On Efficient Multi-Agent Reinforcement Learning for Large Scale Supply Demand Matching

The current technology advances in renewable power generation, fuel cell, and storage devices push forward the growing research interest on supply demand matching in smart grid, which could involve hundreds of thousands of agents making decisions in real time. In this multi-agent reinforcement learning (MARL) problem, the key questions are what to share and how to utilize the shared information. These questions remain open due to the complexity and difference among the subproblems of each agent. We consider this important problem in this work, and make the following major contributions. First, we formulate the problem to maximize the probability of correctly selecting (PCS) the best action for each agent under limited sampling budget. Second, we clarify what to be shared among the agents and quantify the value of such shared information. Third, we present an algorithm to asymptotically maximize the PCS. We hope this work shed light on MARL in general.