Con- 212(S # Fuzes, Proximity Electrical, Part Two (U)
cepts and Analysis of Measurement Data 213{S)# Fuzes, Proximity Electrical, Part Three (U)
111 Experimntal Statistics, Section 2, Analysis of 214(S)# Fuzes, Proximity Electrical, Part Four (U)
Enumerative and Classificatory Data 215(C)# Fuzes. Proximity Electrical, Part Five (U)
112 Experimental Statistics, Section 3, Planning 235 Hardening Weapon Systems Against RF Energy
and Analysis of Comparative Experiments 238 "Recoilless Rifle Weapon Systems
113 Experimental Statistics, Section 4, Special 239 "Small Arms Weapon Systems
ToDies 240(S f Grenades (U)
114 Experimental Statistics, Section 5. Tables 242 Design for Control of Projectile Flight
115 Environmental Series, Part One, Basic Environ- Characteristics (REPLACES-246)
mental Concepts 244 Ammunition, Section 1, Artillery Ammunition--
116 *Environmental Series, Part Two, Basic Environ- General, with Table of Contents, Glossary,
mental Factors and Index for Series
120 Criteria for Environmental Control of Mobile 245(C # Ammunition, Section 2, Design for Terminal
Systems Effects (U)
121 Packaging and Pack Engineering 246 Ammunition, Section 3, Design for Control of
123 Hydraulic Fluids Flight Characteristics (REPLACED BY -242)
125 Electrical Wire and Cable 2471 Amnunition, Section 4, Design for Projection
127 Infrared Military Systems, Part One 248 Ammunition, Section 5, Inspection Aspects of
128(5)* Infrared Military Systems, Part Two (U) Artillery Amnunition Design
130 Oesign for Air Transport and Airdrop of Materiel 249 Ammunition, Section 6, Manufacture of Metallic
132 *Maintenance Engineering Components of Art ill ery Amnunition
133 *Maintainability Engineering Theory and Practice 250 GunsGeneral
134 Maintainability Guide for Design 251 Muzzle Devices
135 Inventions, Patents, and Related Matters 252 Gun Tubes
136 Servomechanisms, Section 1. Theory 253 "Breech Mechanism Design
137 Servomchanisms, Section 2, Measurement and 255 Spectral Characteristics of Muzzle Flash
Signal Converters 260 Automatic Weapons
138 Servomechanisms, Section 3, Amplification 270 "Propellant Actuated Devices
139 Servomechanisms, Section 4, Power Elements and 280 Design of Aerodynamically Stabilized Free
System Design Rockets
Trajectories, Differential Effects, and Data 281(SRD)# Weapon System Effectiveness (U)
for Projectiles 282 +Propulsion and Propellants (REPLACED BY -285)
150 Interior Ballistics of Guns 283 Aerodynamics
160(S)# Elements of Terminal Ballistics. Part One, Kill 284(C)* Trajectories (U)
Mechanisms and Vulnerability (U) 285 Elements of Aircraft and Missile Propulsion
161(S)# Elements of Terminal Ballistics, Part Two, (REPLACES -282)
Collection and Analysis of Data Concerning 286 structures
Targets (U) 290(C)* WarheadsGeneral (U)
162(SRD) # Elements of Terminal Ballistics, Part Three, 291 Surface-to-Air Missiles Part One, System
Application to Missile and Space Targets (U) Integration
165 Liquid-Filled Projectile Design 292 Surface-to-Air Missiles, Part Two, Weapon
170(C)* **Armor and Its Applications (U) Control
175# Solid Propellants, Part Ore 293 Surface-to-Air Missiles Part Three, Computers
176{C)# Solid Propellants, Part Two (U) 294<S)* Surface-to-Air Missiles Part Four, Missile
177 Properties of Explosives of Military Interest Armament (U)
178(C) Properties of Explosives of Military Interest, 295(5)* Surface-to-Air Missiles, Part Five, Counter-
Section 2 (U) (REPLACED BY -177) measures (U)
179# ""Explosive Trains 296 Surface-to-Air Missiles, Part Six, Structures
180 Principles of Explosive Behavior and Power Sources
181 *Explosions in Air. Part One 297(S)# Surface-to-Air Missiles. Part Seven, Sample
182(S)# *Explosions in Air, Part Two (U) Problem (U)
185# Military Pyrotechnics. Part One, Theory and 327 Fire Control SystemsGeneral
Application 329 Fire Control Computing Systems
186 Military Pyrotechnics, Part Two, Safety, 331 Compensating Elements
Procedures and Glossary 335(SRD)#*Design Engineers' Nuclear Effects Manual,
Military Pyrotechnics, Part Three, Properties Volume I, Munitions and Weapon Systems (U)
of Materials Used in Pyrotechnic Compositions 336(SRD)#*Design Engineers' Nuclear Effects Manual,
* M i I it a ry Pyrotechnics, Part Four, Design of Volume II, Electronic Systems and Logistical
Ammunition for Pyrotechnic Effects Systems (U)
189 Military Pyrotechnics. Part Five, Bibliography 337(SRD)#*Design Engineers' Nuclear Effects Manual,
190 *Army Weapon System Analysis Volume III, Nuclear Environment (U)
191 System Analysis and Cost-Effectiveness 338(SRD)#*Design Engineers' Nirclear Effects Manual,
195 *Development Guide for Reliability, Part One, Volume IV. Nuclear Effects (U)
Introduction, Background, and Planning for 340 Carriages and MountsGeneral
Army Materiel Requirements 341 Cradles
196 *Development Guide for Reliability, Part Two, 342 Recoil Systems
Design for Reliability 343 Top Carriages
157 "Development Guide for Reliability, Part Three, 344 Bottom Carriages
Reliability Prediction 345 Equi 1 ibrators
198 "Development Guide for Reliability, Part Four, 346 Elevating Mechanisms
Reliability Measurement 347 Traversing Mechanisms
199 *Development Guide for Reliability, Part Five, 350 Wheeled Amphibians
Contracting for Reliability 355 The Automotive Assembly
200 Development Guide for Reliability, Part Six, 356 Automotive Suspensions
Mathematical Appendix and Glossary 357 Automotive Bodies and Hulls
360 "Military Vehicle Electrical Systems
"UNDER PREPARATION-not available 445 Sabot Technology Engineering
OBSOLETE--out of stock
""REVISION UNDER PREPARATION
#NQT AVAILABLE ffCM NTIS

