Grammar-based genetic programming with dependence learning and Bayesian network classifier

Grammar-Based Genetic Programming formalizes constraints on the solution structure based on domain knowledge to reduce the search space and generate grammatically correct individuals. Nevertheless, building blocks in a program can often be dependent, so the effective search space can be further reduced. Approaches have been proposed to learn the dependence using probabilistic models and shown to be useful in finding the optimal solutions with complex structure. It raises questions on how to use the individuals in the population to uncover the underlying dependence. Usually, only the good individuals are selected. To model the dependence better, we introduce Grammar-Based Genetic Programming with Bayesian Network Classifier (GBGPBC) which also uses poorer individuals. With the introduction of class labels, we further propose a refinement technique on probability distribution based on class label. Our results show that GBGPBC performs well on two benchmark problems. These techniques boost the performance of our system.