Effective strategies for complex skill real-time learning using reinforcement learning

摘要

Following the principle of human skill learning, robot acquiring skill is a process similar to human skill learning. Reinforcement learning is on-line actor critic method for robot to develop its skill. The reinforcement Junction has become the critical component for its effect of evaluating the action and guiding the learning process. A difference form of augmented reward function is considered carefully. In this paper we present a strategy for the task of complex skill learning. Automatic robot shaping policy is to dissolve the complex skill into a hierarchical learning process. Variable resolution discretization of input space is introduced to improve the generalization capability of CMAC-based RL. Conventional e -greedy policy has the shortage of unnecessary randomization. Boltzmann distribution selection is also introduced to the balance of exploration and exploitation. We describe our ideas of reinforcement learning methods and also illustrate with an example the utility of method for learning skilled robot control on line.