Importance-Based Pruning for Genetic Programming Based Symbolic Regression

Abstract

Pruning is a major research field in neural networks, enhancing their efficiency and generalization. The field of pruning approaches in genetic programming (GP) is continually evolving, with researchers actively exploring new techniques and approaches to optimise the performance of GP models. This research introduces a novel pruning algorithm for Genetic Programming-based Symbolic Regression (GPSR). The proposed method employs a weighting mechanism to identify and filter out unimportant subtrees in each generation. To achieve this, the method arranges all subtrees linearly and assigns weights to each subtree and terminal. It then uses Ordinary Least Squares (OLS) to optimize these weights, enabling the identification of unimportant subtrees and terminals for effective pruning. The algorithm’s effectiveness was evaluated using ten regression datasets, including high-dimensional and complex feature sets. Furthermore, comparisons were made with two other algorithms to evaluate its performance. The results indicate that the proposed approach not only achieves better learning and generalisation performance but also generates smaller trees compared to standard GP, thereby improving interpretability.