Design of Air Defence Weapon

Download as pdf or txt
Download as pdf or txt
You are on page 1of 306

A Top-Down, Hierarchical, System-of-Systems

Approach to the Design of an Air Defense Weapon

A Thesis
Presented to
The Academic Faculty
by

Tommer Rafael Ender

In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy

School of Aerospace Engineering


Georgia Institute of Technology
August 2006
c 2006 by Tommer Rafael Ender
Copyright

A Top-Down, Hierarchical, System-of-Systems


Approach to the Design of an Air Defense Weapon

Approved by:
Prof. Dimitri Mavris
Committee Chair
School of Aerospace Engineering
Georgia Institute of Technology

Dr. Kevin Massey


Aerospace, Transportation,
and Advanced Systems Laboratory
Georgia Tech Research Institute

Prof. Daniel Schrage


School of Aerospace Engineering
Georgia Institute of Technology

Dr. E. Jeff Holder


Sensors and Electromagnetic
Applications Laboratory
Georgia Tech Research Institute

Dr. Neil Weston


School of Aerospace Engineering
Georgia Institute of Technology

Date Approved: 03 July 2006

"

for my family

iii

ACKNOWLEDGEMENTS

The work presented in this document could not have been completed without the
love and support of my parents, Judy and Ilan, my best friend and brother Roy, and
my bashert Sonja. I would not have even considered pursuing a Ph.D. without the
encouragement of my advisor Dr. Dimitri Mavris, who has supported my academic
progression since my first undergraduate term at Georgia Tech. He has taught me to
expect more from myself, and certainly more for myself in life. Thanks Doc.
Id also like to thank the other members of my committee who helped me along
the process: Dr. Neil Weston, who was instrumental in the final round of iterations
of editing this document; Dr. Daniel Schrage, who took an inspiring interest in this
topic very early on and provided encouraging feedback throughout; Dr. Jeff Holder,
GTRI, for helping me begin to understand the complex radar concepts I decided
to tackle in this thesis; and finally, a very special thank you to Dr. Kevin Massey,
GTRI, not only for the encouragement to pursue this specific topic for a Ph.D., but
also for working with me every step of the way to formulate a very interesting and
important problem to apply the methodology developed in this thesis. That being
said, the 6-DoF trajectory analysis in this document would not have been possible
without the hard work of Mike Hieges of GTRI, and the co-op students working at
GTRI that Ive had to pleasure to work with over the past few years: Kevin Guthrie,
Patrick OLeary, Ben DiFrancesco, and Kyle French - I am still very impressed how
quickly you guys were able to catch on and contribute to the development of the codes
used in this study. Thank you to the U.S. Army-AMRDEC and GTRI for additional
computational and financial support.
And many thanks to the friends Ive made at ASDL over the years, especially

iv

those that helped out with my thesis review: Andrew Frits, Brian German, Peter
Hollingsworth, Hernando Jimenez, Rob McDonald, Janel Nixon, Holger Pfaender, and
Jack Zentner. Also a special thanks to the other students I worked with in the original
methodology used in this thesis: Bjorn Cole, David Culverhouse, Jared Harrington,
and to Patrick Biltgen for helping shape the story telling aspect of the method, as
well as the members of the AFICE team Ive left out. Thank you to Dr. Michelle
Kirby for always making herself available for answering questions and guidance. I
would finally like to thank Liz and Jason Billman for proving the occasional retreat
from my normal surroundings, giving me the much needed boost to get the thesis
proposal started.

TABLE OF CONTENTS
DEDICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

iii

ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . .

iv

LIST OF TABLES

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xi

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xii

LIST OF FIGURES

LIST OF SYMBOLS OR ABBREVIATIONS

. . . . . . . . . . . . . . xviii

SUMMARY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I

MOTIVATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1.1

Top-Down Systems Engineering . . . . . . . . . . . . . . . .

1.1.2

Initial Research Formulation . . . . . . . . . . . . . . . . . .

1.1.3

Overview of Thesis . . . . . . . . . . . . . . . . . . . . . . .

Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2

1.2.1

II

xx

The Rocket, Artillery, and Mortar (RAM) Threat in Asymmetric Warfare . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2.2

Problem Scope: Extended Area Protection & Survivability .

13

1.2.3

Approaches to the RAM Threat . . . . . . . . . . . . . . . .

19

1.2.4

Guided Projectiles . . . . . . . . . . . . . . . . . . . . . . . .

26

1.2.5

Summary of Discussed Solutions . . . . . . . . . . . . . . . .

32

1.3

Paradigm Shift in Conceptual Design . . . . . . . . . . . . . . . . .

34

1.4

Disclaimer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

36

LITERATURE SEARCH: COMBAT SYSTEM MODELING & SIMULATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37


2.1

2.2

Modeling Military Operations . . . . . . . . . . . . . . . . . . . . .

38

2.1.1

General Requirements of Combat Modeling . . . . . . . . . .

38

2.1.2

Large Scale Combat Computer Simulations . . . . . . . . . .

39

Modeling and Simulation Concepts . . . . . . . . . . . . . . . . . . .

41

vi

2.3

2.4

2.5

2.6

2.7

2.8

2.2.1

Deterministic Models . . . . . . . . . . . . . . . . . . . . . .

44

2.2.2

Probabilistic (Stochastic) Models . . . . . . . . . . . . . . . .

45

Survivability as a Discipline . . . . . . . . . . . . . . . . . . . . . . .

47

2.3.1

Susceptibility . . . . . . . . . . . . . . . . . . . . . . . . . . .

48

2.3.2

Vulnerability . . . . . . . . . . . . . . . . . . . . . . . . . . .

50

2.3.3

Powering-up/Single Sortie Probability of Damage

. . . . . .

51

Air Defense Systems . . . . . . . . . . . . . . . . . . . . . . . . . . .

53

2.4.1

Operational Requirement & System Specification . . . . . . .

54

2.4.2

Types of Air Defense Systems . . . . . . . . . . . . . . . . .

55

Weapon System Effectiveness Parameters . . . . . . . . . . . . . . .

59

2.5.1

Weapon Accuracy . . . . . . . . . . . . . . . . . . . . . . . .

59

2.5.2

Lethality . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

60

Guidance, Navigation, and Control of Related Weapon Concepts . .

62

2.6.1

Guidance . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

62

2.6.2

Navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . .

65

2.6.3

Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

67

Radar Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

69

2.7.1

Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . .

69

2.7.2

Radar Range . . . . . . . . . . . . . . . . . . . . . . . . . . .

70

2.7.3

Signal-to-Noise Ratio . . . . . . . . . . . . . . . . . . . . . .

71

2.7.4

Radar Cross Section . . . . . . . . . . . . . . . . . . . . . . .

72

2.7.5

Tracking Radars . . . . . . . . . . . . . . . . . . . . . . . . .

74

Chapter Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . .

78

III LITERATURE SEARCH: ADVANCED DESIGN METHODOLOGIES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80


3.1

System-of-Systems Approach . . . . . . . . . . . . . . . . . . . . . .

80

3.1.1

Functional Decomposition . . . . . . . . . . . . . . . . . . . .

84

3.1.2

Functional Composition . . . . . . . . . . . . . . . . . . . . .

86

3.1.3

System-of-System Architecture . . . . . . . . . . . . . . . . .

87

vii

3.2

3.3

3.4

3.5

Uncertainty Analysis and Quantification in the Design Process . . .

90

3.2.1

Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92

3.2.2

Monte Carlo Simulation . . . . . . . . . . . . . . . . . . . . .

94

3.2.3

Filtered Monte Carlo Approach . . . . . . . . . . . . . . . .

95

Surrogate Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . .

96

3.3.1

Polynomial Response Surface Equations . . . . . . . . . . . .

97

3.3.2

Neural Network Response Surface Equations . . . . . . . . .

99

3.3.3

Analysis of Variance . . . . . . . . . . . . . . . . . . . . . . . 105

Application of Techniques . . . . . . . . . . . . . . . . . . . . . . . . 105


3.4.1

Response Surface Methodology & Design Space Exploration . 105

3.4.2

Top-Down Capability Based Design . . . . . . . . . . . . . . 107

3.4.3

Design for Affordability . . . . . . . . . . . . . . . . . . . . . 111

Chapter Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

IV RESEARCH QUESTIONS & HYPOTHESES . . . . . . . . . . . . 113


4.1

Problem Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

4.2

Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

4.3
V

4.2.1

Initial Formulation Research Questions Reviewed . . . . . . . 114

4.2.2

Additional Research Questions . . . . . . . . . . . . . . . . . 115

Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

RESEARCH FORMULATION & APPROACH . . . . . . . . . . . 121


5.1

Decomposing the Bounded Problem . . . . . . . . . . . . . . . . . . 121


5.1.1

Narrowing the Scope of Interest . . . . . . . . . . . . . . . . 121

5.1.2

Morphological Matrix . . . . . . . . . . . . . . . . . . . . . . 122

5.1.3

Hierarchical Decomposition of Proposed Solution . . . . . . . 133

5.1.4

Review of Morphological Matrix Selections . . . . . . . . . . 136

5.2

Research Focus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

5.3

Modeling and Simulation Approach . . . . . . . . . . . . . . . . . . 147


5.3.1

Modeling and Simulation Requirements . . . . . . . . . . . . 147

viii

5.3.2

6-DoF Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 147

5.3.3

Radar Subsystem Model . . . . . . . . . . . . . . . . . . . . 150

5.3.4

Combined Effect of Multiple Projectiles . . . . . . . . . . . . 153

5.3.5

Creating Surrogate Models . . . . . . . . . . . . . . . . . . . 154

VI IMPLEMENTATION . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.1

Approach Review & Implementation Introduction . . . . . . . . . . 157

6.2

Preliminary Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

6.3

Initial 6-DoF Study . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

6.4

6.3.1

Variable and Response Identification . . . . . . . . . . . . . . 159

6.3.2

Surrogate Model Creation & Visualization

6.3.3

Introduction of Gun Firing Noise . . . . . . . . . . . . . . . . 166

6.3.4

Response Variability Study . . . . . . . . . . . . . . . . . . . 168

. . . . . . . . . . 162

Addition of Radar & Guidance Update Rates . . . . . . . . . . . . . 189


6.4.1

Control State Slaved to Radar Update . . . . . . . . . . . . . 190

6.4.2

Independent Control and Radar Update Rates . . . . . . . . 193

6.5

Effect of Firing Multiple Projectiles . . . . . . . . . . . . . . . . . . 197

6.6

Subsystem Level Radar Properties . . . . . . . . . . . . . . . . . . . 204

6.7

Detailed 6-DoF Model . . . . . . . . . . . . . . . . . . . . . . . . . . 216


6.7.1

Variable Listing . . . . . . . . . . . . . . . . . . . . . . . . . 216

6.7.2

Surrogate Modeling . . . . . . . . . . . . . . . . . . . . . . . 220

6.8

Assembled Hierarchal Environment . . . . . . . . . . . . . . . . . . 222

6.9

Chapter Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 234

VII CONCLUSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236


7.1

7.2

Review of Research Questions & Hypotheses . . . . . . . . . . . . . 236


7.1.1

First Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . 236

7.1.2

Second Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . 238

7.1.3

Third Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . 240

Summary of Contributions . . . . . . . . . . . . . . . . . . . . . . . 241

ix

7.3

Recommendations for Further Study . . . . . . . . . . . . . . . . . . 242

APPENDIX A

PROBABILITY DISTRIBUTIONS . . . . . . . 245

APPENDIX B

PRELIMINARY RESULTS . . . . . . . . . . . . 248

APPENDIX C

GUN RATE OF FIRE REGRESSION . . . . . . 264

APPENDIX D

SUPPORTING DATA . . . . . . . . . . . . . . . . 267

REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
VITA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284

LIST OF TABLES
Table 1

MK 15 Phalanx Close-In Weapons System Characteristics . . . . .

22

Table 2

Ahead 35 mm Ammunition Characteristics . . . . . . . . . . . . . .

24

Table 3

Summary of Relevant Unguided & Guided Alternatives . . . . . . .

33

Table 4

Model Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

43

Table 5

Probability Factors that Influence Susceptibility . . . . . . . . . . .

49

Table 6

Attrition Kill Levels . . . . . . . . . . . . . . . . . . . . . . . . . .

51

Table 7

Radar Frequency Bands . . . . . . . . . . . . . . . . . . . . . . . .

70

Table 8

Radar Cross Section Magnitudes . . . . . . . . . . . . . . . . . . .

73

Table 9

Morphological Matrix of System Alternatives . . . . . . . . . . . . 124

Table 10

Morphological Matrix of System Alternatives with Selections . . . . 139

Table 11

Radar Noise Errors for Initial Study(1) . . . . . . . . . . . . . . . 160

Table 12

Gun Pointing Bias for Initial Study . . . . . . . . . . . . . . . . . . 161

Table 13

Discrete Variable Settings for Initial Study . . . . . . . . . . . . . . 161

Table 14

Gun Firing Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

Table 15

Fixed Variables for Radar & Control Update Rate Study . . . . . . 190

Table 16

Variables for Radar Update Rate Study with Control State Slaved
to Radar Update . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191

Table 17

Variables for Radar Update Study with Independent Radar & Guidance Updates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

Table 18

Variables Used for Radar Subsystem Study . . . . . . . . . . . . . 205

Table 19

Detailed Study - List of Variables . . . . . . . . . . . . . . . . . . . 221

Table 20

Discrete Variable Settings for Detailed Study . . . . . . . . . . . . 221

Table 21

Area Captured Under a Normal Distribution for Standard Deviation


Multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246

Table 22

Ahead Study - Errors Used to Quantify Miss Distance . . . . . . . 249

Table 23

Historical Data For Gun Rate of Fire . . . . . . . . . . . . . . . . . 265

xi

LIST OF FIGURES
Figure 1

Systems Engineering Process with Top-Down Hierarchical Requirements Flow with Process Applied to a Notional Problem . . . . . .

Figure 2

Green Zone in Baghdad, Iraq . . . . . . . . . . . . . . . . . . . . .

10

Figure 3

60mm, 81mm, & 120mm Fin-stabilized and 4.2 inch Spin-stabilized


Mortars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

12

Figure 4

The M252 81mm Medium Extended Range Mortar . . . . . . . . .

12

Figure 5

Extended Area Protection & Survivability Operational Scenario . .

14

Figure 6

EAPS Engagement Kinematics . . . . . . . . . . . . . . . . . . . .

15

Figure 7

Typical Intercept Velocity Profiles of Guns and Rockets . . . . . .

16

Figure 8

Projectile Timeline Engagement - Typical Mortar . . . . . . . . . .

17

Figure 9

Fire Control Angle Error versus Warhead Lethal Radius . . . . . .

18

Figure 10 MK 15 Phalanx Close-In Weapons System . . . . . . . . . . . . . .

22

Figure 11 The Skyshield 35 Air Defence System . . . . . . . . . . . . . . . . .

23

Figure 12 Ahead 35mm Ammunition . . . . . . . . . . . . . . . . . . . . . . .

24

Figure 13 THEL Rocket Interception Process . . . . . . . . . . . . . . . . . .

26

Figure 14 Potential Projectile Guidance Corrections Effect on Accuracy . . .

29

Figure 15 GTRI Guided Projectile with Pin Firing Mechanism . . . . . . . .

31

Figure 16 Pin-Fin Shock Interaction Guidance Concept . . . . . . . . . . . .

31

Figure 17 Shadowgraph Ensemble Showing Divert of Pin-Controlled Projectile

31

Figure 18 Military Operations Modeling Hierarchy Pyramid . . . . . . . . . .

42

Figure 19 Model Classification Hierarchy . . . . . . . . . . . . . . . . . . . .

44

Figure 20 Example of Powering-Up Rule for an SSPD . . . . . . . . . . . . .

52

Figure 21 A Generic Integrated Air Defense System . . . . . . . . . . . . . .

58

Figure 22 Missile Navigation Laws . . . . . . . . . . . . . . . . . . . . . . . .

65

Figure 23 Close-Loop Control . . . . . . . . . . . . . . . . . . . . . . . . . . .

67

Figure 24 Reduction of Maximum Detection Range with Reduced RCS . . . .

73

Figure 25 Wavefront Phase Relationship for an Interferometer Radar . . . . .

76

xii

Figure 26 Uncertainty Based Design Domains . . . . . . . . . . . . . . . . . .

91

Figure 27 Reliability Versus Robustness in Terms of the Probability Density


Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92

Figure 28 Uncertainty Propagation in the Simulation Model Chain . . . . . .

92

Figure 29 The Forecasting Process . . . . . . . . . . . . . . . . . . . . . . . .

93

Figure 30 Filtered Monte Carlo Approach . . . . . . . . . . . . . . . . . . . .

96

Figure 31 Neural Network Training Process . . . . . . . . . . . . . . . . . . . 100


Figure 32 A Simple Feed-Forward Neural Network . . . . . . . . . . . . . . . 101
Figure 33 The Threshold and Sigmoid Logistic Functions . . . . . . . . . . . 101
Figure 34 Example of Over- and Underfitting a Neural Network . . . . . . . . 103
Figure 35 Example Latin Hypercube Used to Fill 2-D Space . . . . . . . . . . 105
Figure 36 Example Pareto Plot - Gun Accuracy Contributions to Miss Distance 106
Figure 37 Hierarchical, Surrogate Modeling Environment for Systems-of-Systems
Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Figure 38 Top-Down Capability Based Design for a Campaign Scenario . . . 110
Figure 39 Side & Top Views of Gun and Radar Orientations . . . . . . . . . . 132
Figure 40 Moments About Projectile Axes . . . . . . . . . . . . . . . . . . . . 132
Figure 41 Hierarchical Decomposition of Proposed Solution . . . . . . . . . . 134
Figure 42 82mm Mortar Geometry Used in 6-DoF Simulations . . . . . . . . 136
Figure 43 Measures of Merit for Gun System Optimization . . . . . . . . . . 140
Figure 44 Traditional Application of Optimizer on a Design-Structured-Matrix 141
Figure 45 Flow Chart of Research Process . . . . . . . . . . . . . . . . . . . . 142
Figure 46 Top-down Requirements Flow and Bottom-Up Weapon Capabilities
Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Figure 47 Bottom-up Single Design Point vs. Top-Down Multiple Design Point
Hierarchical Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
Figure 48 GTRI 6-DoF Multi-Body Simulation . . . . . . . . . . . . . . . . . 150
Figure 49 Initial Study - Viewing Surrogate Models Using a Prediction Profiler 164
Figure 50 Manipulating the Surrogate Models Using a Bottom-Up GUI

. . . 165

Figure 51 Normal Distribution, =0, =0.25 . . . . . . . . . . . . . . . . . . 167

xiii

Figure 52 Initial Study - Viewing Surrogate Models Using a Prediction Profiler 169
Figure 53 Initial Study - Pareto Plot for Discrete and Continuous Variables on
Variability of 90% Confidence Intercept Miss Distance . . . . . . . 170
Figure 54 Initial Study - Pareto Plots for Fixed Combinations of Range and
Bore on Variability of 90% Confidence Intercept Miss Distance . . . 171
Figure 55 Initial Study - Multivariate Scatter Plot Design Space Exploration

172

Figure 56 Initial Study - Intercept Miss Distance CDF Shape Constraint . . . 173
Figure 57 Initial Study - Bivariate Plots of Radar Noises, For Intercept Miss
Distance CDF Shape Constraint . . . . . . . . . . . . . . . . . . . 174
Figure 58 Initial Study - Bivariate Plots of Gun Noises, For Intercept Miss
Distance CDF Shape Constraint . . . . . . . . . . . . . . . . . . . 175
Figure 59 Initial Study - Bivariate Plots of Showing Maximum Range, For
Intercept Miss Distance CDF Shape Constraint . . . . . . . . . . . 175
Figure 60 Initial Study - Bivariate Plots of Range and Radar Noises, For Intercept Miss Distance CDF Shape Constraint . . . . . . . . . . . . 176
Figure 61 Initial Study - Feasible Designs Separated by Radar Cost . . . . . . 178
Figure 62 Initial Study - Radar Cost by Range & Minimum Miss Distance by
Range, Divided by Radar Cost . . . . . . . . . . . . . . . . . . . . 178
Figure 63 Initial Study - Radar Noises, Divided by Radar Cost . . . . . . . . 179
Figure 64 Initial Study - Range by Radar Azimuth & Elevation Noises, Divided
by Radar Cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Figure 65 Initial Study - Selection of Required 90% Confidence Miss Distance
and Intercept Range . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Figure 66 Initial Study - Radar Noises, For Required 90% Confidence Miss
Distance and Intercept Range . . . . . . . . . . . . . . . . . . . . . 181
Figure 67 Initial Study - Gun Noises, For Required 90% Confidence Miss Distance and Intercept Range . . . . . . . . . . . . . . . . . . . . . . . 181
Figure 68 Initial Study - Multivariate Scatterplot DSE, Sorted by Radar Azimuth and Elevation Noises . . . . . . . . . . . . . . . . . . . . . . 183
Figure 69 Initial Study - Bivariate Plots of Minimum Miss Distances and Range,
Sorted by Radar Azimuth and Elevation Noises . . . . . . . . . . . 183
Figure 70 Initial Study - Bivariate Plot of 90% Confidence Miss Distance and
Bore, Sorted by Radar Azimuth and Elevation Noises . . . . . . . . 184

xiv

Figure 71 Initial Study - Multivariate Scatterplot DSE, Sorted by Radar Roll


and Range Noises . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Figure 72 Initial Study - Bivariate Plots of Minimum Miss Distances and Range,
Sorted by Radar Roll and Range Noises . . . . . . . . . . . . . . . 185
Figure 73 Initial Study - Multivariate Scatterplot DSE, Sorted by Bore . . . . 187
Figure 74 Initial Study - Bivariate Plots of Minimum Miss Distances and Bore,
Sorted by Bore . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Figure 75 Initial Study - Bivariate Plots of Minimum Miss Distances and Range,
Sorted by Bore . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
Figure 76 Initial Study - Bivariate Plot of 90% Confidence and 10% Confidence
Minimum Miss Distance, Sorted by Bore . . . . . . . . . . . . . . . 188
Figure 77 Radar Update Study with Control State Slaved to Radar Update Surface Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
Figure 78 Radar Update Study with Independent Radar & Guidance Updates
- Prediction Profiler . . . . . . . . . . . . . . . . . . . . . . . . . . 194
Figure 79 Radar Update Study with Independent Radar & Guidance Updates
- Surface Profilers . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
Figure 80 Radar Update Study with Independent Radar & Guidance Updates
- Pareto plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
Figure 81 Radar Update Study with Independent Radar & Guidance Updates
- Multivariate Scatter Plot with PH Region of Interest in Red . . . 197
Figure 82 Radar Update Study with Independent Radar & Guidance Updates
- Bivariate Plots of Radar Accuracies and Guidance Update . . . . 198
Figure 83 Multiple Projectiles Study - Multivariate Scatter Plot Design Space
Exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Figure 84 Multiple Projectiles Study - Selection of Required MSPH . . . . . . 200
Figure 85 Multiple Projectiles Study - Feasible Design Space Isolated and Grouped
by Shots Fired . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Figure 86 Multiple Projectiles Study - SSPH vs. Radar Properties, Grouped
by Shots Fired . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Figure 87 Multiple Projectiles Study - Radar Accuracies & Update Rates,
Grouped by Shots Fired . . . . . . . . . . . . . . . . . . . . . . . . 204
Figure 88 Prediction Profiler for Radar Subsystem Study . . . . . . . . . . . 205

xv

Figure 89 Radar Subsystem Study - Multivariate Scatter Plot Design Space


Exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
Figure 90 Radar Subsystem Study - Radar Error Capability Defined . . . . . 209
Figure 91 Radar Subsystem Study - Design Space for Radar Error Capability
Requirement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
Figure 92 Radar Subsystem Study - Regions of Low and High Frequencies are
Selected . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Figure 93 Radar Subsystem Study - Regions of Low and High Frequencies are
Selected . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Figure 94 Radar Subsystem Study - Regions of Low and High Antenna Gains
are Selected . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
Figure 95 Radar Subsystem Study - Capabilities and Properties for Low and
High Antenna Gains . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Figure 96 Warhead Round for Detailed 6-DoF Study . . . . . . . . . . . . . . 218
Figure 97 Thruster Controller for Warhead Round . . . . . . . . . . . . . . . 218
Figure 98 Detailed Study - Prediction Profiler for Three Projectile Kill Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
Figure 99 Assembled Hierarchical Environment - Complete Design Space . . . 224
Figure 100 Assembled Hierarchical Environment - Design Space Sorted by Kill
Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
Figure 101 Assembled Hierarchical Environment - Operational Level Bivariate
Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Figure 102 Assembled Hierarchical Environment - Emphasizing a Discrete Option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Figure 103 Assembled Hierarchical Environment - Selected Desired Phit and
Radar Cost Operational Constraints . . . . . . . . . . . . . . . . . 231
Figure 104 Assembled Hierarchical Environment - Radar Properties Due to Operational Level Constraints . . . . . . . . . . . . . . . . . . . . . . 232
Figure 105 Assembled Hierarchical Environment - Number of Rounds Fired
Trade Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Figure 106 Assembled Hierarchical Environment - Gun Muzzle Velocity Trade
Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
Figure 107 Sample Probability Distributions . . . . . . . . . . . . . . . . . . . 245

xvi

Figure 108 Ahead Study - Interactive Prediction Profiler of Response Surface


Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
Figure 109 Ahead Study - PDF and CDF for Miss Distance and Velocity at
Target Intercept . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
Figure 110 Ahead Study - Miss Distance Point Cloud Dispersion Results . . . 253
Figure 111 Ahead Study - Normalized Submunition Witness Plate Dispersion
Pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Figure 112 Ahead Study - Velocity Degradation Due to Submunition Dispersion 255
Figure 113 Ahead Study - Submunition Dispersion Clouds from Multiple Projectiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
Figure 114 CIWS Study - Target Intercept Ranges . . . . . . . . . . . . . . . . 257
Figure 115 CIWS Study - Prediction Profiler of Surrogate Models . . . . . . . 258
Figure 116 CIWS Study - Monte Carlo CEP Results Single Round Intercept at
250, 500, 750, and 1000 m . . . . . . . . . . . . . . . . . . . . . . . 260
Figure 117 CIWS Study - CDF for 250, 500, 750, 1000 m Ranges . . . . . . . . 261
Figure 118 CIWS Study - Single Round PH and Discretized Single Round PH . 262
Figure 119 CIWS Study - Cumulative PH of Approaching Target . . . . . . . . 263
Figure 120 Firing Rate Curve Fit and Associated Error Distribution . . . . . . 266
Figure 121 Initial Study - Pareto Plots for 40 mm Bore, All Ranges . . . . . . 268
Figure 122 Initial Study - Pareto Plots for 50 mm Bore, All Ranges . . . . . . 269
Figure 123 Initial Study - Pareto Plots for 57 mm Bore, All Ranges . . . . . . 270
Figure 124 Detailed Study - Actual by Predicted and Residual Plots for Neural
Network Fits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271

xvii

LIST OF SYMBOLS OR ABBREVIATIONS


AAA

Anti Aircraft Armament.

ANOVA

Analysis of Variance.

ASDL

Aerospace Systems Design Laboratory (Georgia Tech).

C-RAM

Counter Rockets, Artillery, and Mortars.

CCD

Central Composite Design.

CDF

Cumulative Density Function.

CEP

Circular Error Probable.

CIWS

Close-In Weapons System.

DoD

Department of Defense.

DoE

Design of Experiments.

DoF

Degrees of Freedom.

DSE

Design Space Exploration.

EAPS

Extended Area Protection & Survivability.

FANNGS Function Approximating Neural Network Generation System.


Gr

Receive antenna gain [dB].

Gt

Transmit antenna gain [dB].

GTRI

Georgia Tech Research Institute.

GUI

Graphical User Interface.

IADS

Integrated Air Defense System.

INS

Inertial Navigation System.

IPP

Impact Point Prediction.

LHS

Latin Hypercube Sample.

LRIT

Long Range Inner Tier.

MSPH

Multiple Shot Probability of Hit.

Pavg

Transmitted average power [W].

xviii

PDF

Probability Density Function.

PH

Probability of Hit.

PID

Proportional Integral Derivative (Controller).

PK

Probability of Kill.

PK|H

Probability of Kill if Hit.

PNG

Proportional Navigation Guidance.

PS

Probability of Survival.

RAM

Rocket, Artillery, and Mortar.

RCS

Radar Cross Section.

RoF

Rate of Fire.

RPG

Rocket Propelled Grenade.

RPM

Rounds Per Minute.

RSE

Response Surface Equation.

SAM

Surface to Air Missile.

SEAL

Sensors and Electromagnetic Applications Laboratory (GTRI).

SLID

Small Low-cost Interceptor Device.

SNR

Signal-to-Noise Ratio.

SRIT

Short Range Inner Tier.

SSPD

Single Sortie Probability of Damage.

SSPH

Single Shot Probability of Hit.

THEL

Tactical High Energy Laser.

TIF

Technology Impact Forecasting.

USAF

United States Air Force.

xix

SUMMARY

Systems engineering introduces the notion of top-down design, which involves


viewing an entire system comprised of its components as a whole functioning unit.
This requires an understanding of how those components efficiently interact, with
optimization of the process emphasized rather than solely focusing on micro-level
system components. The traditional approach to the systems engineering process
involves requirements decomposition and flow down across a hierarchy of decision
making levels, in which needs and requirements at one level are transformed into a
set of system product and process descriptions for the next lower level. This top-down
requirements flow approach therefore requires an iterative process between adjacent
levels to verify that the design solution satisfies the requirements, with no direct flow
between nonadjacent hierarchy levels.
This thesis will introduce a methodology that enables decision makers anywhere
across a system-of-systems hierarchy to rapidly and simultaneously manipulate the
design space, however complex. A hierarchical decision making process will be developed in which multiple operationally and managerially independent systems interact
to affect a series of top level metrics. This takes the notion of top-down requirements
flow one step further to allow for simultaneous bottom-up and top-down design, enabled by the use of neural network surrogate models to represent the complex design space. Using a proof-of-concept case study of employing a guided projectile for
mortar interception, this process will show how the iterative steps that are usually
required when dealing with flowing requirements from one level to the next lower
in the systems engineering process are eliminated, allowing for direct manipulation

xx

across nonadjacent levels in the hierarchy.


In many cases, the analysis tools used to model those particular system components can be very complex and require a nontrivial amount of time for a single execution. Neural network based surrogate models are emerging as a viable method to
capturing complex design spaces in the form of surrogate models. Unlike polynomial
based surrogate models which are widely used for data regression, neural network
based surrogate models give the user the ability to model highly nonlinear, multimodal, and discontinuous design spaces. The use of neural network based surrogate
models to represent more complex analysis tools serves as the prime enabler of rapid
cross hierarchical manipulation for the methodology developed in this thesis. This is
done using a Monte Carlo simulation to sample from bounded distributions on the
inputs used to create those surrogate models, populating the design space with the
resulting responses. When designing to a specific goal in an extremely multimodal
and/or discontinuous design space, this method is more efficient than making use of
an optimizer that relies on smooth continuous functions and which often only finds
local optima. This gives the user the ability to apply requirement constraints in real
time, revealing how the feasible design space of every other response and variable
changes, all the while maintaining a family of design options.
As a proof-of-concept, the methodology developed has been applied to the specific
case of designing a gun launched guided projectile to intercept a small, indirect fire
mortar target. It has not traditionally been the function of air defense weapons to
defend against these types of unguided, low radar cross section targets. For this type
of threat, one would ideally defeat the source of fire, however, this may not be a
feasible option in a densely populated area where the attackers can quickly blend in
with civilians, therefore a midair intercept is required. Notional metrics of interest
identified for this type of air defense system include a balance between effectiveness
against the target, the mobility to be moved to new locations as threats arise, all while

xxi

minimizing cost. The 6-DoF trajectory modeling and simulation environment used
to study this weapon will be constructed as part of an effort underway at the Georgia
Tech Research Institute and Georgia Tech Aerospace Systems Design Laboratory to
meet this urgent requirement by the United States Army. This 6-DoF tool gives
projectile performance against a mortar target, as a function of errors introduced
by the gun and radar. In addition, the radar subsystem properties that define the
radar measurement accuracy will be defined as part of the functional decomposition
process.
For this system-of-systems environment comprised of a Monte Carlo based design
space exploration employing rapid surrogate models, both bottom-up and top-down
design analysis may be executed simultaneously. This process enables any response
to be treated as an independent variable, meaning that information can flow in either
direction within the hierarchy. For bottom-up design, the effect of manipulating
weapon design variables on weapon performance (or other metrics of interest, such as
cost), as well as the effect of the weapon performance on higher level metrics within an
operational environment (for example, the combined effect of using multiple weapons
against a single target) can be evaluated. For top-down design, placing limits or
constraints on high-level metrics will drive the required weapon performance, with
the effects cascading all the way down to the radar subsystem level properties, showing
the ability to conduct rapid top-down design among nonadjacent levels in the design
hierarchy. When the design space is completely populated, any design region may be
visualized using a multivariate scatter plot. As constraints are now used to limit the
remaining design space, the identification of not just the variables that are key drivers
to response variability, but also those variables that are dependent on the settings of
other variables is possible.

xxii

CHAPTER I

MOTIVATION
1.1
1.1.1

Introduction
Top-Down Systems Engineering

Systems engineering involves the orderly process of bringing a system into being, and
is meant to provide a structure for solving design problems and tracking requirements
flow through the design effort [35]. A primary function of the systems engineer is to
document the requirements of a system, which involves laying out the initial system
design in such a manner that the system will do what it is intended to do - for as
long as it is required to operate. As Blanchard [40] states:
Broadly defined, system engineering is the effective application of scientific
and engineering efforts to transform an operational need into a defined system configuration through the top-down iterative process of requirements
definition, functional analysis, synthesis, optimization, design, test, and
evaluation.
Systems engineering introduces the notion of top-down design, which involves
viewing an entire system comprised of its components as a whole functioning unit
[40]. This requires an understanding of how those components efficiently interact,
with optimization of the process emphasized rather than solely focusing on micro-level
system components [47]. This drives the need for having a complete identification
of system requirements, and relating these requirements to specific design goals and
developing of appropriate design criteria. However, it is difficult to develop models
that are sufficiently general to be applied to multiple systems [50]. Follow-on analysis

may be required to ensure system effectiveness of early decision making within the
design process. This last point can best be described in this context using the on
going paradigm shift in conceptual design (discussed in further detail later in this
chapter). The current, or traditional way of doing systems design involves very
little, if any, front end engineering analysis effort. This requires much greater design
effort by individual component designers downstream in the system design process,
or even in the life cycle of the system itself. This after-the-fact decision-making
may not necessarily integrate well with other design activities and may require many
modifications, which may prove expensive in both time and money. Therefore, rather
than rely on individual point design solutions, it is desired to keep families of solutions
for a given design space of interest giving decision makers the flexibility to observe
the impacts of various requirements and constraints on the entire solution space.
To summarize, the entire system design and development process requires an interdisciplinary effort to meet all design objectives effectively. This requires a complete
understanding of the various design disciplines, and most importantly for the systems
engineer, how the interrelationships between those disciplines affect overall system capability. The same conclusions can be drawn for the interrelationships of the components within a system-of-systems, where individually operated and managed systems
interact to affect an overall metric (discussed in further detail in Chapter II).
1.1.2

Initial Research Formulation

In the authors opinion, part of developing a novel systems engineering process is in


the investigation of its application to the development of a revolutionary concept,
and then identifying deficiencies that the new process addresses in whichever process
is currently in place, accepted, or in practice. The desire now is to find out how the
related community approaches applying the systems engineering process to a complex
problem, leading to the first research question.

Research Question 1: What is the systems engineering process that can


be applied to set the design requirements of a complex hierarchical system,
and what are the major deficiencies associated with that process?
For example, designing an air defense weapon system to complete a function in
an operational environment not usually associated with air defense weapons, while
accounting for uncertainties throughout the design hierarchy, requires several different systems to function together efficiently, and drives the need to utilize a top-down
design approach and view the entire process as the system. As will be explained
in more detail throughout this chapter, the specific case of launching a gun fired
guided projectile to intercept a small, indirect fire target will serve as the primary
proof-of-concept of the methodology developed. To help answer this question, several
sources were discovered leading to an accepted process. The Department of Defense
defines three characteristics of systems engineering [113]. First, systems engineering
transforms an operational need into a description of system performance parameters
and a system configuration. This process is achieved through an iterative process
of definition, synthesis, analysis, design, test, and evaluation. Second, systems engineering integrates related technical parameters and must ensure compatibility of the
physical, functional, and program interfaces. This is to be done in such a manner
as to optimize the total system definition and design. Finally, systems engineering
integrates reliability, maintainability, safety, survivability, human, and other related
factors into the total effort to meet specific program objectives, such as cost, schedule,
and technical performance goals.
In an effort to apply the general systems engineering process to the design of an
air defense weapon capable of defeating small targets, Baumann [35] breaks down
the systems engineering process into a hierarchy of decision making levels, using

information from multiple sources. The highest level is the overall capability level 1 ,
which identifies the overall need or function that must be addressed. This system
concept description is created with the intent of meeting a requirement at the overall
capability level. Next below is the system level, which produces a system description,
for example a performance requirement. At the lowest level is the subsystem level that
produces a subsystem performance description. Baumann states that the systems
engineering process is applied to each level in the design hierarchy, one level at a
time. This is a top-down, comprehensive, iterative and recursive process and is
applied sequentially through all stages of development. It transforms needs and
requirements into a set of system product and process descriptions. Figure 1 takes
this systems engineering concept, showing the iteration step that must be taken across
each hierarchical level. This figure depicts an important fact about this systems
engineering process: requirements are decided upon and flowed from the top-down in
the design hierarchy, and at each level there must be an iteration to make sure that
the design solution satisfies the requirement, one level at a time.
In addition, Figure 1 applies this top-down structure to a notional air defense
weapon, where the primary weapon system metric of interest is the probability of
hitting a target. This notional weapon system may consist of a projectile, a gun
firing that projectile, and a radar that tracks both target and interceptor trajectory
throughout the engagement. Each one of those systems has an individual performance
capability, for example the gun firing the projectile may experience pointing errors,
or the radar may have some error in the location measurements of the interceptor or
target. Therefore, only certain combinations of the individual system capabilities will
achieve the top level capability requirement, and the iteration process is not required
between that high level goal and each system. Once the system level requirements are
1

Baumann [35] actually refers to this first level as the concept level, however for the purposes of
remaining within the context of this thesis, it is referred to in this document as the overall capability
level.

Produces system level


requirements

Overall Capability Level


Design loop
iteration
Verification that system
properties enable concept
level requirements

System Level

PHit

Produces subsystem
level requirements

Design loop
iteration
Verification that subsystem
properties enable system
level requirements

Radar System
Capability

Gun System
Capability

Projectile System
Capability

Radar
Subsystem
properties

Gun
Subsystem
properties

Projectile
Subsystem
properties

Subsystem Level

Figure 1: Systems Engineering Process with Top-Down Hierarchical Requirements


Flow (adapted from [35]) with Process Applied to a Notional Problem
established, those requirements are then flowed down to the subsystem level properties
that enable that individual system performance capability. For example, the radar
measurement error can be a function of many of the radar properties, such as antenna
gain, operating frequency, transmitted power, etc... Now the next level of iterations
is required between each of the required system performance metrics and the variable
space among the subsystem level properties.
To further address the question posed in Research Question 1, there are clear
deficiencies associated with the traditional systems engineering process as applied
to systems that themselves are composed of several component systems that must
function together, but contribute individually to the overall process. There must be an
iteration between adjacent levels of the hierarchy to make sure that the design solution
satisfies the requirement, therefore there can be no direct flow between nonadjacent
levels. Therefore, the next Research Question asks how an improved process can be
created, encapsulating the overall research objective of this thesis.
Research Question 2: How can a process be created that enables rapid
manipulation across a hierarchy of decision making levels?
This research objective provides the primary motivation for the direction of the
literature search chapters in this document. The example discussed of designing an
air defense weapon requires several different systems to function together efficiently
5

within an operational environment in which air defense weapons are not usually used.
Each of these systems has its own uncertainties associated with it that affect the
combined systems overall capability. In addition, the operational environment may
comprise of several high level requirements, such as some measure of effectiveness
against a target at a given cost. This requires a more holistic approach to the problem,
rather than focusing on the optimization of any of the individual system components.
1.1.3

Overview of Thesis

With the primary research objective introduced in Research Question 2 and the goal
of applying this process to the example problem introduced in mind, Chapter II will
be dedicated to finding general methods to model and quantify air defense weapon
system capability, but do not in themselves allow for a solution to the specific example problem. Chapter III will be dedicated to those advanced design methodologies
necessary to fulfilling the primary research objective, addressing the deficiencies in
the current process. The motivation discussed in Chapter I and literature reviewed
in Chapters II and III led to a another series of research questions and hypotheses
discussed in Chapter IV. The questions will be used to refine the problem of interest
into specific challenges that must be addressed, and the hypotheses will generally
address how to answer those questions, based on information learned throughout the
literature search. Chapter V will outline the research formulation and the modeling
and simulation approach taken to prove the hypotheses presented. This chapter will
also focus all of the information gathered during the literature search into a coherent
problem within the scope of this thesis. Chapter VI will discuss the implementation
of the methodology developed to solve the problem of interest, as well as any key
findings. Finally, Chapter VII will review the research questions with the major elements of the results discussed that prove or disprove the hypotheses. This concluding
chapter will also present some recommendations for further study relating to this

thesis.
First however, the remainder of this chapter will provide the reader with some
background for an example problem of current interest, benchmarking current proposals for a solution to this problem, and identifying those proposals shortfalls which
drive the need to utilize a revolutionary concept. It is the discovery of a novel systems engineering process to investigate this revolutionary concept that motivates this
thesis.

1.2
1.2.1

Background
The Rocket, Artillery, and Mortar (RAM) Threat in Asymmetric
Warfare
The Army is fast-tracking a lot of technology to send it into Iraq because
we are suffering casualties there. Normally, these things would be tested
for one, two, three or four years before they were put into the field. So,
we are going to be putting many of these new systems into the field in Iraq
without complete testing, and some of them are going to work and some
of them are probably not going to work very well, but I think it is worth
our trouble...any American life saved is worth the effort...Although it is
not the ultimate solution, of course, we dont have anything now that will,
with 80 to 90 percent accuracy, take these mortars out.
Brig. Gen. Nick Halley (U.S. Army, Ret.), March 2005 [21]

The 2006 Quadrennial Defense Review [26] compiled by senior civilian and military U.S. Department of Defense (DoD) officials discusses attempts of non-state actors
to choose irregular warfare in an attempt to break the will of the U.S. military in a
protracted conflict. The DoD is therefore refocusing its portfolio of capabilities to
address irregular and disruptive challenges in addition to sustaining the ability to
address traditional challenges.
7

The [National Defense Strategy] acknowledges that although the U.S. military maintains considerable advantages in traditional forms of warfare,
this realm is not the only, or even the most likely, one in which adversaries will challenge the United States during the period immediately
ahead. Enemies are more likely to pose asymmetric threats, including
irregular, catastrophic, and disruptive challenges.
Asymmetric strategies toward warfare are strategies that do not directly engage
the full armed forces of a military power, instead taking advantage of anonymity,
patience, and willingness to suffer casualties to counter an adversarys advantage in
firepower [123]. The belief that the U.S. has a very limited willingness to suffer
causalities may drive an adversary to employ asymmetric strategies. Insurgents usually employ weapons in a war zone that are available in large numbers and simple to
operate, and the threat from rockets, artillery, and mortars (RAM) is ever growing.
This approach is certainly not a new idea, such as the Vietcong guerrilla attacks
on U.S. soldiers during the Vietnam War, however, Vietcong use (and many other
historical uses) of asymmetric warfare was in addition to, not in lieu of, a nationally
organized military.
The collapse of the Soviet Union strengthened the position for the use of asymmetric warfare, as it was a large driver in the creation of many new centers of arms
production in the world making available large numbers of highly sophisticated conventional weapons to anyone willing to pay for them. The Persian Gulf War in 1991
clearly exemplified how the United States, arguably having the most advanced military in the world, likely could not be beaten in a war, in the traditional sense. This
traditional military advantage is likely a major factor that makes asymmetric warfare
strategy more appealing to insurgents.

The U.N. intervention in Somalia is an early example of this premise. The broadcasts of the bodies of American soldiers getting dragged through the streets of Mogadishu on the news had an immense impact on the American public, limiting its willingness to sustain additional casualties. The civil unrest after Operation Iraqi Freedom is an example of how insurgents with minimally capable conventional weapons
are able to successfully stage operations against American troops and interests, as
well as attempt to cripple Iraqs fragile economy. This type of warfare attempts to
show the world that the U.S. may quickly be victorious in a full scale conventional
war, but not when trying to keep the peace in an asymmetric warfare environment.
1.2.1.1

The Baghdad Green Zone

The Green Zone, formally known as the International Zone, is the heavily guarded
area of closed-off streets in central Baghdad, Iraq, where U.S. authorities live and
work, and includes the main palaces of former President Saddam Hussein (Figure 2).
This area houses the civilian ruling authority, as well as the military presence run
by the Americans and British, and includes the main offices of major U.S. consulting
companies. This area is commonly referred to as the ultimate gated community
because of the numerous armed checkpoints, razor wire coiled and chain linked fences,
and numerous reinforced blast-proof concrete slabs. Security is very tight, and American and British officials rarely leave the Green Zone, only doing so with many body
guards and armored vehicles. Civilian traffic is very limited through the Green Zone
because of the many terrorist attacks against both Coalition and Iraqi security forces
and civilians [22].
The Green Zone is considered to be the safest area within Baghdad, however,
terrorists are no longer limiting themselves to planting bombs. On September 27,
2003, insurgents fired three rockets at a Baghdad hotel housing American soldiers and
civilians [22]. The attacks did very minimal damage, as the hotel stands hundreds of

Figure 2: Green Zone in Baghdad, Iraq [22]


yards from high, earth-filled barriers circling that area of the Green Zone. This was
the first terrorist attack of this type on coalition forces in this area, and mortar and
rocket-propelled-grenade (RPG) attacks are now a common daily occurrence. New
defenses against these asymmetric rocket and mortar attacks are sorely needed.
1.2.1.2

The Need for RAM Defense

Historian Major General Robert H. Scales Jr. describes in an interview that in


limited liability wars, most Americans killed in combat were killed by rudimentary
weapons; the greatest killer of Americans on the battlefield is the mortar [77]. He
states that a historical pattern has emerged since the Korean War that will continue:
most American soldiers killed in combat will die while facing an enemy fighting a
close fight on equal terms using simple weapons. General Scales further points out
the irony that a B-2 bomber has the ability to fly transcontinental distances to destroy
a building with one bomb and return home safely, but that a platoon under mortar
fire is relatively helpless.
As a contemporary example if this, the Green Zone is very large and requires

10

many types of defense systems, such as M1 Abrams tanks, Bradley fighting vehicles,
and 0.50 caliber machine gun armed HUMVEEs. However, these defense systems
are not capable of defending against very small targets such as those posed by the
RAM threat. It is the very basic nature of these weapons that make them so difficult
to defend against. For example, once fired, mortars such as those shown in Figure
3 follow a simple ballistic trajectory. Because mortars do not rely on a guidance
package, traditional missile countermeasures can not be used, such as chaffs against
radar guided projectiles and flares against infrared guided projectiles. Fortunately,
the use of these dumb projectiles usually results in a very small amount of direct hits
on an intended targets. However, their effective range and the compactness of their
launchers, as shown in Figure 4, means that a terrorist may stand off and fire from
several kilometers away, making finding the source of fire a difficult task. Mortars
provide a source of indirect fire, meaning that an attacker does not aim directly
at an intended target. These types of unguided weapons provide rapid, high-angle,
steeply dropping fires that have historically proven beneficial to military commanders
fighting against dug-in enemy troops or targets with uncertain location, which are far
less vulnerable to direct fire. Most mortars and launchers are light enough to be
carried by backpack across all terrain.
There is a very urgent need for a weapon system that is able to intercept these
small projectiles in mid-flight. Further, the reality of the situation is that this defense
system needs to defend against several types of attacks from many directions simultaneously. This weapon system needs to be able to detect multiple targets with small
radar cross sections (RCS), establish trajectory track for each target, hand off target
position to the fire control system, engage each target, and successfully intercept each
target with enough kinetic energy to destroy it. Of course, all this must be executed
in a very short time (on the order of several seconds, depending on the threat range).
All these requirements clearly drive the need for high accuracy and fast response

11

M252 81mm Medium Extended Range Mortar

FAS | Military | DOD 101 | Systems | Land Warfare ||||

Figure 3: 60mm (top left), 81mm & 120mm (top right) Fin-stabilized and 4.2 inch
Spin-stabilized Mortars (bottom) (photos from FAS [1])

M252 81mm Medium Extended Range M

The M252 81mm Mortar System


developement agreement with the
M29A1 Mortar. A Blast Attenuati
the muzzle of the cannon assembl
the mortar crew. The M252 is idea
air assault, mountain and light inf

Figure 4:

Length: 56 inches (142.24 centim


Weight:
Mortar Assembly: 35 pounds (15
Bipod: 26 pounds (11.80 kilogram
Baseplate: 25.5 pounds (11.58 kil
Sight Unit: 2.5 pounds (1.14 kilog
The M252 81mm
Medium
Extended
Range
Mortar
(photo from FAS [1])
Total: 89 pounds (40.41 kilograms)
Bore diameter: 81mm
Maximum effective range: 5700 meters
Minimum Range: 80 meters
12
Rates of fire:
Maximum: 33 rounds per minute

at each level of the engagement, however put simply: hitting many small far away
targets is a very difficult task.
1.2.2

Problem Scope: Extended Area Protection & Survivability


The Extended Area Protection and Survivability (EAPS) Science and Technology (S&T) Program Architecture Study will provide concept definition,
requirements analysis, functional analysis, and design synthesis, and establish a baseline architecture to enable stationary/mobile 360 degree hemispherical extended area protection and survivability of future Army combat unit assets from direct and indirect fires, including rocket, artillery
and mortar (RAM), Unmanned Aerial Vehicle (UAV), and Cruise Missile (CM) threats while minimizing cost, weight, and space allowing for
improved deployment and mobility. Currently, no existing or programmed system has the capability to negate unguided rockets and mortar/artillery projectiles with mechanical fuzes after they are launched. Only a limited capability exists to negate UAV
reconnaissance capabilities at sufficient range to preclude detection, identification, and targeting. Protection and survivability capabilities in Army
combat units faced with this threat needs to be demonstrated to deter or
defeat those enemy capabilities. [138]

Bill Nourse, program manager of the U.S. Armys Extended Area Protection &
Survivability (EAPS) program, describes the area air defense operational scenario as
being broken down into an inner tier, comprising of the ranges up to 10 km, and
an outer tier for defended ranges beyond 10 km [120]. The outer tier is focused on
defending against large threats, such as ballistic and cruise missiles, unmanned aerial
vehicles, and rotary wing aircraft. The inner tier is primarily focused on the smaller
threats, such as rockets, artillery, and mortars, and potentially even on outer tier

13

Figure 5: Extended Area Protection & Survivability Operational Scenario [120]


leakers. As shown graphically in Figure 5, the inner tier is further broken down
into a Short Range Inner Tier (SRIT) for ranges between 500-2000 m and a Long
Range Inner Tier (LRIT) for ranges between 2-10 km.
Nourse describes several of the potential advantages of solutions addressing each of
the inner tier options. A solution required to only defend the SRIT has the attributes
of earlier fielding and lower unit cost when compared to a solution designed to defend
the LRIT. In addition, threats that have very short times of flight may only operate
within the SRIT. However, the ability to protect moving assets may only be addressed
by a solution that can defend at the LRIT level. Nourse suggests that a mix of both
short and long range shooters might be the most cost effective solution.
In [35], Dr. Jim Baumann, the Deputy EAPS program manager, lays out the basic
air defense weapon engagement kinematics as given in Figure 6. This is independent
of the type of interceptor solution, for example a gun fired projectile can be launched
to intercept the target, or a directed energy (DE) weapon can be used to heat up
the targets warhead causing detonation. The basic timeline of events is as follows.

