Simultaneous adversarial multi-robot learning

摘要

Multi-robot learning faces all of the challenges of robot learning with all of the challenges of multiagent learning. There has been a great deal of recent research on multiagent reinforcement learning in stochastic games, which is the intuitive extension of MDPs to multiple agents. This recent work, although general, has only been applied to small games with at most hundreds of states. On the other hand robot tasks have continuous, and often complex, state and action spaces. Robot learning tasks demand approximation and generalization techniques, which have only received extensive attention in single-agent learning. In this paper we introduce GraWoLF, a general-purpose, scalable, multiagent learning algorithm. It combines gradient-based policy learning techniques with the WoLF ("Win or Learn Fast") variable learning rate.