Ad 0767826

Uploaded by

Copyright:

Available Formats

Ad 0767826

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ad 0767826

Uploaded by

Copyright:

Available Formats

AMCP-706-192

JC PAMPHLET WICP 706-192

COMPUTER AIDED DESIGN

EADQUARTERS, U.S. ARMY MATERIEL COMMAND JULY 1973

ENGINEERING DESIGN HANDBOOK

CHAPTER 1. ELEMENTS OF COMPUTER AIDED DESIGN

1- -1 Synthesis vs Analysis in Engineering Design 1-1

TABLE OF CONTENTS (CorVt.)

CHAPTER 2. FINITE DIMENSIONAL

2-1 Introduction 2-1

TABLE OF CONTENTS (CorVt.)

CHAPTER 3. LINEAR PROGRAMMING

CHAPTER 4. NONLINEAR PROGRAMMING AND

41 Introduction to the Theory of

TABLE OF CONTENTS (Con't.)

CHAPTER 5. FINITE DIMENSIONAL

TABLE OF CONTENTS (Con't.)

54.5 A Minimum Weight Plate With Frequency

CHAPTER 6. THE CALCULUS OF VARIATIONS AND

6-1 Introduction 6-1

TABLE OF CONTENTS (Con't.)

63.3 Necessary Conditions for the Bolza

CHAPTER 7. OPTIMAL STRUCTURAL DESIGN

TABLE OF CONTENTS (Con't.)

74.1 Statement of the Problem 716

CHAPTER 8. METHODS OF STEEPEST DESCENT

TABLE OF CONTENTS (CorVt.)

85 Optimal Design of an Artillery Recoil

CHAPTER 9. APPLICATION OF STEEPEST DESCENT

APPENDIX A CONVEXITY A-l

APPENDIX B ANALYSIS OF BEAM-TYPE

Hgr. N3. Title Bage

11 A System Engineering Model 12

LIST OF ILLUSTRATIONS (Con't.)

Fig. No. Title Page

5 10 Frame With Side Loading 5 18

LIST OF ILLUSTRATIONS (Con't.)

Eg. No. Eife Bags

538 Iteration vs Weight Curves for Example 5-7,

LIST OF ILLUSTRATIONS (Con't.)

Fig. No. Title Page

712 Simply Supported Beam With Positive

Table No. Title Page

1_1 Weights of Towers 1 13

LIST OF TABLES (Con't.)

Table No. Hue Page

56 Optimal Design Variable bi for

The Engineering Design Handbooks of the U. S. Army Materiel Command

Extensive application of design optimization techniques is made in the

The text is structured so that it can be understood and used by practicing

a. Activities within AMC, DOD agencies and Government agencies other

b. Contractors and universities must forward their requests to:

National Technical Information Service

(Requests for classified documents must be sent, with appropriate "Need to

Comments and suggestions on this Handbook are welcome and should be

US Army Materiel Command

DA Forms 2028 (Recommended Changes to Publications), which are

ELEMENTS OF COMPUTER AIDED DESIGN

material in order to meet strength require- A key challenge to developers of practical

Figure 1-1. A System Engineering Model

In order to verify the sufficiency condition **,, <>> 0and