Академический Документы
Профессиональный Документы
Культура Документы
GENETIC OPTIMIZATION
By S. Pezeshk,1 Member, ASCE, C. V. Camp,2 Associate Member, ASCE, and
D. Chen,3 Member, ASCE
ABSTRACT: In this paper we present a genetic algorithm (GA)-based optimization procedure for the design of
2D, geometrical, nonlinear steel-framed structures. The approach presented uses GAs as a tool to achieve discrete
nonlinear optimal or near-optimal designs. Frames are designed in accordance with the requirements of the
AISC-LRFD specification. In this paper, we employ a group selection mechanism, discuss an improved adapting
crossover operator, and provide recommendations on the penalty function selection. We compare the differences
between optimized designs obtained by linear and geometrically nonlinear analyses. Through two examples, we
will illustrate that the optimal designs are not affected significantly by the P- effects. However, in some cases
we may achieve a better design by performing nonlinear analysis instead of linear analysis.
Ne
minimize W = Ai Li i (1)
i=1
different, say one is 00 and the other one is 11, the Ne Nn Nc
crossover operator is randomly chosen from one-point cross-
over and uniform crossover. C= C i C
d
i C Ii (3)
i=1 i=1 i=1
Since the extra two bits are used to determine which cross-
over to apply, this mechanism should give greater award to where C i , C di , and C Ii = constraint violations for stress, dis-
the crossover operator that produces superior offspring. Note placement, and interaction formulas of the AISC-LRFD spec-
that this mechanism allows the GA to adjust the relative mix- ification. Ne, Nn, and Nc = number of elements, number nodes,
ture of the four crossover operators. For example, the adap- and number of beam columns, respectively. In general, we
tational GA can use multipoint, uniform crossover primarily, express the constraint violation Ci as
or any combination among them.
0 if i 0
Mutation Ci = i if 0 < i 1.0 (4)
2i if i > 1.0
Although mutation is a secondary GA operator, it can play
an important role in the search. Mutation can be an explorative For stress constraints, we define i as
operator by moving the search into regions of the solution
i
space it may have never reached. Mutation is a character-based i = 1 (5)
string operation. The procedure for mutation can be summa- ia
rized as follows: For each character of a solution string, a where i = stress in the ith element and ai = allowable stress
randomly generated number is compared against a mutation for the ith element. For displacement constraints we have
probability. If the random number is less than the mutation
probability, the value of the character at that position is di
changed; otherwise, move to the next character position and i = 1 (6)
d ia
repeat the procedure. Typical values for the mutation proba-
bility are around 0.1%. where di = displacement at ith node and d ai = allowable dis-
JOURNAL OF STRUCTURAL ENGINEERING / MARCH 2000 / 383
placement for the ith node. The penalty terms have been nor- 2. The initial population is randomly generated.
malized by their allowable values. 3. Decode the binary design variables into decimal values
The interaction formula of AISC-LRFD specification for Pu/ and generate an input file for finite-element analysis
(Pn) 0.2 is (FEA).
4. Perform FEA using a suitable software package, check
Pu 8 Mux Muy the given constraints, and calculate the value of the pen-
< 1.0 (7)
Pn 9 b Mnx b Mny alty function.
5. Check the convergence criteria. Terminate the design
Then i becomes
process if it is satisfied; otherwise continue.
i =
Pu
Pn
8
9 Mux
b Mnx
Muy
b Mny 1 (8)
6. Calculate the penalized fitness for every individual of the
population and generate the next generation through re-
production, crossover, and mutation.
where Pu = required axial strength (tension or compression); 7. Repeat steps 3 through 6.
Pn = nominal axial strength (tension or compression); =
resistance factor (tension 0.90, compression 0.85); Mux and Muy DESIGN EXAMPLES
= required flexural strengths in x and y direction, respectively;
Mnx and Mny = nominal flexural strengths in x and y direction, To represent the strength and limitations of the GA opti-
respectively (for 2D structures, Muy is equal to zero); and b mization, we present three design examples. The first two ex-
= flexural resistance reduction factor (b = 0.90). amples are 2-bay, 3-story frames. The third example is a 1-
Using the form of penalty function as explained above, the bay, 10-story frame. The cross sections of all members are
penalty quantity will always be some percentage of the weight assumed to be W shapes. We used 256 available cross sections
of the structure. The larger the violation is, the heavier the for each member according to the AISC-LRFD. We used a
penalty will be. binary coding length of 8, representing these 256 cross sec-
tions. In addition, we used the same GA control parameters
GA-BASED DESIGN for all examples, which are population of 60, crossover prob-
ability of 0.85, mutation probability of 0.01, and ratio of ran-
The proposed design procedure involves a GA, a linear and dom generating portion of population to whole population of
geometrical finite-element analyses for fitness evaluation, an 0.01. As mentioned earlier, we used a group selection scheme
enforcement of code provisions, and a calculation of the pen- in our calculations. The population is divided into two groups.
alty function. Step-by-step operations of the GA procedure The first group occupies 30% of the population and a selection
used in this study can be summarized as the following (also probability of 0.75. The second group occupies the rest of the
see Fig. 2). 70% of the population and will have a select probability of
0.25. The design process may be stopped automatically if the
1. Select the GA control parameters suitable to the given best feasible design is not improved within five successive
problem. These parameters include population size, generations. However, in examples presented in this paper we
string length per individual design variable, crossover, did not follow this convergence criterion.
and mutation rate. The selection of these parameters may Youngs modulus of E = 29,000 ksi and a yield stress of fy
require some experimentation. = 36 ksi are used. For each example, we performed three dif-
ferent analysis and design cycles.