14

Engagement Kinematics
Threat at Interceptor Launch
or DE System Lasing Initiation

Threat at FCR Firm Track


Threat at
SR Firm Track

SR

Threat Kill

FCR

Interceptor Launcher
or DE System

Interceptor Flight Time


or DE Dwell Time

Defendable
Asset

Defendable
Radius (DR)

Keepout
Distance (KOD)
Keepout
Time (KOT)

Threat
Launcher
System Reaction Time (SRT)
(Threat launch to Interceptor Launch)
Threat Flight Time (TFT)

Time Line Relationships:


-For an Interceptor System:
Interceptor
Flight Time
= TFT - SRT
FigureAllowable
6: EAPS
Engagement
Kinematics
[35]- KOT
-For a DE System:
Allowable DE Dwell Time
= TFT - SRT - KOT
Radius
(DR) (SR)
Relationships:
ADefendable
surveillance
radar
detects a threat fired from a certain distance with the
-For an Interceptor with average velocity = Vi:
-For
a DE of
system
withsome
reqd dwell
= f1 (range):
intention
hitting
defendable
asset,

DR = Vi*(TFT - SRT - KOT) - KOD


= f2track
(TFT - is
SRT
- KOT) - KOD A
and DR
firm
established.

flight

-There isradar
a close(FCR)
analogyestablishes
between the akinematic
relationships
for interceptors
andto
DEestablish
systems. a
control
firm track
of the threat
trajectory

-The term (TFT-SRT-KOT) is key to the defense solution for both system types.
Radius is directly related to this term
fire-Defendable
control solution.
When a solution is established, the air defense interceptor is
-Combinations of short TFT, long SRT, or long KOT can make threats unengageable.
Jan 20-21, 2005
Session 1-2 - Systems Engineering
launched (or if a DE weapon is used, system lasing initiation begins). The time14up to

this point is known as the System Reaction Time (SRT). The interceptor flight time,
or DE lasing dwell time, determines the point at which the threat is killed, and the
distance to the defendable asset is the Keepout Distance.

1.2.2.1 Gun vs.


n 1-2 - Systems Engineering

Rocket Interceptor Solutions

Baumann [35] compares the use of both gun and rocket interceptor solutions against
a notional RAM type target in Figure 7. The red curve shows the intercept velocity
versus intercept range for a notional gun firing a 20 mm projectile with a muzzle exit
velocity of about 1100 m/s. The black curve shows the same profile for a notional
70 mm bore, 11 kg hypervelocity rocket (HVR). The straight blue curve gives a
postulated lethality requirement achieved with a kinetic energy delivered at 800 m/s.
Baumann concludes that a gun or rocket system would provide the necessary kinetic
energy to destroy this notional target at the shorter range SRIT requirements (0-2
km). From the notional velocity profiles, note that a gun solution is especially better
than a rocket at delivering kinetic energy at very short range, since the rocket requires
15

Typical Velocity Profiles of


Guns and Rockets
Velocity at Intercept

Average Velocity to

1400

70 mm, 11 kg Hypervelocity Rkt (HVR)

Velocity, m/s

1200
1000

Postulated Lethality
Requirement (KE)

800
600

20 mm Gun

400
200
SRIT Country

0
0

Average Velocity, m/s

1200
1000

600

Postulated Kinematic
Requirement (122-mm
rocket @ 10 deg QE)

400
200

LRIT Country

70

800

SRIT Country

Intercept Ran

Intercept Range, km

Figure 7: Typical Intercept Velocity Profiles of Guns and Rockets [35]


-SRIT requirements can likely be met with either gun or rocket systems.

a certain distance after launch


achieve
maximum
velocity,
where Its
as average
a gun fires
its to 6 km is ab
-HVR isto
typical
of high
performance
rockets.
velocity

- It could meet the intercept velocity (postulated lethality) requirement out t

projectiles at maximum velocity.


the rapid
velocity
decay ofkinematic)
a bullet as
- It couldHowever,
meet the average
velocity
(postulated
requirement out

soon as it leaves the gun-An


muzzle
severelysolution
limits the
of kinetic
energy
can
LRIT missile
mayamount
be required
to exceed
HVRitperformance,
depen
selected defendable radius.

deliver at longer ranges such


as those
(2-10 km).
Jan 20-21,
2005 dictated by the LRIT
Session requirement
1-2 - Systems Engineering
In fact, the 20 mm gun solution appears to be able to deliver that required kinetic
energy up until about 2 km (the boundary of the SRIT/LRIT requirements), and the
rocket delivers the required kinetic energy up until about 3.5 km. A self propelled
solution such as a rocket may be required to maintain a certain velocity profile at
a ceratin range, depending on the amount of kinetic energy required at an intercept

Long Range Inner Tier Shooter


drives the weapon design requirements.
Trades
range. This clearly shows how the required defendable radius for a single solution

1600

80

Radar Accuracy
Requirements
1400

70

1200

Missile Weight, kg

Required Interceptor
Average Velocity, m/s

1.2.2.2

60

1000
The performance of an interceptor
projectile, guided or unguided, highly50depends on
800

40

the accuracy of the fire control


radar to deliver it to the intended target. To
30 defend an
600
400

20

area against the RAM threat,


the radar must have 360 degree operation10in an urban
200
A

0
and mountain environment,
against
low4 RCS 6and low
trajectory
targets 0with short
2
0
2
8
10
12
Defendable Radius, km

35

40

30

16

Relative Power*Aperture

20

16

s per 500 kg Pod

R Firm Track Rng, km


ve Power*Aperture

times of flight (on the order of 12 to 15 seconds) [75]. These


requirements pose several
54
30
25
20
15

Defendable Radiu

Figure 8: Projectile Timeline Engagement - Typical Mortar [75]


dilemmas. Full 360 degree operation requires a very large coverage volume. The short
flight time of the threat requires a high radar search rate. However, finding a low RCS
target would likely mean a longer dwell time before detection. A notional projectile
timeline engagement against a typical mortar using a surveillance/fire control radar
is given in Figure 8. Note a defined minimum kill altitude requirement2 of 75 m.
Accurate handover to the weapon system requires high radar track accuracy, which
in turn drives the need for a narrow antenna beamwidth. Unfortunately, a narrow
antenna beamwidth does not readily allow for the large coverage volume and high
update rate requirements driven by the threat characteristics. Therefore, the surveillance and weapon handover radar design requirements are incompatible, and drive
the design in opposite directions.
The main driver for both surveillance and weapon handover requirements is radar
2

This 75 m minimum kill altitude is assumed for all scenarios studied in the implementation
discussed in Chapter 6.

17

Figure 9: Fire Control Angle Error versus Warhead Lethal Radius [75]
accuracy, which plays an important role as to how close the fired projectile actually
ends up to its intended target and its associated probability of direct hit. The use
of a warhead would increase the projectiles lethal radius, meaning that a projectile
could actually hit its intended target without directly intersecting flight paths. This
leaves room for error in the radar. Holder [75] therefore states that warhead size
will dictate radar accuracy requirements as given in Figure 9. The reader will note
that for a given target size, as the radar error is reduced, the required lethal radius
for a target hit is reduced until a certain point where error reduction does not affect
required lethal radius. But the converse can also be stated in that a given radar
accuracy might drive the need for a warhead of given lethal radius. The probability
of hitting the target with a single projectile will also dictate the number of projectiles
required to actually destroy the target. Therefore there is an important tradeoff space
to be examined between the lethal radius of one projectile, the number of projectiles
fired, and radar accuracy, for a desired probability of hit.

18

1.2.3

Approaches to the RAM Threat

As can be seen by the shift in how warfare is conducted, one of the most pressing
questions facing the modern day war planner is how to defeat the types of weapons
posed by the RAM threat. Ideally, one would try to eliminate the source of fire, but as
stated previously, this is a particularly daunting task. Therefore the only choice left
is to defeat the threat in mid flight, but as the program manager of the U.S. Armys
EAPS program states, currently, no existing or programmed system with a capability
to negate RAM projectiles after they are launched [120]. However, Macfadzean [92],
considered one of the primary authorities on the analysis of surface-based air defense
systems, states:
Surface-launched, direct-fire battlefield missiles are not viable targets for
air defense systems. The launcher must be disabled, but this is not an air
defense function.
Nevertheless, as enemy tactics have evolved, so does the required function of
air defense weapons. As a reaction to the mortar threat facing coalition forces in
the Green Zone within Baghdad, the chief of staff of the U.S. Army has issued a
directive for a near term fielding of a Counter Rocket, Artillery, and Mortar (C-RAM)
system [19]. This system is in direct response to an operational needs statement
from the Multinational Corps-Iraq. Each C-RAM system must identify, track, and
engage an incoming threat munition. Additionally, because this defense system will be
fielded in crowded areas, it must destroy the incoming threat with minimal collateral
damage. The Army has three systems proposed to defeat this threat: the 20 mm
Phalanx Close in Weapon System, a Navy anti-ship missile defense gun system; the
35 mm Skyshield, a Swiss rotary wing air defense system; and the Active Protection
System, developed by the Defense Advanced Research Projects Agency for anti-tank
missile defense. Additionally, the Northrop Grumman Corporation has proposed
19

using the Tactical High Energy Laser (THEL) developed for destroying Katyusha
rockets launched into northern Israel, to counter the total RAM threat. Northrop
Grumman has even conducted actual firing tests to prove that a single system can
destroy multiple mortars in mid air [18].
One main problem with any of these approaches is the rush to field a system.
Retired Army Brig. Gen. Nick Halley, the commanding general of artillery and
rocket forces for the 18th Airborne Corps during the first Gulf War, stresses in the
opening quotation to the background section of this chapter how the U.S. military
would be willing to sacrifice some performance and effectiveness for the ability to
deploy a C-RAM weapon system quickly [21]. The following sections will briefly
introduce these proposed systems.
1.2.3.1

MK 15 Phalanx Close-In Weapons System (CIWS)

The MK 15 Phalanx Close-In Weapons System (CIWS) built by the Raytheon Systems Company, is a fast reaction, rapid-fire 20 mm gun system designed to provide
naval ships a last-ditch defense against incoming antiship missiles or aircraft that
successfully penetrated longer range defensive systems, such as surface-to-air missiles
(SAM). A Phalanx system, shown in Figure 10, includes a fire control that has the
ability to automatically detect, track, and engage targets using an advanced searchand-track radar integrated with a stabilized, forward-looking infra-red (FLIR) detector giving a multi-spectral capability. It can also provide an additional sensor and
fire-control capability to an existing combat system. Finding specific characteristics
on any weapon system used by the U.S. military is very difficult and often conflicting in open-source literature. Therefore, the Phalanx characteristics and description
listed in this section are for the latest Block 1B standard derived from Miller [114],
and summarized in Table 1.
The Phalanx fire-control system consists of two high frequency Ku-band radars

20

(discussed later in Chapter 2), which are beneficial when used to track small objects in
a high clutter environment [96]. One radar is used to track a target, and the second is
used to track its own projectile stream. An internal computer is used to compute the
necessary elevation and azimuth required to eliminate the difference between the two,
and deliver the projectiles fired to the target. The Phalanx employs a Vulcan Gatling
gun element originally designed for use in aircraft. A six barrel cluster rotates at high
speed firing projectiles at either continuous fire (4500 rounds per minute) or at burst
lengths of 60 or 100 rounds. The gun fires a 20 mm armor-piercing discarding sabot
(APDS) projectile consisting of a 15 mm subcaliber tungsten or depleted uranium
penetrator. The Phalanx has been issued at several block standards. The latest Block
1B standard introduced many improvements to the gun system, but most notable to
this thesis is a man-in-the-loop facility, which makes the system more suitable for
use in engagements against small high-speed surface ships, small aircraft, helicopters,
and surface mines.
A land based version of the Phalanx, shown in the right side of Figure 10, was
selected by the U.S. Army as the interim C-RAM interceptor (and given the C-RAM
designation). Tests conducted in December of 2004 showed a 60 to 70 percent shoot
down capability of a live mortar [23], and subsequently two of these C-RAM systems
were deployed to Iraq in August 2005 [39]. Note that these open source performance
reports do not mention the size of the target, the intercept range, and the number of
bullets fired (typical bursts of rounds number in the hundreds). As will be discussed
much later in this thesis, preliminary studies to determine the effectiveness of a land
based Phalanx weapon system using some of the methodologies introduced in Chapter
3 are given in Appendix B.2.

21

Table 1: MK 15
Property
Type
Date deployed
Gun type/caliber
Weight
Range
Guidance system
Type of fire
Elevation/depression
Ammunition
Magazine capacity

Phalanx Close-In Weapons System Characteristics [114]


Description
Fast-reaction, rapid-fire 20mm gun system
Block 0: 1980; Block 1: 1988; Block 1b: 1999
M-61A1 Gatling; 20mm
12,500 lb
4875 ft (horizontal)
Self-contained search-and-track radar with integrated FLIR
4500 rounds/min (Block 1 onwards)
+80/-20 degrees
Armor-piercing discarding sabot, tungsten subcaliber penetrator
1550 rounds (Block 1 onwards)

Figure 10: MK 15 Phalanx Close-In Weapons System [1]

22

and precision-guided weapons (f.i. anti-radiation missiles).

Up
AHEAD System

A fully modular SHORAD system, the Skyshield 35 is based on two


innovative, extremely accurate and reliable 35 mm Revolver Guns and
a Fire Control System consisting of a Sensor Unit and a detached
Command Post.

35/1000 Revolver Gun

The Sensor Unit is b


and employs a mult
precision tracking.
In case of need for a
complemented by in
Due to its small, com
standard platforms f
be rapidly and easil
trailers, railway, airc
Subprojectiles befor
Subprojectiles.

Modular Skyshield 35 Air Defence System

Figure 11: The Skyshield 35 Air Defence System [2]


1.2.3.2

Skyshield 35 Air Defence System

The Skyshield 35 Air Defence System, manufactured by the Swiss company Oerlikon
Contraves [2], was designed for short range air defense against aircraft and cruise
missiles. A fully modular Skyshield 35 shown in Figure 11, is based on two 35/1000
Revolver Guns firing 35 mm projectiles at a rate of 1,000 rounds per minute in 24
round bursts and a fire control system consisting of a sensor unit and a detached
commandhttp://www.oerlikoncontraves.it/Difesa/eng/Skyshield.htm
post. The sensor unit has an X-band search and track radar
(discussed
6/26/2005
later in Chapter 2) that employs a multi-search unit for radar/TV and/or Laser/FLIR
precision tracking. The entire system was designed for rapid deployment by means
of trucks, trailers, railway, or aircraft.
The 35/1000 Revolver gun was designed to fire standard 35 mm ammunition, as
well as the Ahead Ammunition, show in Figure 12. The Ahead round was specifically
developed against small, low, and fast flying targets (here the manufacturers intended
idea of small is about the size of a cruise missile). The Ahead round itself, also
manufactured by Oerlikon Contraves, is a 35 mm caliber projectile that consists
of 152 tungsten submunitions that disperse at a given fuze distance ahead of a
target. Each projectile is programmed by a muzzle based electromagnetic inductor
that sets an electric timer as the projectile exits the gun, to activate and separate
the projectile into its 152 submunitions [20]. Weapon characteristics of interest are

23

# Submunitions
Submunition mass (each)
Payload mass
Projectile mass (total)
Muzzle Velocity

target. The basic characteristics used to model this projectile are


given in Table II. The mass properties and aerodynamic
coefficients were taken from range tests of the weapon system
reported by Oerlikon Contraves. The gun selected for projectile
analysis was the standard 35mm air defense gun used previously
by Oerlikon Contraves weapon systems; this sets the twist rate for
spin-stabilized projectile analysis.

152
3.3
500
750
1050

g
g
g
m/s

Figure 3. Ahead 35mm Ammunition.5

Figure
12: Ahead
35mm
Ammunition
[2]
A simple iterative fire control
solver written
using the
PRODAS
integrated scripting
engine which initially
assumes flat-fire and repeatedly corrects the initial inputs, was used to determine the correct gun elevation angle to
intercept each of the desired points. The time-to-target was also determined by the solver, establishing an ideal,
Table 2: Ahead 35 mm Ammunition Characteristics [2]
error-free trajectory and its fuze time. Since the goal of this study was to examine the effects of system error, all
Property
Description
initial conditions not under investigation were
set to their ideal
values. Thus, there was no initial pitch, yaw, pitch
rate, or yaw rate; the projectile was assumed to leave
the
muzzle
perfectly parallel to the muzzle axis. Realistically,
Calibre 35 mm
this error-free condition is not possible, but the proper analysis of the interior ballistics required to determine the
Length 228 mm
appropriate amount of muzzle error was beyond the scope of the present investigation. The PRODAS trajectory
No.Submunition
analysis used was a six degree of freedom
(6DOF) fixed 152
plane code. The 6DOF codes are the most accurate
trajectory simulations available inSubmunition
PRODAS and still
run extremely
quickly on modern hardware, on the order of
mass
3.3 g (each)
one second per simulation. As previously
indicated,
aerodynamic
were taken from manufacturer
Payload mass 500 g (allcoefficients
submunitions)
provided data, and the ICAO standard atmosphere was employed, setting the gun site at sea level. At each target
Projectile
mass time
(total)
750 by
g the fire control solver to eliminate fuze time as
point, all simulations were stopped
at the fuzing
determined
Muzzle
velocity
1050
a source of error. Each resulting trajectory was output into a file form/s
later processing.
C. Design of Experiments
provided
Table
Preliminary
studies
methodologies
introduced
Chapter
Because
of the in
many
trials2.
required
by the Monte
Carlousing
simulation,
a metamodel of
the modelingin
and
simulation
environment was used to minimize the overall run time. A metamodel is a model of a small segment of a more
sophisticated
analysis toolthe
based
on statisticalofinputs
and response
metrics.
This
model
allows
for a are
quick
3 to determine
effectiveness
the Ahead
round
against
very
small
targets
calculation of responses by employing simple equations to relate independent variables to responses. The
metamodel
used
this study was
a Response Surface Equation (RSE) in the form of a quadratic equation, given in
given
in inAppendix
B.1.
Eq. 1.

1.2.3.3

i =1

i =1

n 1

= bo + Laser
RSEEnergy
bi xi +
bii xi +
Tactical High
(THEL)
2

bij xi x j

i =1 j =i +1

(1)
n
= number of factors
Theb Tactical
High Energy Laser (THEL) is a high-energy laser weapon system de= regression coefficients for linear terms
i
bii
= regression coefficients for pure quadratic terms
signed
short
range for
rockets,
developed
bij to= defeat
regression
coefficients
cross product
terms for U.S. Space & Missile Defense
xi, xj = design variables or factors

where:

Command and the Israeli Ministry of Defense. The THEL uses a high-energy chemi-

4
cal laser that focuses its beam on an incoming
threat, the objective of which is to heat
American Institute of Aeronautics and Astronautics

up the warhead to point of detonation in mid-flight. The laser beam is created by


mixing fluorine atoms with helium and deuterium to generate deuterium-fluoride in
an excited state. A resonator is used to extract the deuterium-fluoride and transform
it into a coherent, monochromatic beam of light. FAS [12] claims a low-cost per kill
(about $3000 per kill), the ability to fire 60 shots with a single load of chemical
fuel, and a probability for kill near 1 at ranges within 5 km.

24

The main driving factor for the development of the THEL was the threat that
Israel was facing from Hezbollah terrorists firing Katyusha rockets at Israeli towns
across the border with Lebanon. These rockets fly very fast and at low altitude ballistic trajectories, and when aimed at crowded towns, can cause considerable damage.
These attacks grew so numerous that Israel could not use interceptor missiles, and
since the Katyushas flew on ballistic trajectories that landed on Israeli towns unless
completely destroyed, they could not deploy an advanced machine gun defense such
as the CIWS developed by U.S. Navy against low-flying cruise missiles. In April 1996,
Hezbollah fired over two dozen Katyusha rockets at Israel within 17 days. In response,
the THEL was designed to handle a large volume of attacks while maintaining a low
cost per kill. [8]
The fire control radar would be positioned near the hostile zone, continuously
scanning the horizon for threats. For a typical engagement scenario shown in Figure
13, a rocket is launched at a defended area. The THEL fire control radar detects
the incoming rocket and establishes trajectory information, which then hands off the
target to the pointer-tracker subsystem that includes the laser beam director. This
process tracks the target optically, then finely tracks the target for the beam director,
placing the high energy beam on the target. The transfer of energy from the highenergy laser causes intense heating of the target, causing the warhead to explode,
causing debris from the target to fall to the ground short of the intended impact area.
However, the weapon system has not progressed far beyond the demonstration
program. The systems lack of mobility limited the weapon systems effectiveness
against targets with various ranges and trajectories. As discussed earlier, a weapon
designed to counter the RAM threat must be able to be rapidly mobilized to protect
an area. Therefore, Northrop Grummen has proposed a that the weapon system
developed beyond the demonstration phase be a mobile version of the THEL, the
MTHEL, which would be able to be air transported on a C-17, C-5, or C-130 transport

25

THEL

Page 2 of 9

Figure 13: THEL Rocket Interception Process [17]


aircraft [17]. Although originally designed to counter rockets, such as Katyushas, the
THEL has successfully shot down multiple mortar rounds in August 2004 during test
firings at White Sands Missile Range, New Mexico, proving that the laser weapon
could be used to protect against common threats [18]. However, funding for the
MTHEL was cut in 2004.
Although no further attention will be paid to laser based weapons in this thesis,
the original motivation for the development of the THEL and the difficulty associated
with destroying small unguided airborne threats further motivates this thesis.
1.2.4

Guided Projectiles

1.2.4.1

Introducing the Concept

The basic objective of a guided projectile is to increase the accuracy of an otherwise unguided gun-fired projectile in hitting a target. Different approaches to the
http://www.israeli-weapons.com/weapons/missile_systems/systems/THEL.html
guided projectile solution have been examined. The projectile may 6/5/2005
carry a seeker, or

on-board tracking sensor that homes in on the targets infrared or radar signature.

26

Another solution may have a ground based radar sending updates to the projectile as
it approaches the target. These updates may include path corrections to adjust for
any uncertainties, such as gun firing errors, wind, target tracking location error, etc...
For example, Buzzett et al. [44] report that a notional unguided projectile with a Circular Error Probable (CEP)3 of 200 m could be reduced to under 10 m, using guidance
technologies for spinning projectiles using a body fixed guidance approach employing
fast impulse thrusters (for control) and a low cost on-board seeker. However, the use
of complex components that must be carried on-board the projectile may seriously
drive up the cost of a single round. Any on board components or moving parts (i.e.
moving fins for control) must be able to stand the high acceleration of being fired
from a gun, sometime in the tens of thousands of gs. Nourse [120] identified several
technology barriers to a guided projectile solution to the RAM threat, which include:
1. Low Cost Sensor : sensors for acquisition of small, short-range, and low trajectory RAM threats
2. Guidance: Miniature sensors and control mechanisms for low cost guidance
3. Lethality: Lethal kill mechanisms for defeating tough diverse RAM threat
4. Fire Control : Algorithms/network for timely impact point prediction/preferential
engagement of RAM threat
5. Shooter trade-off studies: Efficient and mobile hemispheric protection from engagement of multiple RAM threats arriving simultaneously
A good deal of academic work at the undergraduate level has been conducted
at the University of Alabama, including partnerships with other academic, industry,
and government institutions. There were specific efforts to equip students with the
specific technical skills related to the U.S. Army defined EAPS problem [43]. The
efforts of the students to design a guided projectile are discussed in [69] [72] [144].
3

Circular Error Probable (CEP) is the radius of a circle, centered on the desired mean point of
impact such that 50% of the impact points lie within it [57].

27

Owen [121] discusses the use of Response Surface Methodology to represent simple
trajectory calculations within more complicated 6-DoF codes, demonstrating a desire
to expedite the 6-DoF trajectory process. Tournes [136] discusses using Monte Carlo
simulations to quantify uncertainties in guided projectile flight control due to errors
in aerodynamics, actuators used to direct control surfaces, wind, radar properties,
and mass properties. Costello [51] discusses research at the faculty level at Oregon
State University on the dynamic modeling of guided projectiles. Cave [46] introduced
a system effectiveness analysis tool to evaluate guided interceptors (both gun-fired
and missiles). This tool is designed to provide results that may be used to define
defense system architecture and interceptor subsystem requirements.
1.2.4.2

Defense Industry Examples

In a Raytheon presentation explaining the theory of operations of guided projectiles,


Geswender [66] compares the different levels of guidance on impact point accuracy,
shown in Figure 14. The simplest level, the unguided round, clearly has the worst
accuracy (i.e. the largest impact area of uncertainty). As guidance degrees of freedom
are added, the impact area of uncertainty is reduced. In the development of the
Raytheon Extended Range Guided Munition (ERGM), Geswender points out that
the challenge is to build a projectile with all the functionality of a missile but robust
enough to [be] gun fired. The ERGM is a gun-fired, rocket assisted, GPS/INS-guided
127 mm projectile designed to be fired from the MK 45 5-inch gun for the U.S. Navy.
The rocket-assisted round extends the range of the 5-inch gun from 13 nmi to 63 nmi,
and the GPS/INS can achieve at accuracy of 10-20 m CEP at maximum range [14]
[124].
The Boeing Company is developing a defense system which provides protection
from missiles and artillery threats, designated the Small Low-cost Interceptor Device
(SLID) [7]. The SLID is designed to be a small, low-cost, fully self-contained active

28

1-DOF Correction
Controls X

2-DOF Correction
Controls X, Y

7-DOF Correction
Controls time, X, Y, Z, pitch, yaw, roll

Figure 14: Potential Projectile Guidance Corrections Effect on Accuracy [66]


defense system for military or other high value assets. This system comprises of several components, including a small maneuverable hit-to-kill interceptor, high-speed
launcher, passive threat warning sensor, and precision fire control system. Boeing
claims that the SLID is designed to defeat threats at standoff ranges up to 250 m,
which may include anti-tank guided munitions, high explosive anti-tank rounds, mortar rounds, and artillery shells. At a cost of $10,000 per round, the SLID relies on
a number of technologies, notably for this thesis is the use of an optofluidic control
system for maneuvering [36].
When compared to an unguided gun-fired round, the cost of guided munitions currently labeled low-cost would still be too expensive for widespread fielding (about
10-20 times). Even these cost predictions underestimate the cost, as most low cost
claims are based on overestimated production volumes that do not materialize when
actual procurement begins. Therefore, Horwath et al. [78] state that there is a need
to give up some precision to achieve low cost. Understandably, single shot accuracy
and system cost are inversely proportional to each other.
An effort is underway at General Dynamics to study Low Cost Course Correction
(LCCC) technologies [78]. The idea is to use off-the-shelf parts already used from

29

suppliers that build components in high volume. A restriction is also applied to not
use any moving parts, and therefore more suited to high-g gun launch applications.
Therefore, there will be several restrictions on performance, such as not being able
to employ pursuit guidance (i.e. chasing the target; discussed with the guidance
navigation laws in Section 2.6.2). Additionally, it would not likely yield a one shot
one kill solution, but if there is confidence of intercepting within a certain miss
distance of the target, then a multiple shot scenario could be used, the close is good
enough approach. Available literature shows that General Dynamics have applied
the LCCC methodology to such smaller 40 mm and 50 mm bore projectiles.
1.2.4.3

Guided Projectile Research at GTRI

The Georgia Tech Research Institute (GTRI) has been investigating the use of pin
based actuators to generate forces capable to steer small supersonic projectiles, shown
in Figure 15. This section will highlight key findings and information appropriate to
this thesis as interpreted from papers by Massey et al. [99], Massey et al. [101], Silton
[131], and Silton and Massey [132]. Control of the projectile is achieved through taking
advantage of the complex shock-boundary layer interactions produced by mechanical
devices, as shown in Figure 16. The forces created by the shock impingement on
the fins and body surfaces provide control forces through asymmetric lift. Angleof-attack control is achieved by deploying two pins placed close to an adjacent fin,
and rotation is achieved by deploying two diametrically opposed pins. Range tests
of these projectiles fired at Mach 4+ demonstrated the divert required for intercept
[100]. As shown in the shadowgraphs of live test firings in Figure 17, the robust pin
firing mechanism was able to provide control after withstanding the high g load of
being gun fired, and may be a comparatively low cost technology alternative for a
guided projectile.

30

Whyte, R., Hathaway, W. and Steinhoff, M., 2002:


Aerodynamic Analysis of the Hit-to-Kill (HK)
NSWC / ARL Projectile. ARL-CR-501, U.S. Army
Research Laboratory, APG, MD, July 2002.

Figure 4 Surface pressures for pin with proje


angle of attack, roll control configuration.
Figure 1 Guided munition defense of a forward base
against cruise missile attack.

Figure 15: GTRI Guided Projectile with Pin Firing Mechanism [101]

Guidance Pins

Figure 5 Shadowgraph of fired projectile at A


control pins. (CAD image superimposed.)
Shock Interaction
Figure 2 Pin-fin guidance concept (patent pending).

Figure 16: Pin-Fin Shock Interaction Guidance Concept [99]

Figure 6 Wind tunnel model of projectile.

Figure 3 Pins used for roll and directional control.

Figure 17: Shadowgraph Ensemble Showing Divert of 5Pin-Controlled Projectile


[100]
24th Army Sciences Conference

31

1.2.4.4

Concluding Remarks on Guided Projectiles

Clearly, the study of guided projectiles is still in relative infancy compared to mature
operational weapons systems. One issue related to the problem defined in this thesis
is that most of the literature about industry built guided projectiles is that they are
designed to hit fixed surface targets at ranges much farther than those outlined by
the EAPS SRIT requirements discussed in Section 1.2.2, which address the need to
defend against moving airborne targets. This leads to the design of very complicated
and expensive projectiles not suited to defend an area against the RAM threat.
1.2.5

Summary of Discussed Solutions

Table 3 reviews the different alternatives discussed as potential solutions to the RAM
threat, as well as several of the guided projectiles discussed. As discussed in Section
1.2.2.1, both a gun fired projectile and a self propelled rocket (or missile) could
deliver the necessary kinetic energy to destroy a notional RAM threat at distances
within the SRIT requirement (<2 km), and in fact at very short distances (within
the acceleration period of the rocket) a gun fired projectile would deliver more kinetic
energy. A gun fired projectile would certainly be less complex (i.e. less expensive)
than a rocket or missile, and therefore more desirable to meet the SRIT requirements.
However, with their single shot probability of hit being orders of magnitude lower
than that of guided missiles, many gun fired projectiles would have to be used to be
effective, something not desirable in an area trying to defend against many incoming
targets. Additionally, having to fire many projectiles greatly increases the chances
of collateral damage. Therefore, Table 3 includes a review of the guided projectiles
discussed in this thesis.
The only directed energy weapon discussed, the THEL, is not capable of being
relocated with great effort, probably beyond the effort afforded in a combat situation.
For this reason, it is ruled out as a potential possibility for this study.

32

33

Other Gun
Fired Guided
Projectiles

Proposed
Approaches

127

TBD

ERGM

GTRI
Guided
Projectile

n/a

THEL
SLID

35

Command
updates

GPS/INS

onboard
laser
seeker

none

Bore Guidance
(mm)
20
none

Skysheild/Ahead

Phalanx/C-RAM

Weapon System

pin-firing
mechanism

hit-to-kill;
warhead
req. TBD

submunition Large round fired from large


burst
guns; designed to hit targets
well beyond EAPS requirement

Very expensive, immature


and immobile technology
Max range far short of the
EAPS SRIT requirement

Submunition Designed to take out target


burst
larger than RAM threats

Kill
Summary
Mechanism
hit-to-kill
Inexpensive, low single shot
PH

Directed
Energy
optofluidic hit-to-kill
thrusters

none

none

Control

Table 3: Summary of Relevant Unguided & Guided Alternatives

1.3

Paradigm Shift in Conceptual Design

The proposed approaches to defeating the RAM threat discussed in Section 1.2.3
included evolutionary applications to already existing weapon systems that clearly
were not designed with the intention of meeting those needs. Those needs are outside
the described by the problem of interest are outside the traditional required function
of air defense weapons. Therefore, there is a need for a revolutionary solution, such
as using a gun launched guided projectile, which serves as an excellent case study
example for the implementation of a process desired in Research Question 2.
A paradigm shift is under way in the field of conceptual design to bring knowledge
of the system in question up front in the design process, while maximizing design
freedom as long as possible, without committing unnecessary cost [106]. Failure to
consider all of the requirements for producing a product may result in many changes
in the design. Most of the costs associated with developing a product are determined
very early in its life cycle, where uncertainty in the design is highest. Chapman et
al. [47] states that 80% of the system cost is determined by the initial system design,
which is the portion of the system that can no longer be optimized with respect
to cost or performance. With only 20% of the product development costs spent,
80% of the total product cost has already been locked. Future optimization of the
product during its lifetime will only affect the remaining 20% of the products cost.
Therefore, mistakes made in the initial stages of the design of a system will have a
serious negative impact on the total cost and success of the final product. Investing
more effort up front in the design cycle, and ensuring that all the aspects of the
concept are optimized, a more quality product can be delivered.
Dieter [56] addresses the idea of product cost committed during the phases of the
design process as well. He makes the point that true quality must be designed into the
product to produce a competitive product that will not have to be redesigned after it
goes into the market. Deiters major points on product design are summarized below:

34

Decisions made in the design process cost very little in terms of the overall
product cost but have a major effect on the cost of the product.
Quality cannot be built into a product unless it is designed into it
The design process should be conducted so as to develop quality cost-competitive
products in the shortest time possible.
Todays traditional design process commits cost rapidly at the early stages of
design, where knowledge of the system is the lowest. Additionally, design freedom
rapidly decreases as important decisions are made in the initial conceptual design
stages. Bringing more knowledge of the system forward will enable a designer to make
more informed decisions before the final costs are locked in, enabling the freedom to
make design changes without a heavy cost for a longer period of time during the
design process. This concept is known as a paradigm shift, inspired by initiatives
in government and industry to focus the engineering design community on system
affordability as the overall decision making objective. The evaluation of a design is
no longer driven solely by its capability to achieve specific mission requirements or
remain within specific product constraints. Rather, a robust design process, or one
that leads to a design that is least sensitive to influence of uncontrollable factors,
is needed to balance mission capability with other system effectiveness attributes
[106]. This design process would ideally bring more knowledge to the early part of
the product design phase where design freedom is highest. Additionally, the designer
would like to have the ability to keep a family of designs, rather than a point design,
for as long as possible, and keep the design freedom available longer.
Therefore, the following two literature search chapters are dedicated to finding
methods to realizing and furthering the spirit of this paradigm shift, as applied to
the general systems engineering process development desired in the primary research
objective and using the example problem of interest discussed in this chapter as a
35

proof-of-concept case study.

1.4

Disclaimer

This thesis does not intend to reflect the true performance characteristics of actual
weapon systems. It is meant only to show the trends one would expect if actual data
was used to conduct the study, data which is usually of a sensitive nature. Every
attempt has been made to protect any sensitive data by normalizing or eliminating
data on figures and plots, where possible.

36

CHAPTER II

LITERATURE SEARCH: COMBAT SYSTEM


MODELING & SIMULATION
The scientific methods of analysis and optimization have now found acceptance at every level of military operations, starting with the planning
of strategies, tactics, and doctrines through force deployment and employment in actual combat. The President of the United States and the National Command Authorities use the results of the extensive analysis when
various options for action are submitted by the military departments and
the Department of Defense for the final decision. Also, at lower echelons the decision-makers rely heavily on the analytical studies that define
measures of system effectiveness and present a process of determining appropriate criteria for choosing preferred alternatives.
Przemieniecki [126]
The intention of this chapter is to introduce general methods to model and quantify air defense weapon system capability, and the various related aspects of combat
systems modeling. However, there exists is a lack of literature in modeling of the RAM
threat and of the air defense systems designed to defend against this threat, but there
is an abundance in the area of aircraft survivability and large scale combat modeling.
Although there is not absolute correlation between the two subjects, those topics pertinent to this thesis will be introduced in this chapter, and used as a starting point to
identify related topics that are required to be included in any methodology implementation that will address the specific problem. The goal is therefore to find accepted
37

methods that establish total weapon system accuracy and lethality by modeling the
firing of multiple guided projectiles for target intercept, with an additional focus on
the modeling of the radar subsystem variables effect on radar system accuracy and
cost.

2.1
2.1.1

Modeling Military Operations


General Requirements of Combat Modeling

The planning of successful military operations relies on accurate modeling based on


mathematical methods of analysis. Exploring different alternatives for resource allocation or specific force deployments is a useful way to determine the best strategies
for use in actual combat engagements. In the opening quote, Przemieniecki [126]
underscores the importance of using mathematical methods in defense analysis for
decision making, at various levels within the military planning framework.
The basic concept in the mathematical modeling of combat engagements focuses
on the attrition of combat troops, weapons, vehicles, targets, and any other resource
expended in the engagement, according to Przemieniecki. These attrition, or force
reduction rates are expressed as number of individual objects (i.e. targets, resources)
destroyed per unit time. These rates may vary as a function of time, depending on factors such as: resource allocation (i.e. troop location and concentration); availability
of resources (i.e. resupply, or refitting an aircraft returning from a sortie); command,
control, and communications; intelligence information about the target area; other
major human factors including prior training, and the willingness of combatants to
engage in battle to the maximum capabilities constrained only by physical limitations. Much of the data used in these models is generated in terms of probabilities of
event occurrence, varying from 0 to 1. A common example of this is using analysis of
gun or missile test data to determine the Circular Error Probable (CEP), leading to
calculation of probability of hit.

38

2.1.2

Large Scale Combat Computer Simulations

2.1.2.1

Identification of Large Scale Combat Computer Simulations

There are many modeling and simulation tools that have been developed to attempt
to describe the outcome of military encounters at various levels of fidelity. Decision
makers can then use the outcomes of these simulations to base important tactical
and/or strategic conclusions. The different branches of the United States armed forces
have adopted respective theater level analytical campaign simulation tools. The U.S.
Air Force uses THUNDER [68] to simulate large-scale application of air and space
power in a joint warfighting context at the theater level. The Integrated Theater
Engagement Model (ITEM) [129] was developed for the U.S. Navy at the request of
the Defense Nuclear Agency to model naval nuclear simulations. It has since evolved
to model naval surface and subsurface combat, amphibious operations, air combat,
and ground combat, and has gained acceptance by the U.S. Pacific Command to
model joint expeditions in the Pacific Theater [55].
Janus was one of the first codes developed to simulate low level conflict, primarily
used to train platoon and company level commanders to apply tactical doctrines and
combat techniques. It is an event driven simulation that models fighting systems as
entities (such as tanks, helicopter, etc) [13]. The Joint Conflict and Tactical Simulation (JCTS) is an upgrade to Janus giving it an urban combat modeling capability,
and gives the user to control thousands of soldiers, aircraft, and naval vessels simultaneously. Although as with Janus, the focus is primarily on the infantry and soldier
level, giving the ability to simulate hostage rescues and drug interdiction raids. Many
of these simulations or games can last weeks and, sometimes a short game is run
dozens of times so that statistical sampling can be used to evaluate a particular tactic
or weapon [71].
These tools are just several examples of large scale simulation tools. The one
thing they have in common is that they are meant to provide enough scope to make

39

high level decisions, such as the ability to defeat an enemy or estimate asset loss.
2.1.2.2

Modeling RAM Defense Using Janus

Baumann [35] discusses the use of Janus to simulate low to mid intensity conflicts
having intense RAM barrages. From this study, Baumann concludes that the survivability of assets undergoing an intense RAM barrage was found to be dependent upon
asset dispersion, maneuver, concealment and counterfire against only the potentially
threatening RAM sources. Because the various RAM projectiles are unguided, they
can not be steered into a target and usually have very poor probabilities of hit with a
single projectile. It is certainly a waste of resources to engage any RAM threat that is
determined to miss any assets. This means that an accurate Impact Point Prediction
(IPP) must be used, which uses precision tracking of incoming threats to calculate
prediction of threats flight path and impact point in relation to the location of dispersed assets, both stationary and moving. This means that there are two options to
ensure asset survivability on the battlefield: control asset location such that most of
the potential RAM fired pose little threat, or only engage the ones that are assessed
to threaten the assets.
The key findings of Baumanns Janus study was that effective engagement against
RAM considered a threat is dependent on three factors assuming accurate IPP: high
threat kill rate, large shooter magazine, and engagement of multiple targets simultaneously. These statements very likely come with a sizable list of assumptions that
were not clearly stated in [35], however several assessments can be made as applicable to this thesis. The first statement, that a high threat kill rate is an obvious
requirement. For any number of RAM encountered, a high percentage of them must
be destroyed. However, this could also be used to set the confidence requirement
of an individual encounter, saying that the probability of one interception solution
destroying a RAM threat must be high. The second statement implies that the likely

40

case that a single interceptor launched against a RAM threat would not defeat it
and requires firing multiple interceptors. However, the statement itself indicates that
whatever solutions studied in the Janus simulation must have had a very small Single
Shot Probability of Hit (SSPH). Finally, the last statement of engagement of multiple
targets simultaneously could lead to the assumption that a specific defendable area
was defined. In fact one statement Baumann made was the there was a need to
assess other scenarios. Clearly, there is a need to model an interceptor and target
engagement accurately.

2.2

Modeling and Simulation Concepts

It is first necessary to establish definitions to some commonly used terms. The Military Operations Research Society (MORS) [79] explains that according to terminology
established by the Department of Defense (DoD), a model is a physical, mathematical, or otherwise logical representation of a system, entity, phenomenon, or process,
and a simulation is a method of implementing a model over time. Also, a technique
for testing, analysis, or training in which real-world and conceptual systems are reproduced by a model. MORS further breaks down the levels of model aggregation for
warfare modeling into four categories: 1) phenomenological models, which describe
the behavior of individual physical entities (human or otherwise), 2) single engagement models, 3) battle or multiple-unit engagement models, such as force on force
engagements, and 4) campaign or theater level models. These levels may be useful in
constructing a hierarchy of models to serve a common purpose of a system-of-systems.
This hierarchy may be visualized as a pyramid as shown in Figure 18, with the bottom containing the phenomenological models and the top the campain/theater level
functional as the single macro model. This pyramid represents the nesting of models
in one interacting whole, showing how results are integrated from one echelon to another, or simply an abstract representation of a structure relating independent model

41

Campaign/theater level models


Multiple-unit engagement models
Single engagement models
Phenomenological models

Figure 18: Military Operations Modeling Hierarchy Pyramid (adapted from [79])
operations.
[FIGURE
System of Systems Hierarchy Pyramid, adapted from MORS[]]
A more generalized approach to defining model types is useful for describing total
system effectiveness. Dieter [56] states that a model is an idealization of part of the
real world that aids in the analysis of a problem...Simulation is a manipulation of the
model. This manipulation involves subjecting the model to various input conditions
to observe its behavior enabling a user to explore the nature of results that would
otherwise be obtained from an actual system. In general terms, a model may be
either descriptive or predictive. A descriptive model enables an understanding of a
real-world system, such as a model of an engine or complete aircraft. A predictive
model is used to understand and predict the performance of a system.
Dieter further classifies the model types adapted from [65], as 1) static or dynamic,
2) deterministic or probabilistic, and 3) iconic, analog, or symbolic. Table 4 lists
the properties of each of those model classes. The first classification level describes
whether or not model properties are affected with time; i.e. static or dynamic. The
next level describes whether or not the behavior described by the model is known with
certainty; i.e. deterministic or probabilistic. Finally, the physical model is described.
If a physically geometric representation is used, such as a wind tunnel model, the
model is iconic. An analog model is used when the model only needs to describe the
actual behavior without physically representing the actual event, such as a process

42

Table 4: Model Types (adapted from [56])


Classification
Static
Dynamic
Deterministic
Probabilistic
Iconic
Analog

Symbolic

Description
Model with properties that do not change with time
Model in which time varying effects are considered
Model that describes the behavior of a system in which
the outcome of an event is known with certainty
Model that describes situations that have outcomes not
known with certainty
Model that is a physically geometric representation used
to describe the static characteristics of a system
Model that behaves like the real systems they represent,
by either obeying the same physical principles or simulating the behavior of the system
Model with abstractions of the important quantifiable
components of a physical system, providing the greatest
generality leading to a qualitative result

flow chart. Finally, the greatest qualitative results are obtained by using symbolic
models, such as mathematical equations, that are able to model abstractions and
quantifiable components of a physical system. Therefore, a mathematical equation
used to model a system that uses no random variables would be a static-deterministicsymbolic model.
Symbolic mathematical models, which are usually used to treat a process by logical quantitative relationships, are further classified as analytic or simulation models
by Przemieniecki [126]. The analytic model determines exact solutions with deterministic outcomes, whereas the simulation model, as Przemieniecki adopts from [83],
is a series of models that may be used to converge on solutions to very complex
problems involving uncertainty (probabilistic occurrence) and risk. Przemieniecki
then takes that simulation model, and subdivides it as either a deterministic or stochastic (probabilistic) model, as shown in Table 4. Figure 19 graphically shows the
hierarchical relationship between the different model classifications described above

43

Models

Iconic

Analog

Symbolic
Descriptive

Mathematical

Analytic

Simulation

Deterministic

Stochastic

Figure 19: Model Classification Hierarchy [126]


and listed in Table 4, leading down to the model types of interest. As will be discussed in the following subsections, the basic concepts originally developed for the

Figure 12.1 Model Classification (adapted from Przemienircki)

theory of combat were deterministic models. Probability based combat models were
developed to simulate more complex situations that take into account the randomness
associated with engagements.
2.2.1

Deterministic Models

For simple homogenous combat engagements, a deterministic approach can be used


to create a set of differential equations describing the strength of each side of the
engaging forces as a function of time. The deterministic approach can be used to
model combat engagements only when analytical solutions are possible.
Frederic Lanchester [89] is credited with developing the fundamental model for
developing theories of combat and for calculating attrition of forces in military engagements. In 1916, Lanchester developed a mathematical analysis of air combat,
the overall concept being based on a simple expression relating attrition rate to the
product of the opposing force effective firing rate of a single force (or weapon unit) of
time, and the total number of opposing units (or weapons) engaged in combat. The
attrition rate is equal to the product of the effective firing rate and number of the

44

opposing force. The Lanchester equations as applied to notional Blue and Red forces
are given in Equations 1 and 2, respectively. The total number of Blue forces mB is
shown as a function of Red force effective firing rate aR and total number mR , with
similar logic shown for determining the total number of Red forces mR . Note that
the negative sign denotes attrition.
dmB
= aR mR
dt
dmR
= aB mB
dt

(1)
(2)

However fundamental these equations were to the development of modern combat


modeling, Przemieniecki points out that they are bound to the following assumptions
that quickly show their practical limitations: 1) both sides are homogeneous and
are continually engaged in combat; 2) each unit (or individual weapon) is within
the maximum weapon range of all opposing units; 3) collateral damage is considered
negligible within the target area and ignored; 4) the attrition coefficients include the
probabilities of the target being destroyed when hit; 5) the effective firing rates are
independent of the opposite force level; 6) every unit is aware of the location and
condition of all opposing units - fire is directed only at live targets, and when a target
is destroyed, search (or fire direction) is directed at a new target; 7) there is a uniform
distribution of fire over the surviving units (i.e. no prioritization of targets). It is
evident that all of these assumptions severely limit the practical use of these models,
and their very use presumes that an attrition rate can capture all of the pertinent
information in an engagement scenario. This drives the need for other types of models
that take into account the random nature of engagement models.
2.2.2

Probabilistic (Stochastic) Models

Probabilistic models simulate random processes, such as an endgame, a one-on-one


encounter, or a many-on-many scenario. Ball [32] specifies three types of probabilistic

45

models used for survivability assessment: expected value models, Monte Carlo models,
and Markov Chain models.
2.2.2.1

Expected Value Probabilistic Models

Each phase of an engagement may have associated with it a probability that a certain
outcome occurs. For example, the probability of detecting an enemy is 0.6, the
probability of hitting the target with a weapon is 0.4, and probability of kill if hit
may be 0.8. The actual probability of killing the target is the joint probability of
each of the battle phases occurring. The expected value of a random number is
simply its mean [74], therefore an expected value engagement model just multiplies
the probabilities together to obtain the answer. In the very basic example case above,
the probability of a successful kill would be 0.6*0.4*0.8 = 0.192.
2.2.2.2

Monte Carlo Probabilistic Models

When the scenario is too complex because of too many phases in the process that
must account for randomness, or too many conditional probabilities are included in
the process, a Monte Carlo model is often used. Instead of just multiplying the
probabilities of individual events together to determine the joint probability of the
combat outcome such as with the expected value model, a Monte Carlo simulation
draws a random number [0,1] at each event in the process and compares it to the
assigned probability to determine whether that particular event occurs. This process
is repeated many times to give a probability distribution. With increased number
of trials, the Monte Carlo model results will approach the expected value model results. A more detailed discussion of Monte Carlo simulation as applied to uncertainty
quantification will presented in Section 3.2.2 of Chapter 3.

