Evolving teamwork and coordination with genetic programming

Some problems can be solved only by multiagent teams. In using genetic programming to produce such teams, one faces several design decisions. First, there are questions of team diversity and of breeding strategy. In one commonly used scheme, teams consist of clones of single individuals; these individuals breed in the normal way and are cloned to form teams during fitness evaluation. In contrast, teams could also consist of distinct individuals. In this case one can either allow free interbreeding between members of different teams, or one can restrict interbreeding in various ways. A second design decision concerns the types of coordination-facilitating mechanisms provided to individual team members; these range from sensors of various sorts to complex communication systems. This paper examines three breeding strategies (clones, free, and restricted) and three coordination mechanisms (none, deictic sensing, and name-based sensing) for evolving teams of agents in the Serengeti world, a simple predator/prey environment. Among the conclusions are the fact that a simple form of restricted interbreeding outperforms free interbreeding in all teams with distinct individuals, and the fact that name-based sensing consistently outperforms deictic sensing.