Computer-Aided Civil and Infrastructure Engineering 22 (2007) 282–292

Reliability-Based Optimal Design of Electrical

Transmission Towers Using Multi-Objective
Genetic Algorithms
Sachin Mathakari & Paolo Gardoni∗
Zachry Department of Civil Engineering, Texas A&M University, College Station, TX 77843-3136

Pranab Agarwal
RISA Technologies, Foothill Ranch, CA 92610

Anne Raich
Department of Civil and Environ. Engineering, Lafayette College, Easton, PA 18042

&

Terje Haukaas
Department of Civil Engineering, University of British Columbia, Vancouver, BC, V6T 1Z4, Canada

Abstract: A hybrid methodology for performing analysis. Both the wind pressure and the wind direction
reliability-based structural optimization of three- are considered as random variables in the analysis.
dimensional trusses is presented. This hybrid method- The research results presented demonstrate the benefit
ology links the search and optimization capabilities of implementing MOGA optimization as an integral
of multi-objective genetic algorithms (MOGA) with part of a reliability-based optimization procedure for
structural performance information provided by finite three-dimensional trusses.
element reliability analysis. To highlight the strengths
of the proposed methodology, a practical example
is presented that concerns optimizing the topology, 1 INTRODUCTION
geometry, and member sizes of electrical transmission
towers. The weight and reliability index of a tower are Safety and economy are the primary concerns dur-
defined as the two objectives used by MOGA to perform ing the design of any civil engineering structure, be-
Pareto ranking of tower designs. The truss deformation cause the designer, the owner, and the public prefer
and the member stresses are compared to threshold a structure that is safe and cheap. Reliability-based
values to assess the reliability of each tower under wind optimization helps designers in identifying a good bal-
loading. Importance sampling is used for the reliability ance between the safety and economy of a struc-
ture. Although theoretically appealing, the optimization
∗ Towhom correspondence should be addressed. E-mail: gardoni@ of practical structures has several computational and
tamu.edu. modeling challenges. The reliability-based optimization


C 2007 Computer-Aided Civil and Infrastructure Engineering. Published by Blackwell Publishing, 350 Main Street, Malden, MA 02148, USA,
and 9600 Garsington Road, Oxford OX4 2DQ, UK.
 Reliability-based optimal design of electrical transmission towers 283

problem domain is complex, which requires the appli- been proposed by Fonseca and Fleming (1993), Horn
cation of robust search and optimization techniques to et al. (1994), and Srinivas and Deb (1994). The con-
arrive at a near-global optimal solution. Traditional op- cept of Pareto-optimality to capture the trade-off be-
timization methods, which include calculus-based, enu- tween multiple objectives has gained prominence in the
merative or hill-climbing algorithms, work best in find- last decade, although the basic concept was developed
ing a single, locally optimal solution, and are therefore as early as 1906 by Pareto. Cheng and Li (1998) used
ineffective in discontinuous and multi-modal search do- a multi-objective GA (MOGA), which implements a
mains. In comparison, genetic algorithms (GAs) perform Pareto ranking procedure for selection, to address the
a more global, probabilistic directed search, which re- trade-off between weight and strain energy objectives
sults in a greater likelihood of obtaining near-global opti- in the optimization of a 72-bar space truss. Deb (2001)
mal solutions when searching complex problem domains. provides a comprehensive review on the application and
Since their original development by John Holland implementation of MOGA.
(1975) and his colleagues at the University of Michigan, The method proposed in this research uses MOGA
GAs have been used in various fields to find solu- to search for and optimize transmission tower designs.
tions to complicated optimization problems. Rajeev and In any problem where multiple objectives are defined,
Krishnamoorthy (1992) and Jenkins (1992) were among trade-offs exist in the ability to satisfy the objectives.
the first researchers to implement simple GAs in struc- In this research, a tradeoff exists between minimizing
tural optimization problems. The former explored the the total weight of the truss and maximizing the re-
application of GAs to optimize the cross-sectional ar- liability of the truss. MOGA relies on the use of the
eas of truss members, although the latter used GAs to concept of Pareto-optimality to perform multi-objective
optimize the member sizes for plane frames. In both re- optimization. MOGA works with a population of design
search efforts, the fitness of each design was assessed by alternatives and uses the concept of Pareto-optimality
considering how well the objectives were met and any to determine a ranking of current design alternatives
constraint violations were handled by imposing a fitness in order to guide the search process toward an opti-
penalty. mal solution. Solutions are identified as nondominated
Thampan and Krishnamoorthy (2001) performed or dominated by other solutions in the population. The
reliability-based structural optimization of trusses us- nondominated solutions, which are solutions that are
ing a simple GA. Member section sizes and the ge- considered to be equally good, define a Pareto-optimal
ometric truss configuration were considered as design front. A set of Pareto-optimal designs can be evolved by
variables for producing optimal trusses. Thampan and MOGA in a single trial and can be used to clearly define
Krishnamoorthy defined a fitness function with a single the tradeoffs occurring in meeting the stated objectives.
objective of minimizing weight and used normalized re- MOGA does not require assigning weights to each ob-
liability indices as penalties for performing selection in jective as is required to construct single fitness functions,
subsequent generations. In their research the reliability which avoids the solution evolved being dependent on
of a truss is estimated using upper bounds. Dimou and the weights selected.
Koumousis (2003) also evaluated the system reliability In this article, a new approach for the reliability-based
using bounds on the probability of failure. Tolson et al. optimal design of structures that combines MOGA opti-
(2004) optimized a water distribution network by link- mization with finite element reliability analysis is pro-
ing a simple GA with a first order reliability method posed. In general, FORM or second order reliability
(FORM). Instead of using reliability bounds, a single re- method (SORM) may not provide an accurate esti-
liability measure was imposed as a penalty on the cost mate of the system reliability. Therefore in this re-
optimization function to optimize overall network per- search, importance sampling is utilized to perform the
formance. Adeli and Cheng (1993, 1994a, 1994b), Adeli reliability analysis. An application of the proposed ap-
and Kumar (1995a, 1995b), and Adeli and Sarma (2006), proach is presented that optimizes the design of elec-
applied GA to large realistic structures. Adeli and Hung trical transmission towers subjected to wind loading.
(1995) provided a comprehensive review on the applica- The topology, geometry, and member sizes of the tower
tion and implementation of GA. are optimized while keeping the height of the tower
During the last few decades, many different ap- constant.
proaches to multi-objective optimization have been in-
vestigated by researchers. Some of the early attempts to
consider multiple objectives include VEGA (Schaffer, 2 MULTI-OBJECTIVE GENETIC ALGORITHMS
1985), lexicographic ordering (Rao, 1984), and weighted
sum objective functions (Hajela and Lin, 1992). Several A GA is classified as a heuristic optimization method that
different multi-objective GA implementations have works with a population of individual solutions. Genetic
284 Mathakari et al.

