Estratégias de penalização adaptativa para a solução de problemas de otimização com restrições via algoritmo genético

Abstract

The application of metaheuristics on real problems with constraints is not possible without adjustments. This impossibility is due to the fact that they are developed, in their essence, to solve unconstrained optimization problems. These adjustments are made by techniques that address the constraints present in the problem. Penalty Techniques are common, transforming a constrained optimization problem into an unconstrained optimization problem, adding a penalty to the fitness function of infeasible solutions. This thesis considers a technique that adapts the value of the penalty coefficient of each constraint using the information of the population, such as the average of the objective function and the level of violation of each constraint. This technique is known as Adaptive Penalty Method (or simply APM). There are in the literature, several variants for the APM and they can be synthesized as: Sporadic APM which holds the fixed penalty coefficients for a fixed number of generations, a second approach similar to the first, but accumulating values of the violations; the variant entitled Monotonic APM, which is similar to the original APM but not allowing the penalty coefficients be reduced along the evolutionary process and the variant damped APM, which uses a weighted average of the current and previous values of the penalty coefficients. New variants for the APM are proposed in this thesis in order reach improvements in the original APM. The performance of these new variants is examined using test-functions and problems of mechanical and structural engineering. Comparisons are conducted using performance profiles, which allow to identify more clearly the robustness of these variants pointing out the best ones.