Journal of Structural Engineering No.

34-10
Vol. 34, No.2, June-July 2007 pp. 1–6

Logical penalty functions for efficient search in optimisation of discrete

structures using genetic algorithm

P. Sivakumar∗ and C.V. Vaidyanathan∗∗

Genetic algorithms are suitable for discrete optimisation of engineering structures such as trusses and lattice towers.
The efficiency and speed of optimisation depend to a large extent on the number of design variables and constraints.
Many times the solutions obtained are less optimal or infeasible as the probability of constraint violation increases
when total number of constraints increases. This complexity with constraints makes convergence difficult. Simple
strategies, which can account for the complexities and practical considerations, are required to make the algorithm
efficient.
Logically evolved penalty functions are developed and introduced in the algorithm to improve the performance of
the algorithm. These functions direct the search towards feasible and eliminate the participation of inefficient strings.
The purpose of this approach is to eliminate the difficulties that arise in handling constraints in the problem. Improved
search and rapid convergence are obtained using this approach. The function varies according to constraints, type of
problem and trend of results. A typical 10-bar truss benchmark problem is studied with this approach using genetic
algorithm. The results obtained using the above approach is compared with the results reported in literature. The
results obtained in this study are found to be better than the results given in the literature.

Structural design is the process of defining the system itself. ber of possibilities is small. The lack of efficiency leads to
The purpose of structural design is to evaluate the sizes and the non-robustness of these methods. The shortcomings of
location of members necessary to support a prescribed set of calculus-based and enumerative methods have popularised
loads. Optimisation is an automated design procedure giving the Random Search algorithms. But these methods are also
the optimal values of certain design quantities. Trial and error expected to do no better than enumerative schemes in the
procedure is usually followed to obtain the desired solution. long run. The efficiency is not enough to make these algo-
An initial guess is used as a starting point for a systemat- rithms robust. Hence the conventional methods are not so ro-
ic search for better designs. The search is terminated when bust. This does not imply that they are not useful. The scheme
certain criteria are satisfied indicating that the design is suf- with the hybrid combinations and permutations has been used
ficiently close to the optimum. In this process, the structure successfully in many applications. One such algorithm is Ge-
is analysed repeatedly for successive modification in the de- netic Algorithm (GA).
sign. A near optimal design is automatically generated in an
iterative manner by using the numerical methods. Develop- Recent developments in the GA have given thrust to
ments in the programming method have facilitated the appli- structural optimization. Goldberg2 introduced GA based on
cation to find the solution for large real world design prob- the principles of natural evolution as a search technique
lems. Mathematical programming methods have been devel- for engineering optimization. The basic operations of GA
oped to deal with specific optimisation problems. and practical applications were presented by Jenkins3 . A
simple genetic algorithm for optimising structural systems
The engineering design optimisation usually differs from with discrete design variables was presented by Rajeev
problem to problem. There is no single method of for- and Krishnamoorthy4 . Koumousis and Georgiou5 solved the
mulation available for solving all optimisation problems mixed layout and sizing optimization problem of a typical
efficiently1 . Hence a number of optimisation methods have steel roof using GA and a logic program. Galante6 applied
been developed for solving different types of optimisation GA as an approach to optimize real-world trusses. Sivakumar
problems. A large number of optimisation algorithms are et al7 presented an object oriented optimization technique us-
available in the optimisation literature, each suitable to solve ing GA for steel lattice towers. These literatures reveal ap-
a particular type of problem. The search space in reality is of- plication of GA to the structural engineering optimization
ten fraught with discontinuous and multimodal noisy search problems. Earlier studies have already indicated that genet-
spaces in a less calculus-friendly function. ic algorithms are suitable in handling discrete design vari-
Enumerative search schemes look at objective function ables of engineering problems. Formulation of the optimi-
values at every point in the space, one at a time. These types sation procedure needs different strategies in many engineer-
of algorithms are simple and attractive only when the num- ing optimisation problems. Application oriented optimisation
line space pl ∗ Scientist, ∗∗Formerly Director Grade Scientist, Structural Engineering Research Centre, CSIR Campus, Taramani, Chennai-600 113, India
(Discussion on this article must reach the editor before September 31, 2007)