algorithms “combine survival of the fittest among string fq (x) ≤ fq (x∗ ) for all q ∈ {1, 2, . . . , Q} (1)
structures with a structured yet randomized informa-
tion exchange to form a search algorithm with some fq (x) < fq (x∗ ) for at least one q ∈ {1, 2, . . . , Q} (2)
of the innovative flair of human search” (Goldberg,
where, f q (x) is the objective function for the Q objective.
1989). In essence, it is a computational technique that
A rank of 1 is assigned to all nondominated individuals
bases its probabilistic operations on natural evolution
identified in the population. Rank 1 individuals form a
and Darwin’s survival of the fittest theory. GAs provide
trade-off surface that is commonly known as the Pareto-
a balance between exploration for better solutions and
optimal surface. For a two-objective minimization prob-
exploitation of the good solutions found. The main bene-
lem, Figure 1 shows an example of the Pareto-optimal
fit of using GAs over other optimization techniques is its
surface defined by a set of nondominated individuals.
ability to escape local optima and attain the near-global
The ranking procedure used by MOGA continues by
optimum.
excluding Rank 1 individuals from consideration and
The basic operations performed by GAs are the gen-
identifying the set of nondominated individuals in the
eration and evaluation of an initial population. The evo-
remaining population, which are assigned a rank of 2.
lution begins with a randomly generated population of
This process continues until all individuals in the popula-
individuals and advances in generations. In each gener-
tion have been assigned a rank (Srinivas and Deb, 1994).
ation, the fitness of the whole population is evaluated
Tournament selection is then performed to identify pairs
and individuals are selected from the current popula-
of individuals to undergo crossover and mutation. A ran-
tion using selection criteria based on their fitness. The
dom subset of n-individuals from the population com-
selected individuals are then modified using the genetic
pete in each tournament. The winner of a tournament is
operations of crossover and mutation to form a new pop-
the individual with the lowest rank. In case of a tie, one
ulation, which becomes current in the next iteration of
of the individuals is selected at random.
the algorithm. In GAs, design variables are encoded into
Crossover and mutation are then applied probabilisti-
character strings and these strings are combined to form
cally to the pairs of selected individuals and a new pop-
an individual string representing a design solution. This
ulation is defined. During crossover, multiple locations
individual string is a genetic representation of the design
are randomly selected along the binary bits encoded by
solution.
each individual and the sections of strings are swapped
To perform topology and geometry optimization, each
to generate two children as shown in Figure 2. Muta-
individual encodes a set of topology and geometry de-
tion then is performed by randomly flipping selected en-
sign variables associated with the truss. For sizing opti-
coded bits in randomly selected individuals as shown in
mization in this research, each individual also encodes a
Figure 3. Mutation helps prevent information from be-
member cross-section design variable for each member
ing permanently lost and also introduces new informa-
in the truss based on the set of member cross-sections
tion to the search process. An elitism strategy is adopted
defined. These design variables are encoded as binary
numbers in each individual.
A set of integer fitness values f q (·) are assigned to Non-dominated
each individual in the population that describe how f1
Points (Rank 1)
well that individual satisfies the Q stated objectives. An
individual’s fitness plays an important role in determin-
ing whether an individual will be selected for reproduc-
tion in the next generation. In simple GAs, often the
fitness function is highly sensitive to its formulation and
has to be optimized in order to produce good results. In
MOGA, selection of individuals from the current pop-
ulation is performed by applying the concept of Pareto
optimality to the population and assigning each individ-
ual a rank based on their fitness values. The Eschenauer
(1990) criterion is used to evaluate the nondominated
points and can be stated as follows for the minimization
problem: f2
A feasible solution x∗ is a nondominated (Pareto,
1906) optimum if and only if there exist no feasible vector Fig. 1. Illustration of applying Pareto ranking to a population
x such that to identify the nondominated front.
 Reliability-based optimal design of electrical transmission towers 285