46

2.2.2.3

Markov Chain Probabilistic Models

A Markov process has a system which can reside in two or more states or conditions,
and the values associated with those states are contained in a state vector. A Markov
Chain model assumes that a sequence of independent events associated with a system
can be modeled as a Markov process. As an event occurs, each state within the system
transitions to all other possible states with a specific probability for each transition.
The probability of the system existing in each of the possible states is a sequential
process, in which the state it will exist in after events 1, 2, 3, ..., J is based on the
probability the system existed in each of the possible states after the previous events
0, 1, 2, ...J 1 respectively. This sequential process is the Markov chain process. A
transition matrix [T ] transforms the current state vector {S}{j} to the next state
{S}{j+1} as given in Equation 3.

{S}{j+1} = [T ] {S}{j} for j = 0, 1, 2, J 1

(3)

As they are not directly applicable to the type of problem studied in this thesis, Markov Chains will not be used, and no more attention will be paid to them1 .
Interested readers should refer to Appendix B in Balls book [32] for an example
engagement scenario modeled as a Markov Chain.

2.3

Survivability as a Discipline

For an aircraft in a combat environment, Ball [32] defines its survivability as the
capability of an aircraft to avoid or withstand a man-made hostile environment.
The survival of an aircraft on a combat mission is not a deterministic outcome, or
something that can be predicted with absolute certainty. Aircraft survivability is
1

As will be discussed further in the approach chapter, an individual engagement will yield probability of hit as a function of noise on radar measurement and gun firing accuracies. Many executions
sampling from distributions of those noises will be required, driving the need for using a Monte
Carlo simulation.

47

influenced by a number of random variables, and is measured by the probability PS ,


varying from 0 to 1. The closer PS has a value to 1, the more likely an aircraft is
to survive a combat scenario. The probability that an aircraft is killed, PK , is the
compliment of PS , as show in Equation 4, and the two are mutually exclusive; the
aircraft either survives or it is killed.

PS = 1 PK

(4)

The probability that an aircraft is killed is dependent upon two main factors. The
susceptibility of an aircraft describes its inability to avoid being hit by enemy fire,
and measured by the probability of being hit PH . Aircraft vulnerability describes
how easily it is killed if actually hit, and described by the conditional probability of
being killed if hit PK|H . The actual probability of killing an aircraft is measured by
the joint probability of the probability of hit and conditional probability of kill given
a hit. Therefore by substitution, survivability may be written as a function of both
susceptibility and vulnerability, as in Equation 5.

PS = 1 PH PK|H
2.3.1

(5)

Susceptibility

Susceptibility is the inability of an aircraft to avoid being hit by one or more damage
mechanisms in the pursuit of its mission [32]. Defined by the probability of hit PH ,
the higher the susceptibility, the closer PH approaches 1. Therefore, the more susceptible an aircraft is, the more likely it will get hit by an enemy weapon. According to
Ball, examining an aircrafts susceptibility in a given scenario consists of determining
the factors that influence susceptibility, modeling the scenario events to determine the
likelihood that the aircraft is hit by each projectile fired at it (called susceptibility
assessment), and finally, reduce the aircraft susceptibility through design. An aircraft

48

Table 5: Probability Factors that Influence Susceptibility [32]


Term Description
PA
Prob. that a threat weapon is active, searching, and
ready to encounter
PD|A Prob. that aircraft is detected given an active weapon
PL|D Prob. that aircraft is tracked, fire control solution obtained, and weapon is launched, given the threat was
active and detected the aircraft
PI|L Prob. that the threat propagator intercepts the aircraft,
given a launch
PH|I Prob. that aircraft is hit by threat, given an intercept
designer desiring to minimize susceptibility would design to reduce the PH that an
air defense weapon would have against it (or probability of fuze PF for an air burst
weapon). An aircrafts susceptibility is primarily affected by the air defenses detection, tracking, and weapon guidance elements. Ball assigns probabilities to several of
those key factors, listed in Table 5. Calculating PH involves the joint probabilities of
all the factors listed in Table 5, as shown in Equation 6.

PH = PA PD|A PL|D PI|L PH|I

(6)

Each of the probability factors listed in Table 5 correspond to an element in


the EAPS engagement kill chain shown back in Figure 5. One of the main factors
affecting susceptibility is the probability of detection, which is a function of both
the detection capabilities of the air defense system, as well as the aircraft signatures.
Aircraft signatures include radar cross section, infrared, visual, and aural. Much can
be written on each of these types of signatures, however to limit the scope to the
small unguided mortar targets studied in this thesis, special attention will be given
to radar signature in Section 2.7.

49

2.3.2

Vulnerability

Aircraft vulnerability is defined as the inability of the aircraft to withstand the damage caused by the man-made hostile environment, to its vincibility, to its liability to
serious damage or destruction when hit by enemy fire [32]. Vulnerability is described
by the probability of being killed if hit PK|H , and the higher the aircraft vulnerability,
the higher the PK|H as its value approaches 1. It may be thought of as how soft a
target is, or how easily it would be destroyed if hit. Therefore, the more vulnerable
an aircraft is, the more likely it will get destroyed if hit by enemy fire and the higher
its PK|H . The Department of Defense [9] defines aircraft attrition kill categories that
describe the elapsed time between when an aircraft is hit, and when it is rendered incapable of completing its mission. These are adapted by Ball [32], as given in Table 6.
Although these are defined for aircraft, Macfadzean [92] states that these are equally
applicable to air-to-surface missiles engaged by air defense weapons, depending on
the situation.
Macfadzean [92] does not state that these kill levels are applicable to surface-tosurface indirect fire projectiles, however, this thesis takes this basic idea one step
further to describe the attrition to the surface launched RAM threats. Therefore it
is important that an air defense system be designed to hit the critical components
in an incoming threat so that a catastrophic kill occurs sooner, therefore reducing
the probability of the threat to be able complete its mission (i.e., destroy the air
defense system). The more stringent the kill level desired against a target, the fewer
components that will cause that damage if they fail. For example, at extremely short
ranges, a catastrophic kill level (KK) may be the only option to defend a given area,
and this might only be achieved by directly hitting the fuze or warhead of a mortar or
rocket threat. At longer ranges, hits to other sections of the threat might suffice, such
as the stabilizers, that would cause the threat to tumble away from the intended mark.
To defend against very short range threats, such as mortars, it may be necessary to

50

Table 6: Attrition Kill Levels (adapted from [32])


Kill Level
KK kill
K kill
A kill
B kill
C kill

Definition
Immediate disintegration
Out of control within 30 sec of a hit
Out of control within 5 min of a hit
Out of control within 30 min of a hit
Aircraft unable to complete mission after a
hit, but able to return safely to base (mission
kill)

directly impact the fuze and completely destroy it.


2.3.3

Powering-up/Single Sortie Probability of Damage

It may be necessary to use more than one weapon to achieve a desired probability of
damaging a target, which Driels [57] terms the powering-up or survivor rule. This
is an important application of the binomial distribution and is used to determine the
probability that a target will be hit at least once by n weapons, given the probability
of hitting a target with one weapon p. Driels derives the powering-up formula from
the Bernoulli distribution, which is a set of repeated trials of the same experiment in
which each of the trials are independent. This is obtained by first determining the
probability of not hitting the target by substituting 0 for x in the binomial equation,
given in Equation 32 in Appendix A, such that
n!
(1 p)n = (1 p)n
(n x)!x!
and noting that the probability of hitting the target would be the compliment of this
expression where
P (target hit) = 1 P (target not hit)
Therefore, the probability of hitting the target would be given by the Driels poweringup rule given in Equation 7.
P (target hit) = 1 (1 p)n
51

(7)

Total Force Probability of Kill

1.0
SSPD=0.80

0.8

SSPD=0.50
SSPD=0.30

0.6

SSPD=0.05

0.4

0.2

0.0
0

10

12

Number of Weapons

Figure 20: Example of Powering-Up Rule for an SSPD


Driels furthers this technique to a force estimation study, where now PD is the
resulting damage caused by n weapons, and p would be the Single Sortie Probability
of Damage (SSPD) for a single weapon, given in Equation 8. A common quantity
used in force estimation studies is the probability of damage if only a certain limited
number of weapons q is available. An example of powering-up is given in Figure 20.
This plot relates the differences in using four different weapons capable of different
SSPDs. This allows a user to determine the probability that the target is hit once (or
damaged) when 1, 2, etc. weapons are fired. Note that each of the curves approaches
unity for increasing numbers of weapons. Recall that the initial assumption was that
each of the trials are independent of one another.
PD = 1 (1 SSP D)q

(8)

Macfadzean [92] applies the idea behind the powering-up equation for Antiaircraft
Artillery (AAA) fire, where projectiles are fired from the ground at a moving airborne
target. Here it is desirable to determine the cumulative hit probability as a function
of range. This is done to determine the range at which some minimum acceptable

52

probability can be achieved. For a given set of single shot hit probabilities achieved
at given ranges Pssh,i , the cumulative probability of hit Ph,cum at that range is given
in Equation 9.
Ph,cum = 1

N
Y

(1 Pssh,i )

(9)

i=1

2.4

Air Defense Systems

With the aspects of aircraft survivability pertinent to this problem identified, this
section will now describe the subject of defending an area from an airborne attack.
Air defenses are all defensive measures designed to destroy attacking enemy aircraft
or missiles in the earths envelope of atmosphere, or to nullify or reduce the effectiveness of such attack [15]. The general term air defense can be used to describe an
entire air defense order of battle in a theater, or to describe a single sensor-weapon
combination operating autonomously to both detect and engage threats. Air defense
weapons may include large surface based guns known as antiaircraft artillery (AAA),
surface-to-air missiles (SAM), and even the air-to-air missiles (AAM) and guns carried by interceptor fighter aircraft. The air defense forces may be assigned to defend
a specific area from attacking aircraft, which may include manned figheters/bombers,
UAVs/UCAVs, cruise/ballistic missiles, etc...
The resources being protected from air attack may include any land or maritime
asset contributing to the capability of an armed force engaged in a conflict. The
threat may include manned or unmanned aircraft, air-to-surface missiles, and surfaceto-surface missiles. Of course who the enemy is depends on the perspective of which
side is attacking or being attacked. To successfully defeat an airborne threat, it is
necessary to conduct a threat characterization in which probable mission profiles, observability, maneuverability, and vulnerability properties of that threat are identified
[92].

53

2.4.1

Operational Requirement & System Specification

An operational requirement is the identification of the perceived need of an air defense


capability as prepared by the military user, and serves as the basis for acquisition of
a new system. Macfadzean [92] identifies five issues that the operational requirement
must specifically address:
1. Definition of the threat and the specific environment that the threat will be
encountered in.
2. Identification of existing systems that can deal with the threat, and any shortfalls to meet this operational requirement.
3. Specification of what the new system would have to do to counter the threat in
the expected environment, and the military value it would offer.
4. Estimation of development cost and schedule.
5. Identification of how many systems are needed over a given time frame, for a
given cost per system.
The third bullet in the previous list relates directly to the performance requirements of a new system. These generalized statements may be applied to define requirements for the type of threat to engage, the identification of a specific area to
be defended, and even the probability of acceptable fratricide and/or level of collateral damage inflicted. Once a concept has been identified, system specifications can
be created, which are the explicit requirements placed on the systems design and
performance. However, the system requirements do not just restate the operational
requirements, but the performance characteristics that ensure meeting the operational
requirements.

54

2.4.2

Types of Air Defense Systems

This section is dedicated to introducing air defense systems that have attributes
pertinent to a solution required by the problem statement.
2.4.2.1

Antiaircraft Artillery

Antiaircraft artillery (AAA) systems fire gun-launched unguided, ballistic flight path
projectiles at air targets. They are usually found in the role of terminal defense,
most effectively for engaging targets within four kilometers. The only controls over a
gun-launched projectile flight path are the two gun aiming angles: azimuth and elevation. Guns are considered open loop systems; gun aiming angles are calculated for
a predicted target flight path (i.e. predicted future position) such that the projectile
will hit the target without any further corrections. Most AAA gun systems employ
power-driven mounts to elevate the gun to the commanded azimuth and elevation
angles with the smallest possible error. The difference between the commanded angle
and the actual response angle is the position error. Employing a AAA system successfully requires an accurate error analysis. In addition to target tracking and flight
path prediction, Macfadzean [92] lists the following eight major contributions to miss
distance of AAA fire.
Initial velocity
Air density
Air temperature
Range wind
Projectile weight
Gun elevation angle
Crosswind
Gun azimuth

55

Both the gun itself and the ammunition fired from it are designed to achieve
consistent, predictable firing velocities (the velocity at which the projectile leaves the
muzzle) between successive shots. In reality, the initial velocity varies between each
gun and each projectile round. One example that Macfadzean points out is that the
gun barrel wears more with each shot fired, which increases the barrel diameter and
allows some of the gases to escape around the projectile body, thereby reducing the
initial firing velocity. Even variations in the temperature of the gun powder can have
measurable effects on firing velocity. Manufacturing tolerances drive the variation of
the projectile rounds themselves. Macfadzean gives a basic feel for the magnitude of
the possible variations in initial velocity. Ammunition from the same manufacturing
lot may have small variation from round to round, about 2 to 3 m/s 1 error (assuming
the error is normally distributed around the desired value). Poorly manufactured or
mixed lots may have greater than 5 m/s 1 error.
Differentials in air density and temperature only add a significant contribution
to position error when dealing with great variations in altitude over long distances.
Range wind, which is the velocity component of wind parallel to the direction that
the projectile travels, directly modifies the projectile airspeed and the drag force
component in the down range direction (denoted as the X direction). Crosswind is
the airspeed component perpendicular to the range axis (denoted as the Y direction).
The variation of weight of the projectile results in a change in velocity along the entire
trajectory. Macfadzean states that variations from 0.5 to 1 percent are common, and
controlled by the manufacturing process and quality control. Gun elevation controls
the vertical plane trajectory, and therefore elevation angle variations induce vertical
errors that are in the velocity vector normal plane. The azimuth angle controls the
motion in the horizontal X-Y plane, and the effect of error produces a lateral effect.
Error analysis of AAA trajectories may be based on baseline standard conditions,
defined as:

56

Projectile specification initial velocity (i.e. fixed exit muzzle velocity)


Standard atmosphere conditions
Published drag coefficient versus Mach number
No wind
Projectile specification weight
No aiming error (azimuth or elevation)
2.4.2.2

Surface-to-Air Missiles

Surface-to-Air Missiles (SAM) are missiles that are launched from land (or sea) based
platforms [32]. The missiles themselves are self-propelled, and are usually large
enough to carry a warhead for long ranges using some type of onboard guidance
system. They use movable control surfaces that are deflected by guidance system
commands to direct flight, and on board sensors may be used to track the target.
However, the complexity of SAMs usually limits their use against large targets, such
as a piloted aircraft and cruise missiles, and therefore will not be elaborated on further. The SAM topic is simply included here to show how surface to air guided
weapons are generally used.
2.4.2.3

Integrated Air Defense System

A layered air defense is a combination of the different types of air defense where
one or more point or area defense weapons are assigned overlapping areas. The
linking together of the various air defense target sensors, weapon firing platforms,
and command centers is known as an Integrated Air Defense System (IADS). An
IADS consists of 1) surveillance sensors (surface-, air-, and/or space-based); 2) various weapon system installations (including target detection, tracking, and guidance
sensors for fire control); 3) filter centers used for data fusion; 4) decision making command centers, or command, control, communication, and information (C3I) centers;
and 5) communication links between the sensors, weapon systems, and command
centers. [32]
57

Page 1 of 1

Figure 21: A Generic Integrated Air Defense System [32]


Information gathered within the field of view of any of the sensors is passed to
the data fusion nodes. Target location(s) data collected at the data fusion nodes is
correlated to determine the number and location of each target, a process called track
correlation used to create a composite of the defended area sent to the decision node.
There, the targets are prioritized after decisions are made as whether the targets are
hostile, friendly, or unknown, and whether the targets are likely to attack, how soon,
and how much damage they would inflict if they attack. The decision center uses this
tactical picture to assign firing units (defensive weapon systems) to targets based on
prioritization. The assignments are based on firing unit availability and condition.
After the command center assigns the firing units to targets, the individual firing units
then attempt to encounter,
engage, and kill its assigned target. This process of target
file://C:\Classes\Thesis\Ball\Ch03\fig03.jpg
4/18/2005
prioritization and firing unit assignment is called classification [32]. A generic IADS
is shown in Figure 21, and shows a complex example of how the different aspects of
the kill chain are distributed over various systems to defend an area.

58

2.5

Weapon System Effectiveness Parameters

Once a particular air defense system is selected, its particular weapon system effectiveness parameters must be identified. The USAF Intelligence Targeting Guide [11]
summarizes the weaponeering process and its importance in quantification of weapon
system effectiveness:
Weaponeering is the process of estimating the quantity of a specific type
of weapon required to achieve a specific level of damage to a given target,
considering target vulnerability, weapon effects, munition delivery errors,
damage criteria, probability of kill, weapon reliability, etc.
This targeting guide stresses that in general, the results of weaponeering studies result
in probable outcomes, not the outcome of an individual munition delivery. This is
achieved using statistical averages based on many repetitions of models, weapons
tests, and real world deliveries.
The most important characteristic parameters of weapons are the ones that describe the weapon accuracy and effectiveness against a target. Przemieniecki [126]
lists the parameters that characterize weapon performance as 1) accuracy; 2) range;
3) rate of delivery (rate of fire); 4) warhead or munitions round effectiveness against
a given target; 5) time of target acquisition, pointing, and tracking; 6) time of flight
to the target; 7) weight and size; 8) cost; and 9) reliability and maintainability. However, Przemieniecki claims that it is the accuracy, effectiveness against a target, and
reliability that are most important in defining the weapon system effectiveness.
2.5.1

Weapon Accuracy

Weapon accuracy is the location of the point of impact for a given aim point on the
target [126]. The actual trajectory that a projectile (either guided or unguided) will
likely not coincide exactly with the calculated ideal trajectory, resulting in a point of
impact on the target that deviates from the desired aim point. The precise magnitude
59

of the distance between the desired aim point and actual point of impact can not be
determined, however a random function can be used to approximate this distance.
Driels [57] states that ballistic errors are only found in unguided weapons, where as
the delivery accuracy of guided weapons is mainly influenced by the guidance system
of the weapon itself. Common factors that cause this randomness include production
tolerances (i.e. no two projectiles in a weapon system are manufactured exactly
alike), varying accuracy in aiming, maneuvering target (and the errors associated
with tracking a moving target), atmospheric conditions, and human operator error.
Measuring the accuracy of weapon delivery can be quantified by measuring the closest
point of approach between the weapon and target, or miss distance, which is as Ball
[32] states the shortest distance between the aircraft and the missile as the missile
flies by the aircraft with no warhead detonation. Delivery accuracy directly relates
to the probability of direct hit of a weapon intercept with its intended target.
2.5.2

Lethality

Ball [32] uses the term warhead lethality to denote the ability of the warhead on the
weapon to kill the target given a direct hit or a proximity burst. Weapon lethality
against a particular target is basically a measure of the targets vulnerability, or
the probability of being killed if hit (as described earlier in Section 2.3.2). Whereas
weapon accuracy describes the ability to deliver the weapon as close as possible to
the target every time, the lethality describes the probability that the target will be
destroyed given that the weapon was delivered accurately (or within some measure of
accuracy). Macfadzean [92] identifies three types of damage mechanisms that could
be used to destroy a target.
1. Projectiles that directly hit the target
2. Fragments that are ejected by a warhead hit the target
3. Warhead blast creates a pressure wave that is applied to the target skin
60

The first option that Macfadzean points out is when the projectile directly impacts
the intended target, with the transfer of kinetic energy as the main source of lethality.
This is sometimes referred to as a kinetic kill, or kinetic warhead, even though there
is no explosive charge (which is normally associated with warheads). The second item
is where the warhead has an explosive charge that when detonated, ejects many small
fragments at a high speed. The detonation may occur on impact, timed for certain
point after launch, or set to detonate if the weapon comes within in certain distance of
the target. Additionally, the direction of fragment ejection may be controlled. These
fragments may also be submunitions, such as those discussed for the Ahead round in
Section 1.2.3.2. The third item in the list is when the pressure wave created by the
explosive charge hits the target, which in some cases is enough to damage the target.
With each item in this list, the miss distance accuracy can be alleviated, roughly
speaking. For a more thorough discussion of the more exotic types of warheads,
Macfadzean [92] may be consulted.
An effective way to defend an area against an incoming mortar is to launch a projectile and impact it in such a manner as to detonate the mortar warhead. This deals
with the subject of explosive initiation modeling and its correlation to calculation
of probability of detonation (Lloyd [90] [91] are excellent sources on this topic and
should be consulted for further detail). Of course, this relies on the level of detail
one can afford. In Equation 10, Macfadzean [92] gives a very simple relation between probability of kill and the fragmentation spray density (from the warhead) and
vulnerability area of the target. This vulnerability area is the effective area on the
target that if hit, could result in a kill. For example a mortar could be detonated by
impacting its contact fuze, causing a catastrophic kill (as discussed in Section 2.3.2).

Pk = 1 eAv

61

(10)

where
= fragment spray density (fragments/m2 )
Av = vulnerable area (m2 )

2.6

Guidance, Navigation, and Control of Related Weapon Concepts

Guided weapon systems have traditionally been limited to missiles, and many of the
concepts discussed in this section were developed specifically for missiles. However,
with the emergence of guided projectiles as potential alternatives for shorter range
air defense, it is important to review those guidance, navigation, and control concepts
that could be employed if a gun fired guided projectile or missile solution are carried
through the implementation of this thesis. Since these methods were developed for
missiles, this section will use the term missile where it would be found in the
literature, however the reader is encouraged not to bog down on the term used since
it is only the general method that is important. The term guided projectile could
certainly be used in place of missile anywhere in this section.
Guided missiles carry along a guidance package that tries to navigate the weapon
on a specific course to successfully intercept an intended target. The trajectory that
the weapon would follow is directed by the navigation law used. Finally, the missile
itself physically deploys aerodynamic steering surfaces using a specific control logic.
This section will describe several alternatives to selecting appropriate guidance, navigation, and control logic for a guided missile (or projectile).
2.6.1

Guidance

Ball [32] describes several possibilities for missile guidance summarized in this section.

62

2.6.1.1

Inertial Guidance

The inertial guidance method uses gyroscopes and accelerometers to direct the missile
over a predetermined path provided to the missile prior to launch. Guidance updates
may be sent after launch to correct any errors in the onboard equipment. This method
is more suited for long range missiles, such as cruise missiles, where they can cruise
for long periods of time with a certain amount of autonomy with occasional course
correction updates.
2.6.1.2

Command Guidance

Command guidance logic describes the case where a missile gets its guidance instructions from sources outside the missile. This method relies on a tracking system
outside the missile is used to track the missile and target along their trajectories.
The tracking system may use separate tracking units for the missile and target, or
just one to track both objects. The ranges, elevations, and bearing headings are used
to calculate position and position rate, leading to the flight path the missile should
take to intercept the target. The computed and predicted (based on current tracking
information) flight paths of the missile are used to determine how the control surfaces
should move to make the corrections required to change the current flight path to the
new one. The command guidance signals must be sent to a receiver onboard the
missile, using either the tracking system, or some other form of radio command link.
There are missiles that get guidance information transferred via a wire connected to
the launch site, but this will not be studied as part of this work. The main advantage
of command guidance is that there are no sensors on board the missile that have to
track the target. Ball notes that the use of command guidance of a missile all the way
to the target is usually limited to short ranges, due to the relatively large tracking
errors that occur at long range.

63

2.6.1.3

Beam Rider Guidance

With beam rider guidance, an electromagnetic beam located offboard the missile is
used to track the target. A rearward facing antenna senses the tracking beam, and
uses the position of the missile with respect to the center of the target tracking beam
to compute steering signals used to move control surfaces. One of the main advantages
of beam rider guidance is that multiple missiles can be fired at one target, since they
can all rely on the same tracking beam. However, Ball notes that this technique has
been limited to larger missiles that have the space to carry this guidance package.
2.6.1.4

Homing Guidance

The term homing guidance is used to describe a missile using an onboard guidance
package that can determine the targets position and guide itself to interception. The
major advantage of homing guidance over the previous methods is that as the missile
approaches the target, the error is reduced - the exact opposite is true for the previous
methods. There are types of homing guidance that will be discussed here - active,
semiactive, and passive.
Active homing is a guidance scheme where by the target is tracked entirely using equipment onboard the missile. Here, a radar on board transmits pulses that
illuminate the target, and then uses the echo from the target for navigation. The
major advantage of active homing is that the missile is essentially a fire and forget
weapon and does not require guidance or tracking from an outside source. However,
the added weight and expense limit this option to larger missiles going after targets
at very long ranges. Semiactive homing differs from active homing in that the source
of target illumination is offboard the missile, however the missile does have a system
to collect echoes from the target to steer the missile. Passive homing systems are the
simplest of the homing systems, and do not employ any type of target illumination.

64

Page 8 of 9

Fig 3.79: Missile navigation laws and trajectories.

Figure 22: Missile Navigation Laws [32]

Pursuit: In the pursuit trajectory, illustrated in the upper left corner of Fig. 3.79, the missile flies directly
toward the target at all times. Thus, the line-of-sight between the missile and the aircraft is maintained
essentially along the heading of the missile by the guidance system. Missiles flying a pursuit course
usually end up in a tail chase situation, similar to a dog chasing a rabbit. There are two basic objections
to the pursuit method. First, the maneuvers required of the missile become increasingly hard during the
last, and critical, stages of flight. Second, the missile speed must be considerably greater than the aircraft
speed. The sharpest curvature of the missile flight path usually occurs near the end of the flight. At this
time the missile must overtake the aircraft. If the aircraft attempts to evade, the last-minute angular
acceleration requirements placed on the missile could exceed its aerodynamic capability, thereby
causing a large miss distance. Near the end of the flight, the missile is usually coasting because the
booster (and sustainer) motor thrusts last for only a short part of the flight. Consequently, more energy is
required to make short-radius, high-speed turns at a time when the missile is losing speed and has the
least turning capability. The most favorable application of the pursuit course is against slow moving
aircraft, or for missiles launched from a point directly to the rear of the aircraft, or head on toward an
incoming aircraft.

Sensors onboard the missile can only use electromagnetic emissions or natural reflections from the target to establish location. Passive homing is limited to large targets
such as fighter or bomber aircraft that constantly emit electromagnetic radiation, and
have complex shapes ideal for natural reflections. Note that even the simplest homing
method requires thatLead
theangle:
weapon
carry onboard complex senors and the capability
In the lead angle or constant bearing trajectory, shown in the upper right corner of Fig. 3.79,
the guidance system flies the missile along a lead trajectory that results in an eventual intercept with the

to compute course corrections.


2.6.2

Navigation

file://C:\Classes\Thesis\Sources\Ball\Ch03\chap0_26.htm

3/24/2006

A navigation law is the relationship between the missile and target motion used to
generate missile steering commands. If those steering commands are followed such
that the navigation laws are satisfied, an intercept will occur [92]. Several types of
navigation laws are available for a guidance system to use to steer a missile along a
flight path to intercept a target. The four types of navigation laws presented in this
section and shown graphically in Figure 22 are also summarized from Ball [32].

65

2.6.2.1

Pursuit

The pursuit trajectory causes the missile to always fly directly pointing at the target at
all times. The guidance package maintains the line-of-sight (LOS) along the heading
of the missile. A missile trying to intercept a powered aircraft with pursuit navigation
usually ends up in a tail chase, in the same way a dog would chase a rabbit. This
drives the need for the missile to have speed considerably larger than its intended
target. This method is best used against slow flying targets flying head on toward
the missile.
2.6.2.2

Lead Angle

The lead angle navigation routine causes the missile to keep a constant bearing with
the target. For a target flying at a constant speed and heading, the LOS between
missile and target is constant, and the missile flies a straight collision course. If the
target changes direction, a new lead angle is calculated, again assuming a straight
flight path, and the missile is steered along that path.
2.6.2.3

Three-Point

Three-point navigation is a routine where the missile is constantly steered along the
line between target tracker and target. This method is usually used for short range
missiles using a command guidance package where the offboard (i.e. ground or aircraft
based) target tracker maintains line of sight with the target.
2.6.2.4

Proportional

Proportional navigation is the most common navigation law used in the final phase
of target interception. Here, the guidance package must determine the time rate of
change of the LOS between missile and target. The guidance package then tries to
maintain a constant LOS angle by making the rate of change of the heading direction
proportional to the rate of change of the LOS. A constant of proportionality is unique

66

Input
Signal

Error
Signal
+ -

Actuation
Signal
Controller

Comparator

Output
Signal
System

Feedback Signal

Figure 23: Closed-Loop Control (adapted from [119])


for a missile system, and can change as the missile flies the intercept to efficiently
manage the energy used for maneuvers.
2.6.3

Control

There are two basic types of controllers, open-loop and closed-loop. The open-loop
controller uses control action completely independent of the output, and is therefore
the simplest option. However, it lacks the accuracy of a closed-loop controller. In a
close-loop control system (shown in Figure 23), the action that the controller takes
depends on the system output. This type of control system is also called a feedback
control system, because the output is fed back and compared with the input to
generate an error signal. This error signal is used to make the output signal agree
with the desired input. This output signal is then used to send signals to control
surfaces of an aircraft to actually control the flight path. The method to control the
amount of drive signal to use is dependent on the type of controller to use. Two types
of controllers will now be discussed as potential option for missile or guided projectile
control logic.
2.6.3.1

Bang-Bang

A bang-bang controller is also referred to as on-off control, because the controller has
only the option of either applying maximum drive signal, or none at all. A common
example of a bang-bang controller is the thermostat of a household heater. If the
room temperature drops below a certain temperature, the heater turns on. Once that
temperature is reached, the heater turns off, however there is a tendency to overshoot
that desired temperature, causing a ripple, or fluctuation of temperature with time.
67

One attempt to solve this problem is to solve this problem is to set acceptable bounds
around the desired temperature, say 5 degrees around the desired temperature.
Now, the heater will not turn on until the lower bound is reached, and will not turn
off until the upper bound is reach, letting the temperature fluctuate within those
bounds. This is the easier of two control options discussed here, because the only
logic the controller must consider is on or off. Likewise to the thermostat example, if
a missile using a bang-bang control system is veering off course, the control surfaces
would fully deploy until the desired course is reached. Since there is no ability to
partially deploy the control surfaces, there is no distinction made between having to
correct large or small errors in flight path. [33]
2.6.3.2

Proportional Integral Derivative

The Proportional Integral Derivative (PID) controller is designed to generate a control


output signal dependent on the magnitude of the error signal. A PID controller is
actually a combination of three controller types. A proportional controllers feedback
output is proportional to the error signal, and is the simplest controller type. Its
main advantage is its simplicity, however it has the disadvantage that steady-state
error may exist. To eliminate steady-state error, an integral controller may be used
that creates an output that is proportional to the accumulated error. However an
integral controller may cause instability in the system that it is used in. A derivative
controller can then be used to provide large corrections before the error becomes
too large. Because it relies on the change in error, the derivative controller is at a
disadvantage if the error is constant and will not produce control. It also has the
disadvantage that is has difficulty with high frequency components of error. Each
of the disadvantages of each controller are eliminated when they are combined into
a single PID controller. This section is summarized from a section in Nelson [119],
which may be consulted for a more in depth description of the mathematical theory

68

of PID controllers out of the scope of this thesis.

2.7

Radar Concepts

It is evident from studying the problem of interest identified in the motivation chapter
that a radar will play an important role in a solution to the RAM threat. One of
the key drivers of this thesis is to have the ability to flow top-level requirements
directly to subsystem variables (i.e. required operational capability drives required
radar capability, which therefore drives the specific radar variables; the notion of
top-down design will be discussed further in Chapter 3). Therefore, this section will
address the basic radar topics necessary to quantify a solution space relevant to this
problem.
2.7.1

Frequency

A radar works by transmitting an electromagnetic wave and using the returned echo
picked up by the receiver to locate, identify, and track an aircraft, and is very effective for use against targets at long ranges [125]. The word radar itself is the short
form of the phrase RAdio Detecion And Ranging. The number of pulses that a radar
transmits as function of unit time is the radar frequency. As given in Table 7, radar
frequencies were divided into bands during World War II according to similar power
sources, propagation effects, and target reflectivity. The lower frequency bands (HF,
VHF, UHF) are primarily used for long range surveillance, but not effective at determining precise target location unless a very large antenna is used. For searching
and tracking targets, S-, C-, X-, and Ku-bands are used because the narrow beam
widths required can be obtained using much smaller antennas (as compared to the
lower frequency radars).

69

Table 7: Radar Frequency Bands [125]


Band
Frequency Range Wavelength Usage
Designation
(GHz)
(m)
HF
0.003-0.03
100-10
Over the horizon surveillance
VHF
0.03-0.3
10-1
Very long surveillance
UHF
0.3-1
1-0.3
Very long range surveillance
L
1-2
0.3-0.15
Long range surveillance, enroute air
traffic control
S
2-4
0.15-0.075
Medium range surveillance, terminal
air traffic control
C
4-8
0.075-0.037 Long range tracking
X
8-12
0.037-0.025 Short range tracking, missile guidance,
airborne intercept
Ku
12-18
0.025-0.017 High resolution mapping
K
18-27
0.017-0.011 Rarely used due to water absorption
[73]
Ka
27-40
0.011-0.007 Very high resolution mapping
Millimeter
40-300
0.007-0.001 Experimental [73]

2.7.2

Radar Range

A fundamental topic to radar analysis is the maximum range at which a radar can
detect a target is given by the radar range equation in the form given by Paterson [125]
in Equation 11. Note that for a given radar with R0 fixed properties from Equation
11, the maximum detection range of a radar against a 1 m2 target is proportional to
the fourth root of the radar cross section (RCS), given in Equation 12.

4
Rmax
=

Pt G 2 2
(4)3 Pmin

where
Rmax = maximum range at which a target can be detected
Pt = transmitted power
G = antenna gain

70

(11)

= wavelength
= radar cross section of the target
Pmin = minimum level of received signal based on a signal-to-noise ratio

Rmax = R0 1/4

(12)

The antenna gain and wavelength can be related to physical area of the antenna,
A, and the efficiency of the antenna aperture, , given in Equation 13. The quantity
A is a measure of the effective area presented by the antenna to the incident wave
[133].
G=

4A
2

(13)

On the antenna surface of area A, there is an array of antenna elements, usually


spaced at /2 [117]. The number of these elements on the radar antenna is the
primary driver for the cost of the radar [24].
2.7.3

Signal-to-Noise Ratio

Any unwanted signal that contaminates or competes with the desired signal is known
as signal noise. The relative amount of additive noise to the desired signal is the
signal-to-noise ratio (SNR) [16]. Barton [34] derives a useful equation relating SNR
to pertinent radar properties from a form of radar range equation, given below in
Equation 14.

SN R =

Pavg Gt Gr 2 rcs t0
(4)3 R4 kT

where
SN R = Signal-to-Noise Ratio [dB]
Pavg = Transmitted power [W]
Gt = Transmit antenna gain [dB]
71

(14)

Gr = Receive antenna gain [dB]


= Wavelength [m]
rcs = Radar Cross Section of the target [m2 ]
t0 = Observation time [s]
R = Transmitter/reciever range to target [m]
k = Boltzmanns constant [W/(Hz-K)]
T = System input noise temperature [K]

2.7.4

Radar Cross Section

Most of the properties that make up Equation 14 are controlled by the radar designer,
except the only property associated with the target itself, the radar cross section. The
radar cross section (RCS) of an aircraft (or any object) is the cross-sectional area of
a perfectly conducting sphere of radius a that reflects the same amount of energy as
that aircraft [92], where RCS () is related to that radius as given in Equation 152 .
= a2

(15)

The standard unit for RCS is the square meter (m2 ), however it is also widely expressed in decibel square meters (dBsm), where the two are related in Equation 16.

[dBsm] = 10 log10 [m2 ]

(16)

For a more practical definition, Skolnik [133] defines the RCS of a target as the
(fictional) area intercepting that amount of power which, when scattered equally in
all directions, produces an echo at the radar equal to that from the target.
2

The symbol is usually used to denote RCS, however to avoid confusion with the symbol used
for standard deviation, after this section the symbol will only be used to refer to standard devation,
and RCS will be used for radar cross section.

72

Table 8: Radar Cross Section Magnitudes[125]


Ships Large Aircraft Small Aircraft Man
RCS (m )
104
103
102 ,101
100
RCS (dBsm)
40
30
20,10
0

Rmax,2/Rmax,1

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.001

0.01

0.1

Birds
101 ,102
-10,-20

Insects
103 ,104
-30,-40

2/1

Figure 24: Reduction of Maximum Detection Range with Reduced RCS


Table 8 presents the RCS magnitudes associated with common airborne objects,
in the form given by Paterson [125]. This is meant to give the reader an appreciation
for RCS order of magnitude in both dBsm and m2 , against the size of several common
reference objects. In relation to the types of targets discussed in this thesis, a mortar
would likely fit into the bird category in Table 8, with an associated RCS somewhere
between -10 dBsm and -20 dBsm. While mortars are considered a low technology
weapon, their small size and simple smooth shape give them a fairly small RCS that
make them difficult to detect with radars designed to detect conventional aircraft.
This puts mortar RCS on the about the same order of magnitude with stealth aircraft,
however the detection ranges associated with mortar engagements are far shorter than
those associated with stealth aircraft.
Comparatively, a reduced RCS will decrease the radar detection range Rmax . Using the principle that Rmax varies with the fourth root of RCS, Figure 24 shows the

73

comparative ratio of RCS reduction. Minimizing RCS is therefore an important design criterion in the field of military aircraft design. (Knott [86] describes the methods
in which an aircraft designer can reduce RCS.) However, these methods will not be
exploited in this study, due to the low technology nature of the airborne targets in
question, but the information was included to provide the reader with an appreciation
for the radar accuracy required to detect a low RCS target.
2.7.5

Tracking Radars

2.7.5.1

Methods of Radar Tracking

A tracking radar measures location of a target, proving data that may be used to
determine flight path and future position [133]. The available radar data is usually
range, elevation angle, azimuth angle, and doppler frequency shift. Angular error is
the difference in magnitude and direction between the target position and a reference
direction, usually the direction of the radar antenna which is both the transmitting
and receiving device of the radar. During transmission, the antennas function is to
concentrate radiated energy into a shaped beam pointing in a desired direction. The
antenna also collects the reflected energy contained in the echo signal (from a target)
and delivers it to a receiver (these two antenna parameters are the gain variables in
Equations 11 and 14). A tracking radar attempts to line up the antenna with the
target in order to make the angular error zero, and the target is located along the
reference direction.
There are three methods to determine the direction and magnitude of the angular
error. Sequential lobing (also called lobe switching and sequential switching) is a
method in which the antenna beam is alternately switched between two positions
to obtain the angular error in one coordinate. The error is quantified by measuring
the difference in signal from the two positions from the reference, or switching axis.
The antenna is moved to align the switching axis and the target direction. In order to
obtain angular error in the orthogonal coordinate, two additional switching positions
74

are needed. The primary accuracy limitation of a sequential lobing tracking radar is
due to noise caused by mechanical and electronic fluctuations. This method was one
of the first tracking techniques employed in airborne interception radars, and is not
used as often in modern tracking applications.
The conical scan radar has the same basic principle as the sequential lobing
method, however instead of discontinuously stepping the beam between four discrete
positions, the antenna beam is continuously rotating. In both sequential lobing and
conical scanning techniques, measuring the angle error in two orthogonal coordinates
(i.e. both azimuth and elevation) requires multiple pulses to be processed. If the
train of pulses collected over a given time interval contains additional modulation
components, the tracking accuracy will be degraded. If the angular measurements
are made on the basis of one pulse, rather than multiple pulses, fluctuations in the
amplitude of the echo signal between pulses will have no effect on tracking accuracy.
Those tracking techniques that derive angle error information from a single pulse are
known as simultaneous lobing, or more commonly as monopulse. Also unlike sequential lobing and conical scan trackers which use only one antenna beam, monopulse
trackers use more than one antenna simultaneously. This way, the location of the
target relative to the radar is determined by measuring the relative phase or relative
amplitude of the echo pulse received in each beam.
More specifically, a tracking radar that operates with phase information is an
interferometer radar, since it deduces information from the coherent interference
between the two signals [16]. The phase difference between the two signals is the relative displacement between the two waves having the same frequency (or wavelength).
The complete wave cycle has a phase of 2 radians. To illustrate an interferometer
radar, two antennas are shown in Figure 25 separated by a distance D. The line of
sight to the target is offset an angle from the perpendicular bisector connecting the

75

Target

D
sin
2

Wa
vefr
ont

Antenna 2

Antenna 1

D
sin
2

D
2
D

Figure 25: Wavefront Phase Relationship for an Interferometer Radar


two antennas. The wavefront phase at antennas 1 and 2, respectively, are [24]
1 =
2 =

2 D
sin
2

2 D
sin
2

with a negative sign associated with 1 because of the reference origin defined directly
between the two antennas. Therefore the measure phase difference for an interferometer radar is given in Equation 17 [133].
= 2 1 =

2.7.5.2

2
D sin

(17)

Radar Error Analysis

Statistical methods can be used to describe the error of a radar measurement. This
error is simply defined as the difference between the value indicated by the measuring
instrument and the true value of the measured quantity [34]. For the purposes of this
thesis, this is the error in determining the location of an object in three dimensional
76

space. These radar errors may include both systematic and random properties. The
systematic property, or bias, is present when the radar error is constant with time.
For example, the difference in the actual position and radar measurement of an object
would be identical at each sample. This level of predictability can be corrected by a
process of calibration applied to the measuring instrument. The random property of
radar error, or noise, changes with each sample in time.
The most accurate method of error measurement is with using actual test data,
however this might not be very practical when required for use in a simulation with
changing properties. Mathematical approximations can be created that model actual
error components and enable accurate analysis of a particular systems error. In
most cases, a normal distribution may very accurately approximate radar noise. Of
course, the actual shape of the error distribution may not be normal, but Barton [34]
states that the normal distribution is often assumed to represent errors of unknown
characteristics, and it closely approximates the distribution of many actual errors.
Using the interferometer phase difference calculated in Equation 17, Holder [24]
derives an equation relating the measurement error standard deviation for an interferometer radar, as a function of radar properties. The remainder of this section
will summarize that derivation. Taking the expected value of Equation 17 gives the
variance ( 2 ), of which the square root yields the equation in the form of standard
deviation (), given as
=

2D
cos

Next, substituting
=

1
2SN R

and
=

77

which are obtained from [34]3 , results in the following relation


=

2D cos SN R

If the radar is assumed to be pointing directly at the target, the offset angle goes
to 0, and the cos term goes to 1, resulting in Equation 18, relating radar measurement error 1 to wavelength (), the distance between interferometer antennas (D,
a measure of radar size), and signal-to-noise ratio.

2D SN R

(18)

When the radar properties given in Equation 14 are substituted for SNR, the
measurement error can be a direct function of those radar subsystem properties.

2.8

Chapter Conclusions

This chapter reviewed related concepts of combat system modeling and simulation,
and discussed particular topics closely related to the problem of short range RAM
defense. As mentioned in the opening paragraph of this chapter, there is a lack of
literature in modeling of the RAM threat and of the air defense systems designed to
defeat those threats, certainly no direct approach exists to model the threat within
a weapon design environment. One of the closest ideas discussed in this chapter is
the IADS concept, generally showing how defending an area requires the interaction
of several key systems, such as a radar and a weapon system. The topic of short to
medium range RAM defense is a relatively new field of interest within the combat
modeling and simulation community, and the established systems engineering process
in place to address the specific problem does not address satisfy the overall research
objective of this thesis. Ultimately, an integrated hierarchical environment based on
3

These substitutions are based on ideas put forth in Chapter 8 of [34], where Barton derives
several key relations for tracking radars. The final result in Equation 18 is therefore a derivation
specific to an interferometer radar by Holder [24] using many of the relations that can be found in
[34].

78

accepted combat modeling and simulation methods must be created for the various
components that need to interact, and allow for rapid manipulation enabling decision
making at any level within and across the hierarchy. Because the literature on combat
modeling and simulation does not provide enough background to satisfy this need,
several concepts from appropriate Advanced Design Methodologies will be discussed
in the next chapter.

