Performance study of improved genetic algorithm with cooperative genetic operators

摘要

We explore an empirical model for genetic algorithms that puts genetic operators in a cooperative-competitive stand with each other. The main features of the model are (i) two operators applied in parallel to create offspring - Self Reproduction with Mutation (SRM) and Crossover and Mutation (CM) (ii) an extinctive selection mechanism, and (iii) an adaptive mutation schedule that varies SRM's mutation rates based on SRM's own contribution to the population. Adaptive Dynamic-Segment (ADS) and Adaptive Dynamic-Probability (ADP) are investigated for SRM. A rigorous experimental study is conducted using 0/1 multiple knapssack problems. We isolate the contributions of extinctive selection and the parallel formulation of genetic operators and observe that the former causes an increase in search speed and the latter a substantial increase in convergence reliability. We found a difference in performance between ADS and ADP and argue that factors related to this could be epistasis and speed of adaptation.