Teaching Coordination to Selfish Learning Agents in Resource-Constrained Partially Observable Markov Games

Of increasing relevance to engineering systems are problems that include online resource allocation to agents that feature adaptation and learning capabilities. This paper considers the case where a coordinator gets to design a resource allocation mechanism (i.e. a bidding-allocation-rewards protocol) to efficiently allocate a resource to selfish agents that try to gain access by learning to communicate strategically. Towards aligning the agents' incentives with the social objective, a critical-value-based mechanism is proposed. Analytic results are presented for a simple, stylized setting, while simulation results for a use case with reinforcement learning agents controlling flexible loads in the smart grid demonstrate the mechanism's ability to teach coordinated behavior to the distributed learners.