Mike Klaas, Tristram Southey, and Warren Cheung
One approach to creating realistic game AI is to create autonomous agents that can perform effectively with no more knowledge than a human player would have in their place. In a multi-agent setting, it is also necessary to devise a means for communicating among agents in collaborative game scenarios (such as a group of controlled agents that are searching for the player), since agents no longer have access to global knowledge. We present a method for communication using particle filters in the setting of game state estimation. Particle filters are an efficient, nonparametric means of performing inference in complex environments. Their use in game AI is particularly compelling, as they provide an easy way to represent nonlinear, non-Gaussian inferences about the state space, while exhibiting computational thrift. We demonstrate that communication among a group of agents — using particle filters to reason about the state space — can be accomplished in a natural way by sharing particles among the agents’ filters. We also show how a criterion for deciding when to communicate naturally falls out of this framework. We apply this model in the setting of coordinated target detection, and find that agents of heterogenous types and complexities can nevertheless coordinate effectively.