Controlling diversity of evolutionary algorithms

摘要

This paper presents a control system based method for adapting the mutation step-size in order to control the diversity of the genome population. Population diversity is controlled so that it decreases exponentially with time in order to facilitate the linear order convergence that evolutionary algorithms are capable of. The paper restricts its attention to the application of unimodal search since linear order convergence of evolutionary algorithms has only been established analytically for unimodal and not for multimodal search. The case of multimodal search is left as an exercise in implementations of sub-population schemes. The paper also highlights the subtle but important difference between setting of EAs parameters and control of EAs performance.