Extended genetic programming using reinforcement learning operation

Abstract

Genetic programming (GP) usually has a wide search space and a high flexibility, so GP may search for a global optimum solution. But GP has two problems. One is slow learning speed and a huge number of generations spending. The other is difficulty in operating continuous numbers. GP searches many tree patterns including useless node trees and meaningless expression trees. In general, GP has three genetic operators (mutation, crossover and reproduction). We propose an extended GP learning method including two new genetic operators, pruning (pruning redundant patterns) and fitting (fitting random continuous nodes). These operators have a reinforcement learning effect, and improve the efficiency of GP's search. To verify the validity of the proposed method, we developed a medical diagnostic system for the occurrence of hypertension. We compared the results of the proposed method with prior ones.