Strategies for Evolving Diverse and Effective Behaviours in Pursuit Domains

Abstract

Evolutionary algorithms have a tendency to overuse and exploit particular behaviours in their search for optimality, even across separate runs. The resulting set of monotonous solutions caused by this tendency is a problem in many applications. This research explores different strategies designed to encourage an interesting set of diverse behaviours while still maintaining an appreciable level of efficacy. Embodied agents are situated within an open plane and play against each other in various pursuit game scenarios. The pursuit games consist of a single predator agent and twenty prey agents, with the goal always requiring the predator to catch as many prey as possible before the time limit is reached. The predator's controller is evolved through genetic programming while the preys' controllers are hand-crafted. The fitness of a solution is first calculated in a traditional manner. Inspired by Lehman and Stanley's novelty search strategy, the fitness is then combined with the diversity of the solution to produce the final fitness score. The original fitness score is determined by the number of captured prey, and the diversity score is determined through the combination of four behaviour measurements. Among many promising results, a particular diversity-based evaluation strategy and weighting combination was found to provide solutions that exhibit an excellent balance between diversity and efficacy. The results were analyzed quantitatively and qualitatively, showing the emergence of diverse and effective behaviours.