Federated Control with Hierarchical Multi-Agent Deep Reinforcement Learning

We present a framework combining hierarchical and multi-agent deep learning approaches to solve coordination problems among a of agents using a semi-decentralized model. The framework extends the-agent learning setup by introducing a meta-controller that guides the between agent pairs, enabling agents to focus on communicating only one other agent at any step. This hierarchical decomposition of the allows for efficient exploration to learn policies that identify globally solutions even as the number of collaborating agents increases. We show initial experimental results on a simulated distributed scheduling.