Biases Introduced by Adaptive Recombination Operators

The broad goal of adaptive techniques is to acquire knowledge dynamically about the search space and to use this knowledge to bias the evolutionary process. The effectiveness of any adaptive technique is therefore determined by the biases being used. In this paper we identify four key biases introduced by adaptive recombination operators and analyse the relationship between these biases. We use these four biases (directional, credit, initialisation and hitchhiker) to characterise three adaptive recombination operators. We show that the biases introduced by adaptive recombination are not always beneficial to GA performance and we explore methods for minimising the detrimental effects.