Sufficient statistics for multi-agent decision problems

Jeffrey N. Wu, Sanjay Lall

Abstract: Motivated by recent work showing the existence of a separation structure for certain classes of decentralized control problems, we define sufficient statistics for multi-agent team decision problems. We show that these statistics are sufficient for optimality for these problems, and reduce in the single-player case to the well-known construction based on the posterior distribution. We develop results which allow the explicit construction and updating of team-sufficient statistics, and give examples.