Learning Grammar Rules in Probabilistic Grammar-Based Genetic Programming

Abstract

Grammar-based Genetic Programming (GBGP) searches for a computer program in order to solve a given problem. Grammar constrains the set of possible programs in the search space. It is not obvious to write an appropriate grammar for a complex problem. Our proposed Bayesian Grammar-Based Genetic Programming with Hierarchical Learning (BGBGP-HL) aims at automatically designing new rules from existing relatively simple grammar rules during evolution to improve the grammar structure. The new grammar rules also reflects the new understanding of the existing grammar under the given fitness evaluation function. Based on our case study in asymmetric royal tree problem, our evaluation shows that BGBGP-HL achieves the best performance among the competitors. Compared to other algorithms, search performance of BGBGP-HL is demonstrated to be more robust against dependencies and the changes in complexity of programs.