JOURNAL OF STRUCTURAL ENGINEERING VOL. 34, NO.2, JUNE-JULY 2007 1

can be achieved by formulating techniques specific to the de- deflection constraint, the penalty is calculated based on the
sign problem. degree of violation.
Constraint handling in real-world problems such as struc-
tural optimisation problems needs some mechanism to per- FORMULATION OF THE PROBLEM
form an effective search. The objective of the paper is to in-
troduce logically evolved penalty functions to improve the The objective function of the optimisation problem is to min-
performance of the algorithm. This approach tries to elimi- imize the total weight of the ten-bar truss (Fig. 1). Equation
nate the difficulties that arise in handling constraints in the (1) gives the formulation of the weight minimization prob-
problem as these functions direct the search towards feasible lem. Minimize
region.
10

W = ρAi li (1)
GENETIC ALGORITHMS Where
i=1
= 0.271447 ×10−4 /mm3(0.10lb/in3)
Genetic algorithms mimic the principles of evolution and Subjected to
have their basis in Darwin’s theory of Survival of the Fittest. Stress constraint σi ≤ σall i = 1, 2 · · · 10
Genetic algorithm combines the survival of the fittest among Deflection constraint δmax ≤ δall Where
the string structures with a structured yet randomised in- all = 172.37 N/mm (25 ksi
formation exchange to form a search algorithm. The search all = 50.8 mm (2" )
procedure uses random choice as a tool to guide search
through parametric space. Genetic algorithms efficiently
exploit historical information to find new search points with
expected improved performance. These algorithms are com-
putationally simple yet powerful in their search for improve-
ment.
Genetic Algorithms are very effective in finding solu-
tions to a variety of problems. This innovative technique per-
forms especially well when solving complicated real-world
problems. Its simplicity of approach and directness in dis-
crete variable combinatorics makes it attractive in compar-
ison with mathematically complex methods. Genetic algo-
rithms tend to converge on solutions that are globally opti-
mal or nearly so, even in large and complicated search space,
given certain conditions on the problem domain.
FIG. 1 TEN-BAR TRUSS

LOGICAL PENALTY FUNCTIONS The modulus elasticity of the material is assumed as

68947.6 N/mm2 (104 ksi). A list of discrete values of areas
from American Institute of Steel Construction (AISC), used
Real-world structuraloptimization problems with constraints as variables in this example is given in Table 1. Sixteen con-
need some mechanism for handling constraints in the opti- secutive values in the list are taken for selecting the value of
mization methods. The common way for treating constraints the variables.
in the GA is a penalty method. Penalty method converts
a constrained problem to an unconstrained problem. The
penalty term is either added or multiplied. Several meth- TABLE 1
ods based on penalty function such as Death penalty, Static DISCRETE VALUES OF AREAS FROM AMERICAN INSTITUTE
Penalties, Dynamic Penalties, Adaptive Penalties, Annealing OF STEEL CONSTRUCTION MANUAL
Penalties, Co-evolutionary Penalties are used in GAs. The S. No. Area (mm2 ) S. No. Area (mm2 ) S. No. Area (mm2 )
difficulty with penalty methods is to set appropriate penal- 1 1045 15 2342 29 7419
ty. Logical penalty function is also a static penalty function 2 1161 16 2477 30 8710
where penalty values are derived from the level of violation. 3 1284 17 2497 31 8968
When the feasible region is very narrow compared to the de- 4 1374 18 2503 32 9161
sign field these logically derived penalty values are useful. 5 1535 19 2697 33 10000
6 1690 20 2723 34 10323
In discrete structural optimisation problem all members
7 1697 21 2897 35 10903
of the structure have to satisfy some of the constraints inde-
8 1858 22 2961 36 12129
pendently. For example stress constraints have to be satisfied 9 1890 23 3097 37 12839
by all the members independently. In genetic algorithm the 10 1994 24 3206 38 14194
fitness of the string decides the reproduction. A string con- 11 2019 25 3303 39 14774
taining very few invalid members, which do not satisfy the 12 2181 26 3703 40 17097
constraint and many valid members, which satisfy the con- 13 2239 27 4658 41 19355
straint may not find place in the next generation if the to- 14 2290 28 5142 42 21613
tal population is rejected. To accommodate such strings, the
strings are still considered to be valid with some penalty on Fitness values are considered to be inversely proportional
the fitness. These penalty values are logically deduced. The to the weight of the structure obtained in the evaluation, i.e.
degree of violation depends on the number of violation and lower the weight of structure, higher the fitness value. As ge-
quantity. When more members of the structure are violating netic algorithms work on fitness maximisation function, Eq.
stress constraint higher penalty is imposed on the fitness. For (2) is used to convert it to the weight minimization problem.

