Evolving Teams of Multiple Predictors with Genetic Programming

Abstract

This paper reports on the evolution of GP teams in different classiffication and regression problems and compares dierent methods for combining the outputs of the team programs. These include hybrid approaches where (1) a neural network is used to optimize the weights of programs in a team for a common decision and (2) a real-numbered vector of weights (the representation of evolution strategies) is evolved with each team in parallel. The cooperative team approach results in an improved training and generalization performance compared to the standard GP method. The higher computational overhead of coevolving several genetic programs is counteracted by using a fast variant of linear GP. In particular, the processing time of linear genetic programs is reduced signicantly by removing intron code before program execution.