79

CHAPTER III

LITERATURE SEARCH: ADVANCED DESIGN


METHODOLOGIES
One of the great challenges in planning, managing and developing systemof-systems is to develop an ability to handle traceability when dealing with
big pictures of defence capability in various contexts and for different
purposes.
Chen et al. [48]

3.1

System-of-Systems Approach

There are many modeling and simulation methods that have been developed to attempt to describe the outcome of military encounters at various levels. Decision
makers can then use the outcomes of these simulations, and base important tactical
and/or strategic decisions on them as described by the first quote in Section 2.1. However, one common problem with most of these tools is that the models themselves
are only capable of describing events at one level, such as a theater level tool [68]
[129] at the very top level, or an aircraft sizing and synthesis tool [112] at the aircraft
system level. A system-of-systems approach can be used to describe events at every
level by functionally decomposing the entire system in question into a hierarchy of
decision making levels. An attempt now will be made to differentiate between the
terms system and system-of-systems.
Webster [3] defines a system as a regularly interacting or interdependent group
of items forming a unified whole. Chapman et al. [47] simply define a system as

80

any process or product which accepts inputs and delivers outputs. A more specific
definition by Blanchard [41] is that a system may be considered as constituting a
nucleus of elements combined in such a manner as to accomplish a function in response
to an identified need...A system must have a functional purpose, may include a mix
of products and processes, and may be contained within some form of hierarchy.
Blanchard [40] further describes a system in terms of the following characteristics:
1. A system integrates a complex combination of resources, which may be in the
form of data, equipment, software, or even humans, and must be done in an
efficient manner.
2. A system is contained within a hierarchy, and is highly influenced by the performance of the higher level system.
3. A system may be broken down into subsystems, the scope of which depending
on the complexity and function being performed.
4. The system must have a purpose, meaning it must be able to respond to an
identified need.
Despotou [54] defines a system-of-systems as an organised complex unity assembled from distributed autonomous systems (capable of independent provision of
services) collaborating to achieve an overall system purpose, and further making the
important distinction from any complex or legacy system. A Department of Defense
(DoD) study [49] states that a system-of-systems must have both operational and
managerial independence of its individual system components. There also must exist
a relationship among the system components, and a common objective or purpose
for each component, above and beyond any individual objective or purpose for each
component. Chen et al. [48] add that a system-of-systems is concerned with the

81

ability to analyze joint effects of system components, have interoperability of levels and analysis, enable information sharing, and have the capability for planning,
coordinating, and managing of any systems evolution.
Adding to the DoD principles, Maier [93] defines fives principal characteristics
that describe a true system-of-systems:
1. Operational independence of the elements: the system-of-systems is composed
of systems that can be usefully operated independently if the system-of-systems
is decomposed.
2. Managerial independence of the elements: the individual systems that comprise
the system-of-systems are separately acquired, and therefore are actually run
and maintained independently of each other.
3. Evolutionary development: the development of the system-of-systems is evolutionary with functions added, removed, and modified with experience, therefore
never actually appearing to be fully formed.
4. Emergent behavior : the entire system-of-systems performs functions and carries
out purposes that are emergent of the entire process, and cannot be localized
to any component system. The principal purpose of the system-of-systems is
fulfilled by its emergent behavior.
5. Geographic distribution: the independent system components can readily exchange information only, and not substantial items like mass or energy.
The additional dimension of managerial control is critical to identifying the appropriate principles of the system-of-systems. This is due to the fact that not all
system-of-systems, or the component systems within them, are of similar complexity. Maier further classifies three basic categories of system-of-systems by managerial
control: directed, collaborative, and virtual.
82

Directed systems are built and managed to fulfill specific purposes, with centrally
managed operation. The component system maintains the ability to run independently, but their operation within the integrated system-of-systems is subordinate to
central management purposes. An integrated air defense systems (IADS), for example, is centrally managed to achieve a common objective, although its component
systems, such as the individual radars or surface to air missiles, may still operate
independently. An IADS, discussed in Section 2.4.2.3, is an excellent example of a
true system-of-systems.
Collaborative systems have a central management organization that does not have
the authority to run individual systems. Each component system must voluntarily
collaborate to fulfill the central objective. The internet is an example of a collaborative system, where internet component sites exchange information using protocol
regulations, adhered to voluntarily.
Virtual systems completely lack a central management authority, although they
maintain a common purpose for the system-of-systems. Any large scale behavior
emerging from the combined systems must rely on invisible mechanisms for maintainability. National economies are virtual systems, as their long term nature is determined by highly distributed mechanisms, that appear to have no central management
control.
The fact that the system components have a level of autonomy may have adverse
effects on the system-of-systems in that there may be conflicts of responsibility. Each
component within the system-of-systems has its own distinct goals, which Alexander
[28] lists as including self-preservation and maintenance of a certain level of safety,
as well as sub-goals that define the components contribution to achieving the overall
system-of-systems goals. An Unmanned Aerial Vehicle, for example, may assume
responsibility for surveying a section of area intended for overall system-of-systems
purposes. However, a partially or fully autonomous system can essentially have so

83

much freedom to govern its own choices that it becomes a risk to itself and other
elements in a system [54].
One of the greatest advantages of a system-of-systems approach to problem solving is that it is very adaptable, since it allows improvement within any of its layers,
depending on either the evolution of a threat of advancement of an applicable technology. For example, a study of the application of a system-of-systems approach to a
national missile defense strategy was conducted by the RAND National Defense Research Institute [67]. The study found that this approach would offer more distinct
technological options to improve kill probability in any of the hierarchical layers, thus
reducing the required proportion of interceptors required to respond to proportionally
more enemy re-entry vehicles.
3.1.1

Functional Decomposition

A system is made up of many components that interact with each other. These
interactions make it impossible to effectively design a system in which each component
is considered independently. Dividing a system into smaller subsystems allows for a
simpler approach relative to the analysis required for all of the systems functional
interfaces. The system must be viewed as a whole, broken down into its components,
its component interrelationships studied and understood, and only then may the
system be put back together. Therefore, a technique will now be introduced to
structure the component hierarchy of the system (or system hierarchy within a systemof-systems).
A structured way to break down the components and subcomponents of a system
is by utilizing a functional decomposition technique. A functional decomposition
defines subsystems by the system requirements they enable [45]. The original intent
of functional decomposition, developed by Pahl and Beitz [122], and Koller [87], was
to describe the direction of information flow in a system. First, a system function

84

must be defined, which describes the process that takes inputs in and transforms
these inputs into outputs. This function is a transformation which may involve the
changing of the system state, and must contain an activation criterion that describes
the availability of resources for the function to execute, and an exit criterion that
describes when a function has completed. Dieter [56] considers function to be in the
nature of a physical behavior or action, whereby the systems functions are described
as a transformation between an initial state and a desired final state. The function
describes what the product must do, while its form or structure must describe how it
accomplishes it.
A functionality is simply a set of functions required to obtain a desired output.
Decomposition may be thought of as a form of top-down structuring, in which a
defined top-level system function is partitioned into several subfunctions, as Buede
describes [42]. Each function that requires further insight for the production of outputs is decomposed into another level of subfunctions (i.e. not every function must be
decomposed, as long as the desired output is obtainable). The decomposition process
must conserve all of the inputs to and outputs from the systems top-level function.
This methodology can help in understanding the context in which the term system is used, which requires appreciating the perception from which that term is
used. Therefore, what is considered a system from one perspective, may actually
be a system-of-systems in another. For example, the system may be an aircraft,
consisting of engines, wings, landing gear, etc..., that all function to comprise the
aircraft system. However, the engine sub-system may be further broken down into
its own subsystems such as the compressor, combustor, and turbine. The wing could
be broken down to each individual rib and spar. In an engagement, this aircraft may
operate in an environment where it carries a weapon with the objective of destroying
a target, and this target will likely wish to defend itself. This environment is now
the system, comprising of aircraft and missiles, each a system of their own that can

85

be functionally decomposed into the subsystems that make them up. Examining the
effects of manipulation at any given layer of system/subsystem on every other system
in this environment will define the system-of-systems approach taken in this thesis.
One such tool used to decompose a problem of interest is the morphological matrix,
or otherwise described as a matrix of alternatives [137]. A morphological matrix lists
the key parameters or functions of a system on the vertical scale of the matrix, and
alternative methods for satisfying them along the horizontal scale. The purpose of the
method is to uncover combinations of design concepts that might not be intuitively
generated. The elements of each functionally decomposed system level are arranged
in such a manner as to maximize the system effectiveness. Of course, those elements
could be systems themselves. Hines et al. [73] state that system effectiveness has
become the prime metric for the evaluation of military aircraft. As such, it is the
designers goal to maximize system effectiveness. In an engagement scenario, each
side wishes to maximize its system effectiveness such that it is able to, at the very least
survive the conflict, and if possible, destroy its opponent. Therefore, it is necessary to
quantify system effectiveness in an engagement scenario, where in this case it would
be a measure of how successful a system performs.
3.1.2

Functional Composition

Buede [42] also introduces functional composition as the opposite approach to the decomposition of a function. Composition may be thought of as a bottom-up approach.
This method starts with the identification of simple functionalities associated with
scenarios involving only one of the outputs of the system. Each one of these functionalities is a sequence of input, function, output-input, and then the repetition of
function, output-input, function, until a final output is reached. Within the functional
hierarchy, these functions in functionality are relatively low-level. After a number of

86

these functionalities are defined, the functions in the functionalities must be aggregated into similar groups. Next, these groups are aggregated into similar groups, and
the process is repeated until a hierarchy is formed, bottom-up.
The main advantage of the composition approach, over decomposition, is that
the composition process may be executed in parallel with the development of the
actual physical architecture. This way, the functional and physical hierarchies will
match each other. Furthermore, this is a very comprehensive approach in which a
designer will be less likely to omit major functions. However, these many functionalities must be easily accessible during the composition process. When there is little
experience with the system, or working with a radical departure of that system,
Buede recommends this approach.
Depending on the situation, it may be useful to use a combination of functional
composition and decomposition (which Buede refers to as middle-out). The simple
functionalities in the composition approach associated with specific scenarios defined
in the operational concept may be used to establish a sense of the system. Next,
a top-level decomposition may be positioned to match the divisions of the physical
architecture, which can be bolstered to assure totality by recurrent reference to the
functionalities.
3.1.3

System-of-System Architecture

Proper systems-of-systems analysis requires engineering at the architecture level,


which deals with allocation of functionality to components and inter-component interaction, rather than the internal workings of individual components. The architecture
acts as a framework that constrains the way components interact with their environment, receive and relinquish control, manage data, communicate, and share resources
[50]. For a successful design process, the functions or activities that a system has to
perform must be considered. As Buede [42] states,

87

The functional architecture of a system contains a hierarchical model of


the functions performed by the system, the systems components, and the
systems configuration items; the flow of informational and physical items
from outside the system through the transformational processes of the systems functions and on the waiting external systems being serviced by the
system; a data model of the systems items; and a tracing of input/output
requirements to both the systems functions and items.
The individual elements of a system-of-systems operate independently, collaborating only through information exchange. Although integrated modeling can be used
to define a system-of-systems, as Maier describes as a unified approach to systems
engineering [94], those models in themselves do not provide the guidance necessary
for their structuring. The system-of-systems architecture defines the interfaces and
composition which guides the implementation and evolution of the system-of-systems
[48]. Maier et al. [95] define systems architecting as the art and science of designing
and building systems, and state that the architects greatest concerns and leverage
should be with the systems connections and interfaces which distinguish a system
from its components and provide unique system-level functions. Subsystems (or in a
system-of-systems, the component systems) specialists will want to concentrate most
on the core and least on the periphery of their subsystems because they generally
view the periphery as external constraints placed on their internal design. Therefore,
the subsystem specialists concern for the entire system effectiveness would be less
than that of the architect of the systems. The entire system architecture, as Maier
et al. claim, if not managed well, the system functions can be in jeopardy.
Systems architecting differs from systems engineering in that it relies more on
heuristic reasoning and less on use of analytics. There are qualitatively different problem solving techniques required by high and low complexity levels. The lower levels
would certainly benefit from purely analytical techniques, but those same techniques
88

may be overwhelming at higher levels which may benefit more from heuristics derived
from experience, or even abstraction. It is important to concentrate on the essentials
to simply solve the problem. The system should be modeled at as a high a level as
possible, then the level of abstraction should be reduced progressively. Everything
should be made as simple as possible, but not simpler, as Albert Einstein once said
[59].
The management of relations between the system components is an architectural
issue which does not belong to individual systems, but shared by all the involved
components. Chen et al. [48] introduced an architecture-model based approach to
systems and capability relation management. The responsibility of system relation
management cannot be the responsibility of developers of individual systems, but at
the system-of-systems level. This architecture management system was designed to
store and manage knowledge and information of systems capability in their proper
context. The relations are determined by the interfaces between systems, or by the
manner by which they are integrated. Chen et al. define the possible manners by
which systems and capabilities are related in a system-of-systems manner as structure,
function, information, operation, and generation:
1. Structure-related: Two systems are structure-related if one is a component or
basis of the other.
2. Function-related: Two systems are function-related if one system requires certain functions or services by another system to perform its own function.
3. Information-related: Two systems are information-related if requirements or
information is exchanged between the two.
4. Operation-related: Two systems are operation-related if they are both used in
an operation scenario to jointly fulfil a mission.

89

5. Generation-related: Two systems are generation-related if one system will be a


replacement of the other.
As capabilities evolve, the relation between systems can change. Therefore, Chen
et al. state that a requirement to achieve good relation management is the development of an environment or system that serves as a basis for: managing concepts
and objects in context; conducting analysis, synthesis and evaluation of values and
relations of objects; handling complexity; exploring and studying dependency; maintaining traceability; and finally, visualizing concepts, objects and their relations and
dependencies.

3.2

Uncertainty Analysis and Quantification in


the Design Process

A NASA white paper describes those design problems that have a nondeterministic
formulation as uncertainty-based design [145]. This survey of the state of the art
in uncertainty-based design breaks it down into two major classes. The first is the
robust design problem, which seeks a design relatively insensitive to small changes
in uncertain quantities. Dieter [56] defines a robust design as one that has been
created with system of design tools to reduce product or process variability, while
simultaneously guiding the performance toward an optimal setting. A reliabilitybased design seeks one in which the probability of failure is less than a predetermined
acceptable value. The two major factors that drive the domain of applicability of
each case are the frequency of the event and the impact of the event, as Zang et al.
[145] adapts from Huyse [80] shown in Figure 26. A designer would likely design a
system with performance insensitive to everyday fluctuations, or robust, such that
there is not a catastrophic failure when the system encounters everyday fluctuations.
However, a designer would like to minimize the probability of a catastrophic failure

90

that employ stochastic approximations, such as Monte Carlo approximation of integrals, are only of
interest here to the extent that they are brought to bear on a genuinely nondeterministic problem
formulation. Likewise, random search techniques, such as genetic algorithms and simulated annealing, are
not intrinsically of interest in the present context. We use the term uncertainty-based design to describe
those design problems that have a nondeterministic formulation.

Catastrophe

No engineering
applications

Reliability-based
design and optimization

Performance loss

Impact of event

Robust design
and optimization

Reliability is
not an issue

Everyday fluctuations

Extreme events

Frequency of event

Figure 1. Uncertainty-based design domains (from Huyse 2001).


Figure 26: Uncertainty
Based Design Domains ([145] adapted from [80])

The two major classes of uncertainty-based design problems are robust design problems and
reliability-based design problems. A robust design problem is one in which a design is sought that is
occurring,
or maximizing
the designs
reliability.
the design
goal problem
of robustrelatively insensitive
to small changes
in the uncertain
quantities.Practically,
A reliability-based
is
one in which a design is sought that has a probability of failure that is less than some acceptable
based
design
is value.
to maintain
performance
levels
offtodesign
(invariably
small)
The same acceptable
abstract mathematical
formulation
can beat
used
describe conditions,
both robust
design and reliability-based design. However, their domains of applicability are rather different.

whereas reliability-based design aims to prevent catastrophic failure. The design risk
Figure 1 illustrates these domains. The two major factors are the frequency of the event and the impact
the event. No
system is viable of
if everyday
fluctuations
can lead
catastrophe. Instead,
like
is of
therefore
a combination
both the
likelihood
andto consequences
of one
an would
undesired
the system to be designed such that the performance is insensitive, i.e., robust, to everyday fluctuations.
On the other hand, one would like to ensure that the events that lead to catastrophe are extremely unlikely.
event.
This is the domain of reliability-based design. In both cases, the design risk is a combination of the
likelihood
of cases
an undesired
eventbe
andillustrated
the consequences
that event. An
examplefunction
of risk in (PDF),
the robustas
The two
can best
on a ofprobability
density

shown in Figure 27. Robust design is concerned with the event distribution around
3

the mean of the PDF, whereas reliability-based design is concerned with the event
distribution in the tails of the PDF. Zang et al. state that the mathematical methods
for robust design procedures are considerably less developed than those for reliability
based design, and still largely confined to academic studies.
Simulation tools are proving to be useful in providing designers with efficient
means to explore the interrelationships among various disciplines. However, uncertainties associated with these computer codes, arising from errors in input parameters
and internal uncertainties due to simulation inaccuracy, may propagate through the
different codes as shown in Figure 28. Du et al. [58] introduce two methods for
managing uncertainty in simulation-based design that formally mitigate uncertainty
based on principles of robust design. An extreme condition approach uses an interval

91

link function-generatorlinkage design problem is used to illustrate


the bene ts of applying the robust design approach for uncertainty
design
context is the likelihood
that the aircraft
will fail to meet thecan
aerodynamic
performance
mitigation.The
principlesof
thedesign
proposedmethods
be easily
extargets and will consequently lose sales and perhaps even go bankrupt. An example of risk in the
tended
to
more
complicated,real
multidisciplinarydesignproblems.
reliability-based design context is the probability that a critical structural component will fail, which

3) Similar to step
tainty "min
1 (x1 ) and "

could lead to the loss of the vehicle or spacecraft, payload, and passengers, and to potential lawsuits.

II.

Propagation of the Effect of Uncertainties

Probability
In this section,a
simulation-baseddesign model as shown in Fig. 1
density
is used to explain the proposed methodology. The model consists
of a chain of two simulation programs (imagining they are from
Robustness
two different disciplines) that are
connected to each other through
linking variablesrepresentedby the vector y. The input to simulation
model I is the vectorof designvariablesx1 with uncertainty (external
uncertainties, described by a range D x1 or certain distributions).
For simulation model I, the output y can be expressed as

y = F1 (x1 ) + "1 (x1 )

Reliability

Reliability

(1 )

where F1 (x1 ) is the simulation


model
and "1 (x1 ) is the correspondRandom
variable
ing error model of the internal uncertainty. Additive error model is
2. Reliability versus robustness in terms of the probability density function.
to Figure
represent
model
structure in
uncertainty
this
study, though
Figure 27:used
Reliability
Versus
Robustness
Terms of in
the
Probability
Density FuncAsits
figure
2
illustrates,
robust
design
is
concerned
with
the
event
distribution
near the mean of the
real form can be much more complicated.
tion [145]
probability density function, whereas reliability-based design is concerned with the event distribution in
the tails of the probability density function. Obviously, it is much more difficult to accurately characterize
the tail of a distribution than the center of the distribution. An additional consideration in distinguishing
between robustness and reliability is that the mathematical techniques used for solving robust design
problems are considerably different from those used for solving reliability-based design problems. The
mathematical methods for robust design procedures are less well developed than those for reliabilitybased design procedures, and this work is still largely confined to academic studies. Certainly, the
aerodynamic design procedures in use in industry are exclusively deterministic. (Recall that we are
excluding the use of random search methods to solve a deterministic problem.) There has been
considerable work on robust controls, but this work has been limited to using norm-bounded
descriptions of uncertain variables. Although the robust design principles of Taguchi (1987) are used in
aerospace engineering, these are
not1necessarily
the bestmodel
or evenchain.
appropriate methods for many robust
Fig.
Simulation
design problems.
Figure
28: Uncertainty Propagation in the Simulation Model Chain [58]

Traditional design procedures for aerospace vehicle structures are based on combinations of factors of
safety and knockdown factors, as illustrated in figure 3. Factors of safety are numbers greater than 1.0 that
of the output
chain
of simulations.
Theless
second,
thearestatistical
proare appliedfrom
to the a
loads.
Knockdown
factors are numbers
than 1.0 that
applied to theapproach,
strengths.
Both factors are intended to account for uncertainties. They have proven useful during nearly six decades
of design for conventional
airframes.
vides statistical
estimatesmetal
of the
outputs of the various codes, for example, estimating

the probability and cumulative density function outputs, terms that will be defined
in Section 3.2.2.
3.2.1

Forecasting

In an ideal circumstance, a decision maker would like to know what will happen in
the future. The aim of a forecast is to provide a decision maker with the knowledge of
whether a particular event happens. Twiss [137] defines four elements that a forecast
must contain to adequately provide a basis for decision making: 1) qualitative, 2)
quantitative, 3) time, and 4) probability. The qualitative element defines what should
be forecast, which selects the events which need to be considered. The quantitative
aspect requires selecting a measurement for a given metric. Since forecasting relates

92

Fig. 2 Ex

Inputs:
assumptions
data
insight
judgment

Forecasting Techniques:
exploratory
normative
Outputs:
qualitative
quantitative
time
probability

Figure 29: The Forecasting Process (adapted from [137])


future conditions toThe
the time
when theyProcess
occur, the
time element
is necessary. The
Forecasting
Figure
2.5 Twiss
probability element is necessary to capture the likelihood of a particular outcome due
to uncertainty in the system. The uncertainty aspect is important because it also be
used to identify potential causes of discontinuity in the system.
There are two basic forecasting techniques used to achieve the four elements described by Twiss. Exploratory techniques rely on selected attributes of historical data,
such as functional performance, technical and economic parameters, plotted against
time. Progress is assumed evolutionary and not random, enabling the generation of
characteristic forecasting curves made with a measure of certainty. Normative techniques work backwards from a proposed desired state to determine the steps necessary
to reach a required outcome. It possible to have many paths of development to
reach a desired outcome. The basic forecasting process described by Twiss is shown in
Figure 29, where either of the forecasting techniques use assumptions, data, insight,
and judgment to achieve the four desired forecasting elements. [135] [137]

93

3.2.2

Monte Carlo Simulation

A common approach to uncertainty quantification is the Monte Carlo simulation. A


Monte Carlo method is a way to generate information for a simulation when events
occur in a random way, using random sampling in a computer simulation in which
the results are generated repeatedly to develop statistically reliable answers [56]. It
is a useful method used to solve problems too difficult to obtain a numerical solution
analytically, by observing that fraction of the numbers obeying some property or properties as a result of the random number generation input to a computer simulation
[140]. Therefore, the Monte Carlo simulation is employed to approximate solutions
to quantitative problems with statistical sampling [4]. The Monte Carlo method has
found uses in many fields, as Shonkwiler [130] states, Monte Carlo methods are used
to simulate stochastic processes that occur in science, engineering, business and entertainment, to numerically solve problems in mathematics, science and engineering,
and to ensure the security of data.
Variability in input variables, or control variables, and in assumptions, or
noise variables, can be tracked by adding distributions to any point values. A
random number generator is used to sample from a bounded distribution assumed for
each variable, and then uses those variable settings to run a modeling and simulation
environment. Many cases with randomly selected points from these distributions can
then give a statistical indication of the probability with which each response will occur. A histogram, or graphical display of tabulated frequencies, can be created by
tracking the number of responses in discrete bounded bins, or non-overlapping intervals of some variable. The histogram in this case is actually the Probability Density
Function (PDF), which is the probability of observing a particular outcome. The
integral of the PDF allows for the generation of a Cumulative Distribution Function
(CDF), the mathematical function that maps the probability of obtaining a response
to the metric within the given range. Because the sampling of the response into bins

94

is a discrete process, the more samples, the smoother the PDF and CDF curves, and
therefore a confidence can be extracted with more precision.
The distribution shape and bounds for a given variable of interest requires a
working knowledge of the system in question, and is usually based on experimental
results, expert opinion, and most likely several iterations of trial and error for a system
where little knowledge is known. Several of the probability distributions used in this
study are described with more detail in Appendix A.
3.2.3

Filtered Monte Carlo Approach

According to Kuhne et al. [88], Probabilistic Design is the process of accurately


accounting for and mitigating the effects of variation in part geometry and other
environmental conditions while at the same time optimizing a target performance
factor. However, using mathematical methods, the authors state, probabilistic design may prove to be a complex and daunting task. Using logic and graphics along
with Monte Carlo simulation, Kuhne et al. demonstrate an alternate visual approach
called a Filtered Monte Carlo that achieves useful probabilistic design results efficiently and simply. This method assumes the existence of a fast-running simulation
model that can be called on many times.
The Filtered Monte Carlo approach shown as a flow chart in Figure 30 works
by populating a design space with response values obtained by running a simulation
many times with randomly selected values from bounded distributions on input variables. If the output for that particular Monte Carlo simulation trial violates any
response constraints defined a priori, that response is discarded. The outputs that do
not violate the constraints are then plotted in a scatter plot fashion versus any of the
inputs, given the user the ability to visualize sensitivity to variation in the inputs.
However, the authors note the biggest challenge to this approach is with problems
with large numbers of inputs and responses (i.e. >10), which drives the need for

95

X1

X2

Select variable value


from a uniform
distribution sample

Xn

Calculate Y

All Constraints
Satisfied?

Yes

Keep and Plot

No
Discard

Figure 30: Filtered Monte Carlo Approach (adapted from [88])


improved visualization and data mining tools that would enable the user to simultaneously explore the design space while conducting input variation sensitivity. The
authors claim great success with Filtered Monte Carlo approach on many application within General Electric Transportation-Aircraft Engines, including optimization,
robustness studies, and design space exploration.

3.3

Surrogate Modeling

To be able to conduct an uncertainty quantification or design space exploration, a


modeling and simulation tool would have to be executed many times. As stated earlier, the more Monte Carlo simulations run, the smoother the PDF and CDF curves,
increasing the accuracy in the trends and confidence in the predictions made from
those trends. In many cases, a design related simulation environment may be very
large, comprising of many codes running in series, parallel, or both, requiring anywhere from several seconds, to several hours for a single trial run. This may not be
a practical, or even feasible option when thousands of runs are required for a solution. Therefore, the use of a surrogate modeling approach is proposed to represent a

96

modeling and simulation environment, and increase the efficiency of the uncertainty
quantification process many times fold. A surrogate model, also known as a metamodel, is a model of a small segment of a more complex or sophisticated analysis tool
based on a regression of statistical inputs and response metrics. This model allows for
quick calculation of responses by employing equations to relate independent variables
to those responses. This section will cover two types of surrogate models proposed for
use in this thesis 1) polynomial Response Surface Equations, and 2) neural network
Response Surface Equations.
Many publications have been published on the use of polynomial Response Surface
Equations (RSE), and this thesis will not belabor the theory or derivations involved
in the regression techniques to achieve them. Interested readers may visit the Georgia
Tech Aerospace Systems Design Laboratory (ASDL) publication webpage [5] for a vast
amount of publications on the use of and theory behind polynomial RSEs. However,
the use of Neural Networks for data regression is less practised, and therefore this
study will explore the theory behind it.
3.3.1

Polynomial Response Surface Equations

A surrogate model is made by regressing against a set of data. In theory, a surrogate


model may be built around data collected from a large number of runs with random
selection of values within the bounds of input variables. However, for a very complex
system model requiring many time consuming computer codes to run, a structured
method for data sampling with the minimum number of simulation runs is needed. A
statistical approach to experimental design is necessary to draw meaningful conclusions from data. Montgomery [115] defines a statistical Design of Experiments (DoE)
as the process of planning the experiment so that appropriate data will be collected,
which may be analyzed by statistical methods resulting in valid and objective conclusions.

97

The simplest form of DoE is the full-factorial design. Each variable is discretized
into two or three levels, and the modeling and simulation environment is executed
for every combination of possible values for those discretized variables. A common
DoE for creating second order polynomial RSEs with minimum amount of simulation
executions is the Central Composite Design (CCD) [116]. This design combines a twolevel fractional factorial with center points (point at which all of the factor values are
zero, or midrange) and axial or star points (points at which all but one factor are
zero, and one point is at an outer axial value).
The polynomial form of an RSE is given in Equation 19. The coefficients that
make up the RSE are determined by regressing sample data points taken from a
modeling and simulation tool against the input variable used to obtain those data
points. The statistical software JMP [127] can be used to generate the required DoE
and regress the corresponding data in RSEs, acting as both a pre- and post-processor.

R = b0 +

n
X
i=1

bi x i +

n
X

bii x2ii +

i=1

n1
X

n
X

bij xi xj

i=1 j=i+1

where

n = number of factors
b0 = intercept regression coefficient
bi = regression coefficients for linear terms
bii = regression coefficients for pure quadratic terms
bij = regression coefficients for cross product terms
xi , xj = design variables or factors

98

(19)

3.3.2

Neural Network Response Surface Equations

A neural network is computational architecture based on the design of the interconnections of neurons in our brains [64]. Neural networks are emerging as a useful
way creating highly nonlinear regression models, and are a good alternative when
a polynomial RSE representation does not fit the design space well. When used to
create RSEs, a neural network is just a set of nonlinear equations that predict output variables from a set of given input variables using layers of linear regressions and
S-shaped logistic functions. The implementation introduced here is the one used by
JMP [127] to create neural network regressions, with many concepts and explanations
to the equations taken from help guides written by Demuth et al. [53] and Johnson
[81].
Neural networks have been applied to various fields. In the aerospace field, neural
networks have been used to develop aircraft autopilot routines and control systems,
and for defensive uses such as target discrimination and recognition. They have also
been used to create image processing software, predict stock market activity, develop
speech recognition software, and various other uses in the banking, manufacturing,
and medical fields. Many authors that write about neural network applications,
even recent works, cite the 1998 DARPA Neural Network Study [10] report for the
various applications of their use. The application of neural networks used in this
thesis proposal is different from most of the common applications identified above,
since here they will be used to fit, or regress, analytical models to output data from
computer simulations, rather than predict future results of additional real-world realtime experiments.
3.3.2.1

Neural Networks Background

The elements of neural networks were inspired by biological nervous systems, in which
the connections between elements determines the network function. A neural network

99

function is determined largely by the connections between elements. We can


train a neural network to perform a particular function by adjusting the values
of the connections (weights) between elements.
Commonly neural networks are adjusted, or trained, so that a particular input
leads to a specific target output. Such a situation is shown below. There, the
network is adjusted, based on a comparison of the output and the target, until
the network output matches the target. Typically many such input/target pairs
are used, in this supervised learning, to train a network.

Target

Input

Neural Network
including connections
(called weights)
between neurons

Compare
Output

Adjust
weights
Batch training of a network proceeds by making weight and bias changes based
on an entire set (batch) of input vectors. Incremental training changes the
weights and biases of a network as needed after presentation of each individual
input vector. Incremental training is sometimes referred to as on line or
adaptive training.

Figure 31: Neural Network Training Process [53]

Neural networks
have been function
trained to perform
complex functions
in various
can be trained to perform
a particular
by making
adjustments
to the values
fields of application including pattern recognition, identification, classification,
speech, vision and control systems. A list of applications is given in Chapter 1.

of the connections, orToday


weights,
between the elements. This training process involves
neural networks can be trained to solve problems that are difficult for
conventional computers or human beings. Throughout the toolbox emphasis is
on neural
paradigms
that build
up to or
in
adjusting the weightsplaced
so that
a network
particular
input
leads
toareathemselves
specificused
target
output, as
engineering, financial and other practical applications.

shown in Figure 31. By comparison of the output and target values, the weights are
vi
adjusted
until a match is made, within a specified tolerance. The collection of outputs

that is paired to the inputs is called the training set. The algorithm used to find the
values of the regression coefficients is called back propagation, whereby those values
are found in which they minimize the error as compared to the training set. This
can be achieved using any one of a number of optimization techniques (some relevant
techniques will be discussed in Section 5.3.5).
In a feed-forward network shown in Figure 32, the input layer receives values
which it feeds to a hidden layer. The hidden layer provides the networks non-linear
modeling capabilities [27]. The link between each input and hidden layer has a weight
associated with it. Another set of weights are placed on the connections between the
hidden and output layers. Note that each input node has an effect on each hidden
node, and in turn each hidden node has an effect on each output node.
The logistical function is the way the hidden layer handles the input. An Sshaped or sigmoid logistic function works as the weighted sum of the input values
increase, the value of the response increases gradually. An alternate approach would
use a threshold function, where the response is set to 0, and if the weighted sum of the

100

Input
Layer

Hidden
Layer

Output
Layer

X1

h1

Y1

X2

h2

Y2

X3

h3

Y3

Figure 32: A Simple Feed-Forward Neural Network

Figure 33: The Threshold (L) and Sigmoid (R) Logistic Functions [53]
inputs exceeds a threshold value, the response is set to 1. The threshold function does
closer resemble the actual behavior of the biological neurons, however, the S-shaped
function is used for the regression analysis since it is smooth and easily differentiable.
These logistic transfer functions, shown in Figure 33, and many other types used in
the other applications of neural networks are provided by Demuth et al. [53].
3.3.2.2

Neural Network Equations

The logistic function used to create neural networks, as given in Equation 20, is used
to scale the inputs between [0,1], and is referred to in JMP as the squish function.
As described earlier, this creates the S-shaped logistic function for input manipulation
within the hidden layer.
S (z) =

1
1 + ez

(20)

This logistic function is used to calculate the value of each hidden node, as shown in
Equation 21 by its application to a linear term related to the input variable.

Hj = SH aj +

NX
X
i=1

101

(bij Xi )

(21)

where
Hj = the value of hidden node j
SH = the logistic function for hidden node j
aj = the intercept term for hidden node j
bij = the coefficient for design variable i
Xi = the value of design variable i
NX = the number of input variables
The outputs are then calculated as shown in Equation 22. Note that the bracketed
portion fits within an interval of [0,1], where as the coefficients c and d are used to
scale the magnitude of the response.

Rk = ck + dk ek +

NH
X

(22)

(fjk Hj )

j=1

Finally, the steps described can be combined to describe the form of neural network
equation that JMP uses, as shown in Equation 23.

Rk = ck + dk ek +

NH
X

fjk

j=1

1
1 + e(aj +

PN
i=1

(bij Xi ))

!!

where
aj = the coefficient for hidden node j
bij = the coefficient for design variable i
ck = the response scaling intercept term for response k
dk = the response scaling coefficient for response k
ek = the intercept term for response k
fjk = the coefficient for hidden node j and response k
N = the number of design variables
102

(23)

NH = the number of hidden nodes


Xi = the value of design variable i
3.3.2.3

Overcoming Disadvantages of Neural Network Regression

Neural networks are very useful for creating regression equations when the design
space is highly non-linear and/or discontinuous. Any surface can be approximated to
any accuracy with the proper amount of hidden nodes [127]. However, this fact is not
without its price. Neural networks are not as interpretable as polynomial equations
since there is an intermediate (hidden) layer between inputs and outputs, rather than
a direct path. It is very easy to overfit a set of data with too many hidden nodes,
such that it no longer predicts future data well, resulting in an extremely multimodal
response surface, as shown in Figure 34. The fit itself is not parametrically stable,
meaning that many times the objective function may converge before the parameters
settle.

generalization in neural n
ularization, attempts to
3
is a common source of p
nalizing is done with an a
2
objective function. Inste
error, the optimization
1
tion (usually a weighted
term and the sum of s
0
ever, a new free param
The ratio of the error ter
1
penalization function is
or regularization parame
2
new function as an objec
generalization, it is diffi
3
value for the new free pa
6
4
2
0
2
4
6
8
10
12
to use.29 A technique
x
regularization paramete
Fig. 6 Fitting power of NN for g(x) = (x2)(2x+1)
2)
(1+x
has been implemented.
Figure 34: Example of Over- and Underfitting a Neural Network [128]
is based on the Levenb
algorithm, which is very
With a highly1400
multimodal design space, there are many local optima,
therefore
18
mediumthe
size networks b
large
problems.29 This i
16
0
1200
optimizer that finds
the neural network coefficients must 2be
run with
many
starting
to compute the approxim
14
it
in memory for each it
1000
8
1
Number of inputs
2
1
103
Another empirically b
16
800
20
10
generalization is Early
14
8
nique, the validation se
2
8
4

rees of Freedom

g(x)

Underfitting
Overfitting
Data

points. Each one of these starting estimates is called a tour. Of course, there is no
exact science in determining the appropriate number of tours. Too few will likely
not lead to the global optima, where as too many tours may exhaust the available
computational resources.
To counter the overfit problem, neural networks will fit with a holdback sample,
which is data not used in the estimation process. The holdback sample is called the
validation set when used along with the training set. An overfit penalty, or weight
decay, may be used to avoid overfitting the data. This puts a numerical penalty on
the size of the parameter estimates, and usually improves numerical convergence.
Unlike the polynomial RSEs that use a Design of Experiments to define the
cases needed to run to get the data needed to build the regression model, there
is no well defined methodology for neural network regression. Since the model fits
to both a training and validation set, the more random cases run the better the fit.
However, Johnson [81] states that it is important to include cases for a neural network
regression that include all effects that the model needs to simulate. Running only
random cases will lead to a model that does not fit well in the extremes of the design
space. Therefore, Johnson suggests running a combination of a Latin Hypercube
Sample (LHS), which would provide a list of random cases with as little correlation
as possible, along with a two-level Design of Experiments to capture the corners of
the design space.
An LHS is one of several space-filling designs, which are configured to uniformly
spread design points to the maximum distance possible (within variable limits). The
LHS method chooses points to maximize the minimum distance between design
points, but with a constraint that involves maintaining the even spacing between
factor levels and minimizing the correlation in the design [127]. An example of using
an LHS to fill two dimensional space with 1000 design points is given in Figure 35.
Similarly in three dimensional space, a cube would appear.

104

1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
-1

-0.8

-0.6

-0.4

-0.2

0.2

0.4

0.6

0.8

Figure 35: Example Latin Hypercube Used to Fill 2-D Space


3.3.3

Analysis of Variance

One of the major advantages of creating surrogate models is the ability to rapidly run
an analysis of variance (ANOVA), which measures the contribution of each variable
on the variability of each response. A useful way to visualize the results of an ANOVA
is in the form of a Pareto Plot, as shown in Figure 36 for an example of contributions
of gun accuracy to miss distance. Each horizontal bar shows the individual influence
of each error term on the response, listed in descending order, and the cumulative
effect of each of the terms is represented as a line (totaling 100%). This is also a useful
tool for examining the combined effects of variables. For example, the combination
of two variables may affect a response more than the some of each variable.

3.4
3.4.1

Application of Techniques
Response Surface Methodology & Design Space Exploration

Many processes for using surrogate models to expedite a Monte Carlo simulation to
quantify uncertainty and provide a robust design have been developed for aerospace
vehicle design. A major application of this includes creating the ability of forecasting
105

Term
Azimuth
Wind
QE
Initial Yaw
QE*Azimuth
QE*QE
Initial Pitch
Wind*Wind
Initial Pitch*Initial Yaw
Initial Yaw*Initial Yaw
Azimuth*Initial Yaw
Wind*Initial Pitch
Azimuth*Azimuth
Azimuth*Wind
QE*Wind
QE*Initial Pitch
Azimuth*Initial Pitch
Wind*Initial Yaw
QE*Initial Yaw
Initial Pitch*Initial Pitch

The bars represent


variability from
individual variables

20

40

60

Total
variability

80

100

Figure 36: Example Pareto Plot - Gun Accuracy Contributions to Miss Distance
the impact of new or immature technologies on a system to provide increased knowledge to a decision maker in the conceptual and preliminary phases of aircraft design
[85]. Mavris et al. [109] have applied the exploratory forecasting technique to select
from a list of potential technologies ones that have the most positive impact on the
design of a high speed civil transport using a technology impact forecasting (TIF)
method. This TIF method was implemented by applying multipliers, or k factors,
to various disciplinary metrics that are used by an aircraft design tool, using polynomial RSEs for rapid manipulation. This technology impact forecasting methodology
has also been applied to other aerospace systems, such as determining the technology
impact on subsonic transport affordability [84], on the design of an Uninhabited Combat Aerial Vehicle (UCAV) [111], and on a civil tiltrotor [104]. Surrogate models have
also been used to create a parametric environment to integrate disciplinary codes together to size a hypersonic cruise missile [62], as well as quantify the errors associated
with disciplinary codes and evaluate the impacts of those errors on the confidence
of reaching mission requirements and maintaining within given physical constraints
[61]. In addition, one of the major applications of the methodologies introduced in
106

this chapter is in the field of design for affordability, as will be discussed shortly.
3.4.2

Top-Down Capability Based Design

Soban [134] developed a formulation that links variables and responses across a
system-of-systems hierarchy using polynomial based surrogate models. In this formulation, the outputs of one level within the hierarchy feed directly to the next
higher level. Manipulating a subsystem level variable, for example the thrust-toweight ratio of an aircraft, changes its performance capability which in turn affects
the way it capability in the system-of-systems context, for example the way that
same aircraft would affect the course of a many-on-many campaign environment. If
the system level metrics are represented by surrogate models, the outputs of those
surrogate models can then be inputs to the next hierarchical level of surrogate models,
essentially enabling a rapid hierarchical modeling and simulation environment.
A Unified Tradeoff Environment (UTE), developed by Baker et al. [30] [31], allows
for simultaneous trades between design variables, requirements, and technologies by
coupling the customer requirements and technologies to the vehicle concept space.
Once again when polynomial surrogate models are able to represent a design tool,
rapid trade-offs enable real time decision making. Using a UTE enabled by these
surrogate models as a real-time decision making tool, a designer can see the effects
of evolving requirements on the design space of a vehicle.
The author has been part of an effort to expand these efforts to a more complex many-on-many campaign scenario by developing a hierarchical, surrogate model
enabled environment for systems-of-systems analysis [38], shown graphically in Figure 37. This hierarchical modeling and simulation environment consists of vehicle
and missile propulsion subsystem models, missile and aircraft (platform) synthesis
and sizing codes, and top-level military campaign analysis tools that are linked together using neural network based surrogate models into a UTE. Real time trades are

107

Figure 37: Hierarchical, Surrogate Modeling Environment for Systems-of-Systems


Analysis [38]
therefore enabled between design variables, requirements, and technologies at each
hierarchical level.
Using surrogate models to link the different subsystem and system models together
allows for rapid manipulation of the design space. Now, the value of each design
variable can be selected using a Monte Carlo simulation to fill the design space. In
the example shown in Figure 38, weapon performance values are output for each
Monte Carlo simulation selection of design variables. However, in the system-ofsystems context, this one weapon does not operate independently to effect top-level
metrics. Subsystem weapon attributes flow up the hierarchy, and additional variables
that would affect top-level metrics, such as enemy weapon capability are also selected
using a Monte Carlo simulation. In this example, the top-level metrics of platforms
lost and targets killed are observed. When this process is repeated many times, the
entire system-of-systems design space is filled. The results of this type of study are
108

best viewed using a multivariate scatter plot, or collection of bivariate plots shown
in Figure 38, showing the results of 1,000 Monte Carlo samples. Note that the plots
in Figure 38 only represent a fraction of the total design space. Although somewhat
overwhelming, when used interactively this process allows a user to perform both
bottom-up and top-down design.
From the bottom-up, individual weapon designs may be identified to compete
against an enemy capability and the results observed by studying the top-level results. However, with the design space filled at each level in the system-of-systems
hierarchy, the notion of top-down design comes in, as opposed to top-down requirements flow. Here, a user may highlight a capability region of interest, for example
defining desirable and undesirable regions (shown respectively as green and red in
Figure 38). Now, the types of targets that may be prosecuted, as defined by their
characteristic are identified. Taking this one level lower, the weapon performance requirements are defined. From this, it is evident that regions of improving performance
(i.e. higher weapon speed and range capabilities) yield a more desirable outcome, as
were defined at the very top of the decision making hierarchy. Finally, at the very
bottom, individual weapon designs are identified that yield desirable outcomes at any
level in the hierarchy. This top-down design based on desired capability allows for
true hierarchical decision making.
Viewing a hierarchical system-of-systems design space populated using a Monte
Carlo simulation brings in the notion of total versus partial derivatives, first discussed
by previous work conducted by the author [38]. In the field of mathematics, the
partial derivative of a function of several variables is the derivative of that function
with respect to only one of those variables, with all other variables held constant.
Generally speaking, a curve on a two-dimensional plot is a partial derivative, because
it shows how a function varies with respect to one variable. If any of the other variables
that define that function change, the shape of that curve will change. This is like

109

Figure 38: Top-Down Capability Based Design for a Campaign Scenario [110]
a contour plot, where the partial derivative is shown for several fixed values of the
other variables that define the function. A total derivative, however, is the derivative
of that function when all of the variables are allowed to vary 1 . Because every degree
of freedom is exploited, it is difficult to view the total derivative design space. For a
vector valued function, its derivatives are vectors as well. This directional derivative
is known as the full derivative 2 . For a vector function f~ with m functions and variable
vector ~x with n variables, Berglund [37] gives the full derivative Df~ as the Jacobian
1

This also includes indirect dependencies of the function due to intermediate variables. For
example, the function y is dependent on variable x, but variable x is a function of t, therefore y is
indirectly dependent on t [141].
2
The terms total and full derivatives are usually interchangeable in much of the literature, and
many times slightly different definitions are given to describe them. This thesis will assume the form
given in this section.

110

Matrix:

Df~ (~x) =

3.4.3

f1
x1
f2
x1

(~x)

f1
x2

(~x)
..
.

f2
x2

fm
x1

(~x)

(~x) . . .

f1
xn

(~x) . . .
..
...
.

f2
xn

fm
x2

(~x) . . .

(~x)

(~x)

..

fm
xn

(24)

(~x)

Design for Affordability

There has been a shift in the design industry to design for affordability, and several
of the techniques introduced earlier have been applied to this field. Mavris and
DeLaurentis [105] state that the design process must not only address interactions
between the traditional aerospace disciplines, such as aerodynamics and propulsion,
but also include life cycle disciplines, such as economics, reliability, manufacturing,
safety, and supportability. However, these additional disciplines bring with them a
variety of uncertainties into the design problem. The presence of uncertainty demands
a stochastic treatment of the design process, handled with uncertainties.
The term affordability here is used as a measure of cost effectiveness. It is defined
here as the ratio of desired outcome (i.e. performance) over the cost of a given system.
A more affordable system therefore is not always the least expensive one. It can be
thought of as a measure of how much performance each dollar buys. Designing with
affordability concerns in the aerospace industry was driven primarily by military budget cuts and increased competition in the commercial market [108]. So much so, the
Defense Manufacturing Council has requested that cost be treated as an independent
design variable on the same level as traditional design variables [97]. Affordibility can
also be a measurement of product effectiveness against both cost and risk [105]. Affordability itself can be thought of as a multiattribute and multidisciplinary process.
More on implementing life cycle analysis to reduce total cost within the conceptual
design phase of aerospace systems can be found in [98] [107]. However, the notion of
treating cost as an independent variable brings forth the desire to have the ability to
111

treat any response metric, which is a function of independent variables, itself as an


independent variable.

3.5

Chapter Conclusions

This chapter reviewed advanced design methodology concepts that could be used to
enable the creation of a process to set the design requirements of a complex system
across a multilevel decision making hierarchy. The system-of-systems concept was
introduced as a methodology for describing events at every level by functionally decomposing the entire system in question into a hierarchy of decision making levels.
Several uncertainty analysis and quantification techniques were discussed, including
a filtered Monte Carlo approach that uses a Monte Carlo simulation to determine
the combination of bounded design variables that do not violate a series of response
constraints. Two surrogate modeling techniques were discussed as potential methods
for representing bounded segments of complex analysis tools using a series of regression based equations. Finally, an initial study into the topic of top-down capability
based design was introduced as a method for implementing the cross hierarchical flow
of information. Using the information gathered in the literature sections, a set of
additional research questions and hypotheses can now be formed, in an attempt to
expound on the primary research objective of this work.

112

CHAPTER IV

RESEARCH QUESTIONS & HYPOTHESES


4.1

Problem Revisited

The motivation chapter introduced a problem in which a need was established for a
process enabling the rapid manipulation across a hierarchy of decision making levels in
the systems engineering process. An example problem was presented in which an air
defense weapon is required to defeat the threat posed by rockets, artillery, and mortars
(RAM) fired from heavily populated areas. Several issues were brought up throughout
the chapter that impede a quick solution to this problem, and technology barriers of
the proposed solutions that the Department of Defense is studying were reviewed.
Formal requirements have been established by the U.S. Army under the Extended
Area Protection & Survivability (EAPS) program, describing the needs for different
defendable area tiers. Initial studies have been conducted by various organizations
under the EAPS program to define general needs, and to examine potential solutions
to the problem.
The second chapter was dedicated to presenting methods for modeling and simulation of combat systems pertinent to the topic in question. It was determined,
however, that there is no defined method to modeling the RAM threat or air defense
systems that could be used to defeat them. This established the need to create a
modeling and simulation environment that brings together the various interacting
components of this problem. However, this leads to the issue of simulating complex
systems, each with its own uncertainty contributions to capabilities at various levels
of the hierarchy. Many of these simulations may be physics based tools that would

113

exhaust most computational frameworks, especially if many executions would be required such as in any uncertainty quantification process. Therefore, various Advanced
Design Methodologies were presented in Chapter 3 that could be applied to assist in
this problem. Many of these methodologies have been established through many years
of use, and several were more recently developed such as those the author was a part
of developing.
The purpose of this thesis is not to provide a specific solution down to the detail
design level of a specific problem, but a process that may be applied to complex
revolutionary problems in which independent components interact with one another
in a hierarchy to effect an outcome. The goal is to define capabilities anywhere in
a hierarchy, even in many cases where there will be components outside the users
control, and define the requirements of the components within the users control.
The following research questions will formally address this and related issues. First,
the initial formulation research questions introduced during the motivation to this
thesis will be revisited for consistency. Next, more specific research questions will
be presented, inspired by information gathered during the literature search. The
set of hypotheses will propose specific answers to those research questions based on
literature, evidence, and information concerning related studies collected as part of
the literature search conducted for this thesis.

4.2
4.2.1

Research Questions
Initial Formulation Research Questions Reviewed

Section 1.1.2 introduced two research questions that inspired the research objective
of this work, as well as the direction of the literature searches reviewed in Chapters
2 and 3.
Research Question 1: What is the systems engineering process that can
be applied to set the design requirements of a complex hierarchical system,
114

and what are the major deficiencies associated with that process?
Several deficiencies were identified for the systems engineering process discussed in
Section 1.1.2, as applied to a sample complex system. In summary, there must be an
iteration between adjacent levels of the hierarchy to make sure that the design solution
satisfies the requirement, therefore there can be no direct flow between nonadjacent
levels. Therefore, the second Research Question asks how an improved process can
be created
Research Question 2: How can a process be created that enables rapid
manipulation across a hierarchy of decision making levels?
This question drives the need for a specific process to be defined, as opposed to
a point design solution for a given set of rigid requirements and constraints. This
question also motivated the direction taken as part of the literature searches. As an
example problem, a solution designed to satisfy the requirements drawn by the EAPS
program requires several independent systems to function together within an operational environment. Each of these systems has its own uncertainties associated with
it that affect the combined systems overall capability. The operational environment
may comprise of several high level requirements that must be satisfied, as well as any
other constraints within the hierarchy.
4.2.2

Additional Research Questions

Based on the methods and techniques discovered during the literature search phase,
more specific research question were formulated regarding this thesis, which are included in this section.
Research Question 3: An ANOVA/Pareto plot technique can be used
to quantify the contribution of each variable to the variability of a response,
but is there a way to identify which variables drive each other?
115

This question is related to the topic of identifying the variables that have the
most impact on the variability of a response. The contribution of each variable to
the variability of a response can be measured by conducting an Analysis of Variance
(ANOVA) based on main effects (i.e. the bounds of the variables), and visualized
using a Pareto Plot. This has been proven very useful because the variables that
have little or no contribution to response variability can be defaulted to a reasonable
value and left out of any further parametric study. However, this gives only a one
dimensional representation of the results, listing the percentage of response variability
contribution of each variable. This gives no indication of the variability tradeoffs
between the variable bounds, or the tradeoff between the variables themselves. More
specifically, of the variables that do contribute to the variability of the response (i.e.
the variables that drive the response in a particular direction), are there variable
settings that drive any other variable in one direction or another based on response
constraints?
Research Question 4: Can a tradeoff space be identified interchangeably
between variables and responses?
Question 4 is a natural extension of explanation given for the previous question.
Regarding the identification of tradeoff space between variables based on response
constraints, the same question can also be asked regarding placing constraints on
variables as well as responses, leading to the notion of creating a trade space between
any combination of variables and responses based on any constraints placed on any
other variables or responses.
Research Question 5: How can rapid design tradeoffs be assessed in a
hierarchical environment composed of independent time consuming models
if the design space between them is multimodal or discontinuous?

116

The topic of Response Surface Methodology has been well established as a useful
way to capture a small segment of the design space of a complex computer code,
reducing the wait time for the code to run to fractions of a second. However, this
requires that the user bound the design space tightly enough so that it could be
accurately be represented using a series of polynomial equations, or transformations
thereof. An argument can certainly be made that one could constrain any design
space so tightly that even a linear trend would suffice. Across a hierarchy of design
levels for interacting systems, however, it may be desirable or even necessary to
open up the design space to allow for tradeoffs across the hierarchy and between
the systems. This leads to the likely potential that polynomial regressions (or their
transformations) cease to accurately represent the design space.
Research Question 6: How can the effects of uncertainty propagation
at any level within a hierarchical decision making design environment be
assessed?
There is a clearly defined need to be able to quantify the effects of uncertainty of a
systems performance within its operational environment. Sources of uncertainty may
include errors in system attributes or unexpected deviations from expected performance. When dealing with a hierarchical environment in which interacting systems
may be decomposed into the subsystems that describe them, there may exist sources
of uncertainty at any level that may effect the outcome from the bottom-up. It is
important to have the ability to quantify those sources of uncertainty regardless of
position in the hierarchy. Another source of uncertainty may also lie in the top-level
requirements or constraints. For example, the exact requirement may not be known,
but a region of interest may be defined. This drives the need to determine how those
top-level regions of interest affect the design space of lower level metrics and variables.