Pf, jk = P[Cjk (x) ≤ Djk (x)] = P[Cjk (xd , xr ) ≤ Djk (xd , xr )]

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
Parents

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 = P[gjk (xd , xr ) ≤ 0] = fXr (xr )dxr
gjk (xd ,xr )≤0
Location 1 Location 2 Location 3

j = 1, . . . , J and k = 1, . . . , K (3)
1 1 1 0 0 0 0 1 1 1 1 1 1 0 0 0
Children
where, fXr (xr ) denotes the joint probability density func-
0 0 0 1 1 1 1 0 0 0 0 0 0 1 1 1
tion of the random variables xr .
In our application, we consider the following limit
Fig. 2. Illustration of multi-point crossover operation.
states:
r Serviceability limit state: Monitors the maximum lat-
Before Mutation 1 1 1 1 1 1 11 1 1 1 1 1 1 1 1 1
eral displacement of the top of the structure due to
external loads.
After Mutation 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 r Ultimate limit state: Monitors the maximum axial
stress in each member due to external loads.
Fig. 3. Illustration of mutation operation.

3.2 Assessment of structural system reliability

in which the first five individuals are chosen from the In this article, the failure of the system occurs when any
Pareto-optimal set and copied directly into the next gen- of its components fail either in the ultimate limit state
eration. This helps to prevent the loss of fitter individuals or in the serviceability limit state. However, it is noted
from the population. that, for statically redundant structures, a single element
The parameter values used by GAs, such as the num- failure does not necessarily imply that the structure has
ber of elite members, crossover rate and mutation rate, failed. The probability of failure Pf of such a system in
are problem dependent and are typically set by a trial terms of limit state functions gjk (x) is expressed as
and error process. In the MOGA trial performed in this  
research, the following parameter settings were used: J K
Pf = P gjk (xd , xr ) ≤ 0 (4)
population size of 200, string length of individual is 170,
j=1 k=1
tournament size of 5, crossover over rate of 0.8, muta-
tion rate of 0.004, and the number of elite individuals The generalized reliability index (Ditlevsen, 1979) is
selected was five. defined as
β = −1 (1 − Pf ) (5)
where, −1 (·) denotes the inverse of the standard normal
3 STRUCTURAL RELIABILITY cumulative probability function.
3.1 Assessment of structural component reliability
3.3 Reliability analysis methods
Failure of a structural component j can be expressed
in terms of its demand measure Djk (x) and its capac- Given that xr are common to each limit state function
ity measure Cjk (x), where k ranges over all the possible in Equation (4), the failure events are correlated and a
modes of failure, K, (e.g., failure due to excessive stress closed-form solution is not available. Several methods
or deflection). Djk (x) and Cjk (x) are functions of a set of can be used for an approximate solution, including
basic variables x of the model, which include the geom- FORM, SORM, Monte Carlo simulations, importance
etry, material properties, and loads. The variables x can sampling, response surface methods, and numerical in-
be partitioned into deterministic variables, xd , and ran- tegration schemes (Ditlevesen and Madsen, 1996). The
dom variables, xr . Following the conventional notation present case is a series system problem, for which FORM
in structural reliability theory (Ditlevsen and Madsen, analysis combined with bounding formulas constitutes
1996), component failure is described using a limit state one feasible alternative. Another approach is sampling
function gjk (x). The limit state function is defined such analysis, although mean-centered Monte Carlo sampling
that {gjk (x) ≤ 0} indicates a failure of component j in tends to be excessively computationally expensive in
mode k. Probability of failure P f, jk of a component j in the presence of finite element computations to eval-
mode k is given as uate the limit state function. In this study we select
286 Mathakari et al.

to determine the failure probability by a two-step ap- threshold values to assess the reliability of each tower
proach. First, the design point is determined for the under wind loading.
most critical limit state function. Subsequently, an effi-
cient importance sampling centered at this design point 4.1 Structural configuration
is carried out. This approach is an improvement over The structural configuration of a truss is governed by
the traditional Monte Carlo simulations. The approach is its topology, geometry, and member section properties.
particularly well suited when high correlation is present A typical three-dimensional three-story truss is shown
between the limit state functions. in Figure 4. The details of structural configuration used
in the geometry, topology, and sizing optimization are
3.4 Importance measures discussed below:

