Partially Synchronized DEC-MDPs in Dynamic Mechanism Design

In this paper, we combine for the first time the methods of dynamic mechanism design with techniques from decentralized decision making under uncertainty. Consider a multi-agent system with self-interested agents acting in an uncertain environment, each with private actions, states and rewards. There is also a social planner with its own actions, rewards, and states, acting as a coordinator and able to influence the agents via actions (e.g., resource allocations). Agents can only communicate with the center, but may become inaccessible, e.g., when their communication device fails. When accessible to the center, agents can report their local state (and models) and receive recommendations from the center about local policies to follow for the present period and also, should they become inaccessible, until becoming accessible again. Without self-interest, this poses a new problem class which we call partially-synchronized DEC-MDPs, and for which we establish some positive complexity results under reasonable assumptions. Allowing for self-interested agents, we are able to bridge to methods of dynamic mechanism design, aligning incentives so that agents truthfully report local state when accessible and choose to follow the prescribed “emergency policies” of the center.