Feature Encapsulation by Stages Using Grammatical Evolution

This paper introduces a novel mechanism, Feature Encapsulation by Stages (FES), to encapsulate and transfer features as knowledge in a staged manner within the evolutionary process. Encapsulation happens via input space expansion in one or more stages by adding the best-of-run individual as an additional input. This input space expansion is managed by augmenting the grammar. We study the feasibility of dynamically modifying the grammar and reinitialising the population to make way for new individuals which quickly evolve to a better fitness level. Five different approaches to stage management are examined. In addition, three different selection processes, namely, Tournament, Lexicase and Lexi2, are used to investigate which is best suited to use with our encapsulation procedure. We benchmark our procedure on two problem domains, Boolean and Classification, and demonstrate these staging strategies lead to significantly better results. Statistical tests show our FES outperforms the standard baseline in all Boolean problems, with a 4-stage version performing best, obtaining significant differences in all Boolean problems.