There may be several random variables in a limit state. r Geometry optimization: A three-dimensional, sym-
Some random variables have a larger effect on the vari- metric truss is defined. The basic geometry of the truss
ance of the limit state (and thus are more important) and is identified by the bottom width w b , top width w t , and
some have a smaller effect (and thus are less important). the height of the truss h. The trusses are generated so
Following Der Kiureghian and Ke (1985), the measure that wt ≤ w b and h is kept constant, that is, 15 m. The
of importance γ can be defined as wb is selected from a pool that ranges from 2.25 m to
6.0 m, with a step size of 0.25 m. w t is selected from a
αT Ju∗ ,xr∗ SD
γT =  T  (6) pool that ranges from 1 m to 6 m, with a step size of
α Ju∗ ,x∗ SD 
r 0.25 m.
r Topology optimization: The topology of the truss is
where, α is the negative normalized gradient to the limit
state surface at the design point in the standard nor- altered by changing the number of stories, N s , and the
mal space. This vector is also an importance measure height from the ground of each internal floor zr , r =
for the random variables in the case of no correlation 1, . . . , (N s − 1). N s varies from 1 to 6 and zr is selected
although γ is the importance vector for the case of from a pool of values ranging from 0.5 m to 14.5 m
correlated random variables. If variables are uncorre- with a step size of 0.5 m The horizontal members of
lated then this vector is equal to α vector. Ju∗ ,xr∗ is the each internal floor can also be randomly removed.
Jacobian of the probability transformation from the orig- r Size optimization: The discrete vertical, horizontal,
inal space xr to the standard normal space u, with re-
spect to the coordinates of the design point xr∗ . SD
is and bracing members at each floor are grouped
the standard deviation matrix of equivalent normal vari-
ables z defined by the linearized inverse transformation
xr = xr∗ + Jxr∗ ,u∗ (u − u∗ ) at the design point. The elements wt
of SD
are the square roots of the corresponding diag-
onal elements of the covariance matrix Σ = Jxr∗ ,u∗ JxTr∗ ,u∗
of the variables xr . 3rd story
Importance measures are used to reduce the number
of random variables by considering less important ran-
dom variables as deterministic. This selective simplifica-
tion is used to reduce the analysis time required for the
importance sampling. 2nd story
h=15.0 m
z
4 DESCRIPTION OF TRUSS TOWER
PROBLEM DOMAIN
1st story
To highlight the strengths of the proposed methodology,
a practical example is presented that concerns optimizing
the topology, geometry, and member sizes of electrical
transmission towers. The weight and reliability index of
wb
a tower are defined as the two objectives used by MOGA
to perform Pareto ranking of tower designs. The truss
deformation and the member stresses are compared to Fig. 4. Typical three-dimensional three-story truss.
 Reliability-based optimal design of electrical transmission towers 287

Table 1
List of available member sections based on the AISC LRFD
Manual (2001) 0.5Dm 2 ( z )
Truss
Dm 2 ( z )
W4 × 13 W5 × 16 W5 × 19 W6 × 9
W6 × 16 W6 × 25 W8 × 10 W8 × 15
W8 × 18 W8 × 21 W8 × 24 W8 × 28 θ
W8 × 31 W10 × 12 W10 × 15 W10 × 17 0.5Dm1 ( z )
W10 × 19 W10 × 30 W10 × 33 W10 × 39
W10 × 49 W12 × 14 W12 × 22 W12 × 26 Dm1 ( z )
W12 × 40 W12 × 50 W14 × 34 W14 × 38
W14 × 43 W14 × 48 W16 × 36 W16 × 45 Wind Direction

Group 6 Fig. 6. Wind loading distribution.

Group 9 where δ(z) is the solidity of tower face defined as a ratio

Group 3 of area of member sections to the area projected by the
Group 5 members, w(z) is the width of the truss at height z, Cd is
the drag force coefficient assumed to be 2.2, and pm (z)
is the pressure distribution given by
Group 8 
Group 2   z 2α
pm(z) = 0.5ρ ūh 2
+ σε (8)
Group 4 h
where ρ is the air density assumed to be 1.225 Kg/m3 ,
ūh is the mean annual maximum wind speed, assumed as
Group 7 nominal 3-second gust wind speed according to ASCE
Group 1
(2002). The value for the mean wind speed is assumed to
be 60 m/s, α is the power law exponent that depends on
the surface roughness assumed to be 0.28, σ represents
the standard deviation of the pressure such that the co-
Fig. 5. Group assignments for member sizes for an example efficient of variation of pm (z) is 30%, and ε is a random
truss. variable with zero mean and unit standard deviation.
The direction of the wind identified by the angle θ
separately. The member section properties are as- is considered a random variable with uniform distribu-
signed to groups in such a way that heavier sections tion from 0◦ to 90◦ (see Figure 6). The drag forces on
are available for selection for the bottom levels and the two faces of the transmission tower affected by the
lighter sections are available for selection at the top wind, called Dm1 (z, θ) and Dm2 (z, θ ), are a function of z
levels. The available member sections are selected and θ.
from the AISC LRFD Manual (2001) and are listed
in Table 1. As shown in Figure 5, nine groups of truss 4.3 Assessment of structural reliability
members are defined for the three story truss. Groups
1 to 3 are assigned to vertical members, Groups 4 to The finite element platform OpenSees (Haukaas and
6 are assigned to horizontal members, and Groups 7 Der Kiureghian, 2004; McKenna et al., 2005; McKenna
to 9 are assigned to the bracing members. and Fenves, 2005) was used to perform finite element re-
liability analysis. Finite element reliability methods are
characterized by response quantities from a finite ele-
ment solution entering the limit state function. In this
4.2 Wind load model article limit state functions are defined as,
Following Holmes (1996), the mean drag force per unit gjk = threshold − response quantity (9)
height Dm (z) is defined as
where the threshold for the serviceability limit state is
Dm(z) = pm(z)Cd δ(z)w(z) (7) defined as the allowable deflection and the threshold for
288 Mathakari et al.