117

4.3

Hypotheses

The hypotheses presented in this section are an attempt to answer the research questions given in the previous section, using methods collected in the literature search.
These hypotheses will shape the research formulation and approach taken for the
remainder of this thesis. The first hypothesis encapsulates the overall research objective of this thesis.
Hypothesis 1: A system-of-systems approach can be used to create an
environment that will allow for both weapon requirements definition and
system optimization from top-down requirements that take into account
operational environment characteristics, as well as for analyzing the impacts of subsystem level changes on top level metrics from the bottom-up.
Information acquired about system-of-systems and the overall systems engineering
process inspired the approach proposed in this thesis to create a hierarchical decision
making design environment in which multiple independent systems interact to affect
an overall metric. The notion of top-down design in the field of systems engineering
introduces the concept of flowing requirements from high level needs down to system
and subsystems. This thesis takes that notion one step further to propose constructing
a design environment that enables simultaneous bottom-up and top-down design. To
do this, the entire process must be functionally decomposed by degrees of freedom
allotted for each component. The ability to conduct rapid or real-time tradeoffs are
addressed in Hypothesis 2.
Hypothesis 2: A hierarchical environment composed of neural network
based surrogate models will enable multimodal and/or discontinuous design
spaces to be accurately represented, allowing for rapid tradeoffs. These
tradeoffs may be conducted with real time application of constraints at

118

any level of the hierarchy, for both independent variables and dependent
responses.
Assembling a hierarchical design environment as discussed in the first hypothesis
usually requires assembling simulation tools of various complexities. Surrogate models based on predetermined simulation runs have proven to be useful in representing
small segments of design space by providing parametric polynomial equations. As
the complexity of the design space increases, polynomials are less likely to be able
to represent the complexity of the design space, and certainly not any multimodal or
discontinuous effects. Neural network based surrogate models are emerging as viable
alternatives to polynomials, and have the ability to model highly nonlinear, multimodal, and discontinuous design spaces. As noted in the literature search discussing
them, the use of neural networks have several drawbacks that must be overcome.
Special attention must be paid to not over fit the data, leading to a surrogate
model that is not able to predict future points not used to create them. Additionally,
as the number of variables increases, the complexity of the neural networks requires
more computational resource to properly execute the optimizer calculating the neural
network coefficients.
Once the technical hurdles will be overcome, neural network based surrogate models will enable rapid tradeoffs of complex design spaces. In conjunction with Hypothesis 1, these neural network based surrogate models will be the prime enabler of rapid
cross hierarchical manipulation, and with proper design space allotment, requirement
constraints can be applied in real time revealing how the design space of every other
response and variable changes.
Hypothesis 3: A properly constructed system-of-systems environment
will allow for the rapid assessment of the effects of uncertainty at any
level in the hierarchy on any other level in the hierarchy.

119

Many sources of uncertainty may exist among the various interacting components
of a system-of-systems design environment. Using a hierarchical system-of-systems
design environment composed of neural network based surrogate models, the effects
of uncertainty propagation throughout the environment can be rapidly assessed. An
even more profound proposal using the combined conjectures of the hypotheses presented in this section is the ability to bound the acceptable errors or uncertainties
in individual components based on top-level requirements. This gives the ability to
bound an acceptable region of uncertainty anywhere in the system-of-systems hierarchy for a desired level of total system robustness.

120

CHAPTER V

RESEARCH FORMULATION & APPROACH


5.1
5.1.1

Decomposing the Bounded Problem


Narrowing the Scope of Interest

In an effort to provide an example problem to study to specific systems engineering


process identified, Chapter 1 discussed the urgent need to provide an air defense
system to protect coalition assets in the Baghdad Green Zone from the threat of
rockets, artillery, and mortars. However, it is unfeasible as well as out of the focus
of this thesis to propose a complete solution to this specific problem with substantial
technical merit expected of this work. Therefore, the need to defend the Green Zone
only serves as a motivation to shape a sufficiently bounded problem studied in this
thesis. This is an appropriate point to discuss how the problem scope is further
bounded to provide attainable goals suitable to proving the hypotheses presented in
Chapter 4.
One of the hypotheses stated that a system-of-systems approach can be used to
assemble a hierarchical decision making environment composed of interacting systems.
This may bring to mind a generic campaign code such as those discussed in Chapter
2, in which any number of the identical or different systems may interact with many
opposing threats (i.e. a squadron of fighters aircraft destroying an enemy IADS).
This thesis will not study a many-on-many scenario relating to this problem, for
example defending a specific area shape and size with multiple air defense systems
simultaneously intercepting multiple targets. Only single instances of each system
type will be examined, such as one gun firing one interceptor projectile type against
one target. However, once the a suitable architecture is in place to analyze a single

121

weapon system and engagement scenario, statistical methods introduced in Chapter 2


will be used to determine the combined effect of firing multiple projectiles at a single
target.
Several interceptor solutions and target types falling in the RAM threat category
were identified in Chapter 1, and there is obvious benefit to studying each of those in
great detail in addition to any new ideas. This is most certainly out of the scope of this
thesis, and therefore a representative scenario will be defined in which an air defense
weapon intercepts a mortar threat. Additionally, top-down design was introduced as
a method of flowing top-level requirements down to the system and subsystem level.
There is more than one system interacting in this problem scenario, and each one can
certainly be functionally decomposed into any number of desired degrees of freedom.
For this to occur for each of the systems would create a daunting task, especially when
the time comes to show how top-level requirements impact subsystem level variables.
Therefore, only one of the systems will be decomposed to the subsystem level, as will
be discussed later in this chapter.
There may be a concern regarding bounding the problem as such, while maintaining a true system-of-systems context. These restrictions are in fact done to enable a
thorough study of a top-down/bottom-up flow of requirements and constraints within
a hierarchical system-of-systems environment, from subsystems, to systems, to the operational (i.e. system-of-systems) level, as stated by the primary research objective
of this thesis.
5.1.2

Morphological Matrix

Section 3.1.1 discussed methods for functionally decomposing a problem or system


into the subsystems that enable it. In addition, the morphological matrix, or matrix
of alternatives, was introduced as a method for functional decomposition by listing
key parameters or functions of a system and the alternative methods for satisfying

122

them. The morphological matrix of alternatives for the problem of interest is shown
in Table 9. Presented in the morphological matrix are all of the methods discussed in
this thesis relevant to addressing the problem of interest, namely defending against
the RAM threat. This exercise helps identify which degrees of freedom to keep open,
and which ones are to be locked down. As will be discussed, many of the factors
are locked down for subjective reasons, whereas others simply due to computation
limitations. In some cases, more than just one option will be selected in order to
maintain that particular degree of freedom.
This exercise suggests the extent of the magnitude of the problem, as the number
of potential solutions is calculated by multiplying the number of solutions per row
by each other. Some of the selections are not compatible with each other, however
several of the rows have multiple options that could be selected simultaneously. The
amount of perturbations even in this already constrained design space numbers in the
hundreds of billions. Even if a modeling and simulation environment was in place to
handle every one of the combinations, it would take many years of computational run
time assuming even only one second required to complete a single case. Therefore, it is
important to have clear justification of why certain decisions were made about design
freedom, and that the proper functional decomposition took place. What follows is
a description of each of option discovered in the literature search that could satisfy
the functions required of a related air defense system, with justification of concept
selection, and and overall functional decomposition of the entire process.
5.1.2.1

General Alternatives

The general problem is first decomposed into the defended range tiers. The EAPS
problem definition discussed in Section 1.2.2 defines an inner tier broken down into
a short (500 m - 2 km) and long (2 km - 10 km) ranges, and an outer tier for
ranges beyond 10 km. The potential intercept solution types discussed included gun

123

General

Projectile

Lead Angle
PID
40mm
Blast/Fragmentation

Pursuit
Bang-Bang
30mm
Kinetic Kill

Guidance

Navigation Law
Controller
Bore

Warhead

Adjustable
Air Density
Fuction of Specific
Selected Gun
Conical Scan
Azimuth
Random Noise
60 mm Mortar
multiple
low slant (<45o)

Fixed
Muzzle Velocity
Calculated for
Optimum Performance
Sequential Lobing
Elevation
Bias
Artillery
single
head on (0o)

Firing Error

Gun Firing rate

Tracker Type
Radar Error
Radar Error Type

Threat Types
# simultaneous threats
Threat Firing angle

Self-Destruct Projectile

Interital

None

Intercept Solution

None

Rocket/Missile

Gun Fired Projectile

Engagement Range

Collateral Damge
Reduction Consideration

2
Long Range Inner Tier
2km - 10km

Short Range Inner Tier


500m - 2km

high slant (45o-90o)

82 mm Mortar

Monopulse
Range

Function of Bore
(historical database)

Air Temperature

Forced Tumble

Pitch
120 mm Rocket

107 mm Rocket

Projectile Weight

Custom

Passive Homing

Roll

Range/Cross Wind

57mm
Calculated Lethal
Radius Based Need

50mm
Submunition Dispersal

Proportional

Command

Three Point

Beam Rider

Directed Energy/Laser

Outer Tier
10km +

Table 9: Morphological Matrix of System Alternatives


1

Firing Muzzle Velocity

Gun

Radar

Threat

124

Yaw

Elevation

Semi-Active
Homing

Azimuth

Active
Homing

fired projectiles, a self propelled rocket or missile, and a directed energy or laser
weapon. The selection of defendable range as defined by the selected engagement
range will drive the intercept solution type, as discussed in Section 1.2.2.1 there will
be a range at which an unpowered gun fired kinetic kill projectile will not be able
to deliver the required kinetic energy to destroy a notional target, driving the need
for a self-propelled solution. Additionally, a directed energy solution was discussed,
but the immaturity of technology associated with it hinders its use in the desired
environment discussed Chapter 1. However, the simplest solution is desirable for
many reasons (cost, technology limit, etc...), and therefore a gun fired projectile will
only be examined for the SRIT requirement.
5.1.2.2

Projectile Alternatives

One of the major subjects discussed in the EAPS community regarding a gun fired
projectile solution is issue of guidance. Namely, the question asked is how to quantify
the benefit of having a guided versus an unguided projectile. If in fact guidance is
preferred, there are several options discussed as relating to the field of missile design.
The first was the use of inertial guidance that relies on gyroscopes and accelerometers
to direct the missile over a predetermined path provided to the missile prior to launch.
The next option listed is beam rider guidance, where a rearward facing antenna on
the projectile senses an electromagnetic tracking beam and uses the position of the
missile with respect to the center of the target tracking beam to compute steering
signals used to move control surfaces. Command guidance is where a projectile gets
its guidance instructions from a tracking system outside the projectile to track itself
and the target along their trajectories, aiding the projectile to steer itself to intercept
the target. Passive homing uses sensors onboard the projectile that can only use
electromagnetic emissions or natural reflections from the target to establish location.
The next option is semiactive guidance, whereby a source other than the projectile (i.e.

125

the ground radar) illuminates the designated target, and a sensor on the projectile
sensitive to the energy reflected off of the target guides the projectile. Finally, the
active guidance option involves having the target illumination transmitter and target
echo receiver co-located on the projectile itself, however this option has historically
been restricted to larger weapons that have space to carry those components. Since
command guidance only requires a receiver on board to receive commands from a
ground radar, it is the simplest solution and allows for a less complex projectile.
Four types of navigation laws were discussed. The pursuit trajectory causes the
missile to always fly directly pointing at the target at all times by maintaining the
line-of-sight (LOS) along the heading of the missile. The lead angle navigation routine
causes the missile to keep a constant bearing with the target with the intention that
the missile flies a straight collision course with the target. Three-point navigation is
a routine where the missile is constantly steered along the LOS between an offboard
target tracker and target. Finally, a proportional navigation law uses a guidance
package that tries to maintain a constant LOS angle between projectile and target by
making the rate of change of the heading direction proportional to the rate of change
of the LOS.
Assuming that a guided projectile is used, the next option to select from is the
controller type, and two were discussed as potential options. The first discussed was
a bang-bang controller, where the application of control is either full on or off, with
no ability for fine adjustments . The other method discussed was the Proportional
Integral Derivative (PID) controller, where amount of feedback control applied is
proportional to the error in the desired trajectory. Practically, this means that using
PID control, the amount of control surface deployed would be a function of desired
aerodynamic force required, whereas a bang-bang controller would only have the
option of fully deploying the control surfaces. Clearly the PID option is far more
complex component wise, but would provide the best course correction for a guided

126

projectile.
The issue of projectile bore stems from the desire to use a projectile that can
be fired from an existing gun. This would limit the projectile to specific discrete
bore sizes, which has several disadvantages. The first is that in most likely cases,
any one of the existing bore sizes would be either larger or smaller than an ideal
calculated bore size (i.e. if a designer could optimize around bore size), which could
be a function of internal volumetric packing of guidance components, warhead, etc...
and required external aerodynamic forces to steer the guided projectile. However,
limiting the projectile bore to one compatible with an existing gun greatly simplifies
determining optimal bore size in the simulation process, and is based on the actual
practicality that it would be a very expensive venture to design a new gun based
on a non-standard projectile bore size. Therefore, the projectile caliber, although a
variable, will be discrete to one of the following four existing and standard guns bores:
1. 30 mm
2. 40 mm
3. 50 mm
4. 57 mm
The issue of internal packing leads to the next option for the projectile listed
in the morphological matrix. The projectile may or may not carry a warhead. If,
for example, a guided round could be designed to directly intercept a target with
a high enough confidence, and assuming that enough kinetic energy is delivered to
detonate the targets warhead or render it incapacitated, a kinetic kill is sufficient
and a warhead is not needed. Now the issue of confidence in a direct hit comes into
play, and begs the question of how many projectiles would have to be fired at the
target to achieve a confidence of at least one direct hit. If a warhead is desired, two

127

options discussed were the blast/fragmentation warhead and a warhead that simply
disperses submunition ahead of the target flight path. Of course, the desire for a
small inexpensive projectile seriously constrains the ability to use a warhead. The
last option listed for the warhead leaves the need for a warhead open, based on a
lethal radius required around a projectile to destroy the target based on a calculated
miss distance confidence.
Finally is the issue of collateral damage, which is directly stated as an important
issue for consideration this problem. One option is designing a self-destruct projectile
that is fuzed destroys itself after a certain amount of time after fire, such as the Phalanx based C-RAM discussed in Chapter 1. Another option is deploying an external
surface after forcing the projectile into a tumble, decreasing the velocity at which
the projectile hits the ground, as well as limiting the range otherwise taken by an
unaltered ballistic flight path. However, if the projectile already had a warhead this
would not be an issue, as the warhead could be fuzed to detonate after a given time
after launch if there was no target intercept. Of course, an ideal kinetic kill projectile
would be designed to directly impact the target every time and not need to consider
this issue, assuming that both interceptor and target are completely destroyed in the
process.
5.1.2.3

Gun Alternatives

Decomposing the functions of the gun firing the projectile is a matter of selecting the
desired degrees of freedom. As mentioned earlier, it was difficult to find literature on
medium caliber gun design in open literature. However, in a computer simulation,
the gun firing the projectile can in many cases be defined simply by the firing position
(in azimuth and quadrant elevation) and the velocity at which the projectile leaves
the muzzle. The first option in the morphological matrix for the gun is whether to
keep open the degree of freedom to select muzzle velocity, or to simply fix it at some

128

value. The simple option is to assume that a specific gun design firing a specific
unpropelled projectile has a single muzzle velocity associated with it, therefore the
firing position would be a function of determining the azimuth and elevation required
to cause the ballistic flight paths to intersect. For a head on intercept, the problem
further simplifies to only adjusting the firing elevation. However, this limits the ability
to determine the optimal firing muzzle velocity, i.e. a major design variable for the
gun.
Another major function of the gun design would be the pointing error accuracy.
When information is known about the projectile flight characteristics, desired trajectory, and muzzle exit velocity, a fire control solution can prescribe the gun pointing
characteristics (i.e. elevation and azimuth). However, the gun may not be pointing
in the exact direction at the time of fire, which result in error surrounding the elevation and azimuth from the intended firing direction. In addition, the velocity at
which the projectile leaves the gun muzzle may not be exactly what was intended.
These gun firing errors may play a very important role surrounding the trajectory of a
non-propelled projectile, and certainly for an unguided projectile following a ballistic
trajectory.
Actual guns would have specific error distributions associated with those properties, usually given as normal distributions with the standard deviation given around
the desired property. Differing guns would have poorer or better accuracy, depending on that reported standard deviation. Several sources of error for the gun were
introduced, and is the second element listed for the gun in the morphological matrix.
These can be broken down in to three major groups: firing condition (muzzle velocity,
elevation, azimuth), atmospheric condition (wind, air temperature and density), and
projectile manufacturing tolerance (the difference in weight from one projectile to
the next). Because the projectile/target interception is not meant to happen at very
high altitude and at relatively large ranges, the effects of perturbations in density

129

and temperature will not have a great effect on pointing error and can be eliminated.
Sensitivity to randomness in wind is certainly an important aspect in gun design,
however to limit the scope it will not be a factor for this study. The errors based on
firing conditions are the important factors that define the properties and capabilities
of a particular gun system, and can not be eliminated.
Finally, after the capability of one projectile is established, it is possible to statistically determine the capability of multiple projectile fired simultaneously can be
determined (Section 2.3.3). This means that for a given engagement time window,
the amount of projectiles fired (from a single gun) is an important factor, therefore
rate of fire is listed as the third option of the gun in the morphological matrix. The
first option is to calculate based on given condition to number of bullets fired, since
too few would not result in an acceptable probability of hit, but firing too many would
simply waste resources. If a specific gun system is selected, the rate of fire would be
fixed for that gun. Finally, a historical database of gun bores can be used to establish
a simple relation between bore size and rate of fire.
5.1.2.4

Radar Alternatives

The radar plays an important role in the systems hierarchy in that it relays target
information to determine a firing solution (i.e. how to point the gun), and would also
be used to track the projectile so that course correction guidance updates could be
sent. Three types of radar tracking methods were introduced, including sequential
lobing, conical scan, and monopulse. Sequential lobing is a method in which the antenna beam is alternately switched between two positions to obtain the angular error
in one coordinate. Instead of discontinuously stepping the beam between four discrete
positions, the conical scan method continuously rotates the antenna. To determine
angular measurement, both of these methods require receiving multiple pulses, or
the received echo, as the antenna translates. The monopulse method requires only

130

receiving one pulse to make these measurements, however this requires the use of at
least two antennas in each dimension measured. This thesis also reviewed some of the
methods developed to approximate the error in the received signal using the phase
difference in the two signals (per dimension), as is done with an interferometer radar.
A measure of radar capability is in its level of accuracy when taking measurements
of an object. All radars have errors associated with the measurements they take of an
object. These errors include measurement of information about the target in relation
to the radar which identify the key attributes of an object in three dimensional space
including elevation, azimuth, and range relative to the radar. The radar can also
measure orientation properties including the moments about the axes of the object,
including yaw, pitch, and roll orientations. Figures 39 and 40 graphically depict the
orientations described here. The errors may be a fixed bias that is the same for
every measurement, a random noise that is different for each measurement defined
by a distribution, or a combination of the two. As discussed regarding the gun firing
errors, a radar may have measured error distributions associated with measurements
it takes in any of the dimensions. These errors are commonly normally distributed,
with a reported standard deviation about a mean. Radar measurement accuracy can
also be rated subjectively as poor or advanced, and a given radar system would
fall somewhere in that subjective scale depending on the standard deviations that
define its error distributions. However, for this study it is of great interest to leave
these values open for design space definition later on.
5.1.2.5

Threat Alternatives

The first chapter defining the motivation for this thesis discussed the need for an air
defense system capable of defeating the threat due to rockets, artillery, and mortars,
or more commonly referred to as the RAM threat. Ideally, one would design a single
weapon that is optimized between the tradeoffs concerning the different target types.

131

Range
Range

Azimuth

Quadrant Elevation

Figure 39: Side (L) & Top (R) Views of Gun and Radar Orientations
Yaw

Pitch

Roll

Figure 40: Moments About Projectile Axes


For example, a different firing solution would be obtained for rocket target flying a
shallow trajectory than a mortar that is commonly fired at very high trajectories. In
Range

the morphological matrix, specific RAM threats are listed. All Range
gun fired Azimuth
projectiles
Quadrant Elevation
are listed as artillery, and two common mortar
sizes are given. Additionally, two

common rocket sizes are given as well.


In addition to selecting the target type, for the purposes of assembling a decision
making hierarchy the number of simultaneous threats to be encountered must be determined. This is simply listed as either single or multiple. Finally, the potential
threat firing angle can be defined. As mentioned in this section, the problem can be
greatly simplified if a head on intercept is assumed, however a slant angle between
the trajectories of the interceptor and target can be identified. The greater the slant
angle, however, the more difficult it would be to ensure interception.

132

5.1.3

Hierarchical Decomposition of Proposed Solution

5.1.3.1

Identification of Key Metrics

The previous section provided alternatives based on functional decomposition of each


independent system. In order to focus the scope of the problem definition, this section
will hierarchically decompose the problem by function using the systems discussed
in the previous sections. This first requires that certain quantifiable top level requirements be established beyond the qualitative desire to intercept a target with
an inexpensive solution. As a result, several high level metrics are now defined to
address the problem. The first metric defined will be the minimum miss distance
between the interceptor projectile and the target. Since there will be uncertainties
to account for in quantifying this metric, there will have to be confidences associated
with reporting specific values of intercept miss distance. This one metric allows for
direct determination of two key metrics. For one, a lethality radius requirement can
be extracted from the miss distance quantification. Nevertheless, since kinetic kill
is the ideal option and less complicated than using a warhead to destroy the target,
the probability of direct hit must also be quantified. As might be expected, the most
intuitive solution to decrease the intercept miss distance and ultimately increase the
probability of hit would be to increase the radar and gun accuracies. In lieu of some
form of cost metric, the simple answer would be to just design a projectile around
the best radar accuracy available. Therefore a form of a cost metric that penalizes
for increased radar accuracy must be accounted for.
The system-of-systems hierarchy decomposition used in this thesis is given in
Figure 41. At the very top, the capabilities of interest are identified at the operational
level, including probability of direct hit, warhead lethality radius, and a measure of
weapon system cost as relating to the radar. Probability of direct hit and warhead
lethality radius are directly related to miss distance quantification, which is a function
of specific properties of the radar, gun, interceptor projectile, and target as well as a

133

Problem Hierarchical Decomposition

Operational
Level
Properties

Probability of
Direct Hit

Radar
Cost

Subsystem
Level
Properties

Intercept
Range

Miss Distance Quantification

Radar
Properties
System
Level
Properties

Required Lethality radius


for Warhead

Update Rates
(Radar &
Control States)

Projectile
Location &
Orientation
Noises

Gun
Properties

Pointing
Errors

Firing
Rate

Interceptor
Properties

Bore

Scaleable
geometry/
kill
mechanism

Target
Properties

Controller
Type

Fixed
geometry,
trajectory

Subsystem
Radar
Variables

Figure 41: Hierarchical Decomposition of Proposed Solution


specified intercept range. Each of the the systems operating at the system level can
further be decomposed to their own properties. When each of the components work
together in an assembled environment providing the desired outputs as a function
of the desired inputs, the environment is considered functionally composed. The
individual components of the simulation will be discussed in this section. Many of the
assumptions taken to focus the scope of the problem were taken with the advice and
consultation of the thesis committee members in their respective fields of expertise.
5.1.3.2

Radar Decomposition

The radar properties will be decomposed to two separate property types. The first
includes the noises describing the errors in measuring the different projectile location
and orientation properties. The error in the radar measurements of the targets
location directly impacts the position that the gun barrel points to fire the interceptor
(this is in addition to the inherent pointing error due to mechanical errors in the gun,
discussed in the very next section). The second set of major radar properties are the
update rates at which the radar measurements on the position of the projectile and

134

target, and the rate at which the projectile controller state is updated. The radar
accuracy capabilities can be further decomposed to the subsystem level variables that
define the radar capability. Those subsystem level variable directly affect the cost of
the radar.
5.1.3.3

Gun Decomposition

As shown in Figure 41, the gun was not functionally decomposed below the system
capability level, and therefore there are no subsystem variables identified for the gun.
As can be deduced from literature search relating to the modeling and simulation
of combat systems presented in Chapter 2, there is not much open literature on the
design of medium caliber guns. Therefore, this study will not decompose the gun firing
system below the gun firing errors, including noises around firing elevation, azimuth,
and muzzle velocity. As discussed in the preceding radar decomposition section, there
are gun pointing errors inherent to the mechanical properties of the gun itself, and
the error due to the noises in the radar measurement of target location.
5.1.3.4

Interceptor Decomposition

The primary interceptor used in this thesis will be the GTRI guided projectile discussed in Section 1.2.4.31 . The physical geometrical properties of the projectile will
not be modified, as it is presumed from the studies discussed on this projectile that
its design is sufficient optimized to be used properly. Recall that the focus of this thesis is not component design, but the formulation and application of a methodology.
Additionally, the papers discussing the guidance and control mechanisms, i.e. the
actuator controlled pins, show that this guidance concept is validated in both computer simulation and real live gun firings. However, the bore of the of the interceptor
will be a considered a variable, and be scaled accordingly in the 6-DoF simulations.
1

The last portion of this thesis introducing the completed assembled environment will compare
the GTRI kinetic-kill/pin-controlled round to a round specifically designed to carry a small warhead,
as well as an unguided round, however this will not be addressed until Section 6.7.

135

Figure 42: 82mm Mortar Geometry Used in 6-DoF Simulations [101]


For this study, it is assumed that there are only nominal variations in inertial and
aerodynamic characteristics due to different bore sizes [101]. Therefore, any nondimensional inertial and aerodynamics information for the baseline interceptor will
be used for any bore size projectile (i.e. drag coefficient variation with Mach). As
discussed for the morphological matrix of alternatives, the type of controller used is
a primary attribute of the interceptor.
5.1.3.5

Target Decomposition

Examining multiple targets is considered outside the scope for this thesis. Therefore,
a single representative target of interest was selected. This arguably falls short of
proving an all inclusive defense against the stated RAM threat, but it should allow
for one to make educated conclusions of the value of the proposed methodology application. An 82 mm mortar was selected as this representative target, as presented
in Figure 42.
5.1.4

Review of Morphological Matrix Selections

The morphological matrix was introduced in the beginning of this section in an attempt to decompose the problem and potential solutions into the properties that
enable them. What followed was the justification to narrow the design space based

136

on problem definition, initial studies into notional concepts and solutions discussed
in the literature search, or the need to limit the problem within the scope of this
thesis. Table 10 revisits the morphological matrix presented in Table 9, but with
specific selections based on those justifications discussed all throughout this section.
The engagement range was bounded to the Short Range Inner Tier as described by
the EAPS program, and therefore will only focus on intercepting targets between 500
m and 2 km. The intercept solution studied will be a gun fired projectile.
The projectile itself will be command guided, meaning that after launch it can
receive course correction updates and follow a proportional navigation law, the most
commonly used navigation law. At this point, there will be no restriction as the use
of a PID or bang-bang controller. However, for the sake of the simulation, an initial
study will be conducted in which the projectile will use an onboard PID controller, so
that the controller can react according to the magnitude of required control surface
force application. A more detailed study will follow in which a bang-bang controller
will be used to more accurately represent the actual projectile. The projectile bore
will be studied at discrete values based on currently available guns, and not optimized for any measure of merit. Therefore, the four gun bore sizes discussed are
selected. The decision to employ a warhead will be left as part of the design study
requirements. Consequently, a kinetic kill projectile will be studied, with the quantification of intercept miss distance setting the requirement for a lethal radius required
of a warhead, if one were to be used. There will be no safety precautions regarding
collateral damage onboard the projectile itself.
Each 6-DoF trial in the initial study will examine a fixed gun firing muzzle velocity,
as opposed to having the ability to tune the firing velocity based on intercept range.
This however does not eliminate the ability to study error around that fixed muzzle
velocity. For the more detailed study, the muzzle velocity will be variable, however
only to discrete muzzle velocities. Additionally, errors on firing azimuth and elevation

137

will be considered. For reasons discussed in Section 5.1.2, errors on the remaining
factors are not key drivers and will not be examined. The magnitudes of those gun
firing errors will be considered as variable, rather than pick notional gun accuracies
(i.e. a gun with poor accuracy). There is no clear method of determining the
firing rate of a gun based on simple design properties, therefore regressions based on
a historical database will be used. This is needed in order to calculate the effect of
firing multiple rounds at a single target. A rate of fire regression based on historical
database, simply as a function of bore size, is given in Appendix C.
An interferometer radar using the monopulse tracking method will be used, and
as such any methods, equations, or simulations will include that decision when discussing attributes at the subsystem level. As with the gun, the radar accuracies will
be considered variable, bounded between notional poor and advanced radar accuracy capabilities. All projectile location errors (elevation, azimuth, range) will be
examined. For a gun fired projectile that exits the muzzle spinning about its primary
axis, the controller must have accurate knowledge as to the roll orientation in order
to accurately apply the control surfaces for steering. Any errors associated with the
radars ability to measure roll rate will have an effect on the ability to steer the projectile to a desired impact point, and will be studied. The projectile can be assumed
to be flying at negligible angle of attack and sideslip, therefore pitch and yaw will
not be studied in this thesis. Only radar noise will be considered as a form of radar
error, and bias will not be considered for reasons already discussed.
Finally, the only threat type studied will be an 82 mm mortar, considered a
sufficient representative of the RAM threat. The 6-DoF simulation will only consider
one simultaneous threat fired head on to the gun firing the guided interceptor.

138

General

Projectile

Lead Angle
PID
40mm
Blast/Fragmentation

Pursuit
Bang-Bang
30mm
Kinetic Kill

Guidance

Navigation Law
Controller
Bore

Warhead

Adjustable
Air Density
Fuction of Specific
Selected Gun
Conical Scan
Azimuth
Random Noise
60 mm Mortar
multiple
low slant (<45o)

Fixed
Muzzle Velocity
Calculated for
Optimum Performance
Sequential Lobing
Elevation
Bias
Artillery
single
head on (0o)

Firing Error

Gun Firing rate

Tracker Type
Radar Error
Radar Error Type

Threat Types
# simultaneous threats
Threat Firing angle

Self-Destruct Projectile

Interital

None

Intercept Solution

None

Rocket/Missile

Gun Fired Projectile

Engagement Range

Collateral Damge
Reduction Consideration

2
Long Range Inner Tier
2km - 10km

1
Short Range Inner Tier
500m - 2km

high slant (45o-90o)

82 mm Mortar

Monopulse
Range

Function of Bore
(historical database)

Air Temperature

Forced Tumble

Pitch
120 mm Rocket

107 mm Rocket

Projectile Weight

Custom

Passive Homing

Roll

Range/Cross Wind

57mm
Calculated Lethal
Radius Based Need

50mm
Submunition Dispersal

Proportional

Command

Three Point

Beam Rider

Directed Energy/Laser

Outer Tier
10km +

Table 10: Morphological Matrix of System Alternatives with Selections

Firing Muzzle Velocity

Gun

Radar

Threat

139

Yaw

Elevation

Semi-Active
Homing

Azimuth

Active
Homing

Effectiveness

Mobility

Cost

System-of Systems
Decision Making
Architecture

Figure 43: Measures of Merit for Gun System Optimization

5.2

Research Focus
Optimizer

Design variables

Optimizer

Design variables

The research in this study will focus on the application of a system-of-systems topEffectiveness
down Monte Carlo Effectiveness
based approach to the design of an air defense weapon
employing
Mobility

Measures Basic requirements


guided Measures
munitions with the goalMobility
of defeating the RAM threat.
of Merit

Cost

of Merit

Cost

assembled from the literature searches for this type of weapon are that it must be
lethal enough to destroy the incoming threat, mobile enough to be moved to new
locations as threats arise, all while providing an affordable solution2 . These are
Target

Effectiveness

Information
certainly vague, but show the conflicts
and complexity of the issue, and make for an
Tactics

+effectiveness,

excellent example to use the methods developed in this thesis. One of the main goals
is to optimize between three main system metrics as

Other
(collateral
shown in
Figure 43:
damage, etc)

mobility, and cost. Note that each of these three has some dependency on each of the
others, and one can not be improved without some degradation to another metric.
For example, to increase effectiveness, a larger radar may be used to improve the
accuracy in finding and tracking a target. However, this would result in a less mobile
weapon system and certainly would increase cost. On the other hand, a very mobile
weapon system may end up being too small to be effective against a target, and the
necessity to build very small electronics would increase cost.
The different disciplines may be arranged in a design-structured-matrix (or N 2
matrix), as shown in Figure 44, to show the feed-forward and feed-back coupling
of each discipline. In this case, the system is fully coupled. In a traditional design
2

As summarized by Holder [24].

140

Mobility

timizer

asures
Merit

Mobility

Cost

Design variables

System-of Systems
Decision Making
Architecture

Design variables

Optimizer

Effectiveness

Effectiveness
Mobility

Mobility

Measures
of Merit

Cost

Cost

Figure 44: Traditional Application of Optimizer on a Design-Structured-Matrix


Target
Information
optimization

Effectiveness

problem, an optimizer may be used to determine the values for the design
Tactics

Mobility

variables that maximize Other


given measures of merit (i.e. effectiveness
and mobility,
Cost
while minimizing

(collateral
cost).
The
optimizer
damage,
etc)

may even be used to break the feed-back

loops for faster convergence. However, most gradient-based optimizers can only be
used on a completely smooth, continuous, and unimodal design space. To overcome a
complex multimodal design space, the gradient-based optimizer must be coupled with
stochastic techniques, such as Genetic Algorithms or random searches, which consume
a great deal of computing power, even for simple multimodal problems. Additionally,
the traditional approach shown in Figure 44 is best applicable when optimizing on
the weapon system itself (or any single system). The addition of several interacting
systems and the subsystems that compose them, would create a highly combinatorial
and complex problem, in which only local optimal solutions would likely be found.
The generic process created as part of this research is given in Figure 45. The first
step is to collect which ever modeling and simulation tools are necessary to assemble the hierarchical environment. For those tools not capable of rapid manipulation,
surrogate models will be created. To model a highly nonlinear or discontinuous desired design space neural network surrogate models will have to be used, otherwise,
polynomial surrogate models may suffice. The next step is to assemble the models of
the various interacting systems and subsystems, such that the necessary outputs of
subsystems are fed as inputs to the appropriate system level tools. A Monte Carlo

141

Collect modeling & simulation tools for various interacting elements

yes

Are models capable of


rapid manipulation?
no
Is design space highly
nonlinear or multimodal?

no

Create polynomial
surrogate models

yes

Create neural network


surrogate models

Assemble models according to hierarchy


Populate design space with Monte Carlo simulation
sampling for all independent variables in the hierarchy
Examine complete hierarchical design
space on multivariate scatter plot
Select response and/or variable region of interest
Examine remaining design space for trends of interest

Figure 45: Flow Chart of Research Process


simulation will then be used to sample from bounded uniform distributions on the
inputs to the various models. The complete hierarchical design space can then be
viewed collectively on an interactive multivariate scatter plot. Here a user may examine any trends of interest. For example, the plot between random variables should
be filled uniformly, by virtue of the Monte Carlo sampling. Now, a user may interactively select or highlight design space regions of interest between any combination
of responses and variables, and examine the remaining design space for trends that
appear.
This study proposes to use this hierarchical system-of-systems decision-making

142

architecture to examine weapon design requirements based not only on the measures
of effectiveness described above, but also taking into account the entire interacting
environment. A Monte Carlo simulation will be used to quantify uncertainty in both
design tolerances and uncertainties in the scenario as well as be used to explore the
design space. A properly conducted design space exploration will eliminate the need
for running a gradient-based/stochastic combined optimizer that only produces a
point design. Constraints may be placed only after the architecture is fully linked
together, rather than during the optimization process.
For a system-of-systems environment comprised of a Monte Carlo based design
space exploration employing rapid surrogate models, both bottom-up and top-down
design analysis may be executed simultaneously. The general top-down/bottom-up
flow of information is shown in Figure 46. This process enables any response to
be treated as an independent variable, meaning that information can flow in either
direction in the hierarchy. From the bottom-up, the subsystem level properties of a
particular weapon system component define that components capability independent
of the other components (note: as was discussed earlier, only the radar properties are
taken down to the subsystem level). At the next level, each of the weapon system
components capabilities are defined, which collectively yield the weapon system performance. Next, the engagement scenario is defined, in which a weapon of given
performance is fired to intercept a target. Other attributes of the scenario independent of the target and weapon are defined in this level, and for this study the intercept
range is a major factor. Each of these factors in the scenario definition affect the operational level factors. These operational factors are what the decision maker primarily
cares about, including the ability to quantify the ability to hit (and/or destroy) the
target, and at what cost. This bottom-approach amounts to nothing more than a traditional approach of linking together several computer codes in a hierarchical fashion,
i.e. the outputs of one code feed directly as inputs at the next level in a hierarchy.

143

From the top-down, the operational requirements will limit the feasible scenarios
and drive weapon performance. For example, the limit of intercept range can be
identified for a given weapon system capability. Additionally, the weapon performance
that meets those operational level requirements can be identified. However, now
different combinations of intercept range and weapon capability against a given target
may meet the operational requirements. Moving on one level lower, the particular
weapon system capability drives the design space of each of the individual weapon
system components. However, for a given desired weapon capability, there may be
many perturbations of the gun, projectile, and radar capabilities that meet that need.
And finally, further decomposing the radar to its composing subsystems will identify
the feasible options that provide the required radar capability. Therefore at each level
of the hierarchy in the top-down process, a capability is identified, and the feasible
options at the adjacent lower level may be filtered out that enable that next higher
level capability.
Comparing the bottom-up and top-down approaches in Figure 47 shows that for
bottom-up design, selections are made at the lowest level, which then define the
capability at each higher level. This results in one design point flowing up through
the hierarchy. In the traditional design approaches, this is where an optimizer would
use logic to search the design space at each level for options that do not violate
any constraints and minimize (or maximize, depending on the problem) the desired
response. However, the top-down approach that uses a Monte Carlo based design
space exploration linked together with rapid surrogate models will likely define more
than one combination of design variables that meet favorable constraints at higher
levels. Therefore, an efficient way to identify those feasible combinations is needed.
When appropriately populated, the design space can be defined between each of
these variables and responses using multivariate scatter plots, i.e. one plot for each
variable/variable, variable/response, or response/response combination. The idea is

144

Probability of
Hit/Kill

Operational
Level

Operational requirements
limit feasible scenarios and
drive weapon performance

Top-Down
Flow down
requirements
and constraints

Cost
Scenario
Level

Scenario
Definition
(i.e. Range)

Requirements drive entire design


space of individual weapon
system components

Target
Scenario definition, target
information, and weapon
performance
affects operational
environment

Weapon
system
performance

Gun

Weapon System
Component
Level

Projectile

Bottom-Up
Flow up
weapon
system
capabilities

Weapon system design


yields performance

Radar

Subsystem level properties


define a system component

Weapon system component


capability requirement
drives its subsystem level
properties

Radar Subsystem
Level Properties

Figure 46: Top-down Requirements Flow and Bottom-Up Weapon Capabilities Flow

145

Figure 47: Bottom-up Single Design Point vs. Top-Down Multiple Design Point
Hierarchical Flow
to be able to literally carve out the feasible design space as was shown in Figure 38
of Section 3.4.2.
Ultimately, the goal is to optimize a system that strikes the most favorable balance between the metrics identified in Figure 43 by defining the weapon system design
requirements to reach those metrics. Lower level design variables drive system effectiveness, but top level requirements may require performance beyond what is available. Therefore the elements in Figure 43 cannot simply be treated as black boxes
with an optimizer wrapped around them. The elements must be treated as functions
that will be decomposed to whatever systems and/or subsystems that enable them,
as given previously in Figure 41. An ANOVA approach may be used to determine
which variables have the greatest effect on total system-of-systems variability, and
this method may be compared to using the multivariate scatter plot technique to
identify critical design space tradeoffs. A certain application of this would be the
most effective way to apply technologies to narrow the gap between what is currently
available and what is ideally required.

146

5.3
5.3.1

Modeling and Simulation Approach


Modeling and Simulation Requirements

The problem outlined in this chapter drives the need to assemble a specific modeling
and simulation environment. The desire to model a projectile intercepting a target in
midair drives the need to use a six degree-of-freedom (DoF) trajectory analysis tool.
This tool had the requirement to be able to simulate the trajectory of a guided projectile receiving updates from a ground based radar and the trajectory of an unguided
threat projectile. Because it was decided that the radar system will be defined down
to the subsystem level, there was a need to have the ability to evaluate how radar
properties affect radar accuracy. When dealing with capturing the effects of gun and
radar accuracy on interception capability, many executions of the 6-DoF are required
to adequately populate the design space. As such, the need to employ surrogate
models has been well established for this problem, particularly neural network based
representative models able to capture multimodal or discrete effects. Therefore a
tool is required to create those neural network equations rapidly and efficiently. This
section will discuss the various computer based tools required for this thesis study.
5.3.2

6-DoF Modeling

5.3.2.1

PRODAS

PRODAS (PROjectile Design and Analysis System) [29] is a six degree-of-freedom


(DoF) trajectory analysis tool, used to calculate the trajectories of gun fired projectiles. PRODAS is a software package that allows a user to create a three-dimensional
projectile model, calculate mass properties, estimate a full set of aerodynamics coefficients and stability properties, and simulate test firings, including launch dynamics
muzzle velocity calculations, and the ability to fly 4-DoF, 6-DoF, and body fixed
trajectories. The software allows for calculation of penetration of kinetic energy projectiles and shaped charged warheads, and allows for control system simulation of

147

guided projectiles [6]. Although PRODAS has the capability to estimate the ballistic
coefficients of the round, the calculated trajectories are more accurate when experimental aerodynamic coefficients are used, as reported by Silton and Massey [132],
and Silton [131].
There is very little literature available discussing the detailed capabilities of PRODAS, short of acquiring and using the tool itself. However, Whyte [142] [143] provides
an excellent introduction to PRODASs capabilities.
5.3.2.2

GTRI 6-DoF Multibody Simulation

GTRI has conducted several years worth of research on the effects of pin based actuators generating enough force to steer small supersonic projectiles [99] [131] [132].
In most of these studies, the projectile was modeled using PRODAS to carry out
6-DoF simulations to test the stability and simulate pin forces on the projectile. As
researched progressed, it became necessary to implement software that seamlessly
embedded the pin force model into the simulations. In addition, PRODAS currently
does not support two simultaneous trajectory simulations of two separate bodies, an
important aspect of the interceptor/target problem. For this reason, a fully functional
custom 6-DoF program was written at GTRI, which includes the pin force model, and
has the ability to track an interceptor projectile and target with a radar [101]. This
software has been validated using results from actual range tests.
The 6-DoF multi-body simulation code is outlined in the flow chart presented in
Figure 48. The first step involves an initialization step, including loading the physical characteristics of the target and interceptor, loading sensor characteristics, and
initializing the object states and sensor biases. The next step involves computations
regarding the sensors. First, the radar noises are calculated for azimuth, quadrant
elevation, and range, by sampling from a normal distributions defining those errors.
Because this simulation involves the use of a guided projectile, the orientation of the

148

projectile must be known in order to accurately direct the projectile control surfaces.
Therefore, the sensor noise associated with projectile roll is then computed. Next, the
target position with respect to the radar noise is calculated. A filter is then applied
to smooth the radar error with the number of samples taken. The same process for
position calculation with filtered noise is repeated for the interceptor, in addition to
the computing the interceptor roll attitude with respect to noise.
Next the guidance routine uses proportional navigation guidance (PNG) laws to
calculate inertial acceleration commands, which feeds directly to the controller routine. The interceptor roll command is calculated from inertial acceleration commands.
A low pass filter is applied to the roll command and roll attitude. The roll attitude
error signal is calculated, and the error rate and integral are calculated. The roll
control moment is calculated from a PID controller. This is used to calculate the
roll pin (projectile control surface associated with roll control) height from the control moment. The pin vertical acceleration command is then calculated, along with
a vertical acceleration error, and followed by a vertical acceleration error rate and
integral. Next the pitch pin (projectile control surface associated with pitch control)
height is calculated from an open loop acceleration command and PID. A rate limit
is then applied to the roll and pitch pin movements. Finally, the pins are deployed,
taking into account a wake up time from when the deployment command is given.
The pin heights are then converted to forces and moments physically applied to the
interceptor.
In the 6-DoF routine, the aerodynamic and gravity forces and moments are calculated for both the interceptor and target. A Runge-Kutta time integration is used
to determine the trajectory of both bodies. The simulation is performed for a predetermined amount of time, after which the minimum relative distance between the
interceptor and target is calculated. In addition, an actual interception hit is determined, with the two bodies assumed to be cylinders. The entire process after the

149

Initialize

6 DOF

Guidance
Calculate inertial
acceleration cmds
using PNG

Load Target physical


characteristics
Load Interceptor
physical characteristics
Load sensor
characteristics
Initialize:
States
Sensor biases

Sensors

Perform Runge-Kutta
time integration on
interceptor

Controller
Calculate Interceptor
roll command from
inertial accel cmds

Calculate vertical
acceleration error

Low pass filter roll


cmd and roll attitude

Calculate vertical
acceleration error
rate and integral

Calculate roll attitude


error signal

Calculate pitch pin


height from open loop
accel cmd and PID

Compute radar
noise (az, el, R)

Determine error rate


and integral

Compute roll
sensor noise

Calculate roll control


moment from PID

Compute Target
position w.r.t.
radar w/ noise

Calculate roll pin


height from control
moment

Compute Target
position w/ filtered
noise

Calculate vertical
acceleration cmd

Calculate Interceptor
aerodynamic & gravity
forces and moments

Rate limit roll and pitch


pin movement
Delay pin deployment
for wake up time
Convert pin height to
forces and moments

Calculate Target
aerodynamic & gravity
forces and moments
Perform Runge-Kutta
time integration on
Target

Detector
Calculate minimum
relative distance
Calculate hit
Calculate impact
velocity and angle

Compute Interceptor
position w.r.t. radar
w/ noise

Repeat for each


Monte Carlo
Simulation

Compute Interceptor
position w/ filtered
noise
Compute Interceptor
roll attitude w/ noise

Figure 48: GTRI 6-DoF Multi-Body Simulation (adapted from [102])


initialization step is repeated for the number of Monte Carlo simulations desired to
sample from the noises associated with the sensors routine.
5.3.3

Radar Subsystem Model

5.3.3.1

Radar Measurement Error

At the very bottom level in the hierarchy, the subsystem radar variables that result
in a given capability can be identified, using radar properties discussed in Section 2.7
and given in Equation 14. The projectile location noise can then be calculated using
Equation 18.

150

5.3.3.2

Radar Cost

An attempt was made to take some measure of system cost into consideration, however an accurate method for cost estimation was not found. It ultimately would be
beneficial to take total system (or system-of-systems) cost into account, i.e. breaking down the cost of the guided projectile, the gun, the radar, as well as life cycle
costs. This life cycle cost would depend heavily on actual campaign usage, however
a complete campaign model is beyond the scope of this work. Therefore, a good
candidate for cost tracking in this study was the radar, since as described earlier (and
will be proved in the next chapter), radar accuracy is considered a driving factor in
determining the accuracy of a guided projectile.
Based on talks with Holder [24], an accepted radar cost estimation is given in
Equation 25, and is based solely on the number of antenna array elements on the antenna. This very basic relation shows that radar cost is just four times the number of
elements, times the cost of the individual element. From [24], a reasonable estimation
for the cost of an individual element is $1000.
RadarCost = 4 Nelements (cost per element)

(25)

The total amount of elements in one radar antenna array can be simply calculated
as the square of the number of elements along one side of a square shaped antenna
array,
Number of Elements = (# Elements on 1 side of an antenna array)2
The number of elements on one side of the array is just the length of one side of
the antenna array divided by the space between the elements in the array, plus one,
therefore
!2

Number of Elements =

length of 1 side of antenna array


+1
space between elements

151

The length of one side of the antenna array is simply the square root of the calculated
antenna area, and the space between elements can be approximated one half the wave
length (as discussed in Section 2.7), yielding

Number of Elements =

!2

A
+1
0.5

Noting that it takes two antenna arrays for an interferometer to measure angle in
one plane (i.e. either elevation or azimuth), it takes at least three to measure both
elevation and azimuth.