2 JOURNAL OF STRUCTURAL ENGINEERING VOL. 34, NO.2, JUNE-JULY 2007

 The value of K is chosen appropriately such that it is approx- The final penalty factor is obtained by multiplying the
imately equal to the expected weight of the structure to keep penalty factors of stress and deflection constraints. This val-
the fitness value within 0 and 1. ue is used to reduce the fitness of the string. The fitness val-
ue obtained based on the weight of the structure is reduced
by dividing the fitness value with the final penalty value. As
K penalty values are imposed on the fitness values, a structure
Fitness Value = (2)
K +W with less weight can have lower fitness value than a heavier
structure in a few cases where constraints are not satisfied.
Stress and deflection constraints are considered. The crit- A software called ‘Genetic Optimization Algorithm for
icality ratio is defined as the ratio of stress in the member to Towers (GOAT)’ has been developed and this software con-
the allowable stress in the member and this ratio is calculated tains an analytical model generator, analyzer, designer, eval-
for every member of the truss. To satisfy the stress constraint, uator and basic operators of a simple GA. This software has
the criticality ratio should be less than unity. The maximum been used for optimization.
deflection is obtained by comparing the deflection at every A generation history of optimisation of the problem con-
node of the truss and checked for deflection constraint. sidered is shown in Fig. 2, which shows the variation of
weight with the increase in generations. The weight of the
PARAMETERS OF THE GENETIC ALGORITHM truss corresponds to the best fitness obtained till the genera-
tion.
Roulette wheel selection and single point crossover are used
in this genetic algorithm. A population size of 20, a prob-
ability of mutation 0.01 and a crossover probability of 0.8
is considered in this study. The genetic algorithm parameter
values such as population size , probabilities of crossover and
mutation were not studied for different values. It is assumed
that the convergence is obtained when the resulting value or
the minimum weight of the structure is not changed for more
than 100 generations.

STRESS CONSTRAINT PENALTY FUNCTION

Logical penalty functions are introduced in the evaluation

process for both stress and deflection constraint violations.
When members in the structure violate the stress constraint,
criticality ratio of those members. The value of criticality ra-
FIG. 2 GENERATION HISTORY OF A TEN-BAR TRUSS
tio is more than one whenever stress constraint is not sat-
isfied. The average value of the criticality ratio of violating
members is taken as the penalty factor for stress constraint. The graph shows the weight of the best fitness population
obtained till the generation. The kink in the variation indi-

