Improving Performance and Cooperation in Multi-Agent Systems

Abstract

Research has shown that evolutionary algorithms are apromising approach for training agents in heterogeneousmulti-agent systems. However, research in evolvingteams (or ensembles) has proven that commonevolutionary approaches have subtle, but significant,weaknesses when it comes to balancing memberperformance and member cooperation. In addition, thereare potentially significant scaling problems inapplying evolutionary techniques to very largemulti-agent systems. It is impractical to train eachmember of a large system individually, but purelyhomogeneous teams are inadequate. Previously weproposed Orthogonal Evolution of Teams (OET) as a novelapproach to evolving teams that overcomes theweaknesses with balancing member performance and membercooperation. In this paper we test two basicevolutionary techniques and OET on the problem ofevolving multi-agent systems, specifically a landscapeexploration problem with heterogeneous agents, andexamine the ability of the algorithms to evolve teamsthat are scalable in the number of team members. Ourresults confirm that the more traditional evolutionaryapproaches suffer the same weakness with multi-agentsystems as they do with teams and that OET doescompensate for these weaknesses. In addition, the threealgorithms show distinctly different scaling behaviour,with OET scaling significantly better than the two moretraditional approaches.