Efficient Distributed Reinforcement Learning Through Agreement

Distributed robotic systems can benefit from automatic controller design and online adaptation by reinforcement learning (RL), but often suffer from the limitations of partial observability. In this paper, we address the twin problems of limited local experience and locally observed but not necessarily telling reward signals encountered in such systems. We combine direct search in policy space with an agreement algorithm to efficiently exchange local rewards and experience among agents. We demonstrate improved learning ability on the locomotion problem for self-reconfiguring modular robots in simulation, and show that a fully distributed implementation can learn good policies just as fast as the centralized implementation. Our results suggest that prior work on centralized RL algorithms for modular robots may be made effective in practice through the application of agreement algorithms. This approach could be fruitful in many cooperative situations, whenever robots need to learn similar behaviors, but have access only to local information.