Table 2
Uncertain model parameters for the three-dimensional truss problem domain

Parameter Distribution Mean COV Corr.

Cross-sectional area Aj of members Lognormal Assigned by MOGA 5.0% 0.0

Young’s modulus Ej for bilinear material Lognormal 210,000 N/mm2 5.0% 0.3
Yield strength σ y j for bilinear material Lognormal 300 N/mm2 10.0% 0.3
Second stiffness ratio bj for bilinear material Lognormal 0.01 10.0% 0.3
Nodal coordinates in x, y and z Lognormal Assigned by MOGA St. Dev. = 10 mm 0.0
ε Standard normal 0 St. Dev. = 1 –
θ Uniform 0.78539816 0.57735 –

the ultimate limit state is defined as the design strength on the value of gamma (γ ) from importance measures,
(tension or compression) of a member as given by the only epsilon and theta turned out to be important and
AISC LRFD Manual (2001). The corresponding re- they are the only two random variables considered in the
sponse quantities are derived from the finite element reliability-based optimization.
analysis. When the uncertain response quantity exceeds
the specified threshold, the limit state function takes on
4.4 Implementation of reliability-based optimal design
a negative value and failure is implied.
Initially, all of the cross-sectional member areas (Aj ), In the proposed methodology, MOGA initially
material properties, including the Young’s modulus of generates a random population of N individuals. The
elasticity (Ej ), yield strength (σ y j ), and second stiff- reliability of each individual, i, is then computed by
ness ratio (bj ) for bilinear materials, and nodal coordi- importance sampling. The weight of each individual,
nates in the x, y, and z directions were considered as w i , and its reliability index, β i , are used directly as ob-
random variables to capture manufacturing imperfec- jectives in the Pareto ranking method used by MOGA
tions, uncertainties in the material properties, and fabri- to evolve a new population. As shown in Figure 7,
cation and construction errors, respectively. The details this process is repeated for each generation until the
of the uncertain model parameters in the three dimen- MOGA population converges to a near-optimal Pareto
sional trusses are given in Table 2 (Haukaas and Der set of design solutions.
Kiureghian, 2004). Each random variable has a specified An in-house MOGA software developed at Texas
distribution, mean, and coefficient of variation (COV). A A&M University is used in this research. The program
correlation coefficient (Corr.) of 0.3 is assumed among generates truss designs and performs a reliability analy-
the Ej . The same assumption is made to characterize sis of each truss by calling the OpenSees finite element
the correlation among the σ y j , and among the bj . Based reliability module.

MOGA Rel i abi l i ty Anal ysi s

Random Initialization of Interpretation of Binary Finite Element Model

Population Representation Generation

Initialization of GA Finite Element Reliability

GA Crossover and Mutation
Parameters Analysis

Selection based on Ranking Calculation of Weight and

and Elitism Reliability Index

MOGA Pareto Surface Encoding Binary

Ranking Representation

Fig. 7. Flow-chart of the hybrid optimization methodology.

 Reliability-based optimal design of electrical transmission towers 289

-1 Table 3
Number of stories, base and top width, total weight and
0
reliability index for trusses found in the Pareto-optimal set
1 after Generation 9
2
Reliability Index, β