N sv cates a string containing higher fitness with low weight but
P Fs = CRi/Nsv (3) with some constraint violation. Subsequently higher fitness
i=1
with higher weight due to the penalty applied on the fitness.
The density of feasible solutions is shown in Fig. 3 for the
This factor is increased further depending upon the num- first thousand populations. A solution is considered to be in-
ber of members of the structure that violate the constraint. feasible when the constraints are not satisfied. A value of ‘1’
The penalty factor calculated from the average value as above is assigned for infeasible design and ‘0’ is assigned for feasi-
is multiplied further by a stress violation factor to discour- ble design. All the population in all the generation are plotted
age the reappearance of strings that contains more number of with the values for feasibility and are given in Fig. 3.
stress violations. The population in every generation have both feasible and
infeasible solution sets. The number of population containing
Check font Stress Violation Factor = 1 + (Nsv/Nm ) (4) feasible solution set increases with generation increase. The
best fitness value in each generation is shown in Fig. 4. Even
This violation factor is unity when there is no violation though there are fluctuations in the fitness value, it can be
and the value is two when all the members violate. observed that the trend of fitness value increases with gener-
ation. There is a gradual increase in the average fitness value.
The weight of the truss corresponding to the solution set hav-
DEFLECTION CONSTRAINT PENALTY FUNCTION ing the best fitness value in each generation is shown in Fig.5.
There is also a gradual reduction in the average weight of the
For deflection constraint, a value proportional to the extent of truss over generation and is indicated by the trend line.
violation to the allowable limit is added to unity and the value The penalty values are assigned when maximum deflec-
is applied as a penalty factor, when the maximum deflection tion of the truss exceeds the allowable limit and are shown
of the structure violates the deflection constraint. in Fig. 6. penalty value is ‘1’ when there is no violation. De-
pending on the quantity of violation, this value is increased.
Similarly, the penalty values assigned for violation of stress
P Fd = 1 + [(δmax − δall )/δall ] (5) constraint for the population are shown in Fig. 7. It can be

JOURNAL OF STRUCTURAL ENGINEERING VOL. 34, NO.2, JUNE-JULY 2007 3

 observed from the variation that, the number of constraint vi- the results reported by Rajeev and Krishnamoorthy4 , who
olations due to stress is more than that of deflection. have used the same discrete variables. Even though a similar
genetic algorithm with discrete optimisation is used, possi-
bly the penalty function used in this study has contributed to
achieve a better solution.
There is an improvement in the weight reduction with dis-
crete values for variables and optimisation with genetic algo-
rithms. The improvement can be attributed to the method of
handling of the constraints. The method adopted to eliminate

Use same font

repeats

FIG. 3 DENSITY OF FEASIBLE SOLUTIONS FIG.4 BEST FITNESS OF GENERATIONS

FIG.5 BEST FITNESS WEIGHTS FOR GENERATIONS

FIG.4 BEST FITNESS OF GENERATIONS

ANALYSIS OF RESULTS

The solution obtained in this study using GOAT software is

compared with different solutions available in the literature
and this comparison is shown in Fig. 8. The minimum weight
of the truss obtained along with area of cross section for each
variable are compared with the results given in the literature
and are given in Table 2. The results indicate that the solu-
tion obtained through this algorithm is better than all others,
except those by CONMIN and OPTDYN, where the solu-
tions are obtained through continuous optimisation methods,
which have less practical feasibility. Discrete optimisation is FIG.6 PENALTY VALUES FOR DEFLECTION CONSTRAINT
practically feasible and the solution obtained is better than

4 JOURNAL OF STRUCTURAL ENGINEERING VOL. 34, NO.2, JUNE-JULY 2007

 TABLE 2
OPTIMIZATION RESULTS OF 10-BAR TRUSS
Method Weight (N) A1 (mm2 ) A2 A3 (mm2 ) A4 A5 (mm2 ) A6 A7 (mm2 ) A8 A9 (mm2 ) A10
(mm2 ) (mm2 ) (mm2 ) (mm2 ) (mm2 )
LINRM 27797 13916 7084 14245 9645 65 7084 12200 11884 11871 8716
SUMT 26387 19800 1529 20400 7523 63 2394 14006 13484 9013 2103
CRP-UI 25475 15987 2690 15987 9323 65 2690 11264 12426 12432 3394
M-5 25466 16664 1858 17064 8226 65 2432 12497 12374 12110 2826
M-3 25439 16671 1981 17045 8239 65 2213 12477 12368 12103 2852
Rajeev(1) 24999 21613 1045 14194 10000 1045 1161 9161 12839 12839 1690
Rajeev(2) 24972 21613 1045 14194 10000 1045 1045 9161 12839 12839 1690
Rajeev(3) 24972 21613 1045 14194 10000 1045 1045 9161 12839 12839 1690
GOAT 24918 19355 1045 12839 10903 1045 1045 5142 17097 14774 1284
CONMIN 24746 16258 1219 16045 10213 65 1129 10813 12729 13535 1619
OPTDYN 24341 16581 65 16200 12510 65 65 9935 13110 13381 735

