A General Method for Handling Constraints in Genetic Algorithms

Susan E. Carlson University of Virginia Charlottesville, VA Email: carlson@virginia.edu

Introduction

emerged. These approaches can be categorized as follows: Recently, more attention has been given to Eliminating infeasible solutions the use of genetic algorithms for solving constrained optimization problems; falling into Repairing infeasible solutions this class are engineering design problems. These problems are generally highly conUsing modi ed genetic operators strained, non-linear, and multi-objective, resulting in a complex search space with regions Applying penalties for infeasibility of feasibility and infeasibility. This paper explores the development of a generalized con- The elimination of infeasible solutions has straint handling method that does not require been shown by several authors to be a poor parameter tuning. approach 1, 2, 3]. This approach is an espeWithin engineering design problems, con- cially poor performer for those problems havstraints can be classi ed as explicit or implicit. ing a small feasible to infeasible ratio of soExplicit constraints can be checked without re- lutions. Further, by eliminating those popquiring any type of simulation. For example, ulation members violating any constraint no in the traveling salesman problem (TSP) each matter how close it is to being feasible, good city must be visited once and only once. This genetic material is likely being eliminated. constraint can be checked without calculating Repair schemes can be applied to explicit the tour distance. An implicit constraint is and implicit constraints, but are generally one that requires a simulation to check for a used for explicit constraints as in combinatoviolation, and may have limits that are solu- rial problems such as the TSP. For implicit tion dependent. For example, in a hydraulic constraints repair can be more di cult. In system many di erent components can be cho- more complex design problems it may be difsen, and their selection is tightly coupled. If a cult to pinpoint the gene or group of genes pump is chosen which is too large for the other causing a problem. Further, once located it components in the system, the pressure limits may be di cult without several tries to nd a in the other system components may be ex- feasible solution, thereby wasting simulation ceeded causing a failure that can be quite dan- time on a solution that may not be worth gerous. These pressure limits are dependent repairing. As with repair schemes, modi ed on the component characteristics speci ed by genetic operators are most useful for probthe manufacturer. In this paper, we will ad- lems having explicit constraints. Generally, dress only those techniques that can handle crossover and mutation operators are develboth explicit and implicit constraints. oped to produce only feasible children, making the application of this technique problem speci c. Handling Infeasibility The fourth approach is to apply some type Many approaches have been proposed for of penalty to those solutions violating one or handling solutions that violate one or more more constraints. This approach is the most constraints, but no one general method has useful for handling implicit constraints. This 1

paper explores three di erent penalty application methods: multiplying by a factor to reduce the tness, separating the infeasible solutions from the feasible ones by remapping, and subtracting the solution's distance from feasibility. Each of these methods can be applied statically or dynamically. The rst approach detailed in 4] multiplies the objective function of a design violating a constraint by a number that generates a harsher penalty with time. This approach is shown below for a maximization problem: ? Obj X = f(X) where: X 2F = 1?ln(1+t) if X 3 F e if where t is the iteration number, and F is the feasible set of solutions. Powell and Skolnick 5] proposed mapping feasible solutions into the interval (1, 1) and infeasible solutions into (-1, 1). Joines and Houck 6] modi ed Michalewicz and Attia's 7] approach of subtracting the sum of the violated constraints from the tness function to occur within one run of a genetic algorithm rather than over many runs. They have used several parameters in their method: Obj X = f X + (C t)
? ?

The component selection problem has been described fully in the literature 2, 4]; therefore, only a brief overview will be given here. In this design problem, a con guration of generic components is given. A genetic algorithm is used to select an optimal set of manufacturers' components from a database. A set of components is judged optimal if the system meets all of the constraints and best meets the optimization criteria. To evaluate a design generated by the genetic algorithm, a simulation of the system is run, and an objective function is calculated based upon how well the design speci cations are met. The example system used is a hydraulic lift system with six generic components. The database consists of ve choices for each generic component resulting in a total of 15,625 possible designs. An enumerative search was performed to nd the best system. An integer coding scheme was used for this discrete problem. The second problem was taken from Keane 8] and is de ned as: max for

