Asynchronous Execution in Multiagent POMDPs: Reasoning over Partially-Observable Events

This paper proposes a novel modeling approach to problems of multiagent decision-making under partial observability, building on the framework of Multiagent Partial Observable Markov Decision Processes (MPOMDPs). Unfortunately, the size of MPOMDP models (and their solutions) grows exponentially in the number of agents, and agents are required to act in synchrony. In the present work, we show how these problems can be mitigated through an eventdriven, asynchronous formulation of the MPOMDP dynamics. We introduce the necessary extensions to the dynamics and solution algorithms of standard MPOMDPs. In particular, we prove that the optimal value function in our EventDriven Multiagent POMDP framework is piecewise linear and convex, allowing us to extend a standard point-based solver to the event-driven setting. Finally, we present simulation results, showing the computational savings of our modeling approach.