Implementation of a probabilistic model-building co-evolutionary algorithm

Abstract

We propose an extended co-evolutionary algorithm (CA) with probabilistic model building (CA-PMB) in order to improve the search performance of the CA. This article specifically describes an implementation of CA-PMB called a co-evolutionary algorithm with population-based incremental learning (CA-PBIL), and analyzes the behavior of the algorithm through computational experiments using an intransitive numbers game as a benchmark problem. The experimental results show that desirable co-evolution may be inhibited by the over-specialization effect, and that the algorithm shows complex dynamics caused by the game’s intransitivity. However, further experiments show that the intransitivity encourages desirable co-evolution when a different learning rate is set for each population.