n P i=1

cos4 (xi ) ? 2
s

n Q i=1

cos2 (xi )j

n P

i=1

ix2 i

m X j =1

fj X

n n where C, , and are constants, m is the X Y xi < 15n=2 xi > 0:75; number of constraints, and t is the iteration i=1 i=1 number. This paper uses a modi ed form of the three approaches and tests them on two engineering starting from: problems described in the next section. xi = 5; i = 1; : : :; n where the xi are the variables and n is the Experimental Approach number of dimensions. The same genetic algoTwo di erent engineering problems were used rithm described above was used on this probfor testing the penalty approaches: a compo- lem with two notable exceptions. A real numnent selection problem, and a non-linear pro- ber coding scheme and arithmetic crossover gramming problem. Both use the same ba- were used. sic genetic algorithm each time. This algo- Four penalty approaches were tried on both rithm uses a rank-ordered population, uniform problems. The multiplier approach was used crossover, variable mating and mutation rates, as described above and will be referred to as a merged population of parents and o spring, approach M. A modi ed version of Powell and and a Russian roulette scheme for reducing the Skolnick's method was used, and will be referred to as approach PSN. The modi cations combined population.

0 < xi < 10; i = 1; : : :; n subject to:

included a remapping of feasible and infeasible solutions to accomodate maximization problems, and violated constraints were normalized to account for their di ering scales. This modi ed approach is shown below: f(X) X 2F m Obj X = : 0 ? P dj X 3 F
? 8 <

j =1

where dj is a normalized distance from a feasible space. For example, given the following constraint: g(X) h(X) then: dj = h(X) ? g(X) g(X) + h(X) 2 The third approach used Joines and Houck's method with several modi cations including: setting all the parameters to 1, normalizingthe constraint distances as shown above, and subtacting the sum of the constraint distances for these maximization problems. This approach will be referred to as JHN (Joines-Houck Normalized). The fourth approach used Joines and Houck's method with the above modi cations and one additional change. Instead of using the generation number as a multiplier to the sum of the violated constraints, the multiplier was tied to the feasibility of the best member of the current population. For example, if the best member of the population is infeasible, then t is increased by 1; if the best population member is feasible, then t is left alone. This method will be referred to as JHD (Joines-Houck Dynamic). In both problems, the algorithms were run 1,000 times and started with the same set of random number seeds. In the component selection problem, a population size of 30 was used. Each algorithm was run until the optimal set of components was found or for 300 generations, whichever occurred rst. These results are shown in Table 1. In this problem, both modi cations of the Joines-Houck approach worked about equally well nding the best set of components within a con dence interval of about 850 to almost 980. The modi ed Powell-Skolnick found the best solution in an average of about 1200 simulations. The

M PSN JHN JHD % Optimal 0.995 1.000 0.998 0.999 Avg. No. of Simulations 1901 1199 912 920 % Designs Simulated 0.122 0.078 0.058 0.059 95% Conf. 1822, 1126, 847, 861, Interval 1980 1272 977 979 Table 1: Results for Selection Problem M PSN JHN Max. 0.5834 0.7569 0.7589 Min. 0.3549 0.5515 0.4125 Avg. 0.4551 0.6659 0.6519 Std. Dev. 0.0359 0.0336 0.0589 95% Conf. 0.4529, 0.6638, 0.6482, Interval 0.4574 0.6680 0.6556 Table 2: Results for Keane's Problem JHD 0.7400 0.4096 0.5606 0.0819 0.5555, 0.5657

multiplier method required the most simulations to nd the best design. This result adds weight to Richardson et al.'s 3] statement that penalty methods based on the distance from feasibility perform better than other methods. In Keane's problem a population size of 100 was used, and the algorithms were stopped after 20,000 simulations. The best reported solution for 20,000 simulations and n = 20 is f(X) = 0:76. The results from these runs are shown in Table 2. In Keane's problem, the multiplier approach performed poorly as compared to the other algorithms. Both the modi ed Powell-Skolnick and the JoinesHouck normalized methods found a value close to the best reported value of 0.76. The PSN approach has the highest reported minimum value, the highest average value, and the lowest standard deviation.

