A New and Effective Reinforcement Learning based on Tentative Q learning and Knowledge Transfer
Abstract— Aiming at the problem of slow learning speed of reinforcement learning, TQL-RWKT-RRL algorithm is put forward, which is based on tentative Q learning and knowledge transfer. Tentative Q learning increases times of exploration in each iteration, and improves updating method of Q value function. Knowledge transfer algorithm realizes knowledge transfer under different state space based on the method of rolling windows. The path planning experiences in the simple small environment is transferred to more complex and larger state space, which speeds up robot path planning learning speed in large and more complex environment.
Index Terms— reinforcement learning, knowledge transfer, rolling windows, robot path-planning
Duan Jun-Hua, Zhu Yi-An, Zhong Dong, Zhan Tao, Luo Shuyan
School of Computer, Northwestern Polytechnical University, CHINA
Cite: Duan Jun-Hua, Zhu Yi-An, Zhong Dong, Zhan Tao, Luo Shuyan, "A New and Effective Reinforcement Learning based on Tentative Q learning and Knowledge Transfer," Proceedings of 2018 the 8th International Workshop on Computer Science and Engineering, pp. 276-281, Bangkok, 28-30 June, 2018.