Many classes of cooperative multi-agent systems require a diversity of behavior among the agents in order to optimize their performance as a team in the system. Conventionally the control policies for the agents in such systems are programmed or trained so that individual agents are hard-coded to adopt specialized roles within a team.
However, customized collections of specialists can be brittle when they are not deployed in the optimal ratio for a context, when the context changes after they are deployed, or when individual specialists break down. Thus for many multi-agent tasks a useful alternative would be to deploy a collection of identical general-purpose agents that are able to organize themselves with a division of labor appropriate to the current context and number of agents in the team.
In this project we examine how neuroevolution
can be used to train artificial neural networks to be used for the controllers for sets of identical agents in systems where diversity of behavior is required. At present we are working with autonomous agents in strategy games and simulated construction tasks for our application domains.
See movies of agents in the Legion II strategy game.