Investigação de novas abordagens em sistemas imunes artificiais para otimização

Abstract

The computational cost of the optimization process of electromagnetic devices is directly related to the number of objective function evaluations. This has motivated the study of new methods that are capable of determining efficient results with a fewer number of function evaluations. This dissertation proposes two new immune algorithms for mono and multi-objective optimization. The mono-objective version, named Distributed Clonal Selection Algorithm - DCSA, implements a main operator called distributed somatic hipermutation, while the multi-objective version, named Multi-Objective Clonal Selection Algorithm - MCSA, implements in addition a receptor editing operator. The somatic hypermutation, composed of different probability density functions, Gaussian, uniform and chaotic, performs a balancing local searcharound the high affinity solutions, and also facilitates the best distribution of the solutions throughout the extension of the Pareto-optimal front in the MCSA. The receptor editing operator, based on the differential evolution technique, implicitly performs a dynamic search over the feasible region, ensuring the best local refinement of the optimal solutions, and helping the increase of the convergence speed of the method. The optimization parameters of the algorithms have been subjected to sensitivity analysis,which has provided a range of acceptable values for them. Furthermore, the suggested immune operators have been assessed in order to determine the effect of each one in the performance of the methods. The proposed immune algorithms have been validated through the solution of analytical problems with different optimization features, such as, strong smoothness, multimodality, high dimensions and constraints, presenting efficient solutions when compared to other known evolutionary methods. Finally, tests with electromagnetic problems of high computational cost have been performed, resulting in very good solutions with less machine effort, regarding the number of function evaluations.