Skip to Main Content
This paper presents a novel concept of "localization" for a class of infinite horizon two-team zero-sum Markov games (MGs) with a minimizer team of multiple decision makers that competes against nature (a maximizer team) which controls the disturbances that are unknown to the minimizer team. The minimizer team is associated with a general joint cost structure but has a special decomposable state/action structure such that each pair of a minimizing agent's action and a random disturbance to the agent affects the system's state transitions independently from all of the other pairs. By localization, the original MG is decomposed into "local" MGs defined only on local state and action spaces. We discuss how to use localization to develop an efficient distributed heuristic scheme to find an "autonomous" joint policy such that each agent's action is based on only its local state.