Number of Elements =

!2

A
+1
0.5

(26)

where antenna area is found by manipulating Equation 13 from Section 2.7


A=

G r 2
4

Finally, to determine radar cost as a function of radar subsystem level properties,


the number of elements calculated in Equation 26 is substituted into Equation 25.
However, a portion of the implementation chapter (following this chapter) will
not study the radar properties down to the subsystem level, but there still needs
to be some cost measure that penalizes using a more accurate radar. During this
initial implementation study, radar cost will actually be a normalized measure of
radar complexity based on a simple relation given in Equation 27. The normalization
is based on the bounds used for the radar noises (shown as max and min).
This simple equation assumes that the most accurate (i.e. less noisy) radar has a
complexity of +1, and the least accurate (i.e. most noisy) radar has a complexity
of -1. The equation is a summation of the separate and independent contributions
to radar noise, i.e. elevation noise, azimuth noise, etc..., each weighted according to
total contribution to radar complexity. The purpose of the weightings is to normalize
the total radar cost between -1 (all radar noises at highest setting) and +1 (all radar
noises at lowest setting).
152

RadarCost =

"
#
# noise contributions
X
(N oisei,mid N oisei,actual ) 2

N oisei,max + N oisei,min

i=1

W eighti
(27)

5.3.4

Combined Effect of Multiple Projectiles

The methods defined to this point aim to quantify the ability of a single guided projectile to intercept a target as a function of various radar and gun accuracy capabilities.
This ability will be measured in the quantification of minimum miss distance between
interceptor and target, as well as the probability of a direct hit of the target by a
single projectile. Section 2.3.3 discussed the possibility that more than one weapon
may be required to achieve a desired probability of damaging or destroying a target.
Equation 7 was presented as a means to determine the probability that a target will
be hit at least once by n weapons, given the probability of hitting a target with one
weapon PH . This equation was derived using the binomial distribution and the assumption that each trial is independent of every other trial. The same method was
also applied in Equation 8 used to determine that damage caused by a multiple sorties, given the single sortie probability of damage (SSPD). Equation 28 uses the same
idea given in Equations 7 and 8, with nomenclature more suitable to this problem.
Here, the probability of hitting the target at least once using n multiple shots, or the
Multiple Shot Probability of Hit (MSPH), is a function of the Single Shot Probability
of Hit (SSPH).

M SP H = 1 (1 SSP H)n

(28)

The next step is to determine the number to substitute for n in Equation 28.
The maximum number of projectiles that can be fired from a single gun in a given
amount of time is a function of the firing rate. Assuming that the firing rate is given
using the relation discussed in Appendix C, the number of projectiles that can be
153

fired at the target is a function of an engagement window time that assumes that
all of the projectiles arrive at the target plane at the same time. This is a major
assumption necessary to implement this portion of the study in the absence of a 6DoF simulation capable of modeling a single gun firing multiple projectiles at a single
target. Therefore, the maximum number of projectiles fired at the target is simply
the product of the rate of fire [projectiles/second] and engagement window [seconds].
The probability of hitting the target will be determined using the SSPH (determined using methods discussed earlier) substituted in Equation 28, and any integer
up to the maximum number of projectiles fired. This should result in a trend such
as the one given in Figure 20, with the curve bounded by the SSPH for the first
projectile up the maximum attainable MSPH at the maximum possible number of
projectiles fired.
5.3.5

Creating Surrogate Models

Any surrogate models of more complex computer codes used in this thesis will be in
the form of a Response Surface Equation (RSE). As discussed in Section 3.3, the most
common form of RSE used to represent a complex design space is the polynomial RSE.
However, as the design space becomes high nonlinear, multimodal, or discontinuous,
simple polynomials can no longer be used for accurate model representation. Neural
networks are emerging as an excellent alternative to polynomial RSEs to model
those complex design spaces that often arise when capturing effects across a design
hierarchy, such as those that arise in a system of systems formulation.
5.3.5.1

JMP

The statistical package JMP [127] allows for the creation of both polynomial and
neural network response surface equations. If during the implementation portion of
this work, a design space will be simple enough to be modeled with a polynomial,
JMP will be used to create those surrogate models. However, if polynomials are

154

not suited to approximate a given response, neural network RSEs will have to be
used. As noted, JMP has the capability to create neural network RSEs, but it does
have its limitations. The only back propagation method used by the JMP software
is a momentum-based steepest descent optimization method. The steepest descent
is a gradient-based optimization method that uses a search direction within a design
space in the direction of the negative of the gradient of the objective function [139].
The use of momentum means that if the optimizer moves in the same direction more
than once, it increases its step size. This process helps overshoot local minima and
speed up convergence time. This training process is deterministic, however, the initial
weight space (or set of regression coefficients) is very highly dimensional, and therefore
predicting the outcome of training the randomized initial neural network is extremely
unlikely [64]. This back propagation method may work in many cases, but there may
arise instances when the design space is so complex that other optimization methods
might be required. Additionally, the user has to hand iterate the number of hidden
nodes used. As discussed in Section 3.3.2, there is an optimum number of hidden
nodes within the neural network topology that would fit the training and validation
data with the least error. Recall that too few hidden nodes would cause an underfit
model, and too many hidden nodes would cause an overfit mode.
5.3.5.2

FANNGS

The Function Approximating Neural Network Generation System (FANNGS) by


Johnson [82], is program written in MATLAB created to automate the generation
of neural network regressions. FANNGS allows the user to select a range of hidden
nodes to fit data to, and then selects the number of hidden nodes that has the least
error when compared to both the training and validation data. Because FANNGS
uses the MATLAB Neural Network toolbox [53] for many of its functions, it has a
number of training algorithms available to it. All of these algorithms are line search

155

methods that require many starting points to try to achieve a global optimum, just as
JMP. Of the many optimization methods MATLAB has to offer, Johnson [82] notes
that the Levenberg-Marquardt with Bayesian Regularization algorithm offers the best
generalization, or the ability to predict future data not used in the training process
such as the validation data. The MATLAB Neural Network toolbox documentation
[53] may be consulted for further detail on the different neural network training algorithms. For the most complex design spaces, FANNGS will be employed to create
neural network surrogate models.

156

CHAPTER VI

IMPLEMENTATION
6.1

Approach Review & Implementation Introduction

The previous chapter outlined a methodology to attempt to answer the research questions outlined in Chapter 4. A specific problem was defined based on the motivation
introduced in Chapter 1. To focus the design space, a morphological matrix was used
to present alternative approaches to address each decomposed function necessary to
provide a problem solution. Options were eliminated for each function based on several factors, including expert opinion, level of technology maturity, and relativity to
the scope of this Ph.D. thesis. An approach was outlined that creates an environment that enables top-down design throughout a system-of-systems hierarchy, using
a Monte Carlo simulation to populate the design space employing surrogate models
for rapid manipulation. Tools were then selected to address the functions necessary
to complete the system-of-systems hierarchy.
This chapter is composed of several main sections. The first part will very briefly
introduce preliminary results based on related work that served as a proof of concept to begin work on this thesis. The next part shows the implementation of the
methodology proposed for individual portions of the systems hierarchy, with the aim
of proving its use in each section. The first individual part will be an initial study
using the GTRI Multibody 6-DoF code to simulate a guided projectile and mortar
interception, using gun and radar noise variables of interest decided on a priori, which
will show the benefit of using the top-down decision making approach to identify variable tradeoffs based on various constraints on measures of effectiveness. This same

157

approach will also identify the variables that contribute the most to variability of the
constraints placed on those measures of effectiveness, as well as those variables that
contribute little if any. The radar noise variables identified as the main contributors
to the metric sensitivity will then be carried through to study the effects of varying
radar and guidance update rates. Afterwards, the combined effect of firing multiple projectiles at one target will be studied using the surrogate models created that
include the update rates and the radar noise variables of interest.
This is followed by a section that is independent of all of the other work presented
up to that point in the chapter, and will apply this method to study the subsystem
variables of the radar. These subsystem variables enable the radar accuracy capability
that is carried through as the variables at the next higher level in the hierarchy (i.e.
the level where the radar, gun, projectile, and target interact to cause a particular
top-level effect). The design space defined in this section will be carried through to
the final section, which will be the final assembled hierarchical environment. Before
this final environment is demonstrated, a more detailed 6-DoF model of the interception scenario based on knowledge acquired during the initial interception study will
be presented. Finally, the assembled hierarchical model will be demonstrated with
relevant discussion about any discoveries made during this process.

6.2

Preliminary Studies

Preliminary work leading to this thesis was conducted to establish the effectiveness
of using surrogate models in the uncertainty quantification of the ability of gun fired
projectiles to intercept small aerial targets at various distances. The specifics of these
preliminary studies are given in Appendix B, which overviews studies relating to two
of the systems discussed in Section 1.2.3. This is meant to be understood as results
based on simple test cases relating to systems currently being studied as potential
solutions to the RAM defense problem. The successful application of a number of

158

the more basic methods discussed in Chapter 3 to the problem of mortar interception
with an unguided, unpowered gun launched projectile served as the preliminary proof
of concept for the remainder of the effort undertaken to conduct the implementation
portion of this thesis study.

6.3
6.3.1

Initial 6-DoF Study


Variable and Response Identification

This initial study is designed to examine the effects of gun firing and radar accuracies
on the confidence of projectile/target interception. As will be addressed several times
in this chapter, many code executions are required to create surrogate models of
the 6-DoF code that uses an internal Monte Carlo simulation to simulate gun and
radar noises. Limitations in computing capabilities and scope restrict the number
of the degrees of freedom discussed for the gun and radar accuracies that can be
carried to the more detailed analysis level. However, other variables that need to be
considered in describing the intercept problem are range from projectile launch to
target intercept, and the projectile bore.
The initial analysis studies four radar measurement noises involving error in identifying the location of an object: elevation, azimuth, range, and projectile roll orientation (also referred to by its variable name from the field of dynamics as phi
for brevity as was shown in Figure 40 for the moments about a projectiles axes).
Table 11 shows the bounds for radar noises, shown as the standard deviation (1) of
a normal distribution around the reported location (elevation, azimuth, and range)
and roll orientation of the projectile as it travels along its trajectory. The bounds
were selected with consultation with disciplinary experts in the Sensors and Electromagnetic Applications Laboratory (SEAL) of the Georgia Tech Research Institute
(GTRI) [24]. The lower bounds of each of the radar noises is intended to represent
an advanced technology radar, which represent the lowest noise and therefore highest

159

Table 11: Radar Noise Errors for Initial Study(1)


Property
Elevation
Azimuth
Range
Projectile Roll Orientation

Minimum
0.015
0.015
0.00005
0.000045

Maximum
1.000
1.000
0.0001
0.00017

units
mrad
mrad
merror /mdownrange
rad/m

accuracy. The upper bound is intended to represent a low cost, low technology radar,
which has the highest amount of noise, and therefore also has the least accuracy.
The bounds selected for gun pointing biases for the initial study are given in Table
12. Note that the values given are not standard deviations of a distribution around a
desired value. They are static deviations from the intended firing position and muzzle
velocity. This initial 6-DoF study does not assume the use of a fire control computer
that calculates the gun firing position (i.e. in terms of elevation, azimuth) required
for the projectile to intercept the target, assuming an unguided ballistic trajectory.
For each simulation, the gun firing position must be manually set for each desired
intercept range. This gun firing position must be calculated depending on the desired
intercept range, which is not done automatically in the 6-DoF simulation. Therefore,
for each predetermined intercept range, a separate 6-DoF code input file must be
created using a precalculated gun firing position. This means that range can not
be treated as a continuous variable within the 6-DoF code, and must be treated as
a discrete variable. Surrogate models created around these static deviations of gun
firing position (rather than stochastic noises) leave the flexibility of sampling from
a distribution within the bounds used to create those surrogate models, as well as
actually pointing the gun away from an ideal firing position. This is the same
approach used in the preliminary studies explained with more detail in Appendix B.
The two other variables also noted as a necessary component for the initial study
were intercept range and projectile bore. The design of the gun firing the projectile is

160

Table 12: Gun Pointing Bias for Initial Study


Property Minimum Maximum units
Elevation
-5
5
mrad
Azimuth
-5
5
mrad
Muzzle Velocity
-10
10
m/s

Table 13: Discrete Variable Settings for Initial Study


Bore (mm) Range (m)
30
250
40
500
50
1000
57
1500
2000

not within the scope of this thesis, and the projectile bore itself is constrained to be
compatible with an existing gun. Table 13 lists the discrete settings for the projectile
bore and intercept range used for this portion of the study.
Because this portion of the implementation does not examine the radar subsystem
variables, radar cost will be based on the complexity relation given in Equation 27
in Section 5.3. Each of the four radar noise components is assumed to contribute
equally to radar complexity, and therefore is equally weighted in Equation 29; each
noise factor has a weighting of 0.25, summing to 1.0. These weightings could just as
easily have been any other value, and can be changed even after the decision making
environment is constructed.

161

"

RadarCost =
+
+
+
6.3.2

(AzimuthN oisemid AzimuthN oiseactual ) 2


0.25 (29)
AzimuthN oisemax + AzimuthN oisemin
"
#
(ElevationN oisemid ElevationN oiseactual ) 2
0.25
ElevationN oisemax + ElevationN oisemin
#
"
(RangeN oisemid RangeN oiseactual ) 2
0.25
RangeN oisemax + RangeN oisemin
"
#
(P hiN oisemid P hiN oiseactual ) 2
0.25
P hiN oisemax + P hiN oisemin

Surrogate Model Creation & Visualization

The previous sections described seven continuous and two discrete variables needed
to describe gun pointing errors, radar noises, intercept range, and projectile bore.
Therefore, a seven variable Design of Experiments (DoE) was created to dictate the
settings required to model the continuous variables, and was then run for each combination of the two discrete variables. A DoE consisting of 178 runs was created by
combining a Central Composite Design (CCD) consisting of 79 runs to capture the
corners of the design space, a Latin Hypercube Sample (LHS) consisting of 79 runs
to capture multimodal effects within the bounds, and 20 uniformly selected random
points to be used as a hold back sample to validate the model fit. Recall that each
trial of the DoE executes a Monte Carlo simulation within the 6-DoF code. The
more Monte Carlo samples, the greater the resolution of the statistics compiled from
those runs. However, the DoE was then required to run 15 times at each combination
of the two discrete variables, therefore limiting the number of Monte Carlo runs to
100 to maintain within a feasible amount of computer run time. This means that
running the DoE for one range/bore combination executed the 6-DoF 17,800 times,
and therefore a total of 267,000 6-DoF executions.
The radar error is represented by random noise, meaning that at each sweep, the
reported radar measurement is off by a different amount (as described in Section 2.7).
This noise is approximated by a normal (Gaussian) distribution. Therefore, each run
162

of the DoE input settings results in a different set of radar noises, each one requiring
a Monte Carlo simulation to sample from the normal distributions describing each
noise variable. Multiple runs are required to quantify the confidence of achieving
a response; here the responses of interest are the minimum miss distance of the
interceptor and target, and probability of direct hit. Consequently, each DoE run will
have as many outputs as Monte Carlo simulations per run. A frequency distribution
can be created for each response as a function of each input, and useful statistics
can be extracted from these distributions to help analyze the data. A useful way to
analyze these responses is by constructing a CDF using the 90%, 75%, 50%, 25%, and
10% confidences due to random radar noise, and calculated using the Monte Carlo
simulation results for each DoE run. This required creating a simple MATLAB script
to read in the data and run the statistics. At this point, neural network surrogate
models were created using the confidences as the responses, and may be visualized
using a prediction profiler as shown in Figure 49. Note that results are only shown
for the 40, 50, and 57 mm projectile bores. This is because the scaling laws used to
model the smaller 30 mm bore did not allow for sufficient control authority, resulting
in insufficient data to be used in the surrogate model regression.
Rather than just using a prediction profiler, a more useful way to visually analyze
these responses is by constructing a CDF using the neural network surrogate models
in a graphical user interface (GUI), shown in Figure 50. This CDF is formed by
using the neural network surrogate model fits to the 90%, 75%, 50%, 25%, and 10%
confidences of the minimum distance interception due to the randomness caused by
radar noise. Using the GUI, a user may use either the slider bars or type in the value
of each variable and instantly see the CDF update. When using surrogate models, it is
clear to see the benefit of instant results rather than having to wait for a complicated
6-DoF code to execute. It is important to understand that in this case, manipulating
the surrogate models is the equivalent of running a Monte Carlo simulation within

163

90% Conf
Miss Dist
50% Conf 75% Conf
Miss Dist Miss Dist
25% Conf
Miss Dist
10% Conf
Miss Dist

Bore

Range

Gun
Muzzle
Velocity

Gun
Elevation

Gun
Azimuth

Radar
Roll
Noise

Radar
Range
Noise

Radar
Azimuth
Noise

Radar
Elevation
Noise

Figure 49: Initial Study - Viewing Surrogate Models Using a Prediction Profiler
Discrete/Operational
Variables

Gun Pointing Error

Radar Noise Error

the 6-DoF code. In addition to the CDF, the neural network surrogate model created
to capture the actual probability of a direct hit is simply displayed on the GUI as
well.
One of the more valuable uses of having the minimum miss distance CDF is to
be able to determine the confidence of hitting the target with a lethal radius. This
can be directly applicable to setting the requirements for a warhead, if needed, by
typing a lethal radius value in the GUI that simply interpolates the confidence of
achieving that lethal radius. For the example case shown in Figure 50, for a radar
with the particular properties specified by the user, the probability of a direct hit is
about 0.22, whereas the same projectile employing a warhead with a lethal radius of
1.0 m increases the confidence of hitting the target to 0.40. Creating a tool like the
GUI shown in Figure 50 is a very useful way to visualize how radar noise affects the
interceptor/target miss distance confidence. However, it still leaves open the question
of how constraints on top-level metrics affect the design space on the lower levels of
the design hierarchy. The section after next will show how the surrogate models used

164

CDF formed by
fitting Neural Nets
to 90%, 75%, 50%,
25%, and 10%
confidence of
Minimum Distance
interception due to
radar noises

Phit regression results


based on inputs

Can read off the CDF


plot the confidence of
being within a radius
around the target.
Answers the question
how many
projectiles will hit
within X m of the
target?

Figure 50: Manipulating the Surrogate Models Using a Bottom-Up GUI

165

to create the bottom-up GUI can be used to fill the design space throughout the
hierarchy, and then constrained from the top-down. First, the introduction of gun
firing noise to the gun pointing error will be addressed.
6.3.3

Introduction of Gun Firing Noise

Because the use of a fire control computer was not available to point the gun in
the proper direction for a given intercept range, the correct gun firing elevation and
azimuth had to be decided on a priori to running the DoE. This eliminated the
possibility of using gun noise as a variable, as was used for the radar, therefore
within the 6-DoF code, gun error was introduced as a controlled bias from the ideal
firing condition (gun elevation, azimuth, and muzzle velocity). For this initial 6-DoF
study, the introduction of gun firing noise, or randomness on the gun pointing and
muzzle velocity was conducted by creating a new set of surrogate models by running
a Monte Carlo simulation on the gun biases of the surrogate models already in place,
and fitting the new set to the same confidence intervals as those discussed in Section
6.3.2.
Normal distributions extend to positive and negative infinity around the mean, so
placing upper and lower bounds on the distributions comes in the form of selecting
an appropriate standard deviation. The goal is to ensure that the random number
generator in the Monte Carlo simulation selects values within the desired bounds
most of the time. The number of standard deviations away from the mean value
can describe the amount of area under the distribution curve captured. For example,
as discussed in Appendix A there is a 99.9994% probability of selecting a random
number within 4 of the mean of a random variable. Using a random variable X
with normalized bounds -1X1 and mean at 0 as an example shown in Figure 51,
a standard deviation of 0.25 maintains that 99.9994% of any samples taken from a
Monte Carlo simulation to be within the normalized bounds.

166

Frequency

Figure 51: Normal Distribution, =0, =0.25


Table 14: Gun Firing Noise
Property Min Max units
Elevation
0
1.25 mrad
Azimuth
0
1.25 mrad
Muzzle Velocity
0
2.50 m/s

Using the presumption that 4 is a suitable upper bound, by conjecture the absolute value of the limits discussed in Table 12 presented earlier can simply be divided
by 4. A standard deviation centered about a mean of zero results in equal spread
among positive and negative values, therefore one standard deviation value can account for the positive and negative bounds presented in Table 12. As given in Table
14, a standard deviation of zero implies no gun firing noise for a particular variable,
and the maximum value is the value selected to stay within the absolute bounds of
the gun bias surrogate models.
The problem is now set up to make minimum intercept miss distance quantification
a function of both radar and gun noises, as well as intercept range and projectile bore.
An eight variable DoE was created using the three variable bounds given in Table
14 in addition to the four radar noise variables of Table 11. For the first DoE run

167

(discussed in Section 6.3.2), range was run as a discrete variable because the 6-DoF
code could only accept it as such. However, enough discrete values of intercept range
were executed to capture the design space of range and all other variables, and the
analysis conducted in Section 6.3.2 resulted in an equation that allows range to now
be treated as a continuous variable for the remainder of this initial 6-DoF study.
To simplify the problem within the scope of this study, bore was left as a discrete
variable and not included in the DoE, however the DoE was executed once for each
value of bore: once each using a 40 mm, 50 mm, and 57 mm bore (recall that the 30
mm bore results were invalid and therefore insufficient for creating surrogate models).
The results may be visualized using a prediction profiler as shown in Figure 52. Note
the relatively flat curves that describe the individual contributions of the gun noises,
as compared to those of the radar noises. Those gun errors now have a std suffix
in many of the graphics in this section, and are bounded from zero to the maximum
value in Table 14. This indicates that the responses are not very sensitive to gun error
(this will be expanded on in the coming sections). The surrogate models created in
this section that quantify intercept miss distance as a function of both radar and gun
noises will be used for the remainder of this initial 6-DoF study.
6.3.4

Response Variability Study

One of the main intentions of this thesis is to have the ability to discover not only
which variables are the key drivers to the variability of the responses, but also which
variables drive each other. This section will compare the use of Pareto plots and the
proposed multivariate scatter plot method to conduct a response variability study.
6.3.4.1

ANOVA with Pareto Plots

Figure 53 shows a Pareto Plot created using an Analysis of Variance (ANOVA) as


discussed in Section 3.3.3. This figure lists the variables contributing to the variability
of 90% confidence minimum miss distance, in order of most contribution. Range has

168

90% Conf
Miss Dist
75% Conf
Miss Dist
50% Conf
Miss Dist
25% Conf
Miss Dist
10% Conf
Miss Dist

Bore

Range

Gun
Muzzle
Velocity
Noise

Gun
Elevation
Noise

Gun
Azimuth
Noise

Radar
Roll
Noise

Radar
Elevation
Noise

Radar
Azimuth
Noise

Radar
Range
Noise

Figure 52: Initial Study - Viewing Surrogate Models Using a Prediction Profiler
the highest contribution with about 42%, and next is bore size at about 18%. Next,
the four radar noises are listed, however there is clear step in the contribution of
elevation and azimuth noises (both about 17%), and of range noise (4%) and roll (phi)
orientation (3%). The contributions to miss distance variability from gun firing error
are essentially too small to measure. As indicated by the cumulative contribution to
response variability, roughly 90% of the total variability is due to range, bore, radar
elevation noise, and radar azimuth noise.
Clearly the discrete variables evaluated dominate the gun and radar noises. One
of the problems in evaluating a Pareto plot is that it alone does not indicate which
variables drive each other. For example, the question can be asked at which intercept
ranges do the radar noises have more of an impact? One way to answer this is to
conduct the ANOVA for each combination of the discrete variables set to a fixed
value, which would allow one to examine the prioritization of the gun and radar
properties as range increases. A complete set of Pareto plots for each bore/range
combination is given in Figures 121, 122, and 123 in Appendix D.1. Studying those
169

Cumulative
variability
curve

20%

40%

60%

80%

Figure 53: Initial Study - Pareto Plot for Discrete and Continuous Variables on
Variability of 90% Confidence Intercept Miss Distance
figures one can extract important information from examining the intercept range
extremes, summarized in Figure 54. The first observation made from studying those
Pareto plots is that the radar azimuth and elevation noises are the key drivers in each
case.
For the 50 mm and 57 mm bore sizes, the gun errors play a much more important
role in miss distance at the shorter ranges than at the longer ranges. This is because
as range increases, the controller has more time to correct the pointing errors. Interestingly enough, for each of the three bore sizes at 2000 m, the order of the variables is
the same. The only difference is that as bore size increases, the amount of variability
contributed by radar range and roll noises, and the gun noises reduces. In summary,
as range increases, miss distance is less affected by the gun errors due to the controller
having more time to correct the trajectory, and as bore size increases, the relative
contribution of the radar noises is less due to the increased dynamic pressure on the
larger control surfaces giving more control authority.
A Pareto plot showing the results of an ANOVA study is a useful way to prioritize
the variables of interest, but it is a very one-dimensional way to make assessments.
Additionally, showing how the variables drive each other proves to be a cumbersome
process. As this section showed, to do just that, the discrete variables had to be fixed
at each of the various settings, and only then could the ANOVA be conducted to see

170

Bore

Range: 250m

Range: 2000m

40mm

50mm

57mm

Figure 54: Initial Study - Pareto Plots for Fixed Combinations of Range and Bore
on Variability of 90% Confidence Intercept Miss Distance
the way the priority of the other variables changes.
6.3.4.2

Multivariate Design Space Exploration

Another useful way to analyze the data is to conduct a design space exploration using
a multivariate scatter plot. A Monte Carlo simulation sampling from a uniform distribution selected 1,000 samples from each variable within the bounds used to create
the surrogate models. This includes the three discrete bores, the intercept range,
the three gun noises, and the four radar noises. For each combination of variable
values selected, the responses of interest are tracked. Therefore, each corresponding
Monte Carlo simulation has a unique array of variable and response values, and can
be visualized on a multivariate scatter plot, such as the one used for this initial study
as given in Figure 55. Note the multivariate plot is actually just a collection of bivariate plots. Because the independent variable values in the blue box are samples
from a uniform distribution, the points evenly fill the design space between them.

171

Responses

Variables

Figure 55: Initial Study - Multivariate Scatter Plot Design Space Exploration
The dependent responses are outlined with a red box, and are calculated using the
variable values selected by the Monte Carlo simulation. Note that the individual bivariate plots within the red box show the design space between the responses. Each
of the remaining scatter plots show the design space of each of the response/variable
combinations, and is synonymous with the total derivative given in Equation 24 of
Section 3.4.2.
As an example for using this method, a region of interest can be defined on the
multivariate matrix as shown in Figure 56. Here, a region of acceptable performance
is defined: the projectile must achieve a 90% confidence minimum miss distance of 7
m, and a 10% confidence minimum miss distance of 0.9 m. The other CDF settings
(i.e. 75%, 50%, 25%) could have also been included to refine the shape, however
172

Figure 56: Initial Study - Intercept Miss Distance CDF Shape Constraint
they are omitted for clarity on the graphic. The points that lie within this region
of interest are highlighted red, and the corresponding points are then immediately
highlighted in every other dimension. This means that every red point in each of the
other dimensions corresponds to a design that meets (or exceeds) the requirement set
by the initial design space constraint. One can spend a good deal of time examining
the tradeoffs in each dimension, but an attempt will be made to highlight some
interesting results in the form of bivariate plots taken directly off of the multivariate
plot in Figure 56.
For example, as shown in the left plot in Figure 57, note that the miss distance
requirement drives a higher concentration of red design points at the lower radar

173

Figure 57: Initial Study - Bivariate Plots of Radar Noises, For Intercept Miss Distance CDF Shape Constraint
elevation and azimuth noises. This does not mean that certain specific values of
elevation and azimuth noises must be used, since it is not immediately apparent
which red points in this dimension correspond to a red point in another dimension.
Although, what is apparent is that these two radar noises have a great effect on
minimum miss distance. The remaining radar noises, range and roll (phi) noises
(the right plot in Figure 57), and the three gun noises shown in Figure 58 are not
concentrated in any specific region, and therefore are not as driven by the top-level
constraint. This is the same conclusion drawn using the ANOVA and Pareto plots,
however that method only identifies key variables as a function of total response
variability, where as this new methodology enables a user to actually specify the
CDF shape to discover the variable trade space, shown in the next few examples.
Figure 59 shows the maximum intercept range possible to achieve the minimum
miss distance confidence constraint, for both 90% and 10% confidences. Here, the
highest intercept range achievable is roughly 1700 m, evident from the red design
point with the highest range on both plots. Another key finding from the miss
distance constraints placed on the multivariate plot in Figure 56 is the variability of
required radar accuracy with range. This is highlighted in Figure 60, where range is
plotted against all four of the radar noises. Note that the black unconstrained points

174

Figure 58: Initial Study - Bivariate Plots of Gun Noises, For Intercept Miss Distance
CDF Shape Constraint

Figure 59: Initial Study - Bivariate Plots of Showing Maximum Range, For Intercept
Miss Distance CDF Shape Constraint
from the previous figures are hidden for clarity, leaving only the red points that meet
the miss distance constraint. Clearly, at the very low ranges, all radar accuracies are
tolerated (i.e. meet the constrained goal), and as range is increased, the upper bound
allowable radar azimuth and elevation noises are driven lower (top two graphics in
Figure 60). A similar statement can not be made for the effect of increasing range on
radar range and roll noises (bottom two graphics in Figure 60), where it is clear that
for increased range there are fewer points that meet the range requirement, but it is
not clear that increasing range requires higher radar range and roll accuracies.
The design space constrained by the miss distance requirements can further be
constrained by other requirements, such as cost. Shown in Figure 61 are the points
175

Figure 60: Initial Study - Bivariate Plots of Range and Radar Noises, For Intercept
Miss Distance CDF Shape Constraint

176

within the miss distance constraints divided up into two regions in the radar cost
space. The points colored red can be considered the more complex, or expensive
options, and the blue points can be considered the less expensive or complex options.
It is important to remember that both the red and blue points still satisfy the miss
distance requirements identified in Figure 56. Several significant trends appear comparing feasible radar solutions by cost. Figure 62 shows radar cost (left) and 90%
confidence miss distance (right) both versus range. Again, the solutions that do not
meet the miss distance constraint are hidden for clarity. On the left graphic, a notional frontier has been drawn in to exemplify the nondominated solutions providing
maximum range capability for minimum cost. This is a very useful way to determine
at what range one would have to switch from using a cheaper radar to a more
expensive one, depending on how complexity was valued from Figure 61 (in this case
at about 850 m intercept range). The right graphic in Figure 62 shows how the more
expensive radar (red points) enables a lower intercept miss distance at higher ranges
when compared to the less expensive option (blue points).
Figure 63 shows bivariate plots of the radar noises just as shown in Figure 57,
except with the distinction made between the two radar cost regions. The less expensive (blue) points dominate the upper right corner of both plots, and the more
expensive (red) points dominate the bottom left corner. This is clearly intuitive since
the closer a solution is to the bottom left corner of each plot in Figure 63, the more
accuracy it has. However, an interesting observation can be made in the plot showing
elevation and azimuth noises (left), where blue points do not fill the space completely
in the upper right corner (as was pointed out in Figure 57). This shows that the
miss distance constraint sets the lowest tolerable radar accuracy, even using the less
complex radar. However one of the most useful plots extracted from the original
multivariate (Figure 61) is studying radar azimuth and elevation noises versus range,
given in Figure 64. Starting at the lower ranges, it is clear that both sets of radar

177

Feasible space divided within Radar Cost


Red = High Cost
Blue = Low Cost

Original miss distance constraint

Figure 61: Initial Study - Feasible Designs Separated by Radar Cost

Red = High Cost


Blue = Low Cost

Figure 62: Initial Study - Radar Cost by Range (L) & Minimum Miss Distance by
Range (R), Divided by Radar Cost

178

Figure 63: Initial Study - Radar Noises, Divided by Radar Cost

Red = High Cost


Blue = Low Cost

Figure 64: Initial Study - Range by Radar Azimuth & Elevation Noises, Divided
by Radar Cost
complexities provide solutions to the minimum miss distance requirements. However,
at about 800 m intercept range there is a clear shift from blue to red points, and at
about 1000 m there are only red points, indicating the need for more radar accuracy
at high ranges.
Other types of constraints can be flowed down to the radar capability level. As
shown in Figure 65, a certain maximum allowable 90% confidence miss distance can
be defined for a desired intercept range. Immediately, a clear tradeoff Pareto Frontier
emerges between the radar elevation and azimuth noises, shown in the left plot of
Figure 66. However, there is no apparent sensitivity due to radar roll (phi) and radar
range noises, shown in the right plot of Figure 66. And just as discussed for the earlier

179

Figure 65: Initial Study - Selection of Required 90% Confidence Miss Distance and
Intercept Range
examples in this section, there is no particular shift of feasible points in any direction
with regard to the gun errors, shown for gun firing muzzle velocity by gun firing
azimuth on the left, and gun firing quadrant elevation by azimuth on the right side
of Figure 67. Now when compared to the Pareto plots, this method shows not only
the key driveers, but also how the variable trade space varies for multiple response
constraints.
Another useful application of this methodology is to use the interactive multivariate scatter plot from the bottom-up. Please note that the minimum miss distance
constraints discussed earlier no longer apply to the following examples. In Figure

180

Figure 66: Initial Study - Radar Noises, For Required 90% Confidence Miss Distance
and Intercept Range

Figure 67: Initial Study - Gun Noises, For Required 90% Confidence Miss Distance
and Intercept Range

181

68, regions of good and bad radar accuracies are chosen by selecting radar azimuth and elevation noises, in this case notionally. Here, the blue points represent
solutions with very low radar elevation and azimuth noises, and red points indicate
designs with high noise. Now, the impact can be traced to the upper levels of the
hierarchy. In Figure 69, the 90% and 10% confidence minimum miss distances are
plotted against intercept range. Especially examining the 90% confidence plot on
the left, it is clear that as range is increased, the better radar (blue points) enables
solutions staying closer to the x-axis than solutions using the less accurate radars
(red points). This is clearly an intuitive result, however this method allows a user to
rapidly quantify a statement made that a more accurate radar drives the intercept
miss distance down, and with a measurable certainty. Note that both the red and
blue points diverge as range is increased (i.e. there is a clear trend, but the points do
not form a straight line). This is due to the fact that each variable value is selected
from Monte Carlo simulation used to populate the multivariate scatter plot, and each
point in the diverging trends will have various values selected for the other variables.
However, this can be further examined by studying another trend observed from Figure 68. Figure 70 shows a plot of 90% confidence miss distance and gun bore. Notice
that at each discrete bore setting, a majority of the less accurate radars have higher
miss distances. The interesting trend here is that the intersection of the majority of
blue and red points shifts lower with increased bore size, indicating that a guided
projectile with a higher bore can tolerate more radar noise for a given intercept miss
distance.
The previous example showed that selected radar azimuth and elevation noise
accuracies drive the weapon capability, and define a clear tradeoff space with projectile
bore size and intercept range. Now, a similar example is shown to determine whether
the other two radar noises have a similar impact. Figure 71 shows regions of higher
(blue) and lower (red) radar roll rate (phi) and range noises. In a similar fashion

182

Red = High Cost


Blue = Low Cost

= High
RadarNoises
El/Az Noises
Red = HighRed
Radar
El/Az
= Low Radar El/Az Noises
Blue = LowBlue
Radar
El/Az Noises

Figure 68: Initial Study - Multivariate Scatterplot DSE, Sorted by Radar Azimuth
and Elevation Noises

Red = High Radar El/Az Noises


Blue = Low Radar El/Az Noises

Figure 69: Initial Study - Bivariate Plots of Minimum Miss Distances and Range,
Sorted by Radar Azimuth and Elevation Noises

183

Red = High Radar El/Az Noises


Blue = Low Radar El/Az Noises

Figure 70: Initial Study - Bivariate Plot of 90% Confidence Miss Distance and Bore,
Sorted by Radar Azimuth and Elevation Noises
Red

= High Radar El/Az Noises

Low Radarbivariate
El/Az Noises
to the previous example, FigureBlue
72 =shows
samples of the resulting design

space, showing how minimum miss distance varies with range for both 90% and 10%
confidences. There is no evident benefit to selecting more accurate radar roll and
range noises with increased range, especially when directly compared to Figure 69.
In a final graphical example exploiting the surrogate models used to create the
multivariate scatter plot based design exploration, a study was conducted to isolate
the effects of the various bore sizes. Figure 73 shows the design space sorted by bore
size, with the 40 mm solutions colored red, the 50 mm ones colored blue, and 57 mm
bore size solutions colored black. Even without the color difference, it is evident that
increased bore size has a clear advantage of decreasing the miss distance. Figure 74
shows both the 90% and 10% confidence miss distances by bore size. A decreasing
worst case scenario miss distance decreasing with increased bore size is shown, the
effects of which are more prevalent in the 10% confidence miss distance. Figure 75
shows both miss distance confidences by range. Note that the 40 mm bore (red)
points rapidly increase in miss distance as range increases for both the 90% and 10%
miss distance confidences. However, the advantage of 57 mm over 50 mm bore is
not as evident in the 90% miss distance confidence, since with increased range the
dispersion of blue and black points is about the same. Looking at the 10% confidence
miss distance by range plot for contrast, there is a clear advantage of the 57 mm bore

184

Red = High Radar Roll/Range Noises


Blue = Low Radar Roll/Range Noises

Figure 71: Initial Study - Multivariate Scatterplot DSE, Sorted by Radar Roll and
Range Noises

Red = High Radar Roll/Range Noises


Blue = Low Radar Roll/Range Noises

Figure 72: Initial Study - Bivariate Plots of Minimum Miss Distances and Range,
Sorted by Radar Roll and Range Noises

185

over the 50 mm bore, where the dispersion of 57 mm (black) points is not as great
with increased range when compared to the 50 mm (blue) points. A more pronounced
benefit of increased bore at the lower miss distance confidences may give an indication
of the the shape of the CDF, whereby its slope is more gradual with increased bore.
This is more evident by studying Figure 76, where the three bores can be compared
on a bivariate plot of 90% and 10% miss distance confidences. For the same 90%
confidence, a lower 10% confidence is achievable with a larger bore. This has a very
clear practical impact. If designing a solution for a single shot kill, 57 mm vs 50 mm
bore size does not have as much of an impact, because there is not much difference
in their 90% miss distance confidences. However, for a defense system allowing for
multiple shots, there is an advantage using the 57 mm over the 50 mm bore projectile,
since the value added of each additional bullet has more impact. This methodology
clearly enabled the discovery of these tradeoffs.
6.3.4.3

Response Variability Study Concluding Remarks

This purpose of this section was to compare the use of the ANOVA method visualized
with Pareto plots and a Monte Carlo simulation design space exploration visualized
with a multivariate scatter plot as methods for discovering the variable trade space
that drives response variability. The ANOVA method that uses Pareto plots shows
a one-dimensional portrayal of the contributions of each variable to the variability
of a response. Variables that drive each other can be identified only when certain
variables are fixed at their lowest and highest values, with an ANOVA run on the
remaining variables. This not only proved to be a very cumbersome process, but
also did not yield the tradeoff space between those variables. Running a Monte Carlo
simulation, and visualizing the design space on a multivariate scatter plot showed how
constraining requirements can identify where design tradeoff spaces lie, and which
variables drive the values of other variables. The same variables that were identified

186

Red = 40mm
Blue = 50mm
Black = 57mm

Figure 73: Initial Study - Multivariate Scatterplot DSE, Sorted by Bore

Red = 40mm
Blue = 50mm
Black = 57mm

Figure 74: Initial Study - Bivariate Plots of Minimum Miss Distances and Bore,
Sorted by Bore

187

Red = 40mm
Blue = 50mm
Black = 57mm

Figure 75: Initial Study - Bivariate Plots of Minimum Miss Distances and Range,
Sorted by Bore

Red = 40mm
Blue = 50mm
Black = 57mm

Figure 76: Initial Study - Bivariate Plot of 90% Confidence and 10% Confidence
Minimum Miss Distance, Sorted by Bore

188

as being key drivers on a Pareto plot were also identified on the multivariate scatter
plot, however clear tradeoffs were identified using the latter method.
Additionally, a clear benefit of this methodology is the ability to manipulate random functions as independent variables. These random functions result in probabilistic outcomes that can be quantified and bounded in ways previously unavailable. For
example, the manipulation of the response CDF can be used to bound the random
function independent variables.

6.4

Addition of Radar & Guidance Update Rates

With the methodology showing its usefulness in the previous section, it will now be
applied in an attempt to add a degree of freedom to the problem. The next part
of the implementation focused on the tradeoff between radar noise and the rate at
which the radar and guidance updates occur, With the primary radar noise variables
contributing to the variability identified. The radar update rate is the rate at which
projectile and target position updates are sent to the fire controller, and the guidance
update rate is the rate at which guidance updates are sent to the projectile to allow for
course correction. This tradeoff is an important step in determining whether the radar
and guidance updates should be included in the parametric capability of the decision
hierarchy. Eliminated variables that do not significantly contribute to variability in
intercept end game were fixed to default values, shown in Table 15. The gun errors
were shown to contribute little compared to the radar noises, and even within the
radar noises, two of the four examined proved to provide the majority of the response
variability. Therefore these values are fixed, however they are still treated as noises
in the 6-DoF code, and therefore still provide an element of variability. In addition
for this study, only the 50 mm bore is used.
These restrictions on the design space were necessary to allow for design study
to be completed with a limited amount of computing resources, and to focus the

189

Table 15: Fixed Variables for Radar & Control Update Rate Study
Property Fixed Value
Muzzle Velocity Error (1)
8.25
Firing Azimuth Error (1)
0.001
Firing QE Error (1)
0.001
Radar Roll Noise (1)
87.3
Radar Range Noise(1)
0.001
Bore
50

units
m/s
rad
rad
rad
rad
mm

resources available on fine tuning the parametric models created. For example, with
fewer variables needed in the surrogate models, fewer DoE runs would be required
when compared with the initial study. With fewer DoE runs than the previous study,
more Monte Carlo runs per DoE run could be afforded computationally. This allows
for a more precise examination of the statistically based responses. Five hundred
Monte Carlo simulations were used for the radar update rate study. For the probability of hit metric which counts the number of hits for each Monte Carlo simulation
executed for each DoE run, the resolution is much needed.
This section will be broken down into two sections. The first will simply presume
that the projectile controller is slaved to the radar update, in other words the guidance
update is set to the radar update. The second part of this section will treat the radar
and guidance update rates as independent. In both cases, the radar azimuth and
quadrant elevation noises are treated as variable.
6.4.1

Control State Slaved to Radar Update

As stated earlier, this portion of the study focuses on the effects of slaving the projectile controller to the radar update, as opposed to using guidance updates for control.
The variable bounds used to examine the effects of radar update rate are given in
Table 16. A DoE was created for the 3 continuous variables, including a 15 run CCD,
a complimentary 15 run LHS, and 15 random points for a total of 45 runs per each

190

Table 16: Variables for Radar Update Rate Study with Control State Slaved to
Radar Update
Property Minimum Maximum units
Radar Elevation Noise (1)
15
150
rad
Radar Azimuth Noise (1)
15
150
rad
Radar Update Rate
10
100
Hz
Intercept Range
250
2000
m

discrete range setting given in Table 13.


Figure 77 shows a series of surface plots illustrating the study conducted to analyze
the effects of radar update rate on required radar accuracy. The three graphics
comprising Figure 77 show the required radar update rate varies with intercept range
to achieve a desired probability of direct hit, using three notional radar accuracies:
advanced, or low radar noise (top), medium radar noise (center), and high radar
noise (bottom). A notional PH of 0.5 is shown as an example to constrain the feasible
design space.
For the low radar noise case (top graphic), PH is not very sensitive to radar update
rate. As range increases, there exists a region where only radars with a high radar
update rate satisfy the PH constraint. Also, note how at the very upper bounds of
range, even the highest radar update rate does not meet the PH constraint. With an
increased radar noise shown in the center graphic, note how the feasible space above
the PH constraint has diminished. At even higher values of radar noise (bottom
graphic), the only feasible design space is with very low intercept ranges. Even a
cursory examination of the graphics in Figure 77 should provide an appreciation of
using neural network based surrogate models. Capturing this complex design space,
with all of its multimodal peaks and discontinuities would have been unimaginable
using polynomial fits.

191

Feasible Radar Update


Rate and Slant Range
values above the Phit
constraint

Define Phit
constraint plane

Low radar noise

At very low
ranges, Phit
not very
sensitive to
Radar Upate
rate

Feasible Radar Update


and Slant Range values
above the Phit constraint
note less feasible space

Define Phit
constraint plane

Increased radar noise

Only feasible Radar


Update and Slant Range
values above the Phit
constraint at very low
range

Define Phit
constraint plane

High radar noise

Figure 77: Radar Update Study with Control State Slaved to Radar Update Surface Plots
192

Table 17: Variables for Radar Update Study with Independent Radar & Guidance
Updates
Property Minimum Maximum units
Radar Elevation Noise (1)
0.025
0.500
mrad
Radar Azimuth Noise (1)
0.025
0.500
mrad
Radar Update Rate
10
50
Hz
Control Update Rate
90
200
Hz
Intercept Range
500
2500
m

6.4.2

Independent Control and Radar Update Rates

With the importance of radar update established, this section isolates between the
effects of the rate of radar position updates and that of the guidance update rate.
The higher the frequency of guidance updates, the more corrections the projectile can
make to attempt to stay along its intended trajectory. As in the previous section, the
radar azimuth and elevation noises and radar update rate are included as variables,
but now guidance update rate is included. The bounds selected for the parametric
models created for this portion of the study are given in Table 17. An important note
to mention at this point is that the originally desired lower bound for control update
was 10 Hz, however the 6-DoF simulation tool was not properly set up to handle that
lower bound with all other conditions. The value of 90 Hz was selected simply as a
lower bound for the capability of the simulation.
6.4.2.1

Surrogate Modeling and Design Space Exploration

To create the neural network surrogate models, a DoE was created for the 4 continuous variables, including a 25 run CCD, a complimentary 25 run LHS, and 25 random
points for a total of 75 runs per each discrete range setting. A prediction profiler is
given in Figure 78, showing an example setting for how radar capability drives probability of hit. The individual contributions of each variable can be identified. As noted
earlier, there will be a minimum distance at which the controller on the projectile

193

Phit
0.268915

0.4
0.2

1011
Range

0.000105
AzimuthNoise

9.33e-5
ElevNoise

130.3
GuidanceUpdate

50

40

30

20

10

90
110
130
150
170
190

.0005

.0004

.0003

.0002

.0001

.0005
0

.0004

.0003

.0002

.0001

2000

1000

25.44
RadarUpdate

Figure 78: Radar Update Study with Independent Radar & Guidance Updates Prediction Profiler
can engage and have a steering effect beyond the unguided ballistic path. Clearly
reducing the amount of radar noise has a positive effect of increasing the probability
of a direct hit, but the impact of azimuth noise clearly has more impact than the
elevation noise. The prediction profiler also shows the importance of independently
studying the guidance and radar updates. Given the settings of the other variables in
this one example, it is just barely evident of the impact of increased on radar update
rate on PH , however increasing the guidance update rate would greatly increase PH .
This graphic gives another appreciation for capturing the multimodal nature of the
design space. For instance, holding all other variables constant, there is a benefit to
PH to increasing the guidance update rate up until about 170 Hz, where increasing
past that does not provide any additional benefit.
The graphics in Figure 79 show the benefit of using the surface plots to view the
effects of two variables at the same time on a response. The surface plot on the left
shows azimuth noise and range on PH . This plot is a good example of showing the
minimum steerable intercept range. At poorer (i.e. higher) azimuth noises, there
is less variability in PH with range. However at the much more accurate azimuth
noises, effective intercept range bounds can be identified. Also in the left surface plot
of Figure 79 a notional PH constraint is placed roughly at 0.15, and the 3-D surface
above that indicates the feasible tradeoff between radar azimuth noise and intercept
range. An interesting note on this is that at the intercept range providing the greatest

194

Independent control & radar updates

Phit
0.268915

Figure 79: Radar


Update Study with Independent Radar & Guidance Updates 0.4
Surface Profilers
0.2
50

40

30

20

10

90
110
130
150
170
190

.0005

.0004

.0003

.0002

.0001

.0005
0

.0004

.0003

.0002

.0001

2000

1000

PH (about 1500 m),1011a clear0.000105


tradeoff curve
showing
exactly
how much radar azimuth
9.33e-5
130.3
25.44
Range

AzimuthNoise

ElevNoise

GuidanceUpdate

RadarUpdate

noise can be tolerated given the response constraint. This is of course assuming that
the other variables are held constant - manipulating those variables would certainly
shift the surface. The surface plot on the right side of Figure 79 shows the effects of
both radar and guidance update rates on PH . For the settings of the other variables
held constant, there is a benefit to having high radar and guidance update rates. That
is certainly the obvious conclusion, however this shows the importance of including
these two degrees of freedom in this study.
6.4.2.2

Response Variability Contribution of Variables

A similar response variability study conducted earlier to define the CDF shape is now
conducted for this new problem, however here only the contribution of variability to
PH is included as the measure of merit. Figure 80 shows a Pareto plot listing the
variables for this portion of the study in order of importance with associated measure
of contribution. The two radar noises top the list (radar azimuth noise provides about
22% variability, followed by radar elevation noise at about 10%). Next is the guidance

195

A Pareto Plot is the traditional way to identify the contribution of each


variable to the variability of a response
Only a 1-D list of each variable, but does not provide tradeoffs
Term
AzimuthNoise
ElevNoise
GuidanceUpdate
Range
RadarUpdate

Phit Variability

FromFigure
the Phit
the previous
thereUpdates
is a 80: constraint
Radar Updateidentified
Study with on
Independent
Radar chart,
& Guidance
Paretobetween
plot
tradeoff
Radar azimuth noise and guidance update rate
0.0005

update rate providing about 8%, intercept range at about


6%, and radar update rate
0.5

AzimuthNoise

0.0004

at 2% contributions to PH variability. This is a useful0.4way to prioritize the variables


Phit

0.0003

0.3

of interest, but it is a very one-dimensional way to make assessments, and begs an

0.0002

0.2

important question. For example, in the Pareto plot, the number one driver of PH
0.1

0.0001

variability is radar azimuth noise, but is there another 0key variable that drives radar
0

90 100110120130140150160170180190200
azimuth
noise? The answer can be found by utilizing the multivariate
scatter plot
MIN-DIST90
GuidanceUpdate

approach to design space exploration.


