Reasoning about joint beliefs for execution-time communication decisions

Just as POMDPs have been used to reason explicitly about uncertainty in single-agent systems, there has been recent interest in using multi-agent POMDPs to coordinate teams of agents in the presence of uncertainty. Although multi-agent POMDPs are known to be highly intractable, communication at every time step transforms a multi-agent POMDP into a more tractable single-agent POMDP. In this paper, we present an approach that generates "centralized" policies for multi-agent POMDPs at plan-time by assuming the presence of free communication, and at run-time, handles the problem of limited communication resources by reasoning about the use of communication as needed for effective execution. This approach trades off the need to do some computation at execution-time for the ability to generate policies more tractably at plan-time. In our algorithm, each agent, at run-time, models the distribution of possible joint beliefs. Joint actions are selected over this distribution, ensuring that agents remain synchronized. Communication is used to integrate local observations into the team belief only when those observations would improve team performance. We show, both through a detailed example and with experimental results, that our approach allows for effective decentralized execution while avoiding unnecessary instances of communication.