Policy Reuse in Reinforcement Learning for Modular Agents

摘要

We present reusable policy method for modular reinforcement learning problem in continuous state space. Our method relies on two-layered learning architecture. The first layer partitions the agent's problem space into n-folds sub-agents that are inter-connected with each other with dexterity identical to original problem. It further learns a local control policy for standalone 1-fold sub-agent. The second layer learns a global policy to reuse `already learnt' standalone local policy over each n sub-agents by sampling local policy with global parameters for each sub-agent-parameterizing local policy independently to approximate non-linear interconnections between sub-agents. We demonstrate our method on simulation example of 12-DOF modular robot that learns maneuver pattern of snake-like gait. We also compare our proposed method against standard single-policy learning methods to benchmark optimality.