SystemSystem-ofof-Systems Approach to the Design of an AirAir-Defense Weapon

19

Tommer R.
Endermatrix populated using a Monte Carlo
Figure 81 shows a multivariate scatter
plot

simulation on the neural network surrogate models created for this portion of the
study. To illustrate, all of the points in the design space with a PH above about
0.22 are highlighted red. The points are highlighted in the first row, but the design
points corresponding to that value show up red in each of the other bivariate plots.
Figure 82 points out two important findings using this multivariate method. These
two bivariate plots are taken directly off of the multivariate plot, with the design
points not meeting the PH requirement hidden for clarity. On the left side, elevation
noise is shown on a bivariate plot along with guidance update rate, and the right
graphic shows azimuth noise compared with guidance update rate. Assuming that
less accurate radar accuracy (i.e. higher elevation and azimuth noises) and a lower
guidance update rate is desirable for cost and complexity purposes, a black dotted
line shows the non-dominated design points. Now it is possible to assess the relative
importance of variables to the variability of a response. Azimuth noise was shown

196

Figure 81: Radar Update Study with Independent Radar & Guidance Updates Multivariate Scatter Plot with PH Region of Interest in Red
in the Pareto plot in Figure 80 to be the biggest driver for PH variability, but using
the multivariate method, guidance update proves to be a very important driver in
the radar accuracy. The radar accuracies do not stand alone, and their impacts on a
guided projectiles performance capability clearly depend on the rates of both radar
and guidance updates.

6.5

Effect of Firing Multiple Projectiles

This section is dedicated to taking the results presented for quantifying the ability
to hit a target with a single round, and determining the probability of hitting the
target with at least one projectile if multiple are fired using the methodology discussed

197

0.0005

0.0004

0.0004
AzimuthNoise

ElevNoise

0.0005

0.0003
0.0002
0.0001

0.0003
0.0002
0.0001

0
90 100 110 120 130 140 150 160 170 180 190 200

90 100 110 120 130 140 150 160 170 180 190 200

GuidanceUpdate

GuidanceUpdate

Figure 82: Radar Update Study with Independent Radar & Guidance Updates Bivariate Plots of Radar Accuracies and Guidance Update
in Section 5.3.4. This section applies the methodology in this thesis to answer the
specific question, what does the single shot probability of hit (SSPH) have to be if
many bullets are fired ? Taking this down a level in the design hierarchy, the question
can then be asked, what is the tradeoff in the single projectile design space when using
more than one shot fired to achieve a multiple shot probability of hit (MSPH)?
Equation 28 from Section 5.3.4 is used in this section to quantify the probability
of directly hitting the target with at least one projectile if multiple are fired, assuming
that the probability of hitting the target is the same for each projectile (a very common assumption made when dealing with bursts of projectiles [92]). The maximum
number of projectiles that can be fired is a function of the firing rate, and the length
of time of fire. Equation 33 (from Appendix C) is used to calculate firing rate as
a function of gun bore. In order to use the assumption that each of the projectiles
intercept the target with the same probability of hit, a suitable engagement window
must also be assumed. This section arbitrarily assumes an engagement window of 2
seconds, meaning that any projectiles fired within that window will cross the target
plane at (essentially) the same time. This engagement window could have any other
value.
For the sake of example, this section will use the surrogate models generated in
Section 6.4.2 to show how the methodology used in this thesis can be applied to

198

Phit
(Multiple Shot
at Max Firing Rate)

Phit
(Multiple Shot
at Various #
Projectiles Fired)

Rounds Fired

Phit
(Single Shot)

Range

Guidance Update

Radar Update

AzimuthNoise

ElevationNoise

Figure 83: Multiple Projectiles Study - Multivariate Scatter Plot Design Space
Exploration
showing tradeoffs between number of projectiles fired and the required accuracy of
each projectile. This includes using the variables and bounds given in Table 17.
However, now not only is the probability of hit of a single projectile tracked, but
the probability that a single shot will hit if many are fired. Using the number of
rounds fired at the maximum firing rate as the maximum number of rounds possible,
part of the design space exploration includes every integer from 1 to that maximum
number fired. As was previously done, a Monte Carlo simulation was used to conduct
a design space exploration, using a multivariate scatter plot as shown in Figure 83.
Note that for the assumptions used to create these results (i.e. the variable bounds,
the engagement window, etc...), the maximum number of shots fired is 5. Therefore,
this design space exploration uses the same radar variable bounds using between (and
including) 1 and 5 projectiles, and then tracking the responses.

199

Phit
(Multiple Shot
at Various # Projectiles Fired)

Phit
(Multiple Shot
at Max Firing Rate)

Phit
(Multiple Shot
at Various #
Projectiles Fired)

Rounds Fired

4
3
2
1

Phit
(Single Shot)

# projectiles

Phit
(Single Shot)

Range

Guidance Update

Radar Update

AzimuthNoise

ElevationNoise

Figure 84: Multiple Projectiles Study - Selection of Required MSPH


Figure 84 shows how an a desired multiple shot probability of hit can be defined,
this example showing an MSPH of 0.70. Note that the solution using only 1 projectile
does not meet the MSPH requirement cutoff; the maximum attainable SSPH is about
0.42 using only one shot. Another interesting observation made is that as the number
of shots is increased, the benefit in MSPH increases less. This figure is reminiscent of
Figure 20 (Section 2.3.3), in which the probability of hitting or destroying the target
with one weapon (or sortie) is a function of number of weapons (or sorties), for varying
SSPH. Figure 85 then isolates the feasible designs that meet the MSPH requirement,
hiding all other infeasible points, and color grouping the remaining design points by
number of shots fired. Interesting findings appear in the multivariate plot, shown as
bivariate plots for clearer examination in the remaining figures in this section.
The tradeoff between the SSPH and the radar properties is given in Figure 86.
For both radar accuracy plots (the top two), with fewer number of shots fired the

200

Phit
(Multiple Shot
at Various # Projectiles Fired)

Phit
(Multiple Shot
at Max Firing Rate)

Phit
(Multiple Shot
at Various #
Projectiles Fired)

Rounds Fired

# projectiles

Phit
(Single Shot)

Phit
(Single Shot)

Range

Guidance Update

Radar Update

AzimuthNoise

ElevationNoise

Figure 85: Multiple Projectiles Study - Feasible Design Space Isolated and Grouped
by Shots Fired

201

radar noises are driven lower. This means that the more shots are fired, more noise
error can be tolerated in the radar accuracy. The 2 projectile shot solution occupies
only the very top left corners of the plots, with more shots fired spreading to higher
tolerated radar noise with decreased SSPH. Note that as was pointed out in earlier
sections, the design space is more sensitive to variations in radar azimuth noise than
of elevation noise, which is indicated in Figure 86 by the fact that higher elevation
noises are tolerated than azimuth noises. This is of course more evident with an
increased number of projectiles fired. The bottom two plots in Figure 86 show SSPH
as a function of both guidance and radar update rates. As the guidance update rate
is reduced, note that more projectile shots are required. This same conclusion can
not be made of the radar update rate. The reader may have already noticed that
one of the 2 shot (red) points is actually shown as a small square. This is shown to
illustrate that when using this method, one can easily select one of the design points
in one of the dimensions, and immediately see where it is in all other dimensions. The
point selected here is the one with the highest required SSPH in the azimuth noise
plot (top left). Note that this same design point has the lowest azimuth and elevation
noises, and also requires a very high guidance update rate. A high radar update is
not required for this design point, however this is not surprising as the radar update
rate was found to contribute less to performance variability in previous sections (see
Figure 80).
As in previous sections, the radar properties are shown in Figure 87. The left
plot shows the two radar accuracies together. With few projectiles, the design space
constrains itself to the origin (i.e. fewer projectile shots requires less radar noise and
more accurate projectile). With increased number of shots, the feasible design space
opens up to allow less accurate radars. The right plot shows the guidance and radar
update rates. Here, it is only evident that fewer projectiles constrain the guidance
update rate, while not affecting the radar update rate as significantly. The two plots

202

Phit
(Single Shot)

Phit
(Single Shot)

Red = 2 projectiles
Green = 3 Projectiles
Blue = 4 Projectiles
Black = 5 Projectiles

ElevationNoise

Phit
(Single Shot)

Phit
(Single Shot)

AzimuthNoise

Radar Update

Guidance Update

Figure 86: Multiple Projectiles Study - SSPH vs. Radar Properties, Grouped by
Shots Fired

203

Guidance Update

AzimuthNoise

Red = 2 projectiles
Green = 3 Projectiles
Blue = 4 Projectiles
Black = 5 Projectiles

ElevationNoise

Radar Update

Rounds Fired

Rounds Fired

Figure 87: Multiple Projectiles Study - Radar Accuracies (L) & Update Rates (R),
Grouped by Shots Fired
in Figure 87 show that there is not much distinction between the design spaces of
Black = 5 Projectiles

= 4 Projectiles
using 4 versus 5 projectile shots. This is where the decision maker canBlue
decide
whether
Green = 3 Projectiles
Red = 2 projectiles

the increased cost of firing another shot (i.e. making use of the maximum possible
rate of fire) is worth a very minimal increase in MSPH (as first shown
in Figure 85).
AzimuthNoise
ElevationNoise

6.6

Subsystem Level Radar Properties

The goal of this section is to capture the design space of the radar system properties
that drive the radar noise errors, as well as radar size and cost. The ultimate goal is to
incorporate these elements into the completed system-of-systems hierarchy, and have
the ability to study how radar subsystem level properties affect radar performance,
which directly affect the defense systems overall performance capability. The radar
properties considered for the design space exploration are the ones discussed in Section
2.7 for signal-to-noise given in Equation 14, and the measurement error standard
deviation for an interferometer radar given in Equation 18. Table 18 lists the variables
used for the radar subsystem level study. The bounds selected were based on a study
conducted by Holder [76] to select the radar properties for a command guided bullet.
Using those variable bounds, the equations used describe radar capability and can be
manipulated using the prediction profiler shown in Figure 88, similar to the type used

204

Table 18: Variables Used for Radar Subsystem Study


Property Minimum Maximum units
Power (average)
100
250
W
Gain (transmit)
12
50
dB
Gain (receive)
15
40
dB
Frequency
14
18
GHz
Range
500
2000
m
RCS
-25
-15
dBsm
Dwell Time
0.001
0.005
sec
Distance between Antenna
0.5
10.0
m
Element Arrays

Figure 88: Prediction Profiler for Radar Subsystem Study


to manipulate surrogate models. In this case however, direct equations are used in
lieu of a modeling and simulation tool, and surrogate models are not necessary. Note
that many of the radar properties are used to calculate signal-to-noise ratio, which
in turn is used to calculate the measurement standard deviation error (referred to as
sigma theta in many of the graphics in this section). Linking the equations together
in a prediction profiler environment allows for direct calculation of measurement error
using the radar properties, with the in between step of calculating signal-to-noise ratio
handled internally.

205

The design space for the radar subsystem is explored using a Monte Carlo simulation, and visualized using the multivariate scatter plot technique shown in Figure
89. Note that the dependent responses in this design space are included in the red
box, and are the radar signal-to-noise ratio, the angular measurement error standard
deviation (sigma theta on the multivariate plot), the antenna aperture area, and
the radar cost. The dependent variable values outlined in the red box are calculated
using the independent variables in the blue box, which are samples from a uniform
distribution. Each of the remaining scatter plots show the design space of each of the
response/variable combinations. Even at this level in the hierarchy, the same topdown design methodology can be applied, as is shown in Figure 90 where a maximum
acceptable angular measurement error capability is defined at 0.05 mrad, and then
a very tight cost constraint is applied at $4M (blue points) to differentiate from the
more expensive design that meet the accuracy constraint (red points). The remaining
design points that do not meet these standards are hidden for clarity.
Figure 91 shows some of the interesting findings from this example, comparing the
design points of two capability definitions (low angle measurement at low and high
acceptable cost). The top left plot shows the variation of radar cost and aperture
area. There is a clear trend of limiting aperture area when limiting cost. The top
right plot shows the variation of the distance between antennas and the signal-tonoise ratio. For feasible design points that meet the radar noise constraints, a lower
signal-to-noise ratio is tolerable only with an increase in the distance between antenna
element arrays (i.e. a larger radar), with a dashed line included to accentuate the
trend. However, comparing the two cost constraints, we see clearly that having a lower
cost constraint limits how close the antennas can be, in this case there are no blue
(low cost) points below about 6 m antenna array separation. The bottom right plot
shows the desired capability in terms of antenna transmit and receive gains. For all
solutions shown there is a clear tradeoff frontier between the gains, however limiting

206

cost places an upper limit on the receive gain, leaving the remaining blue trend as the
feasible space (recall from Section 5.3.3 that receive gain directly impacts the number
of radar antenna element arrays, therefore also the cost of the radar). The last plot
on Figure 91 on the bottom left shows the distance between antennas tradeoff with
antenna receive gain. The desire to limit cost clearly limits the receive gain, however
this drives the distance between antennas higher. By comparing the bottom left and
top left plots in Figure 91, it is clear that the difference between the high and low cost
constraints in the distance between antennas is very minimal (about 1 m), but would
come a very large cost (the low cost constraint was set at $4M, but the remaining
design points range up $50M).
This example shows the important tradeoff between radar capability and radar
size; i.e. the larger the distance between radar antenna array elements, the larger the
radar, hence the more accurate the radar. This example is another good argument for
this approach: when the design space is filled using a Monte Carlo simulation, using
the multivariate scatter plot to visualize the design space allows any combination of
variables and responses to be studied simultaneously - and every term, both variable
and response can be treated as independent variables.
Regions of low and high radar frequencies are selected in the design space in Figure
92, with all intermediate frequencies hidden for clarity. Interesting tradeoff spaces are
shown in bivariate selections in Figure 93. The top left plot shows that for a given
radar cost, a higher frequency tends toward a small antenna aperture area. The top
right plot shows that for a given measurement accuracy, a higher frequency allows
for a smaller distance between antenna element arrays. On the bottom left plot, for
a given receive gain, a higher frequency reduces the aperture size. The bottom right
plot serves almost as a histogram, showing the maximum aperture area for the two
frequencies.
Another example of exploring the effects of selecting radar properties is shown

207

Figure 89: Radar Subsystem Study - Multivariate Scatter Plot Design Space Exploration

208

Red = High Cost


Blue = Low Cost

Both low

Figure 90: Radar Subsystem Study - Radar Error Capability Defined

209

Red = High Cost


Blue = Low Cost

Both low

Figure 91: Radar Subsystem Study - Design Space for Radar Error Capability
Requirement

210

Blue = High Frequency


Red = Low Frequency

Figure 92: Radar Subsystem Study - Regions of Low and High Frequencies are
Selected

211

Blue = High Frequency


Red = Low Frequency

Figure 93: Radar Subsystem Study - Regions of Low and High Frequencies are
Selected

212

in Figure 94 in which regions of low and high antenna gains (both transmit and
receive) are selected. Important findings that emerged from the multivariate scatter
plot are shown as bivariate plots in Figure 95. The top left plot shows the tradeoff
between radar noise and the distance between antennas. At the lower bound of the
distance between antennas, a higher gain allows for less measurement noise, however
this impact is less evident as antenna separation increases. Therefore, higher gains
allow for a smaller radar with same capability - however cost is not shown here, but
examining the entire multivariate plot in Figure 94, there is a clear order of magnitude
increase in cost for the higher gains selected. The top right plot of Figure 95 shows the
variation of radar noise with range. Looking at all of the solutions, as range increases
there is more spread of radar measurement error, i.e. increased potential for more
radar noise. It is evident that with increased antenna gain, there is less variation of
radar noise with range. This means that higher gains enable a more robust solution
- a radar with accuracy that operates independent of the range of the projectile it
is tracking. The bottom left graphic shows the tradeoff space between radar noise
and signal-to-noise ratio. Here it is evident that radar noise is more sensitive to
signal-to-noise at low antenna gains, and clearly higher gains provide less error and a
higher signal-to-noise ratio. The group of low gain design points was identified at the
extreme corner of high noise and low signal-to-noise, which identifies a clear trend
in the plot in the bottom right corner, showing the variation of RCS with frequency.
Without the high noise/low signal-to-noise designs identified, there does not seem to
be any trends. However with those points identified, a clear trend emerges showing
how with decreased target RCS, the design space opens to lower radar frequencies
(this again is for the low gain, high noise, low signal-to-noise designs).

213

Red = High Gains


Blue = Low Gains

Figure 94: Radar Subsystem Study - Regions of Low and High Antenna Gains are
Selected

214

Red = High Gains


Blue = Low Gains

Figure 95: Radar Subsystem Study - Capabilities and Properties for Low and High
Antenna Gains

215

6.7
6.7.1

Detailed 6-DoF Model


Variable Listing

This section will attempt to refine the parametric model using the findings from the
initial study, in which the key variables were identified. The parameters for this study
were defined with the help of appropriate committee members [24] [25]. The previous
studies did not assume the use of a fire control solution. For a given intercept range
and gun muzzle velocity, the gun firing elevation necessary to cause an intercept had
to be determined iteratively (assuming head on shot with no error anywhere in the
system). Therefore for the initial study, intercept range was treated as a discrete
variable along with bore, and the DoE created for the remaining variables had to be
run for every combination of range and bore. This detailed study will employ the use
of a fire control solution, allowing the treatment of intercept range as a continuous
variable. Therefore, this portion of the study will examine intercept ranges between
500-1500 m. All of the variables used for the remainder of the implementation portion
of this thesis will be discussed in the following sections and are summarized in Tables
19 and 20.
6.7.1.1

Projectile

Regarding the projectile, this study will examine three different kill techniques by
comparing an unguided projectile, a guided hit-to-kill projectile, and a guided projectile carrying a small warhead. The guided hit-to-kill projectile is the same one
discussed in the initial study, using a pin firing mechanism to steer into the target.
The main difference in this portion of the study is that the projectile will now employ
a bang-bang controller that is more representative of how an actual guided projectile
would be implemented, as opposed to the PID control employed in the earlier study;
a PID would provide better control, but is far more complex to actually implement
than a bang-bang controller, especially in a small projectile with clear volumetric

216

constraints. The unguided projectile will effectively be the same as the guided hitto-kill projectile, except it will not receive any radar or guidance updates once it
leaves the gun, and the pin firing mechanisms will therefore not be used. This is
done in order to compare the use of a guided round to an unguided round. Ideally,
a spin-stabilized round would be used (as opposed to the guided fin-stabilized hitto-kill round). However, for the sake of simplifying the simulation this assumption
was made. For these first two projectile types, a direct hit is required to destroy the
target, and the probability of hit is the measure of merit for both.
The third kill mechanism involves using a guided projectile with a small warhead,
and will require using a different aerodynamic shape conducive to carrying a warhead,
shown graphically in Figure 96. For the sake of example, an assumption is made that
the warhead has a 0.25 m lethal radius against the target type studied (any other
value could have been used, using the CDF of the minimum miss distance as was
done for the initial study). This effectively means that the measure of merit for
this kill mechanism of the probability of hitting within 0.25 m of the target. For
control, this projectile uses a small thruster that can fire only once to divert the
projectile, graphically shown in Figure 97. Once an acceleration threshold is reached,
the controller checks to see where the thrust action line is relative to the desired angle
of acceleration. If the thruster line of action is outside of the desired acceleration arc,
then the thruster is armed. The thruster is fired once the thruster line of action is
inside the desired acceleration arc. It is assumed that the thruster is located at the
projectile center of gravity, and that it fires for a very short duration, resulting in
thrust force uniform in time [25].
One of the purposes of conducting this more detailed study was to reintroduce
different bore sizes, which directly affect the rate of fire of the gun (discussed in
the next section). For the two guided rounds (kinetic-kill/pin-control and warheadkill/thruster-control), only the two larger bore sizes examined in the initial study will

217

Figure 96: Warhead Round for Detailed 6-DoF Study [25]

Thruster Line
of Action
Desired
Acceleration

Thruster Line
of Action

Arm
Thruster

Arc of Desired
Acceleration
Fire
Thruster

Figure 97: Thruster Controller for Warhead Round [25]

218

be used, 50 mm and 57 mm. For the unguided round, it was desired to include the
use of a smaller bore, 30 mm, that enables a high rate of fire, in addition to the two
bore sizes used for the guided rounds. The 40 mm bore examined in the initial study
will not be examined here in an effort to provide a more realistic depiction of a gun
firing small unguided projectiles at a high rate of fire. Because there is no control
logic implemented on the 30 mm round, there were no issues with control authority
such as that encountered in the initial study.
6.7.1.2

Gun

The initial study showed very little sensitivity of intercept miss distance to gun firing
accuracy. When this occurs, there are two basic options when progressing to a more
detailed study. The first option would be to fix the gun accuracies, and not allow
any parametric variation. The other option is to open the bounds to see if values
beyond those studied in the initial study would in fact contribute to the variability
of performance. However, this was only studied for a guided round. For an unguided
round, the gun firing error may play a more important role, and will therefore be
studied for all three kill mechanisms. The original 6-DoF tool used in the initial
study did not run a Monte Carlo simulation internally on the gun noises. Rather,
for the initial study the gun error was treated as a controlled bias from the ideal
firing position. A Monte Carlo simulation was conducted on surrogate models of the
deviations from the ideal firing position to make intercept miss distance a function of
gun noise, as described in Section 6.3.3. For this study, the 6-DoF tool will actually
run the Monte Carlo simulation internally, and just as was done for the radar noises,
the parametric surrogate models will be a direct function of gun noise. A more
simplified approach will be taken to parameterize gun firing accuracy. Here, the gun
firing azimuth and quadrant elevation will be coupled, meaning that the same value
will be used for both. Muzzle velocity error will not be varied, and will be fixed at

219

a specific value of 0.4% (1) of the muzzle velocity. Because the 6-DoF tool only
simulates a one-on-one intercept, the gun rate of fire will be taken into account after
the fact in the assembled hierarchical model.
Also different from the initial study, this detailed study will examine three discrete
gun firing muzzle velocities: 1000, 1325, and 1650 m/s. This is done to determine
if one of the kill mechanisms has an advantage over another at a different muzzle
velocity, and to determine how much of an effect the gun itself has on the entire
interception process. Because most large caliber guns are designed to fire at specific
muzzle velocities, there was no need to treat it as a continuous variable in this study.
6.7.1.3

Radar

The radar accuracy will be represented by coupling the azimuth and elevation noises
(i.e. both use the same value). From the previous studies, there was little contribution
to miss distance variability from the radar roll and range noises. For this study, roll
noise will be fixed, however, range noise will vary. As shown in the previous study,
the radar update rate contributes heavily to the performance of a guided projectile,
and will remain a variable for this final study. The guidance update rate, however,
will remain fixed at 1kHz. Much attention has been given to the radar properties
in the previous sections, therefore there is no need to elaborate further about radar
properties here.
6.7.1.4

Summary of Variables for Detail Study

Each of the variables discussed in this section along with their bounding values is
given in Table 19 for the continuous variables, and Table 20 for the discrete variables.
6.7.2

Surrogate Modeling

A five variable DoE was created for the continuous variables listed in Table 19. An
81 run DoE was assembled using a 27 run CCD to capture the corners of the design

220

Table 19: Detailed Study - List of Variables


Property Minimum Maximum units
Gun Pointing Error (Az, QE)
0.5
1.5
mrad
Radar Noise (Az, QE)
0.05
1.0
mrad
Radar Range Noise
0.05
1.0
m/km
Radar Update Rate
10
100
Hz
Range
500
1500
m

Table 20: Discrete Variable Settings for Detailed Study


Kill Mechanism
Kinetic-Kill/Pin-Control
Warhead/Thruster-Control
Kinetic-Kill/Unguided

Bore (mm) Muzzle Velocity (m/s)


30
1000
50
1325
57
1600

space, a 27 run LHS to capture the space throughout the design space, and 27 random
runs for model validation. This 81 run DoE had to be repeated for each combination
of kill mechanism, bore, and muzzle velocity desired, shown in Table 20. For each
of the three kill mechanisms, three gun muzzle velocities were studied, and each of
the guided rounds were studied at two different bore sizes, and the unguided at three
bore sizes. Additionally, as with the initial studies conducted, each DoE run executed
a Monte Carlo simulation within the 6-DoF code. There was a desire to increase the
number of internal 6-DoF Monte Carlo simulations in order to increase the fits of
the surrogate models. For each of the two guided rounds, 750 Monte Carlo samples
were run for each DoE case - for an 81 run DoE at three discrete bores and two
discrete muzzle velocities - resulting in a total of 364,500 6-DoF runs for each. For
the unguided round, 250 Monte Carlo runs were used, resulting in 182,250 6-DoF runs.
Therefore, a total of 911,250 6-DoF executions were used in the data presented in this
section and used in the remainder of this thesis. The 10 seconds per 6-DoF execution
required about 2500 computational run hours distributed over several computers.

221

Figure 98: Detailed Study - Prediction Profiler for Three Projectile Kill Mechanisms
A prediction profiler used to manipulate the surrogate models is shown for each
of the three kill mechanisms in Figure 98. Note again that the measure of merit
for the kinetic kill weapons is probability of hit, where as for the warhead round,
the probability of hitting within 0.25 is used. The analysis used to determine the
goodness of these fits is given in Appendix D.2

6.8

Assembled Hierarchal Environment

This section will show how the various models described up to this point can be
assembled in to a hierarchical decision making environment. The ultimate goal of
this section is to be able to manipulate the entire hierarchy, from the highest level
measures of merit, all the way down to the lowest subsystem level variables.

222

The complete hierarchical design space is populated using a Monte Carlo simulation and visualized using the multivariate scatter plot in Figure 99. At the bottom
right corner within the red box are the radar properties, including its subsystem level
variables, accuracy capability, and update rate. The gun properties are shown in
the orange box, and are simply the pointing accuracy error, the muzzle velocity, and
bore. However, the gun pointing error is also a function of the radar measurement
accuracy, therefore there is an overlap at the radar measurement for both radar and
gun properties. In the green box are the projectile properties, including the discrete
bore sizes and kill mechanisms. Note the other overlap between projectile and gun
bore; recall how gun rate of fire is directly a function of bore. Finally at the top
left corner in the blue box are the operational level properties. Here one can view
the intercept range and corresponding single shot probability of hit for all variables
selected below the hierarchy. Using that single shot probability of hit and the number
of rounds fired (based on rate of fire and bore), the multiple shot probability of hit
is given. Additionally, as a measure of cost, the radar cost is given at the very top of
the hierarchy as an operational level variable.
This is the assembled hierarchy that gives operational level properties decomposed
into the systems that drive those properties, with one of those systems decomposed
down to the subsystem level. Here a decision maker can manipulate the design space
at any level of the hierarchy. For example, a high level decision maker or war planner
can sit at the very top and decide he or she is only interested in solutions that meet
a given multiple shot probability of hit, and must stay within a given budget. By
eliminating the unwanted design space in those metrics of interest, the remaining
design space will appear at each and every other level in the hierarchy.
The design space in Figure 100 is sorted by kill mechanism in order to help differentiate between them. Here very obvious trends appear in many of the individual
plots, some of which are given for clarity as bivariate plots in Figures 101 and 102. In

223

Radar
Cost
($M)
Multiple
Shot
Phit
Rounds
Fired
Single
Shot
Phit

Operational

Range
(m)
Kill
Mechanism

Projectile

Bore
(mm)
Gun
Muzzle
Velocity

Gun

Gun
El/Az
Error
Radar
El/Az
Noise
Radar
Update
Rate
Signal/
Noise
Ratio
Dist.
Between
Antennas

Gt
Gr

Frequency

Radar

Figure 99: Assembled Hierarchical Environment - Complete Design Space

224

RCS

Figure 101, the operational level properties are shown, all versus MSPH. Note that
for each plot, a line is drawn to differentiate the design space meeting or exceeding a
MSPH of 0.90. Looking at the top right plot, it is clear that it would require at least
two of the kinetic/pin rounds to achieve a MSPH of 0.90, five of the warhead/thruster
rounds, and at least 34 of the unguided rounds1 . Clearly, the kinetic/pin round is
capable of reaching a higher MSPH using fewer shots than the other two rounds. The
corresponding plot on the top left shows MSPH versus the SSPH, for curves of corresponding number of rounds fired. For the desired MSPH line drawn on the plot, one
can determine the required SSPH for each projectile type, and for number of shots
fired. The plot on the bottom left shows the tradeoff between MSPH and intercept
range. Very clear trends appear for each of the three kill mechanisms (with dashed
lines added for emphasis). The unguided round (black points) are capable of reaching
high MSPH values at the low ranges, but as range is increased, the MSPH rapidly
decays. Actually the highest range that the unguided round can achieve a MSPH of
0.90 is about 600 m. The warhead/thruster round also has a MSPH decaying with
range, however not as rapid as the unguided round, and can achieve a MSPH of 0.90
up until a range of 700 m. The almost even distribution of kinetic/pins points on
the lower left plot of Figure 101 shows that even at high ranges, a very high MSPH
is achievable. Finally, the plot on the bottom right shows radar cost versus MSPH.
Clearly the most desirable region is the high MSPH, low cost region (further discussed
later).
One thing that was noted even just by grouping the design space by kill mechanism
is that one of the bore/muzzle-velocity/kill-mechanism combinations stood out from
the others. From manipulating the environment, it was evident that a 50 mm bore,
1

Recall from the relation given in Appendix C that the maximum number of rounds fired is a
function of bore size; for the assumption used in this example, the 50 and 57 mm guns can fire off a
maximum of 6 and 7 rounds respectively, and the 30 mm bore, which applies only to the unguided
round, can fire off a maximum of 37 rounds.

225

1650 m/s gun muzzle velocity (the fastest of the three), kinetic-kill/pin-control option
provided the greatest SSPH and MSPH at every intercept range, as shown in Figure
102. These are bivariate plots from the same completed design space given back
in Figure 100, just with this particular bore/muzzle-velocity/kill-mechanism option
emphasized, and will be emphasized in many of the remaining graphics.
One of the greatest benefits of using an assembled hierarchical environment is to
study the impacts of high level decision making on even the lowest level parameters.
For example, at the operational use level, a decision maker may only care about having
a given capability within a given cost. Shown in Figure 103, an example is given where
a decision maker wants a MSPH of at least 0.90 and radar cost under $5 million. The
remaining undesired designs are eliminated, including all of the unguided and most
of the warhead/thruster options. Taken from the larger multivariate plot, Figure 104
shows the impact on the radar subsystem level variables. The top left plot with radar
update rate and RCS clearly shows a trend for the warhead/thruster round. With
higher radar update rate, the warhead/thruster round is able to achieve the desired
MSPH with smaller projectile RCS. There are no noticable trends for the kinetic/pin
round. The top right plot showing distance between antenna element arrays and
intercept range shows several interesting trends. Clearly the warhead/thruster round
is very sensitve to the distance between the antenna arrays in its ability to intercept
the target with increased range.
Of course, a smaller radar is desirable for mobility reasons, and minimizing that
distance between antenna arrays is therefore also desirable. Several designs stand out
at the low distance between antenna arrays, and high intercept range corner - the
green circles indicating the 50 mm bore, 1650 m/s gun muzzle velocity, kinetic/pin
round. The bottom right plot shows radar frequency and average power. Again, there
are no noticeable trends for the kinetic/pin round, however a clear trend showing that
as frequency is reduced, a higher power is required for the warhead/thruster round to

226

Radar
Cost
($M)
Multiple
Shot
Phit
Rounds
Fired
Single
Shot
Phit
Range
(m)
Kill
Mechanism

Bore
(mm)
Gun
Muzzle
Velocity
Gun
El/Az
Error
Radar
El/Az
Noise
Radar
Update
Rate
Signal/
Noise
Ratio
Dist.
Between
Antennas

Gt
Gr

Blue:
Red:
Black:

Kinetic/Pin-Control
Warhead/Thruster Control
Kinetic/Unguided

Frequency

RCS

Figure 100: Assembled Hierarchical Environment - Design Space Sorted by Kill


Mechanism

227

Blue:
Red:
Black:

Kinetic/Pin-Control
Warhead/Thruster Control
Kinetic/Unguided

3
2
1
# Rounds Fired

Figure 101: Assembled Hierarchical Environment - Operational Level Bivariate


Plots

228

Green:
Blue:
Red:
Black:

50mm, Kinetic/Pin, Fastest Muzzle Vel


Kinetic/Pin-Control
Warhead/Thruster Control
Kinetic/Unguided

Figure 102: Assembled Hierarchical Environment - Emphasizing a Discrete Option


achieve the desired MSPH and cost requirements. Finally for the bottom left corner
of Figure 104, SNR and the distance between antenna arrays are shown. There is a
clear trend for all the design points where a higher SNR is required as the distance
between antenna arrays is reduced, and low SNR are tolerable at the higher antenna
array separations. However, now a radar expert can decide that any SNR values
above a given value, say 35, are unreasonable for any given reasons, and a line can be
drawn to differentiate between the designs that do and do not meet this SNR limit.
Clearly the warhead/thruster rounds require a higher antenna array separation than
the kinetic/pin rounds. This even allows the ability to isolate one particular design,
such as the kinetic/pin rounds that achieve the highest MSPH at highest range (the
green circles), and noting that the lowest antenna array separation achievable with
those rounds is about 2.25 m.
Another desired attribute on the battlefield would be to minimize the number of
rounds required to be fired against a target. Even looking back at Figure 101, if
enough unguided rounds are fired at the target, it would be hit. However, the war
fighter would rather not carry many ammunition stores, or wear out the gun barrel

229

with excessive firings. Figure 105 examines the rounds fired trade space on the 0.9
MSPH and $5 million constraints. Clearly as more rounds are fired, a greater intercept
range is achievable. For one or two rounds, only the kinetic/pin round achieves the
requirements, and beyond that, the 50 mm bore, 1650 m/s muzzle velocity, kinetic/pin
round has the clear range advantage. The top right plot shows the radar accuracy
tradeoff with rounds fired, and it is apparant that with more rounds fired, higher
radar noise is tolerable. The bottom right plot shows how the required SSPH can be
reduced with increased shots fired, and note how the warhead/thruster round is at
the lowest SSPH for the rounds fired in which they appear. Finally, the bottom left
plot of Figure 105 shows how the distance between antenna arrays varies with rounds
fired. As the antenna arrays are brought in closer, more shots are required to achieve
the operational level requirements.
The final tradeoff examined in this chapter is the gun muzzle velocity selection,
given in Figure 106. Note that now the green circles are used to differentiate between
the 50 and 57 mm kinetic/pin rounds. There were three muzzle velocities included
in the detailed study, and the effects of which one used has a very defined impact.
For the given high level requirements, note that at the slowest muzzle velocity, more
rounds fired are required, and are only able to achieve intercept ranges up to about 700
m. For minimizing the number of rounds fired to achieve the high level requirements,
a 57 mm kinetic/pin fired at the middle muzzle velocity option is the best. However
as discussed earlier, the 50 mm kinetic/pin fired at the highest muzzle velocity had
the greatest high range MSPH. Looking at the right side plot in Figure 106, there is
not that much range benefit of using the smaller kinetic/pin bore, but this of course
might assume firing more rounds. However, it may be more difficult to design the
electronics and internal packaging for a smaller guided bullet, making the large bore
more desirable.

230

Radar
Cost
($M)
Multiple
Shot
Phit
Rounds
Fired
Single
Shot
Phit
Range
(m)
Kill
Mechanism

Bore
(mm)
Gun
Muzzle
Velocity
Gun
El/Az
Error
Radar
El/Az
Noise
Radar
Update
Rate
Signal/
Noise
Ratio

Blue:
Red:

Dist.
Between

Kinetic/Pin-Control
Warhead/Thruster Control

Antennas

Gt
Gr

Frequency

RCS

Figure 103: Assembled Hierarchical Environment - Selected Desired Phit and Radar
Cost Operational Constraints

231

Green:
Blue:
Red:

50mm, Kinetic/Pin, Fastest Muzzle Vel


Kinetic/Pin-Control
Warhead/Thruster Control

Figure 104: Assembled Hierarchical Environment - Radar Properties Due to Operational Level Constraints

232

Green:
Blue:
Red:

50mm, Kinetic/Pin, Fastest Muzzle Vel


Kinetic/Pin-Control
Warhead/Thruster Control

Figure 105: Assembled Hierarchical Environment - Number of Rounds Fired Trade


Space

233

Green:
Blue:
Red:

Kinetic/Pin-Control (57mm)
Kinetic/Pin-Control (50mm)
Warhead/Thruster Control

Figure 106: Assembled Hierarchical Environment - Gun Muzzle Velocity Trade


Space

6.9

Chapter Conclusions

The assembled environment in Section 6.8 showed how the qualitative metrics of interest for an air defense system identified in Section 5.2 - lethality, mobility, and
cost - can be represented across a system-of-systems hierarchy, using a top-down design approach to directly manipulate any property in that hierarchy. The measure
of lethality was represented by the probability of directly intercepting the target using one or more projectiles, or using the measured confidence of impacting within a
certain distance from that target. That measured lethality was a function of every
property of the functionally decomposed environment discussed in Chapter 5, however as first discussed in Section 1.2.2.2, the radar capability would drive a guided
projectiles effectiveness. Several of the radar subsystem properties that define the
radar accuracy would then be used to represent the other two qualitative measures of
effectiveness. The desired lethality capability would drive the radar accuracy, which
in turn would drive the distance between antenna element arrays, which is a measure
of the weapon system size, or mobility. That radar accuracy would also drive the
radar properties that greatly affect its cost, which was shown to be of significant
234

value. This chapter first proved this approach one hierarchical level at a time with
detail, and finally showed how the entire environment can be manipulated, flowing
top level requirements directly to system and subsystem level properties. Examples
were given throughout this chapter showing how requirements at any level in the hierarchy affect all other properties, as well how this can be used to discover interesting
tradeoff spaces that exist between any combination of variables and responses that
might have not otherwise been known.

235

CHAPTER VII

CONCLUSIONS
7.1

Review of Research Questions & Hypotheses

The very first research question was asked in order to determine the deficiencies the
systems engineering process as applied to complex hierarchical systems. With those
deficiencies identified, such as the inability to flow information across nonadjacent
levels, the second research question expressed the desire to create a process in which
those deficiencies could be overcome. The research questions asked in Chapter 4
helped focus the scope of this thesis and were based on the problem introduced in the
motivation chapter, as well as information learned throughout the literature search
effort. The hypotheses presented attempted to directly answer those questions using
only knowledge gained during the literature search. This section will show how the
approach laid out in Chapter 5 and associated methodology implementation from
Chapter 6 proved the hypotheses, therefore answering the questions presented in
Chapter 4.
7.1.1

First Hypothesis

The first hypothesis laid out a basic approach to answering Research Question 2, and
served as the primary research objective for this thesis.
Research Question 2: How can a process be created that enables rapid
manipulation across a hierarchy of decision making levels?
Hypothesis 1: A system-of-systems approach can be used to create an
environment that will allow for both weapon requirements definition and

236

system optimization from top-down requirements that take into account


operational environment characteristics, as well as for analyzing the impacts of subsystem level changes on top level metrics from the bottom-up.
The motivation for this thesis discussed in Chapter 1 led to the second research
question, which was asked before any literature search was conducted. This question
was meant to address the deficiencies associated with the systems engineering process
discovered in Section 1.1.2. This desired process needed to include more than just generalized requirements for a method to design this type of weapon, but also a process
to assemble an integrated hierarchical environment, based on accepted modeling and
simulation methods, for the various components that need to interact and allow for
rapid manipulation to enable decision making at any level within the hierarchy.
Therefore, the first hypothesis proposed creating a system-of-systems inspired
approach for a hierarchical decision making design environment, where multiple interacting and independent systems affect an overall measure of effectiveness. This
first hypothesis was realized in both the approach (Chapter 5) and implementation
(Chapter 6) portions of this thesis. In Chapter 5, the problem of interest was hierarchically decomposed to the functions that enable the entire process, which served two
purposes. First, the scope of the problem of interest was focused such that a sufficient
methodology can be applied to address depth and breadth. Second, the hierarchical
interdependencies were established showing which properties of the independent systems interact to drive operational measures of effectiveness. This process also showed
the breakdown of the degrees of freedom across the entire system-of-systems hierarchy, from the lowest level subsystem properties all the way up to the highest level of
decision making. The idea here was that once this environment is constructed, the
second hypothesis can be realized, answering the next set of research questions.

237

7.1.2

Second Hypothesis

The second hypothesis addressed the issues relating to Research Questions 3, 4, and
5.
Research Question 3: An ANOVA/Pareto plot technique can be used
to quantify the contribution of each variable to the variability of a response,
but is there a way to identify which variables drive each other?
Research Question 4: Can a tradeoff space be identified interchangeably
between variables and responses?
Research Question 5: How can rapid design tradeoffs be assessed in a
hierarchical environment composed of independent time consuming models
if the design space between them is multimodal or discontinuous?
Hypothesis 2: A hierarchical environment composed of neural network
based surrogate models will enable multimodal and/or discontinuous design
spaces to be accurately represented, allowing for rapid tradeoffs. These
tradeoffs may be conducted with real time application of constraints at
any level of the hierarchy, for both independent variables and dependent
responses.
Research Questions 3, 4 and 5 were asked in ascending detail, however methods
had to be applied to answer the questions in reverse order as implied by the second hypothesis. In order to identify the variable combinations that drive each other
(Question 3), a tradeoff space must be identified between all of the variables and
responses (Question 4). Although polynomial surrogate models have emerged as a
useful way to capture the design space of very complex codes and enable rapid manipulation and tradeoffs, they severely limit the topology of the design space captured
in that that they are appropriate only when the design space is continuous and not
238

highly multimodal. Research Question 5 was addressed using neural network based
surrogate models that were able to capture very complex design spaces required for
exploring across a system-of-systems hierarchy. For example, neural networks were
used to capture the discontinuous design space surrounding the quantification of the
probability of hit (i.e. a successful collision between interceptor projectile and target). This required sampling from an error distribution for every noise variable set
prescribed by a Design of Experiments (DoE), and fitting to discontinuous cases were
sometimes there was a successful target intercept, and many times when there was
not an intercept.
To create the tradeoff space discussed in Question 4, a Monte Carlo simulation was
used to uniformly fill the design space in the system-of-systems hierarchy using those
neural network surrogate models. Using an interactive multivariate scatter plot, the
design space of every variable and every response were viewed simultaneously, and
interchangeably. This brings back the notion of viewing the design space in the context of the total (or full) derivative. Rather than extract single curves to show the
pertinent tradeoff space between responses and variables for all other variables fixed,
as with partial derivatives, this method keeps all degrees of freedom open throughout
the design hierarchy. This allows for constraints to be placed anywhere in the hierarchy, regardless of whether they were variables or responses, clearly identifying the
feasible design options and tradeoffs between any combination of variable/variable,
variable/response, and even response/response. This method enables treating every
variable and response as an independent variable. In doing this, the remainder of
the second hypothesis was realized, answering Research Question 3. As constraints
or requirements were imposed on metrics of interest within the hierarchy, the independent variables that were key drivers to the variability of those metrics could be
identified, as well as the tradeoff spaces between those variables. The tradeoff space

239

that emerged as a result of placing constraints on responses was important in identifying which variables are dependent on the settings of other variables. This was the
realization of the hierarchical top-down environment desired as a result of the first
hypothesis.
7.1.3

Third Hypothesis

The third hypothesis addressed Research Question 6.


Research Question 6: How can the effects of uncertainty propagation
at any level within a hierarchical decision making design environment be
assessed?
Hypothesis 3: A properly constructed system-of-systems environment
will allow for the rapid assessment of the effects of uncertainty at any
level in the hierarchy on any other level in the hierarchy.
This thesis approached the subject of uncertainty in the design hierarchy in a
number of ways. The first was from the bottom-up, where it was desired to determine the effects of uncertainty in a design variable on the ability to achieve a certain
response. The traditional way to achieve this is to run a Monte Carlo simulation
sampling from a distribution on the design space of those given variables, and quantifying the confidence of achieving a given response using a cumulative distribution
function (CDF). The most straight forward method to achieve this in this thesis was
the quantification of the probability of direct interceptor/target hit and of the CDF
curve resulting in variation of the noise variables of various systems in the hierarchy.
This enables the quantification of the bounds of variables that are random functions
to an acceptable CDF
However, having the methodology in place for constructing a hierarchical decision
making environment (as was realized in support of the first two hypotheses), a design

240

variable region of interest could easily be defined. This region of interest may encompass a region of uncertainty within a variable design space, or even as shown in many
cases, several regions. With the interactive multivariate scatter plot, the resulting
design space for every other variable and response are identified. The second major approach to uncertainty was done by identifying regions of acceptable responses
and seeing how that constrains the remaining design space from the top-down. In
other words, for a given region of uncertain desired response, this thesis showed the
acceptable dependent variable design space.

7.2

Summary of Contributions

As discussed in the motivation for this thesis, in order to evaluate complex systems
such as revolutionary applications for air defense weapons, there is a need for rapid
manipulation of the systems engineering process across nonadjacent levels of a design hierarchy. This thesis did not try to provide a specific design solution to the
general example problem presented in the motivation chapter. Rather, a portion of
this problem was addressed to show how the methods developed could be applied to
a problem of current interest for which there is no existing or programmed system
with the ability to negate these threats after they are launched. One of the major
contributions of this thesis to the field of aerospace systems design is the introduction
of a methodology that enables decision makers anywhere across a system-of-systems
hierarchy to rapidly and simultaneously manipulate the design space, however complex. This process eliminates the iterative steps that usually occur when dealing with
flowing requirements from one level to next lower in the systems engineering process.
By removing the internal constraints local to one level in the systems hierarchy, the
design space could be manipulated from the top-down of a design environment directly, without having to deal with the iterations involved in optimizing a system or
subsystem one level at a time.

241

The application of the process developed to the example problem enabled the
identification of interesting phenomena that might not have been possible otherwise.
This was due to the fact the data representing the design space was presented in
bulk, rather than just static curves showing a response as a function of one variable,
with all other variables (and assumptions) held constant. For example, the effects
of the application of overall capability requirements at the highest decision making
level could be instantly flowed down to carve out the feasible design space not only
at the lower system capability level, but directly to the subsystem property level
which enable that particular system capability. Using this new methodology, regions
emerged among those variables where either subsystem design properties were either
insensitive to high level requirements, or where trends among those variables could
be easily identified.
Additionally, this process allows different decision makers to all explore the design space simultaneously each with their own set of assumptions. This is primarily
possible because the neural network surrogate models allow for freedom in the variable design space, while still accurately representing a complex computer simulation.
Therefore, as was shown in the implementation chapter, a user has the ability to
define or eliminate regions of assumptions of variable design space, and observe the
effects on each and every other variable and response in this interacting hierarchical
environment. This also includes the ability to bound variables that are random functions. Prior to this new formulated methodology, these capabilities were not available
on such a rapid manner.

7.3

Recommendations for Further Study

One of the most beneficial improvements to the actual methods developed in this
thesis would be to mathematically define the Pareto frontier discovered during the

242

real time application of requirements and constraints on the interactive multivariate scatter plot. Recall that in many cases where a region of interest was defined
on a multivariate scatter plot, several tradeoff spaces emerged in the remaining response/variable design space. There have been many studies into the definition of a
response versus response Pareto frontier for given variable design space. This defines
the region where an improvement in one response degrades another response. What
is suggested for future work is the definition of the tradeoff space for the variables,
given a response design space, such that any combination of variables along that
frontier result in a feasible space. This should be done in real time while the user is
constraining the response design space. Additionally, in order to further develop the
systems engineering process, the author recommends a more thorough mathematical
understanding of the multivariate analysis nature of the process.
There is much potential to expand the scope of the specific problem addressed in
this thesis to deal with more issues relating to the EAPS issue. The system-of-systems
methodology can be expanded to include multiple projectiles fired at multiple targets
simultaneously. This may even be expanded to include engaging multiple target types
simultaneously. Here the problem can become more complex due to the different
errors associated with each target type. Additionally, this thesis examined targets
flying at a zero angle trajectory in relation to the projectile interceptor. In essence,
a two dimensional problem was set up in the 6-DoF code, and was only perturbed
perpendicular to the direction of forward flight due to the errors in gun firing and
radar azimuth. One could certainly apply this to off-axis trajectory targets, i.e. design
a single gun to defend an area from targets coming in from many directions - and
setting up a true three dimensional problem.
In the same manner discussed where top level requirements are flowed down to
radar subsystem variables, having access to the gun attributes that result in a certain
gun accuracy would allow for the subsystem manipulation of the gun accuracy that

243

is used in the 6-DoF simulation. However, the problem to overcome here would be
the same encountered during the literature search for this thesis - there is not much
open source information available on medium or large caliber gun design. A much
deeper investigation into the field of gun design would be required.
The role of the tracking radar may be expanded to include the role of target
recognition, and several useful sources were discovered that describe the techniques
and methods, such as that by Nebabin [118], but were not used in this thesis as they
were beyond the scope of this thesis. However, this is a topic that is pertinent to the
problem discussed in the motivation chapter in that the remainder of the upstream
kill chain could be quantified, rather than just assume that a valid target has been
identified and tracked. Another excellent application of the power of capturing very
complex design spaces with neural network based surrogate surrogate models is to
capture the actual radar cross section of target in three dimensions, and even as a
function of radar frequency. This way the the fluctuation in target RCS relative to
the radar can be taken into consideration as the target travels along its flight path,
and the associated radar noise errors which are a function of RCS, can be accounted
for.

244

APPENDIX A

PROBABILITY DISTRIBUTIONS

This Appendix is meant to introduce some background for several of the important
probability distributions discussed in this study. There are certainly many different
types of probability distributions, such as those given in Figure 107, and each one has
its strengths and weaknesses when used for a particular application. Here the uniform
and normal continuous distributions, as well as the discrete binomial distribution is
introduced. A good statistics book [70] [74] should certainly be consulted for a more
in depth explanation.