the efficiency of search in genetic algorithm. Analysing the

solution sets over generation can lead to identification of pos-
sible improvement in the algorithm such as problem spe-
cific penalty functions. Improved evaluation using logical-
ly evolved penalty functions is very useful to eliminate the
participation of inefficient strings in the successive genera-
tion. This approach has been applied for a typical bench mark
problem and more efficient results are obtained.

ACKNOWLEDGEMENTS

The authors express their gratitude to Dr. N. Lakshmanan,

Director, Structural Engineering Research Centre for his con-
FIG.7 PENALTY VALUES FOR STRESS CONSTRAINT stant encouragement and support. This paper is published
with the kind permission of the Director, Structural Engineer-
ing Research Centre, Chennai.

NOTATIONS

Ai = Cross sectional area of the members

CRi = Criticality Ratio of violating member
K = An arbitrary constant
Nm = Number of members
Nsv = Number of stress violations
P Fd = Penalty Factor for Deflection
W = Total weight of the truss
li = Length of the members.
ρ = Material Density = 0.271447 ×10−4
N/mm3 (0.10lb/in3 )
σi = Stress in ith member
FIG.8 COMPARISON OF TEN-BAR TRUSS OPTIMIZATION σall = Allowable Stress = 172.37 N/mm2
RESULTS (25 ksi)
δmax = Maximum deflection of the truss delete
infeasible solution sets may be changed and better solution δall = Allowable deflection of the truss
can also be obtained. Higher penalty values may be used to = 50.8 mm (2 )
to eliminate infeasible solutions, but sometimes it may lead to δmax = Maximum deflection of the truss
the search in the infeasible design space. The genetic param- δall = Allowable deflection of the truss
eters such as seed for random number, population size and
number of generation also influences the result, but not to a REFERENCES
considerable extent. The convergence occurs at near optimal
value, but not at any specific value for different parameter
when the search space and the problem size are large. 1. Deb, K., “Optimization for Engineering Design: Algo-
rithms and Examples”, Prentice-Hall Private Ltd., New
Delhi, 1998.
CONCLUSIONS
2. Goldberg, D. E., “Genetic Algorithms in Search, Op-
Formulation of design specific approach for dealing the dis- timization, and Machine Learning”, Addison-Wesley
crete design optimisation problem is necessary to improve Longman, Inc., Reading, MA, USA, 1989.

JOURNAL OF STRUCTURAL ENGINEERING VOL. 34, NO.2, JUNE-JULY 2007 5

3. Jenkins, W. M., “Towards Structural Optimization via the 6. Galante, M., “Genetic Algorithms as an Approach to Op-
Genetic Algorithm”, Technical Note, Comp. and Struct., timize Real-World Trusses”, Int. J. for Num. Meth. in
40(5), 1991, 1321–1327. Engg., 39, 1996, 361–382.
4. Rajeev, S., and Krishnamoorthy, C. S., “Discrete Opti-
mization of Structures using Genetic Algorithms”, J. of 7. Sivakumar, P., Rajaraman, A., Samuel Knight, G. M.
Struct. Engg., ASCE, 118(5), 1992, 1233–1250. and Ramachandramurthy, D. S., “Object-Oriented Opti-
5. Koumousis V. K., and Georgiou P. G., “Genetic Algo- mization Approach Using Genetic Algorithms for Lattice
rithms in Discrete Optimization of Steel Truss Roofs”, J. Towers”, J. of Comp. in Civil Engg., ASCE, Vol. 18, Issue
of Comp. in Civil Engg., ASCE, 8(3), 1994, 309–325. 2, 2004, pp. 162–171

6 JOURNAL OF STRUCTURAL ENGINEERING VOL. 34, NO.2, JUNE-JULY 2007