Total weight Reliability

3
1 Truss Ns w b (mm) wt (mm) (Kg) index β
4 2
1 2 3,250.0 1,500.0 2,919.98 2.9773
5
GENERATION 1 2 2 3,750.0 3,750.0 3,258.86 3.6549
6 3
GENERATION 4 3 2 3,750.0 3,750.0 3,366.35 5.9701
7 4 GENERATION 7 4 2 4,000.0 1,500.0 3,785.63 6.5441
5
13 14 GENERATION 9 5 2 3,750.0 3,750.0 4,189.85 6.8884
8 6 2 3,500.0 2,750.0 4,630.59 7.3252
6
7
9 8-12 7 3 4,500.0 2,750.0 5,136.41 8.1514
2000 4000 6000 8000 10000 12000 14000 8 3 4,000.0 3,750.0 5,558.79 8.2929
Total Weight (Kg)
9 3 4,000.0 3,750.0 5,569.94 8.2929
10 3 4,000.0 3,750.0 5,628.34 8.2929
Fig. 8. Comparison of distribution of individuals in MOGA 11 3 4,000.0 3,750.0 5,639.48 8.2929
population with respect to weight and reliability objective 12 3 4,000.0 3,750.0 5,685.34 8.2929
values in Generations 1, 4, 7, and 9. 13 3 4,000.0 3,750.0 5,828.92 8.2929
14 3 4,000.0 3,750.0 6,910.76 8.2929

5 DISCUSSION OF RESULTS

The present reliability based multi-objective problem

factors in the field of reliability-based structural op-
has two objectives, which are minimizing the structural
timization. Specific issues related to the proposed
weight, w, and maximizing the reliability index, β. The
reliability-based optimization procedure are discussed
results of one of the MOGA optimization trials are pro-
in the following:
vided in Figure 8. Figure 8 compares the population
distributions in terms of objective values for the truss r It is recognized that the quality and diversity of the
designs after Generations 1, 4, 7, and 9. As the number initial population may have a prominent effect on the
of generations progresses, the Pareto-optimal curve de- direction of GA optimization and on the results ob-
fined by Rank 1 individuals in each population shifts tained, particularly in the initial generations. It is en-
toward the left (lower weight) and down (higher re- couraged to spend an effort to ascertain a good initial
liability index) on the graph. After nine generations, population.
the MOGA population converges to the Pareto-optimal r Convergence of the finite element reliability analy-
curve. The Pareto-optimal curve defines a set of near- sis may itself be challenging, although recent devel-
optimal truss designs that show the tradeoff that is oc- opments (Haukaas and Der Kiureghian, 2004) have
curring in trying to optimize the conflicting objectives. enabled the present study. It is noted that changing
The solutions found along the Pareto-optimal curve af- the shape and geometry of a structure during the
ter Generation 9 are shown in Figure 8. The solutions optimization process results in different design vari-
obtained can be graphically compared to a discontinu- able values and loading patterns, which may make
ous, idealized Pareto front. The truss designs obtained the evaluation of the system reliability more difficult
in this research effort on the Pareto-optimal curve are (Murotsu and Shao, 1990).
numbered from 1 to 14 in Figure 8. Tables 3–5 present the r Two important aspects of any reliability-based opti-
design detail information for each of the truss designs de- mization procedure are the accuracy and time effec-
fined in the Pareto-optimal set, including the number of tiveness of the reliability analysis. In this article, the
stories, dimensions, and member sizes selected. Figure 9 authors implement importance sampling to improve
shows graphically the range of topology and geometry the accuracy and to reduce the computational time re-
of the truss towers defined in the Pareto-optimal set of quired for the system reliability analysis, which would
designs evolved by MOGA. be prohibitive if crude Monte Carlo simulations were
pursued.
6 DISCUSSION AND FURTHER WORK r In the present study, the authors assume that fail-
ure of the truss occurs when any component fails
The quality of the results obtained and the computa- either in the ultimate limit state or in the service-
tional time required for the analysis are two important ability limit state. It is noted that, for statically
290 Mathakari et al.

Table 4
Members for trusses found in the Pareto-optimal set after Generation 9

Groups
Vertical members Horizontal members Cross-bracing members
Truss 1 2 3 4 5 6 7 8 9

1 W4 × 13 W8 × 10 – – W8 × 10 – W6 × 9 W6 × 9 –
2 W8 × 10 W6 × 9 – W12 × 14 W8 × 10 – W6 × 9 W6 × 9 –
3 W4 × 13 W8 × 10 – W12 × 14 W8 × 10 – W6 × 9 W6 × 9 –
4 W4 × 13 W10 × 12 – W8 × 15 W10 × 12 – W12 × 14 W8 × 10 –
5 W12 × 14 W4 × 13 – W8 × 18 W8 × 18 – W10 × 12 W6 × 9 –
6 W10 × 17 W12 × 14 – W10 × 15 W10 × 12 – W6 × 9 W6 × 9 –
7 W10 × 12 W6 × 9 W6 × 9 – W5 × 16 W10 × 12 W8 × 21 W5 × 16 W12 × 14
8 W10 × 39 W6 × 25 W6 × 25 W6 × 9 W6 × 9 W6 × 9 W6 × 16 W6 × 9 W6 × 9
9 W14 × 43 W6 × 25 W6 × 25 W6 × 9 W6 × 9 W6 × 9 W6 × 16 W6 × 9 W6 × 9
10 W10 × 39 W6 × 25 W6 × 25 W10 × 12 W6 × 9 W6 × 9 W6 × 16 W6 × 9 W6 × 9
11 W14 × 43 W6 × 25 W6 × 25 W10 × 12 W6 × 9 W6 × 9 W6 × 16 W6 × 9 W6 × 9
12 W14 × 43 W6 × 25 W6 × 25 W4 × 13 W8 × 10 W6 × 9 W6 × 16 W6 × 9 W6 × 9
13 W14 × 43 W6 × 25 W6 × 25 W4 × 13 W8 × 10 W6 × 9 W10 × 19 W6 × 9 W6 × 9
14 W10 × 39 W6 × 25 W6 × 25 W10 × 12 W6 × 9 W6 × 9 W14 × 43 W6 × 9 W6 × 9

