Multiagent Cooperation through Egocentric Modeling

We consider a scenario in which interacting agents cooperate through an iterative process of 1) forming empirical models of the behavior of other agents and 2) selfishly optimizing a local strategy based on these models. In each iteration, an agent revises its models of other agents. Selfish optimization according to these revised models alters the behavior of a each agent. This, in turn, leads to a new round of revised models of other agents. The implication of convergence is a consistency condition. Namely, each agent\´s behavior is consistent with how the agent is modeled by others. Furthermore, each agent\´s local strategy is optimal with respect to how it models other agents. We consider a particular instance of this framework that is motivated by the "roboflag drill" coordination scenario. This paper derives conditions for convergence, provides illustrative simulations, and establishes a connection to related work in evolutionary games