Indexed by:Conference Paper

Date of Publication:2018-01-01

Included Journals:CPCI-S

Volume:11012

Page Number:191-203

Key Words:RL; Robotic control; Multiagent learning; CG

Abstract:Reinforcement learning is widely used to learn complex behaviors for robotics. However, due to the high-dimensional state/action spaces, reinforcement learning usually suffers from slow learning speed in robotic control applications. A feasible solution to this challenge is to utilize structural decomposition of the control problem and resort to decentralized learning methods to expedite the overall learning process. In this paper, a multiagent reinforcement learning approach is proposed to enable decentralized learning of component behaviors for a robot that is decomposed as a coordination graph. By using this approach, all the component behaviors are learned in parallel by some individual reinforcement learning agents and these agents coordinate their behaviors to solve the global control problem. The approach is validated and analyzed in two benchmark robotic control problems. The experimental validation provides evidence that the proposed approach enables better performance than approaches without decomposition.