redundant structures, a single element failure does analysis is successfully implemented and applied to a
not necessarily imply that the structure has failed. practical engineering problem. The design of an elec-
r It is also noted that, in this study, wind load is applied trical transmission tower subject to wind loading is op-
as a static load. The dynamic aspect of wind load and timized. The geometry and topology of the truss and
cable loading along the transmission lines are ignored. groups of member sections are optimized for the con-
flicting objectives of total weight and reliability of the
7 CONCLUSIONS

A hybrid methodology that combines a MOGA search

as optimization method and finite element reliability

Table 5
Story heights for trusses found in the Pareto-optimal set after
Generation 9

Truss Ns z1 (mm) z2 (mm) z3 (mm)

Truss 1 Truss 2 Truss 3 Truss 4
1 2 4,000 15,000 –
2 2 3,000 15,000 –
3 2 4,500 15,000 –
4 2 4,000 15,000 –
5 2 9,000 15,000 –
6 2 7,000 15,000 –
7 3 4,500 8,000 15,000
8 3 3,500 7,500 15,000
9 3 3,500 7,500 15,000
10 3 3,500 7,500 15,000
11 3 3,500 7,500 15,000 Truss 5 Truss 6 Truss 7 Truss 8-14
12 3 3,500 7,500 15,000
13 3 3,500 7,500 15,000
Fig. 9. Topology and geometry of the truss towers obtained
14 3 3,500 7,500 15,000
in the Pareto-optimal set.
 Reliability-based optimal design of electrical transmission towers 291