Conclusions
The goal of this paper was to explore the possibility of using a constraint handling method that is capable of treating both explicit and implicit constraints. An ideal method would not require a tuning of parameters to ensure that good population members remained in the population, and those that violated con-

straints were removed. Tuning parameters are undesirable because repeated experimental runs of the algorithm are needed; a thorough understanding of the solution space is helpful; and repeated runs of a complex system simulation can be computationally expensive. A penalty approach has been the best method found to manage implicit constraints which require a simulation to be detected. For many problems, implicit constraints require a few calculations, but for engineering design problems such as component selection and shape optimization, large simulation codes must be run. Therefore, it is desirable to nd an good if not optimal solution with few required simulations and no algorithm tuning. Three approaches were tested in this paper: a multiplier penalty, a remapping scheme, and a method that uses the calculated distance from a feasible space. All associated parameters were set to one. The calculated distances were all normalized to account for the di erences in scale that can occur. Two engineering design problems were used as test cases: a component selection problem and a nonlinear programming problem. Both are common types of engineering design problems. The modi ed Powell-Skolnick approach which relies on the sum of the constraint violations performed reasonably well in the rst problem, and was the best performer in the second problem. This approach is the recommended method because of the following characteristics: it handles implicit constraints, it has no parameters to be tuned, it requires few simulations to nd an optimal answer, and it does not require a ranking of constraints. In its original form, the sum of the violated constraints is multiplied by a scaling factor r. However, Powell and Skolnick state that this factor is not needed if a rank-ordered population is used. Nonetheless, rank-ordering still does not account for the di erences in scales that the constraints may have. Using a normalization process can eliminate this problem. This method is also the safest approach for a problem about which little is known because it separates the feasible and infeasible solutions into di erent intervals. Therefore, a solution that violates a constraint can never be ranked above one that violates no constraints.

Acknowledgements
This work has been supported in part by the National Science Foundation, grant number DMI-9400114.

References
1. Michalewicz, Z., \A Survey of Constraint Handling Techniques in Evolutionary Methods," submitted for publication, 1995. 2. Carlson, Susan E., Shonkwiler, R., Ingrim, M., \Component Selection Using Genetic Algorithms," Advances in Design Automation 1993 Proceedings. Vol. 1, ASME, New York, (1993) 471-476. 3. Richardson, J., M. Palmer, G. Liepins, M. Hilliard, \Some Guidelines for Genetic Algorithms and Penalty Functions," Proceedings of the Third International Conference on Genetic Algorithms, Morgan Kaufman, San Mateo, CA, (1989), 191-197. 4. Carlson, Susan E., Component Selection Optimization Using Genetic Algorithms, Doctoral Thesis, Department of Mechanical Engineering, Georgia Institute of Technology, Atlanta, Georgia, August (1993). 5. Powell, David and M. Skolnick, \Using Genetic Algorithms in Engineering Design Optimization with Non-Linear Constraints," Proceedings of the Fifth International Conference on Genetic Algorithms, Morgan Kaufmann, San Mateo, CA, (1993), 426-431. 6. Joines, J. and C. R. Houck, \On the Use of Non-Stationary Penalty Functions to Solve Nonlinear Constrained Optimization Problems With GAs," Proceedings of the Evoluationary Computation Conference { Poster Session, IEEE World Congress on Computational Intelligence, (1994), 579-584. 7. Michalewicz, Z. and N. Attia, \In Evolutionary Optimization of Constrained Problems," Proceedings of the 3rd Annual Conference on Evolutionary Programming, World Scienti c Publishing, (1994), 98-108. 8. Keane, A. J., \Genetic Algorithm Optimization of Multi-Peak Problems: Studies in Convergence and Robustness," Arti cial Intellignece in Engineering, Vol. 9, No. 2, (1995), 75-83.