Adversarial bandit gene expression programming for symbolic regression

Abstract

Gene expression programming (GEP) is a commonly used approach in symbolic regression (SR). However, GEP often falls into a premature convergence and may only reach a local optimum. To solve the premature convergence problem, we propose a novel algorithm based on an adversarial bandit technique, named AB-GEP. AB-GEP segments the mathematical space into many subspaces. It leverages a new selection method, AvgExp3, to enhance the population jump between segmented subspaces while maintaining the population diversity. AvgExp3 dynamically estimates a subspace by rewards generated from AB-GEP without any assumption about the distribution of subspace rewards, making AB-GEP choose the appropriate subspace that contains the correct results. The experimental evaluation shows that the proposed AB-GEP method can maintain the population diversity and obtain better results than canonical GEPs.