Adaptive Genetic Algorithms Based onCoevolution with Fuzzy Behaviors

Adaptive genetic algorithms dynamically adjust the genetic algorithm connguration during the course of evolving a problem solution in order to ooer an appropriate balance between exploration (overall search in the solution space) and exploitation (localized search in the promising regions discovered in that space). One promising way followed for building adaptive genetic algorithms involves the application of fuzzy logic controllers for tuning genetic algorithm control parameters. In this paper, a general model based on fuzzy logic controllers is presented for adapting parameters that control the application of any genetic operator. Our proposal is called coevolution with fuzzy behaviors. A fuzzy behavior is a vector with the linguistic values of the fuzzy rule consequent of a fuzzy logic controller, that encodes its fuzzy rule base. Control parameter values are computed for each set of parents that undergo the genetic operator, using a fuzzy logic controller that considers particular features associated with the parents as inputs. On the other hand, fuzzy behaviors (fuzzy rule bases) are implicitly learnt by means of an additional genetic algorithm that coevolves with the main one (coevolution). The goal of coevolution with fuzzy behaviors is to obtain fuzzy rule bases producing suitable control parameter values for allowing the genetic operator to show an adequate performance. In order to analyze the eeectiveness of the model, an instance is implemented for the adaptation of the fuzzy recombination, a crossover operator proposed for real-coded genetic algorithms. An empirical study of the instance is made from two diierent points of view, one of performance and one of adaptation itself (based on the distributions of fuzzy behaviors appearing during the runs). The results show that the instance has an adaptation ability which allows signiicant performance to be achieved for test functions with diierent diiculties.