A.1

Uniform Distribution

The uniform distribution is the simplest of all continuous distributions, in that any
number within the sample bounds is equally likely to be selected in a random number
draw. The uniform distribution has a constant probability density function (PDF)
between its only two parameters, the minimum a and the maximum b bounds, as
given in Equation 30. The standard uniform distribution has the parameters a = 0

Figure 107: Sample Probability Distributions [52]

245

Table 21: Area Captured Under a Normal Distribution for Standard Deviation
Multipliers
Standard Deviations Area Captured Under Curve
1
68.27%
2
95.45%
3
99.73%
4
99.9994%

and b = 1, i.e. selecting a random number between 0 and 1 with equal likelihood.

f (x) =

1
ba

= 0

A.2

axb

(30)

otherwise

Normal Distribution

The normal distribution, sometimes referred to as the Gaussian distribution in honor


of Carl Freidrich Gauss who developed it around 1820, is considered one of the most
important of all continuous probability distributions, and is used frequently in statistics [70]. The PDF for the normal distribution is given in Equation 31,
(x)2
1
f (x) = e 22
2

(31)

where the random variable x is bounded - x and depends on the mean


and variance 2 (where is the standard deviation). This distribution is symmetric
about the mean , and for this reason occasionally referred to as a bell curve. The
number of standard deviations away from the mean value can describe the amount
of are under the distribution curve captured, as given in Table 21.

A.3

Binomial Distribution

Many types of problems have only two types of outcomes for a single trial, or can
have one of two types of states, for example the success or failure of a projectile to

246

intercept a target. A set of these repeated trials of the same experiment in which each
of the trials are independent is called a Bernoulli trial [74]. The binomial distribution
given in Equation 32, was derived by James Bernoulli in 1713 to model the total
number of successes in repeated trials from an infinite population, given the following
conditions [103]:
Only two outcomes are possible on each of n trials
Probability of success for each trial p is held constant
Each of the trials is independent of each other

n x
f (x) =
p (1 p)nx
x

(32)

where
n
x

n!
(n x)!x!

x = 0, 1, 3, , n
It is known as the binomial PDF because the coefficients in the fraction are successive terms of the binomial expansion. Note that the PDF has a dependence on
both p and n, and that it is a discrete, rather than continuous, distribution since the
required number of trials, x can only take integer values.

247

APPENDIX B

PRELIMINARY RESULTS

This chapter will introduce some relevant research conducted by the author that
serves as a proof-of-concept of the application of the methodologies introduced in
Chapters 2 and 3, as applied to the problem of interest introduced in Chapter 1.

B.1
B.1.1

Skyshield/Ahead Ammunition Study


Introduction

A study by Ender et al. [60] was conducted to measure the effectiveness of firing an
air bursting projectile munition against very small targets by quantifying the weapon
system uncertainties inherent to a medium caliber weapon. The goal was to determine
how accurate the projectile must be fired to burst close enough to the inbound target
such that enough submunitions hit the target with the necessary kinetic energy to
defeat it. Additionally, only those submunitions that hit an even smaller portion of
the target, such as its fuze, would count as effective hits. A Monte Carlo simulation
was used to quantify the uncertainties associated with arriving at a desired point in
three dimensional space with a certain velocity as a result of error associated with
gun firing elevation, azimuth, muzzle velocity, and crosswind. The values used for
error are listed in Table 22. The target intercept locations defined for this preliminary
study were between 500 m and 2000 m downrange and between 1000 m and 3000 m
in altitude.
The parameters in the design space consisted of a 35 mm unguided round containing submunitions, based on the Ahead round described in Section 1.2.3.2. This
round is presently designed to defeat larger and softer targets than the target type of

248

Table 22: Ahead Study - Errors Used to Quantify Miss Distance


Variable
Elevation
Azimuth
Muzzle Velocity
Crosswind Gusts

Error (1 values)
1.25
1.25
2.50
2.50

units
milliradians
milliradians
m/s
m/s

interest in this proposal study, thus its effectiveness against the smaller and harder
targets, such as mortar rounds, was the focus of this study. The munition under
investigation disperses 152 small tungsten cylinders where the dispersion density of
the submunitions depends upon several factors, such as the distance the projectile is
fuzed ahead of the target, the velocity degradation due to the projectile burst, and
the dispersion pattern maintained by the submunitions as they cross the target plane.
Quantifying the uncertainty due to the combined effects of the error in the weapon
system and the change in dispersion density will yield the confidence in hitting the
target with a certain number of submunitions, as well as the resulting kinetic energy
delivered.
PRODAS [29] was used as the six degree-of-freedom trajectory analysis tool, with
variations on firing conditions used to model the effects of several types of error. A
Monte Carlo analysis was used to quantify uncertainty, where variability in input
variables and/or assumptions was tracked by adding distributions to the nominal
values. Many cases with randomly selected points from these distributions provided
a statistical indication of the probability with which each response occurred. For this
study, the probability of arriving at a position in three-dimensional space, and the
resultant velocity at impact of the altered trajectory, was quantified as a function of
error in gun elevation, azimuth, muzzle velocity, and crosswind. The best way to run a
Monte Carlo simulation is to sample the actual analysis tool many times, preferably
thousands of times. But with codes that require several seconds to run a single

249

Responses

Current value
(with error)
Current value
(normalized to maintain
order-of-magnitude
variability in responses)

Variables

Bounds

Figure 108: Ahead Study - Interactive Prediction Profiler of Response Surface


Equations
case, this may not prove feasible, especially with limited processor capability and
the desire to try many different conditions. Therefore, polynomial based surrogate
models of the PRODAS simulation tool were created in order to reduce the run time
of the thousands of necessary simulation trials, using a Design of Experiments to
create representative polynomial response surface equations. A static example of
these surrogate models are graphically shown using a JMP [127] prediction profiler
in Figure 108.
B.1.2

Miss Distance Quantification

The first use of the advanced design methodology approach in this study was in
the quantification of the miss distance between the intended trajectory to intercept
the target, and the trajectory resulting from the errors in the weapon system. This
step only examines the flight trajectory of the entire projectile without submunition
dispersion. A Monte Carlo simulation was run in which the error distributions placed
on the variables were sampled 10,000 times, each time using those values as the
250

inputs to the surrogate models that calculate the distance from actual target position
and the resultant projectile location in three dimensional space, and the velocity
of the projectile when it reaches the target point as a result of the system errors.
Figure 109 shows the results of a miss distance Monte Carlo study run on the closest
notional target trajectory (500 m down range; 1000 m altitude), where X is the
downrange distance, Y is the cross range distance, and Z is altitude. For each of
the 10,000 samples, the results are tracked and plotted on a histogram, the peaks
of which form a Probability Density Function (PDF) that closely follows a normal
distribution with the intended target (zero miss distance) as the mean. This shows
that the input error does not bias the projectile in any particular one-dimensional
direction. The integration of the PDF gives the Cumulative Density Function (CDF),
and the probability of reaching within a certain distance of the intended target can
be quantified. The CDF plots the confidence of reaching within a certain distance in
1-D space around the intended target. Because there is no skewness in the PDF, the
likelihood of missing the target the same amount on either side is just as likely. Final
velocity is presented to show the effects that position error has on final velocity. This
plot shows that with these errors, there is still a 90% confidence that the velocity
with which the projectile would hit the intended target would be within about 1.5
m/s of the mean value of the PDF. This small variation may not significantly affect
the kinetic energy delivered if the target is hit.
Here, the CDFs have a use analogous to the circular error probable (CEP) metric
used in the field of missile design, which is the radius of a circle within which 50%
of the flight trajectories with the closest miss distance are expected to occur. An
assumption used by Fleeman [63] in the conceptual design of missiles is that these
miss distances are normally distributed around the intended target (as was the case
in Monte Carlo results in Figure 6). In terms of standard deviation, the probability
that the miss distance would be less than 1 value of the normal distribution is 39%

251

Figure 109: Ahead Study - PDF (top) and CDF (bottom) for Miss Distance and
Velocity at Target Intercept

252

Figure 110: Ahead Study - Miss Distance Point Cloud Dispersion Results
. Similarly, the probability of being within a 2 miss distance is 86% , and a 3 miss
distance is 99%. For example, a blast fragmentation warhead design may be set so that
the warhead lethality radius is equal to the guidance estimate of the 3 miss distance
to provide a comfortable margin for miss distance uncertainty. Fleeman states that
a common assumption in the conceptual design of missiles is to interchangeably use
the CEP and 1 terminology, consistent with design accuracy. In this study, the
Monte Carlo simulation was used to create point cloud dispersion plots as shown
in Figure 110, but instead of two-dimensional circles, the point clouds are threedimensional ellipses. The point clouds relative sizes indicate an increase in dispersion
with increasing range.

253

B.1.3

Fuze Distance

A submunition dispersion model was needed to describe how the submunitions propagate through space after the fuze sets off the airburst. Using information from the
manufacturer Oerlikon Contraves, several simplifying assumptions were made. After
fuzing, the submunitions disperse at a constant 10 angle. Additionally, the submunitions disperse according to a standard witness plate trial, shown in Figure 111. This
dispersion pattern was dimensionally scaled by radius to fit into the target plane striking area shown in Figure 112, depending on how far the projectile was fuzed ahead of
the target. The main assumption is that the witness plate submunition distribution
maintains its normalized shape regardless of fuzing distance, and that all the submunitions cross the target plane at the same time. Additionally, a random spin was
added, meaning that the entire dispersion pattern was rotated at an arbitrary angle.
The degradation in kinetic energy delivered by the individual submunition hits was
taken into account by the velocity profile shown in Figure 112. For example, if the
submunitions travel 10 m after the projectile is fuzed, their velocity is degraded to
93% of the velocity at fuzing. This is a simple drag model based on tests conducted
by Oerlikon Contraves. Another simplifying assumption was made in that all of the
projectiles that were fired in a burst to engage the inbound target arrived at the same
time. In other words, the entire engagement occurred in a single plane.
Figure 113 shows a sample dispersion of 10 projectiles fuzed several meters from
a 14 mm diameter target, projected in the plane of the target located at the origin.
Note the different sizes of the dispersion patterns due to the various errors described
earlier. The zoomed in portion of the figure shows the dispersion around the target
and highlights the difficulty in hitting such a small target, as will be addressed in
the accuracy requirement section. Many of the submunition dispersion patterns from
each of the projectiles overlap. Therefore, if several submunitions from different projectiles hit the intended target, each will deliver a different kinetic energy due to each

254

gure 7. Point Cloud Dispersion Results for Miss Distance Study.

IV.

Weapon System Accuracy

sion Model
sion model was needed to describe how the submunitions propagate through space after the
Using information from Oerlikon Contraves, several simplifying assumptions were made.
nitions disperse at a constant 10o angle. Additionally, the submunitions disperse according
late trial, shown in Figure 8. This
mensionally scaled by radius to fit into
area shown in Figure 9, depending on
s fuzed ahead of the target. The main
witness plate submunition distribution
shape regardless of fuzing distance,
ions cross the target plane at the same
dom spin was added, meaning that the
was rotated at an arbitrary angle. The
energy delivered by the individual
en into account by the velocity profile
r example, if the submunitions travel
is fuzed, their velocity is degraded to
fuzing. This is a simple drag model
ed by Oerlikon Contraves. Another
was made in that all of the projectiles
to engage the inbound target arrived at
Figure
8. Normalized
Submunition
words, the entire engagement
Figureoccurred
111: Ahead Study
- Normalized
Submunition
Witness Plate Dispersion Pature plans include accounting
for the
tern
Witness Plate Dispersion
Pattern.
Engagement
Assumptions
or each round in a burst.

Several simplifying assumptions made to the model (continued)


8
All projectiles arrive at the same time and are aimed at the same physical point in
American Institute of Aeronautics
and Astronautics
space
Really a 0.1 s or 100 m separation between rounds

After fuzing, submunitions spread at constant 10 angle

Figure Source:
www.oerlikoncontraves.it

For Official Use Only

15

Figure 112: Ahead Study - Velocity Degradation Due to Submunition Dispersion


[2]

255

Height

Lateral Distance

Figure 113: Ahead Study - Submunition Dispersion Clouds from Multiple Projectiles
projectiles final velocity as it crosses the target plane due to the errors applied in
the Monte Carlo simulation. Monte Carlo trials conducted at various fuze distances
ahead of the target will result in the optimal fuze distance required to hit the target with a required number of submunitions (i.e. kinetic energy) with quantifiable
confidence.

B.2

CIWS study

The goal of this study was to determine the probability of hitting an 81 mm moving
target as it approaches the defended area, while accounting for miss distance caused
by random gun errors. The target was to be intercepted between 250-1000 m, and as
shown in Figure 114, the range bounds were divided to four discrete cases that the
simulation was run. Just as for the previous study, PRODAS [29] was used as the six
degree-of-freedom trajectory analysis tool, with variations on firing conditions used

256

Incoming mortar trajectory

Intercept Ranges
1000m
750m
500m
250m

Figure 114: CIWS Study - Target Intercept Ranges


to model the effects of several types of error. The ultimate goal was to determine the
probability of hitting the target due to the cumulative effect of tracking the target
along its flight path, given the randomness in gun and crosswind error. For this
study, randomness was introduced into five variables, the bounds of which will not
be presented. These variables include:
1. Gun firing elevation
2. Gun firing azimuth
3. Crosswind gusts
4. Initial pitch
5. Initial yaw
Although a three dimensional problem, this particular study only focused on the
two dimensional plane around the circular cross section of the target. This method
was used since PRODAS can not model the trajectory of two objects simultaneously,
i.e. a target and interceptor, and determine whether a hit occurred in three dimensional space. Polynomial RSEs were created for the actual point in space when it
crosses the cross range/altitude plane of the target. Therefore, the prediction profiler
in Figure 115 does not include down range (X) responses, but only cross range (Y)
and altitude (Z) at each of the intercept ranges.
257

Figure 115: CIWS Study - Prediction Profiler of Surrogate Models

258

A Monte Carlo analysis was run to determine the probability of target intercept by
a single round at the four notional ranges. A scatter plot showing where each round
crosses the target plane can be used to calculate the 50% and 90% CEPs. The ratio
of rounds crossing the target plane within the radius of the target (from the center
of the target) to the total rounds examined in the Monte Carlo study indicates the
probability of hitting the target, for a single round. Figure 116 shows the results of
a Monte Carlo simulation run for each of the target intercept ranges. The blue circle
represents the diameter of the target, the green circle represents the area comprising
50% of the projectiles, and the red circle represents the area comprising 90% of the
projectiles. Note that as intercept range increases, the number of projectiles within
the target area diminishes, as expected.
To quantify the confidence of arriving within a certain distance of the target, the
CDF was calculated using the Monte Carlo simulations. The plots given in Figure
117 show the confidence of a single round arriving at a distance away from the center
of the target, or the probability of a single hit. The red line represents the target
radius, therefore the intersection of the red line with the CDF yields the probability
of directly hitting the target, PH . The curve may also be discretized by the total
number of desired projectiles fired as the target travels, assuming that the single
shot probability of hit at any range in between the four ranges listed earlier can be
determined by simple linear interpolation. For an accurate cumulative hit probability,
Equation 9 should be used. However, for simplicity Macfadzean [92] states that for
very small single shot probabilities of hit, the cumulative probability of hit can be
simply approximated by summing the individual single shot probabilities. Figure 118
shows the curve discretized for 100 rounds, meaning that some firing rate was assumed
that resulted in 100 rounds being fired, the first one cross the target plane at 1000 m,
and each subsequent round evenly spaced. Note that the study was only conducted at
four discrete ranges. Certainly more test points would be required to define the space

259

260

Figure 116: CIWS Study - Monte Carlo CEP Results Single Round Intercept at 250, 500, 750, and 1000 m

Figure 117: CIWS Study - CDF for 250, 500, 750, 1000 m Ranges

261

Figure 118: CIWS Study - Single Round PH (L) and Discretized Single Round PH
(R)
between the 250 m and 500 m ranges, where the greatest difference between PH values
is. The cumulative probability of hitting the target with at least one projectile, or
the confidence of intercepting the target as it approaches, is quantified in Figure 119.
This simple curve now shows how the change in PH varies as the target approaches
the defended area.

262

Figure 119: CIWS Study - Cumulative PH of Approaching Target

263

APPENDIX C

GUN RATE OF FIRE REGRESSION

The historical data shown in Table 23 is based on a compilation of information collected by Massey [25]. The goal in assembling this database of historical information
was to find a basic correlation between a guns rate of fire and bore size. With the
framework in place to create neural network surrogate models in place for most of
the other work presented in this thesis, the author decided to fit a regression using
the neural network methods. Equation 33 represents the method used to calculate
rate of fire in this thesis. This equation calculates Rate of Fire (RoF) in rounds per
minute, as a function of Bore size (in mm). Figure 120 shows the regression used
with the data points used to create them overlaid and the distribution of error. This
fit is certainly not excellent by most measures of merit, but considering that more
than one data point (rate of fire) exist for many of the bore values input, this fit is
an improvement over using a simple polynomial or logarithmic regression. Additionally, as overly complex as this equation make seem, it was very simple to create and
implement in the simulation.

264

265

Gun System
Flak 38
XM301
Phalanx C-RAM
m61 Vulcan
ZSU-23 Shilka
GAU-12 Equalizer
Mk44
Flak 103/38
GAU-8 Avenger
M230
Aden Mk5
Bushmaster III
Skysheild
Flak 18
Flak 36/37
M1939 Type 55
Flak 43
Bushmaster IV
Mk44
mk47
mk19
Flak 41
Bushmaster III
Great 58
S-60
Bofors 57
8.8cm Flak 18,37,37
8.8cm Flak 41
10.5cm Flak 38/39
12.8cm Flak

Caliber (mm)
20
20
20
20
23
25
30
30
30
30
30
35
35
37
37
37
37
40
40
40
40
50
50
55
57
57
88
88
105
128

Rate of Fire (RPM)


450
1500
4500
6000
4000
4200
200
400
4200
625
1600
160
1000
160
160
180
250
160
200
300
350
130
160
140
120
220
15
20
15
12

Barrels
1
3
6
6
4
6
1
1
7
1
1
1
2
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1

RoF/Barrel (RPM)
450
500
750
1000
1000
700
200
400
600
625
1600
160
500
160
160
180
250
160
200
300
350
130
160
140
120
220
15
20
15
12

Table 23: Historical Data For Gun Rate of Fire [25]

1934
1942
1939
1942

Super 40
mk19 replacement
grenade launcher 1960s
1941
New
1941
1950s Soviet

Comment
1939 German
Commanche gun
Gatling gun
1960s
1965 Soviet Gun
A-10 gun
new
1944 German
F-18 gun
1988 Apache gun
1954 Aden Mk1
New
Growth version to 50mm
1935
1937
1960s
1937

Figure 120: Firing Rate Curve Fit (L) and Associated Error Distribution (R)

log10 RoF = 2.48490665 + 6.214608098


"

0.0300307957854 +

(33)
5
X
i=1

Ai

1
1 + e(Bi +Ci (

Bore7.62
120.38

where
A1 = 0.0606368589412 B1 = 27.9929448352691 C1 = 28.0070451148204
A2 = 0.0295660006372 B2 = 21.0101869886458 C2 = 27.9880284206632
A3 = 0.3500044429334 B3 = 14.1935862262056 C3 = 27.9214623686777
A4 = 0.4375050999628

B4 = 5.6595893903830

C4 = 28.3736546079568

A5 = 0.409428487125

B5 = 0.2901708320125

C5 = 27.9617886232176

266

))

APPENDIX D

SUPPORTING DATA
D.1

Initial Study Pareto Plots

This Appendix provides the complete set of Pareto plots referenced in Section 6.3.4.
Figures 121, 122, and 123 include a Pareto plot for each discrete bore/range combination examining the response variability in the initial 6-DoF study.

D.2

Detail Study Surrogate Model Fit Results

Figure 124 shows the fit results for the neural network surrogate models created in
the Detail Study in Section 6.7 and used in the Assembled Hierarchical Environment
in Section 6.8. Note that for the two guided rounds (the kinetic-kill/pin-control and
warhead-kill/thruster-control), the fits are fairly good. However, the unguided round
has a relatively poor fit. The main reason for this is that unlike the two guided rounds
that used 750 Monte Carlo simulations per 6-DoF execution, the unguided trials only
executed 250 runs. With fewer Monte Carlo trials per DoE run, the data was very
coarse, as can be seen in the bottom left plot of Figure 124, showing the actual data by
predicted results for the unguided fits. As discussed in Section 6.7, close to one million
6-DoF trials were necessary to generate the fits shown in Figure 124, and towards
the end of this process, the computational resources were less available. However,
as the unguided results were used only for the purposes of comparison against the
guided rounds, and the unguided trends made sense, these fit results were considered
sufficient.

267

Bore: 40mm
Range: 250m

Bore: 40mm
Range: 500m

Bore: 40mm
Range: 1000m

Bore: 40mm
Range: 1500m

Bore: 40mm
Range: 2000m

Figure 121: Initial Study - Pareto Plots for 40 mm Bore, All Ranges

268

Bore: 50mm
Range: 250m

Bore: 50mm
Range: 500m

Bore: 50mm
Range: 1000m

Bore: 50mm
Range: 1500m

Bore: 50mm
Range: 2000m

Figure 122: Initial Study - Pareto Plots for 50 mm Bore, All Ranges

269

Bore: 57mm
Range: 250m

Bore: 57mm
Range: 500m

Bore: 57mm
Range: 1000m

Bore: 57mm
Range: 1500m

Bore: 57mm
Range: 2000m

Figure 123: Initial Study - Pareto Plots for 57 mm Bore, All Ranges

270

Kinetic-Kill/Pin-Control
Hidden Nodes: 5
R2: 0.98509
R2 (excluded): 0.98042

Warhead-Kill/Thruster-Control
Hidden Nodes: 6
R2: 0.96892
R2 (excluded): 0.93918

Kinetic-Kill/Unguided
Hidden Nodes: 3
R2: 0.88235
R2 (excluded): 0.69103

Figure 124: Detailed Study - Actual by Predicted and Residual Plots for Neural
Network Fits

271

REFERENCES

[1] Federation of American Scientists. http://www.fas.org. [Date accessed: June,


2005].
[2] Oerlikon Contraves. http://www.oerlikoncontraves.com.
June, 2005].

[Date accessed:

[3] Webster online dictionary. http://www.webster.com. [Date accessed: June,


2005].
[4] Risk Glossary. http://www.riskglossary.com. [Date accessed: June, 2005].
[5] Aerospace Systems Design Laboratory, School of Aerospace Engineering, Georgia Institute of Technology. http://www.asdl.gatech.edu/. [Date accessed:
July, 2006].
[6] ArrowTech Associates. http://www.prodas.com. [Date accessed: June, 2006].
[7] Small Low-cost Interceptor Device (SLID). Boeing Website.
http://www.boeing.com/defense-space/missiles/slid/slid.htm. [Date
accessed: January, 2006].
[8] Tactical High Energy Laser (THEL). MissileThreat.com, A Project of the
Claremont Institute, http://www.missilethreat.com/systems/thel.html.
[Date accessed: June, 2005].
[9] Survivability, Aircraft, Nonnuclear, General Criteria. Vol. 1, Dept. of Defense,
MIL-HDBK-336-1, Washington, DC, October 25, 1982.
[10] DARPA Neural Network Study, tech. rep., M.I.T. Lincoln Laboratory, Lexington, MA, 1988.
[11] USAF Intelligence Targeting Guide. Air Force Pamphlet 14-210, 1 February,
1998. http://afpubs.hq.af.mil. [Date accessed: January, 2006].
[12] Tactical High Energy Laser ACTD. Federation of American Scientists., November 1998. http://www.fas.org/spp/starwars/program/thel.htm. [Date
accessed: June, 2005].
[13] Battle Command Training Program, tech. rep., United States Army Training and Doctrine Command, Fort Monroe, Virginia, 1 April 1999. TRADOC
Regulation 350-50-3.

272

[14] MK
45
5-inch
/
54-caliber
(lightweight)
Federation
of
American
Scientists,
Nov
http://www.fas.org/man/dod-101/sys/ship/weaps/mk-45.htm.
accessed: January, 2006].

gun.
1999.
[Date

[15] Department of Defense Dictionary of Military and Associated Terms. Joint


Chiefs of Staff, U. S. Dept. of Defense, Joint Publication 1-02, April 2001.
www.dtic.mil/doctrine/jel/doddict/. [Date accessed: June, 2005].
[16] ASAR Product Handbook. Issue 1.2, European Space Agency, September
2004. http://envisat.esa.int/dataproducts/asar/CNTR.htm. [Date accessed: June, 2005].
[17] Mobile / Tactical High Energy Laser (M-THEL) Technology Demonstration
Program. Defense Update: An International Online Defense Magazine, May
7 2004. http://www.defense-update.com/directory/THEL.htm. [Date accessed: June,2005].
[18] THEL Mortar Shootdown. Northrop Grumman Press Release, 2004.
http://www.st.northropgrumman.com/media/VideoInfo.cfm?VideoID=34.
[Date accessed: June, 2005].
[19] AMSAA Supports Current Operations Through Counter Mortar Collateral
Damage Estimation. RDECOM Magazine, submitted by U.S. Army Materiel
Systems Analysis Activity, February 2005.
[20] AHEAD Air defense Gun System, Oerlikon Contraves (Switzerland). Defense Update: An International Online Defense Magazine, January 2005.
http://www.defense-update.com/products/a/ahead.htm. [Date accessed:
June, 2005].
[21] Pentagon Fast-Tracks Mortar Defense System. Fox News, March 29 2005.
http://www.foxnews.com/story/0,2933,151712,00.html. [Date accessed:
June, 2005].
[22] Green
Zone.
Global
Security,
June
http://www.globalsecurity.org/.../baghdad-green-zone.htm.
accessed: June, 2006].

2005.
[Date

[23] Counter Rocket, Artillery, and Mortar (C-RAM). Global Security, June 2005.
http://www.globalsecurity.org/military/systems/ground/cram.htm.
[Date accessed: June, 2005].
[24] Discussions with thesis committee member Dr. E. Jeff Holder, Chief Scientist,
Sensors and Electromagnetic Applications Laboratory, Georgia Tech Research
Institute, Jan-June 2006.

273

[25] Discussions with thesis committee member Dr. Kevin Massey, Senior Research
Engineer, Aerospace, Transportation, and Advanced Systems Laboratory, Georgia Tech Research Institute, Apr 2005 - June 2006.
[26] Quadrennial Defense Review Report. United States Department of Defense,
February 2006.
[27] Neural
Networks.
StatSoft,
Inc.,
Copyright
1984-2003.
http://www.statsoftinc.com/textbook/stneunet.html. [Date accessed:
June, 2005].
[28] Alexander, R., Hall-May, M., and Kelly, T., Characterisation of Systems of Systems Failures, Proceedings of the 22nd International System Safety
Conference, University of York, 2004.
[29] ArrowTech Associates, South Burlington, VT, PRODAS Version 3 Technical
Manual, 2002.
[30] Baker, A. P., Assessing the Simultaneous Impact of Requirements, Vehicle Characteristics, and Technologies During Aircraft Design, 39th Aerospace
Sciences Meeting & Exhibit, Reno, NV, 8-11 January, 2001. AIAA-2001-0533.
[31] Baker, A. P., The Role of Mission Requirements, Vehicle Attributes, Technologies and Uncertainty in Rotorcraft System Design. PhD thesis, Georgia
Institute of Technology, 2002.
[32] Ball, R. E., The Fundamentals of Aircraft Combat Survivability Analysis and
Design. AIAA, 2003.
[33] Barr, M., Closed-Loop Control, Embedded Systems Programming, pp. 55
66, August 2002.
[34] Barton, D. K., Modern Radar System Analysis. Artech House, 1988.
[35] Baumann, J., Systems Engineering, Session 1-2. Presented at the 2nd AIAA
Tactical Interceptor Technology Symposium, Huntsville, AL, 20-21 January,
2005.
[36] Benson, D., Ryan, B., and Drzewiecki, T., An Optofluidic Control System for Tactical Missiles and Guided Munitions, AIAA Guidance, Navigation,
and Control Conference, New Orleans, LA, 11-13 August, 1997. AIAA-19973834.
[37] Berglund, N., The Full Derivative Matrix & Jacobian. Calculus III Lecture Notes, Georgia Institute of Technology, Spring, 2006.
http://www.math.gatech.edu/\~berglund/teaching.html. [Date accessed:
April, 2006].

274

[38] Biltgen, P. T., Ender, T. R., and Mavris, D. N., Development of a Collaborative Capability-Based Tradeoff Environment for Complex System Architectures, 44th AIAA Aerospace Sciences Meeting and Exhibit, Reno, Nevada,
Jan. 9-12, 2006. AIAA-2006-0728.
[39] Bird,
J.,
Air and Missile Defense Transformation:
Where
We Were, Where We Are Today, and Where We Are Going.
Air Defense Artillery (ADA) Magazine Online,
January 2006.
http://airdefense.bliss.army.mil/.../WhereWeAreGoing.htm.
[Date
accessed: January, 2006].
[40] Blanchard, B. S., System Engineering Management. John Wiley & Sons,
Inc., 1991.
[41] Blanchard, B. S., Logistics Engineering and Management. Prentice Hall
Inc., 5th ed., 1998.
[42] Buede, D. M., The Engineering Design of Systems. John Wiley & Sons, Inc.,
2000.
[43] Burnett, D., Brekke, K., Locke, J., and Frederick, R. A., Shark
Workshop Report, Session 4-2. Presented at the 2nd AIAA Tactical Interceptor
Technology Symposium, Huntsville, AL, 20-21 January, 2005.
[44] Buzzett, J., Nixon, D., and Horwath, T., Investigation of Guidance
Technologies for Spinning Projectiles, presented at the 36th Annual Gun
& Ammunition Symposium & Exhibition, San Diego, CA, 9-12 April, 2001.
http://www.dtic.mil/ndia/2001gun/2001gun.html. [Date accessed: January, 2006].
[45] Cantor, M., Thoughts on Functional Decomposition. IBM online magazine
The Rational Edge, 2003.
[46] Cave, D. R., Systems Effectiveness Analysis Tool for Evaluating Guided
Interceptors, Aerospace Sciences Meeting and Exhibit, 32nd, Reno, NV, 10-13
January, 1994. AIAA-1994-210.
[47] Chapman, W. L., Bahill, A. T., and Wymore, A. W., Engineering Modeling and Design. CRC Press, 1992.
[48] Chen, P., Gori, R., and Pozgay, A., Systems and Capability Relation Management in Defence Systems-of-System Context, (Copenhagen, Denmark), Proceedings of The Ninth International Command and Control Research
and Technology Symposium, 2004.
[49] Clare, T. A., The Engineering and Acquisition of Systems of Systems, in
Proceedings from the 3rd Annual Systems Engineering & Supportability Conference, Department of Defense, Research Development and Acquisition Office,
Oct 2000.
275

[50] Clements, P., Krut, R., Morris, E., and Wallnau, K., The Gadfly:
An Approach to Architectural-Level System Comprehension, Fourth IEEE
Workshop on Program Comprehension, Berlin, March, 1996.
[51] Costello, M., Rigid Body Projectile Exterior, Session 3-2. Presented at the
2nd AIAA Tactical Interceptor Technology Symposium, Huntsville, AL, 20-21
January, 2005.
[52] Decisioneering, Inc., Denver, CO, Crystal Ball User Manual, 2004.
[53] Demuth, H. and Beale, M., MATLAB: Neural Network Toolbox Users
Guide Version 4. The MathWorks, Inc., 2004.
[54] Despotou, G., Alexander, R., and Hall-May, M., Key Concepts and
Characteristics of Systems of Systems, Tech. Rep. DARP-HIRTS Strand 2,
University of York, 2003.
[55] Dew, J., Computer Models: Current Uses and Future Requirements. Presentation given at Systems Integration Branch, Defense Threat Reduction Agency,
18 November, 2003.
[56] Dieter, G. E., Engineering Design: A Materials and Processing Approach.
McGraw Hill, 2000.
[57] Driels, M. R., Weaponeering: Conventional Weapon System Effectiveness.
Reston, VA: AIAA Education Series, 2004.
[58] Du, X. and Chen, W., Methodology for Managing the Effect of Uncertainty
in Simulation-Based Design, AIAA Journal, vol. 38, pp. 14711478, August
2000.
[59] Einstein, A., Collected Quotes from Albert Einstein. Copyright Kevin Harris, 1995.
[60] Ender, T. R., Mavris, D. N., Massey, K. C., Holder, E. J., OLeary,
P., and Smith, R. A., Predicting the Effectiveness of an Air Bursting Munition Using Uncertainty Quantification, Presented at the AIAA 2004 Missile
Sciences Conference, Monterey, CA, November 16-18, 2004.
[61] Ender, T. R., McClure, E. K., and Mavris, D. N., A Probabilistic
Approach to the Conceptual Design of a Ship-Launched High Speed Standoff
Missile, Presented at the AIAA 2002 Missile Sciences Conference, Monterey,
CA, November 5-7, 2002.
[62] Ender, T. R., McClure, E. K., and Mavris, D. N., Development of an
Integrated Parametric Environment for Conceptual Hypersonic Missile Sizing,
Presented at the 2nd AIAA ATIO Forum, Los Angeles, CA, October 1-3, 2002.

276

[63] Fleeman, E. L., Tactical Missile Design. Reston, VA: American Institute of
Aeronautics and Astronautics, 2001.
[64] Frame,
M.,
Mandelbrot,
B.,
and
Neger,
N.,
Fractal
Geometry.
Yale
University,
April
2005.
http://classes.yale.edu/fractals/CA/NeuralNets/NeuralNets.html.
[Date accessed: June, 2005].
[65] Gajda, W. J. and Biles, W. C., Engineering Modeling and Computation.
Boston: Houghton Mifflin Company, 1978.
[66] Geswender, C., Guided Projectiles Theory of Operations, presented at the
36th Annual Gun & Ammunition Symposium & Exhibition, San Diego, CA,
9-12 April, 2001. http://www.dtic.mil/ndia/2001gun/2001gun.html. [Date
accessed: January, 2006].
[67] Gompert, D. and Isaacson, J., Planning a Ballistic Missile Defense System
of Systems: An Adaptive Strategy, Tech. Rep. Paper IP-181, RAND National
Defense Research Institute, 1999.
[68] Group W Inc., Fairfax, VA, THUNDER: Multi-Service Campaign Analysis Tool,
http://www.group-w-inc.com/. [Date accessed: June, 2005].
[69] Hartlage, B., Owen, M., Dimeler, H., Robert A. Frederick, J., and Davis, C., Enhanced Counter Air Projectile, 40th
AIAA/ASME/SAE/ASEE Joint Propulsion Conference and Exhibit, Fort
Lauderdale, Florida, 11-14 July, 2004. AIAA 2004-4086.
[70] Hayter, A. J., Probability and Statistics for Engineers and Scientists. Boston:
PWS Publishing Company, 1996.
[71] Heller, A., Simulation Warfare Is No Video Game, Science & Technology
Review, pp. 411, January/February 2000. National Nuclear Security Administrations Lawrence Livermore National Laboratory.
[72] Hewitt, D. J., Notional System - Shark I, Session 4-1. Presented at the
2nd AIAA Tactical Interceptor Technology Symposium, Huntsville, AL, 20-21
January, 2005.
[73] Hines, N. R. and Mavris, D. N., A Parametric Design Environment for
Including Signatures Analysis in Conceptual Design, 2000. AIAA-2000-015564.
[74] Hines, W. W. and Montgomery, D. C., Probability and Statistics in Engineering and Management Science. John Wiley & Sons, Inc., 3rd ed., 1990.
[75] Holder, E. J., Fire Control Radar Requirements for RAM Defense, Session
2-5. Presented at the 2nd AIAA Tactical Interceptor Technology Symposium,
Huntsville, AL, 20-21 January, 2005.
277

[76] Holder, E. J., Command vs Semi-Active Projectile Guidance. Review Presentation for U.S. Army Customers, Georgia Tech Research Institute, Atlanta,
GA, 2006.
[77] Hollis, P. S., Transforming the Force - From Korea to Today. Robert H.
Scales, Jr. - Major General Retired - Interview, Field Artillery Journal, July
2001.
[78] Horwath, T. G. and Barnych, G., Low Cost Course Correction Cost
Course Correction (LCCC) Technology Applications, NDIA International Armaments Technology International Armaments Technology Symposium, 14-16
June, 2004.
[79] Hughes, Wayne P., J., Military Model for Decision Makers. Military Operations Research Society, 1997.
[80] Huyse, L., Solving Problems of Optimization Under Uncertainty as Statistical
Decision Problems, in AIAA-2001-1519, 2001.
[81] Johnson, C., Introduction to Neural Networks in JMP, Version 2.0. May
2004.
[82] Johnson, C., Function Approximating Neural Network Generation System
(FANNGS) Documentation Version 1.2. June 2005.
[83] Kapper, F. B., Wargaming, Defense, 1981. Department of Defense.
[84] Kirby, M. R. and Mavris, D. N., Forecasting the Impact of Technology
Infusion on Subsonic Transport Affordability, 1998. AIAA-98-5576.
[85] Kirby, M. R. and Mavris, D. N., Forecasting Technology Uncertainty in
Preliminary Aircraft Design, 1999. AIAA-99-01-5631.
[86] Knott, E. F., Tactical Missile Aerodynamics. AIAA Tactical Missile Series,
1992.
[87] Koller, R., Konstruktionslehre fuer den Maschinenbau. Berlin: SpringerVerlag, 1985.
[88] Kuhne, C., Wiggs, G., Beeson, D., Madelone, J., and Gardner, M.,
Using Monte Carlo Simulation for Probabilistic Design, Proceedings of the
2005 Crystal Ball User Conference, 2005.
[89] Lanchester, F. A., Aircraft in Warfare: The Dawn of the Fourth Arm.
London: Constable, 1916.
[90] Lloyd, R. M., Conventional Warhead Systems Physics and Engineering Design, vol. 179 of Progress in Astronautics and Aeronautics Series. AIAA, 1998.

278

[91] Lloyd, R. M., Physics of Direct Hit and Near Miss Warhead Technology,
vol. 194 of Progress in Astronautics and Aeronautics. Reston, VA: AIAA, 2001.
[92] Macfadzean, R. H. M., Surface-Based Air Defense System Analysis. Boston:
Artech House, 1992.
[93] Maier, M. W., Architecting Principles for Systems of Systems, in Proceedings of the Sixth Annual International Symposium, International Council on
Systems Engineering, Boston, MA, 1996.
[94] Maier, M. W., Integrated Modeling: A Unified Approach to System Engineering, Journal of Systems and Software, vol. 32, Feb 1996.
[95] Maier, M. W. and Rechtin, E., The Art of Systems Architecting. Boca
Raton, FL: CRC Press, 2nd ed., 2000.
[96] Martin,
S.
G.,
Phalanx Block 1B CIWS: Ready For
The Fleet. Defense Technical Information Center,
May 2000.
http://www.dtic.mil/ndia/ammo/martin.pdf [Date accessed: June, 2005].
[97] Marx, W. J., Mavris, D. N., and Schrage, D. P., A Hierarchical Aircraft
Life Cycle Cost Analysis Model, 1st AIAA Aircraft Engineering, Technology,
and Operations Conference, Los Angeles, CA, September 19-21, 1995. AIAA95-3861.
[98] Marx, W. J., Mavris, D. N., and Schrage, D. P., Effects of Alternative
Wing Structural Concepts on High Speed Civil Transport Life Cycle Costs,
1996. AIAA-96-1381.
[99] Massey, K. C., McMichael, J., Warnock, T., and Hay, F., Design and
Wind Tunnel Testing Of Guidance Pins For Supersonic Projectiles, Presented
at the 24th Army Sciences Conference, 2004. CO-04.
[100] Massey, K. C. and Silton, S. I., Testing the Maneuvering Performance
of a Mach 4 Projectile, Presented at the 24th AIAA Applied Aerodynamics
Conference, San Francisco, California, June 5-8,, 2006. AIAA-2006-3649.
[101] Massey, K., Heiges, M., DiFrancesco, B., Ender, T., and Mavris,
D., The Use of Simulations and System-of-Systems Design to Investigate a
Guided Projectile Mortar Defense System, Presented at the 36th AIAA Fluid
Dynamics Conference and Exhibit, 24th Applied Aerodynamics Conference,
San Francisco, CA, June 5-8, 2006. AIAA-2006-3652.
[102] Massey, K., Holder, J., Ender, T., and Heiges, M., 6-DOF and Guided
Projectile System Requirements. Review Presentation for U.S. Army Customers, Georgia Tech Research Institute, Atlanta, GA, February 2006.
[103] The MathWorks, Inc., MATLAB: Statistics Toolbox Users Guide Version 5,
2004.
279

[104] Mavris, D. N., Baker, A. P., and Schrage, D. P., Implementation of a


Technology Impact Forecast Technique on a Civil Tiltrotor, Presented at the
American Helicopter Society 55th Annual Forum, Montreal, Quebec, Canada,
May 25-27, 1999. AHS-99-AB.
[105] Mavris, D. N. and DeLaurentis, D. A., A Stochastic Design Approach
for Aircraft Affordability, Presented at the 21st Congress of the International
Council on the Aeronautical Sciences (ICAS), Melbourne, Australia, September,
1998. ICAS-98-6.1.3.
[106] Mavris, D. N., DeLaurentis, D. A., Bandte, O., and Hale, M. A.,
A Stochastic Approach to Multi-disciplinary Aircraft Analysis and Design,
Presented at the 36th Aerospace Sciences Meeting and Exhibit, Reno, NV,
January 12-15, 1998. AIAA-98-0912.
[107] Mavris, D. N., DeLaurentis, D. A., Hale, M. A., and Tai, J. C., Elements of an Emerging Virtual Stochastic life Cycle Design Environment, Presented at the 4th World Aviation Congress and Exposition, San Francisco, CA,
October 19-21, 1999.
[108] Mavris, D. N. and Garcia, E., Affordability Assessment for a Subsonic
Transport, Presented at the 2nd Joint ISAP/SCEA International Conference,
San Antonio, TX, June, 1999.
[109] Mavris, D. N., Kirby, M. R., and Qiu, S., Technology Impact Forecasting
for a High Speed Civil Transport, 1998. AIAA-98-5547.
[110] Mavris, D. N., Biltgen, P. T., Ender, T. R., and Cole, B., Technology Assessment and Capability Tradeoff Using Architecture-Based Systems
Engineering Methodologies, 1st International Conference on Innovation and
Integration in Aerospace Sciences, Queens University Belfast, Northern Ireland,
UK, 4-5 August, 2005. CEIAT 2005-0048.
[111] Mavris, D., Soban, D., and Largent, M., An Application of a Technology
Impact Forecasting (TIF) Method to an Uninhabited Combat Aerial Vehicle,
1999. AIAA-1999-01-5633.
[112] McCullers, L. A., FLOPS Users Guide Ver. 5.94. NASA Langley Research
Center, January 1998.
[113] MIL-STD-499A, Engineering Management, May 1974. Military Standard
published by the Department of Defense.
[114] Miller, D., The Illustrated Directory of Modern American Weapons. Salamander Books Ltd., 2002.
[115] Montgomery, D. C., Design and Analysis of Experiments. New York: John
Wiley & Sons, Inc., 1976.

280

[116] Myers, R. H. and Montgomery, D. C., Response Surface Methodology:


Process and Product Optimization Using Designed Experiments. New York:
John Wiley & Sons, Inc, 1995.
[117] Naval Air Systems Command, Avionics Department, AIR-4.5, EW Class Desk,
Washington, D.C. 20361, Electronic Warfare and Radar Systems Engineering
Handbook, April 1999. NAWCWPNS-TP-8347.
[118] Nebabin, V. G., Methods and Techniques of Radar Recognition. Boston:
Artech House, 1995.
[119] Nelson, R. C., Flight Stability and Automatic Control. McGraw-Hill, 2nd ed.,
1998.
[120] Nourse, B., Extended Area Protection and Survivability, Session 1-1.
Presented at the 2nd AIAA Tactical Interceptor Technology Symposium,
Huntsville, AL, 20-21 January, 2005.
[121] Owen, M., Experimental Design and System Analysis using the Response
Surface Method, Session 4-3. Presented at the 2nd AIAA Tactical Interceptor
Technology Symposium, Huntsville, AL, 20-21 January, 2005.
[122] Pahl, G. and Beitz, W., Engineering Design. London: Springer-Verlag, 1996.
[123] Papp, D. S., Johnson, L. K., and Endicott, J., American Foreign Policy:
History, Politics, and Policy. Pearson Education, Inc., 2005.
[124] Parsch,
A.,
Raytheon ERGM (5 Projectile MK
Directory
of
U.S.
Military
Rockets
and
Missiles,
http://www.designation-systems.net/dusrm/app4/ergm.html.
accessed: January, 2006].

171).
2003.
[Date

[125] Paterson, J., Overview of Low Observable Technology and its Effects on
Combat Aircraft Survivability, Journal of Aircraft, vol. 36, March-April 1999.
[126] Przemieniecki, J., Mathematical Methods in Defense Analysis.
3rd ed., 2000.

AIAA,

[127] SAS Institute Inc., Cary, NC, JMP, Computer Program and Users Manual,
1994.
[128] Scharl, J. and Mavris, D., Building Parametric and Probabilistic Dynamic
Vehicle Models Using Neural Networks, Presented at the AIAA Modeling and
Simulation Conference and Exhibit, Montreal, Canada, August 6-9, 2001. AIAA
2001-4373.
[129] Science Applications International Corporation, San Diego, CA, Integrated Theater Engagement Model (ITEM), 1999.

281

[130] Shonkwiler, R., Math 4225 Class Notes. School of Mathematics, Georgia
Institute of Technology, 2004.
[131] Silton, S. I., Comparison of Predicted Actuator Performance for Guidance
of Supersonic Projectiles to Measured Range Data, 2004. AIAA-2004-5195.
[132] Silton, S. I. and Massey, K. C., Integrated numerical and experimental
investigation of actuator performance for guidance of supersonic projectiles,
Army Sciences Conference, 2004.
[133] Skolnik, M. I., Introduction to Radar Systems. McGraw-Hill Publishing Company, 1980.
[134] Soban, D. S., A Methodology for the Probabilistic Assessment of System Effectiveness as Applied to Aircraft Survivability and Susceptibility. PhD thesis,
Georgia Institute of Technology, 2001.
[135] Tidd, J., Bessant, J., and Pavitt, K., Managing Innovation: Integrating Technological, Market, and Organizational Change.
John Wiley and Sons, 2nd ed., 2001. Online Innovation Management Toolbox
http://www.wiley.co.uk/.../forecastingtools.htm [Date accessed: June,
2005].
[136] Tournes, C., Shark-Flight Control, Session 4-5. Presented at the 2nd AIAA
Tactical Interceptor Technology Symposium, Huntsville, AL, 20-21 January,
2005.
[137] Twiss, B. C., Forecasting for Technologists and Engineers: A Practical Guide
for Better Decisions. London: Peter Peregrinus, Ltd., 1992.
[138] United States Department of Defense (DOD), Department of the Army, U.S. Army Aviation and Missile Command (AMCOM), Extended Area Protection and Survivability Architecture Study. U.S. Government Printing Office FedBizOpps, Research and Development Procurement, September 2004.
http://fedbizopps.cos.com/cgi-bin/getRec?id=20040913a5.
[Date
accessed: January, 2006].
[139] Vanderplaats, G. N., Numerical Optimization Techniques for Engineering
Design. Colorado Springs, CO: Vanderplaats Research & Development, Inc.,
3rd ed., 1999.
[140] Weisstein, E. W., Monte Carlo Method. MathWorldA Wolfram Web Resource, http://mathworld.wolfram.com/MonteCarloMethod.html. [Date accessed: June, 2005].
[141] Weisstein, E. W., Total derivative. From MathWorldA Wolfram Web Resource, http://mathworld.wolfram.com/TotalDerivative.html. [Date accessed: April, 2006].
282

[142] Whyte, J., Projectile Rocket Ordnance Design and Analysis Software (PRODAS Version 3) Capabilities - Part 1, Session 3-3. Presented at the 2nd AIAA
Tactical Interceptor Technology Symposium, Huntsville, AL, 20-21 January,
2005.
[143] Whyte, J., Projectile Rocket Ordnance Design and Analysis Software (PRODAS Version 3) Capabilities - Part 2, Session 3-4. Presented at the 2nd AIAA
Tactical Interceptor Technology Symposium, Huntsville, AL, 20-21 January,
2005.
[144] Williams, J., Brekke, K., Patrick, S., Crosswy, R., Hahna, P.,
and Frederick, R. A., Conceptual Design of a Guided Interceptor, 41st
AIAA/ASME/SAE/ASEE Joint Propulsion Conference & Exhibit, Tucson,
Arizona, 10-13 July, 2005.
[145] Zang, T. A., Hemsch, M. J., Hilburger, M. W., Kenny, S. P., Luckring, J., Maghami, P., Padula, S. L., and Stroud, W. J., Needs
and Opportunities for Uncertainty-Based Multidisciplinary Design Methods
for Aerospace Vehicles, Tech. Rep. TM-2002-211462, NASA Langley Research
Center, Hampton, VA, 2002.

283

VITA

Tommer Ender was born on September 28, 1978 in Englewood, NJ. He was raised in
South Florida where he attended Cooper City High School, graduating in 1996. He
then went on to enroll at the Aerospace Engineering program at the Georgia Institute
of Technology in Atlanta, GA, where he earned a Bachelors degree in 2001. He worked
as an undergraduate research assistant at the Aerospace Systems Design Laboratory
through which he eventually earned a Masters degree in 2002, and subsequently began
pursuing a Ph.D. His professional experience includes working as an undergraduate
cooperative student at Pratt & Whitney Government Engines and Space Propulsion in
West Palm Beach, FL, and has had the opportunity to work with various government
entities during his graduate school tenure. Beyond his technical experience, Mr.
Ender was selected as a Pre-Doctoral Fellow for the 2004-2005 Sam Nunn Security
Program, through the Sam Nunn School of International Affairs at Georgia Tech.

284

You might also like