Commitment Abstraction for Efficient Planning and Coordination in Stochastic Multi-agent Systems (Extended Abstract of Doctoral Thesis)

This thesis is motivated by the vast complexity of cooperative stochastic multi-agent planning, where agents can affect the transitions and rewards of one another, but in order to coordinate their interactions effectively, must account for the uncertainty in these actions. To combat this complexity, I exploit interaction structure in weakly-coupled problems to compute coordinated agent policies. I contend that when the degree of inter-agent dependence is sufficiently limited, the multi-agent problem can be solved more efficiently if it is broken up into (partially) decoupled subproblems: formulation of individual agent policies and coordination of abstract interactions. In support of this thesis, I develop an approach by which individual agents plan with local behavioral models that incorporate only those portions of negotiated nonlocal behavior that are needed for effective coordination.