Evolutionary reinforcement learning and its application in robot path tracking

摘要

The control policy of robot path-tracking based on adaptive heuristic ctritic(AHC) reinforcement learning is researched. The adaptive critic element(ACE)of AHC is composed of a multi-layer feedforward network. TD(λ) algorithm and gradient descent algorithm are integrated, which is used to update the weights of network. The output of the ACE generates the secondary reinforcement signal which can direct the learning of the action select element (ASE). ASE can be implemented by the fuzzy inference system (FIS) which is optimized by using the genetic algorithms. Finally, the method is used for learning the robot behavior. The experiment shows that the scheme can effectively solve the problem of the robot path-tracking.