Dynamic Multi-agent Reinforcement Learning for Control Optimization

In this paper we analyze the use of Reinforcement Learning (RL) in control optimization within dynamic multiagent systems. RL is an effective algorithm for single agent optimization but performs less well in dynamic multi-agent environments. We investigate this principle based upon three of the most common RL algorithms. We also introduce a novel RL algorithm that excels in both single agent optimization and adaptation within multi-agent environments. This algorithm takes into account not only its own current state but also the current states of each of its significant neighbor agents so as to significantly increase performance within multi-agent systems. It employs a model driven approach to facilitate effective adaptation as well as policy-based methods to enable efficient action selection.