Multi-objective topology and weight evolution of neuro-controllers

Abstract

Evolutionary multi-objective optimization has been employed in studies concerning evolutionary robotics, and in particular for the evolution of neuro-controllers. To allow the simultaneous multi-objective evolution of topology and weights, tailored search algorithms should be developed. Here, a modification to the well-known NEAT algorithm is suggested. The proposed algorithm, which is termed NEAT-MODS, involves a specialized selection process that aims to ensure both genotypic diversity and elitism in the context of Pareto-optimality. NEAT-MODS constitutes a generic Multi-objective Topology and Weight Evolution of Artificial Neural-Networks (MO-TWEANN) algorithm. The suggested NEAT-MODS is found to be statistically superior to NEAT-PS, when applied to solve complex multi-objective navigation problem.