Particle Swarm Optimization Based Tuning of Genetic Programming Evolved Classifier Expressions

Abstract

Genetic Programming (GP) has recently emerged as an effective technique for classifier evolution. One specific type of GP classifiers is arithmetic classifier expression trees. In this paper we propose a novel method of tuning these arithmetic classifiers using Particle Swarm Optimization (PSO) technique. A set of weights are introduced into the bottom layer of evolved GP classifier expression tree, associated with each terminal node. These weights are initialized with random values and optimized using PSO. The proposed tuning method is found efficient in increasing performance of GP classifiers with lesser computational cost as compared to GP evolution for longer number of generations. We have conducted a series of experiments over datasets taken from UCI ML repository. Our proposed technique has been found successful in increasing the accuracy of classifiers in much lesser number of function evaluations.