Pareto-Dominance Based MOGP for Evolving Soccer Agents

Robot behaviour generation is an attractive option to automatically produce robot controllers. Most high-level robot behaviours comprise multiple objectives that may be conflicting with each other. This research describes experiments using two Pareto-dominance based algorithms together with a Multiobjective Genetic Programming (MOGP) framework to evolve high-level robot behaviours using only primitive commands. The performance of hand-coded controllers are compared against controllers evolved using the Non-dominated Sorting Genetic Algorithm II (NSGA-II) and Strength Pareto Evolutionary Algorithm 2 (SPEA2) algorithms. An additional comparison is also performed against controllers evolved using the weighted sum fitness function. The experiment results show that the Paretodominance based MOGP performed better than the hand-coded and the weighted sum evolved controllers.