Learning strategy for continuous robot visual control: A multi-objective perspective
作者：
Highlights：
• We formulate the robot visual control as a MOP by taking three conflicting general objectives into consideration and proposing a DRL framework with improved learning efficiency and convergence.
• To improve learning efficiency, a Takagi–Sugeno (T–S) fuzzy tile state representation is designed to construct discrete state variables from the continuous observations.
• In this paper, we adopt the cerebellar model (CM) as the policy evaluation model, given its proved performance of fast learning and low computation to approximate an arbitrary nonlinear function. Moreover, we develop an effective strategy to automatically tune the learning rate of the FCAC method.
摘要
•We formulate the robot visual control as a MOP by taking three conflicting general objectives into consideration and proposing a DRL framework with improved learning efficiency and convergence.•To improve learning efficiency, a Takagi–Sugeno (T–S) fuzzy tile state representation is designed to construct discrete state variables from the continuous observations.•In this paper, we adopt the cerebellar model (CM) as the policy evaluation model, given its proved performance of fast learning and low computation to approximate an arbitrary nonlinear function. Moreover, we develop an effective strategy to automatically tune the learning rate of the FCAC method.
论文关键词：Robot visual control,Deep reinforcement learning,Multi-objective optimization problem,Fuzzy Cerebellar Actor–critic
论文评审过程：Received 12 April 2022, Revised 27 June 2022, Accepted 8 July 2022, Available online 23 July 2022, Version of Record 2 August 2022.
论文官网地址：https://doi.org/10.1016/j.knosys.2022.109448
