Evolving Robotic Neuro-Controllers Using Gene Expression Programming

Abstract

Current trends in evolutionary robotics (ER) involve training a neuro-controller using one of the various population based algorithms. The most popular technique is to learn the optimal weights for the neural network. There is only a limited research into techniques that can be used to fully encode a neural network (NN) and therefore evolve the architecture, weights and thresholds as well as learning rates. The research presented in this paper investigates how the chromosomes of the gene expression programming (GEP) algorithm can be used to evolve robotic neural controllers. The designed neuro-controllers are utilised in a robotic wall following problem. The ensuing results show that the GEP neural network (GEPNN) is a promising tool for use in evolutionary robotics.