Case 1. Linear analysis ignoring the P- effects of the
AISC-LRFD specification.
Case 2. Linear analysis considering the P- effects in ac-
cordance with the AISC-LRFD specification.
Case 3. Geometrically nonlinear analysis in lieu of the
AISC-LRFD specifications P- effects magnifi-
cation factors.
Example 2
To highlight stability as an important design criterion Ex-
ample 1 is redesigned with additional vertical loadings, as il-
lustrated in Fig. 7. Optimal designs for three cases are given
Case 3. Geometrically Nonlinear Analysis in Lieu of AISC- TABLE 2. Summary of Optimal Design Results for Example 1
LRFD Specifications P- Effects Magnification Factors
Analysis Weight
In this case, we performed a geometrically nonlinear anal- procedure Beam Column (lb)
ysis in lieu of the AISC-LRFD specification. Results of 30 (1) (2) (3) (4)
runs are summarized in Table 1. From Table 1, we can observe Case 1 W2462 W1060 18,792
that using W1068 for columns and W2462 for beams re- Case 2 W2462 W1060 18,792
sults in the best design. The corresponding weight of this de- Case 3 W2462 W1068 19,512
sign is 19,512 lb. This is a little heavier than the previous
optimal design (4%).
One typical convergence history is illustrated in Fig. 5. We
note that the optimized design is obtained within 18 genera-
tions. However, the design process keeps fluctuating compared
with the previous linear design.
APPENDIX. REFERENCES
Adeli, H., and Cheng, N. T. (1994). Concurrent genetic algorithms for
optimization of large structures. J. Aero. Engrg., 7(3), 276296.
Camp, C. V., Pezeshk, S., and Cao, G. (1996). Design of 3-D structures
using a genetic algorithm. Proc, 1st U.S.-Japan Sem. on Struct. Op-
timization, April, Chicago.
Chen, D. (1997). Least weight design of 2-D and 3-D geometrically
nonlinear frame structures using a genetic algorithm, PhD disserta-
tion, University of Memphis, Memphis, Tenn.
Gallagher, R. H., and Zienkiewicz, O. C. (1973). Optimum structural
design: Theory and Applications, Wiley, New York.
Geist, A., Beguelin, A., Dongarra, J., Jiang, W., Mancheck, R., and Sun-
deram, V. (1991). PVM: Parallel virtual machine: A users guide and
tutorial for networked parallel computing, The MIT Press, Cambridge,
Mass.
Goldberg, D. E., and Samtani, M. P. (1986). Engineering optimization
via genetic algorithms. Proc., 9th Conf. on Electronic Comput.,
ASCE, New York, 471482.
Hall, S. K., Cameron, G. E., and Grierson, D. E. (1989). Least-weight
design of steel frameworks accounting for P- effects. J. Struct.
Engrg., ASCE, 115(6), 14631475.
Hillier, F. S., and Lieberman, G. J. (1990). Introduction to mathematical
FIG. 10. Distribution of Final Design Variables and Typical programming, McGraw-Hill, New York.
Convergence Histories of Example 3 Homaifar, A., Qi, C. X., and Lai, S. H. (1994). Constrained optimization
via genetic algorithms. Simulation, April, 242254.
From Table 4 and Fig. 9 we observe that the optimal designs Manual of steel construction: Load and resistance factor design. (1994).
2nd Ed., American Institute of Steel Construction.
obtained from three different design procedures are not sig- Pezeshk, S. (1992). Optimal design of structures with kinematic nonlin-
nificantly different. Because the frame is slender it is the dis- ear behavior. J. Engrg. Mech., ASCE, 118(4), 702720.
placement constraint, not the strength requirement, that con- Pezeshk, S. (1998). Design of framed structures: An integrated nonlinear
trols the design. The design obtained by geometrically analysis and optimal minimum weight design. Int. J. Numer. Methods
nonlinear analysis is the heaviest. From Figs. 10(ac), we ob- in Engrg., 41, 459471.
serve that the same design variable for different runs varies in Rajeev, S., and Krishnamoorthy, C. S. (1992). Discrete optimization of
structures using genetic algorithms. J. Struct. Engrg., ASCE, 118(5),
a small range of available space. In addition, Figs. 10(ac) 12331250.
show the average, , and average plus and minus one standard Rajan, S. D. (1995). Sizing, shape, and topology design optimization of
deviation, , of all 30 runs. This demonstrates the present trusses using genetic algorithms. J. Struct. Engrg., ASCE, 121(10),
optimal design methods ability to converge. From Figs. 10(d 14801487.
f) we can conclude that the optimal design can be achieved Richardson, J. T., Palmer, M. R., Liepins, G., and Hilliard, M. (1989).
within 40 generations. Some guidelines for genetic algorithms with penalty functions.
Proc., 3rd Int. Conf. on Genetic Algorithms, Morgan Kaufmann, San
Mateo, Calif., 191197.
CONCLUSIONS Spears, W. M. (1994). Adaptive crossover in genetic algorithms. Ar-
tificial Intelligence Center Internal Rep. #AIC-94-019, Naval Research
The writers have presented a design procedure using a ge- Laboratory, Washington, D.C.
netic algorithm for the design of 2D framed structures. Both Zienkiewicz, O. C. (1982). The finite element method, 3rd Ed., McGraw-
geometrically linear and nonlinear analysis were performed to Hill, London.