truss. The reliability is assessed considering failure due Adeli, H. & Hung, S. L. (1995), Machine Learning—Neural
to excessive deflection of the top floor of the truss and Networks, Genetic Algorithms, and Fuzzy Sets, John Wiley
excessive axial stress in each member. and Sons, New York.
Adeli, H. & Kumar, S. (1995a), Distributed genetic algorithms
Importance measures were used to reduce the num- for structural optimization, Journal of Aerospace Engineer-
ber of random variables by considering less important ing, ASCE, 8(3), 156–63.
random variables as deterministic values. This selective Adeli, H. & Kumar, S. (1995b), Concurrent structural opti-
simplification was used to reduce the computational time mization on a massively parallel supercomputer, Journal of
in the reliability analysis, which allows importance sam- Structural Engineering, ASCE, 121(11), 1588–97.
Adeli, H. & Sarma, K. C. (2006), Cost Optimization of
pling to be applied to estimate the reliability of trusses Structures—Fuzzy Logic, Genetic Algorithms, and Parallel
with multiple limit states. Computing, John Wiley & Sons Ltd., West Sussex, England.
The research results demonstrate the benefit of using AISC LRFD Manual - 325-01 (2001), LRFD Manual of
MOGA to perform reliability-based structural optimiza- Steel Construction, 3rd Edn., American Institute of Steel
tion. The ability of MOGA to obtain the Pareto-optimal Construction.
ASCE Standard 7. (2002), Minimum Design Loads for Build-
set of designs efficiently in a single trial is a significant ings and Other Structures, SEI/ASCE, Reston, VA 20191.
feature of the proposed method. The set of designs ob- Cheng, F. Y. & Li, D. (1998), Genetic algorithm develop-
tained define a range of equally optimal designs that ment for multi-objective optimization of structures, Amer-
highlight the tradeoff occurring in trying to meet the con- ican Journal of Aeronautics and Astronautics, 36(6), 1105–
flicting weight and reliability objectives. A design engi- 12.
Deb, K. (2001), Multi-Objective Optimization Using Evolu-
neer can use this information to make informed decisions tionary Algorithms, John Wiley & Sons Ltd., West Sussex,
concerning how much weight savings may be gained by England.
accepting a lower level of reliability, or what increase in Der Kiureghian, A. & Ke, J.-B. (1985), Finite-element based re-
reliability can be expected by accepting a heavier truss. liability analysis of frame structures, in Proceedings of ICOS-
The designer can also use structural configuration found SAR ’85, 4th International Conference on Structural Safety
and Reliability, Kobe, Japan, 1, 395–404.
along the Pareto curve as guidelines in the truss design Dimou, C. K. & Koumousis, V. K. (2003), Competitive genetic
process. algorithms with application to reliability optimal design, Ad-
Solving complex, realistic problems, such as the re- vances in Engineering Software, 34, 773–85.
liability based optimal design of transmission towers Ditlevsen, O. (1979), Generalized second moment reliability
subjected to wind loading as performed in this re- index, Journal of Structural Mechanics, 7(4), 435–51.
Ditlevsen, O. & Madsen, H. O. (1996), Structural Reliability
search, illustrate the potential of the proposed hybrid Methods, Wiley, New York.
methodology to allow the consideration of other prac- Eschenauer, H., Koski, J. & Osyczka, A., eds. (1990), Multicri-
tical aspects as part of the design process. This ap- teria Design Optimization, Springer-Verlag, New York, NY.
proach could be used to develop Pareto-optimal curves Fonseca, C. M. & Fleming, P. J. (1993), Genetic algorithm for
for reliability-based optimization of large-scale build- multi-objective optimization: Formulation, discussion and
generalization, in Proceedings of the Fifth International Con-
ing and bridge structures or other complex design ference on Genetic Algorithms, Morgan Kauffman, San Ma-
problems. teo, CA, 416–23.
Goldberg, D. E. (1989), Genetic Algorithms in Search, Opti-
mization, and Machine Learning, Addison-Wesley, Reading,
ACKNOWLEDGMENTS MA.
Hajela, P. & Lin, C. Y. (1992), Genetic search strategies in mul-
We acknowledge Soumya S. Pathi (Graduate Student, tioptimal design, Strucutral Optimization, 4, 99–107.
Haukaas, T. & Der Kiureghian, A. (2004), Finite element
Texas A&M University) and Sangwook Paik (Gradu- reliability and sensitivity methods for performance-based
ate Student, Texas A&M University) for their assistance earthquake engineering, Report No. PEER 2003/14, Pacific
during this research effort. Earthquake Engineering Research Center, University of
California, Berkeley.
Holland, J. H. (1975), Adaptation in Natural and Artificial Sys-
tems, The University of Michigan Press, Ann Arbor.
REFERENCES Holmes, J . D. (1996), Along-wind response of lattice towers-
III. Effective load distributions, Engineering Structures,
Adeli, H. & Cheng, N.-T. (1993), Integrated genetic algorithm 18(7), 489–94.
for optimization of space structures, Journal of Aerospace Horn, J., Nafpliotis, N. & Goldberg, D. E. (1994), A niche
Engineering, ASCE, 6(4), 315–28. pareto genetic algorithm for multiobjective optimization, in
Adeli, H. & Cheng, N.-T. (1994a), Augmented lagrangian Proceedings of the First IEEE Conference on Evolutionary
genetic algorithm for structural optimization, Journal of Computations, IEEE Service Center, Piscataway, New Jer-
Aerospace Engineering, ASCE, 7(1), 104–18. sey, 82–87.
Adeli, H. & Cheng, N.-T. (1994b), Concurrent genetic al- Jenkins, W. M. (1992), Plane frame optimum design environ-
gorithms for optimization of large structures, Journal of ment based on genetic algorithm, ASCE Journal of Struc-
Aerospace Engineering, ASCE, 7(3), 276–96. tural Engineering, 118(11), 3103–12.
292 Mathakari et al.

McKenna, F. & Fenves, G. L. (2005), http://opensees.berkeley. Rao, S. S. (1984), Multi-objective optimization in structural
edu/. The OpenSees User Manual. Department of Civil design with uncertain parameters and stochastic processes,
and Environmental Engineering, University of California, American Institute of Aeronautics and Astronautics, 22(11),
Berkeley, CA. 1670–78.
McKenna, F., Fenves, G. L. & Scott, M. H. (2005), Open sys- Schaffer, J. D. (1985), Multi-objective optimization using vec-
tem for earthquake engineering simulation, http://opensees. tor evaluated genetic algorithms, in Proceedings 1st In-
berkeley.edu/, accessed January 15, 2005. Pacific Earth- ternational Conference on Genetic Algorithms, Lawrence
quake Engineering Research Center, University of Califor- Erlbaum, Pittsburgh, PA, 93–100.
nia, Berkeley, CA. Srinivas, N. & Deb, K. (1994), Multi-objective optimization
Murotsu, Y. & Shao, S. (1990), Optimum shape design of truss using nondominated sorting in genetic algorithms, Evolu-
structure based on reliability, Structural Optimization, 2(1), tionary Computation, 2(3), 221–48.
65–76. Thampan, C. K. P. V. & Krishnamoorthy, C. S. (2001), System
Pareto, V. (1906), Manuale d’Economia Politica. Società Ed- reliability-based configuration optimization of trusses, Jour-
itrice Libraria, Milano. nal of Structural Engineering, 127(8), 947–56.
Rajeev, S. & Krishnamoorthy, C. S. (1992), Discrete Tolson, B. A., Maier, H. R., Simpson, A. R. & Lence, B. J.
optimization of structures using genetic algorithms, (2004), Genetic algorithms for reliability-based optimiza-
Journal of Structural Engineering, 118(5), 1233– tion of water distribution systems, Journal of Water Re-
50. sources Planning and Management, 130(1), 63–72.