Reducing code bloat in Genetic Programming based on subtree substituting technique

Abstract

Code bloat is a phenomenon in Genetic Programming (GP) that increases the size of individuals during the evolutionary process. Over the years, there has been a large number of research that attempted to address this problem. In this paper, we propose a new method to control code bloat and reduce the complexity of the solutions in GP. The proposed method is called Substituting a subtree with an Approximate Terminal (SAT-GP). The idea of SAT-GP is to select a portion of the largest individuals in each generation and then replace a random subtree in every individual in this portion by an approximate terminal of the similar semantics. SAT-GP is tested on twelve regression problems and its performance is compared to standard GP and the latest bloat control method (neat-GP). The experimental results are encouraging, SAT-GP achieved good performance on all tested problems regarding to the four popular performance metrics in GP research.