An efficient reinforcement learning algorithm for continuous actions
Fu Bo, Xin Chen, Yong He, Min Wu · 2013
In this paper a fast and effective reinforcement learning algorism named Dyna-CA in which the learning agent or agents can get the continuous action has been proposed to get the generalization of reinforcement learning methods to large-scale or continuous space. Firstly, the set of k states around the current state will be observed and the probability distribution of k states on condition current state can be calculated by functional mapping. Secondly the selection action-making in the current state for agent is recommended the weighted sum of best actions taken in the neighbor states to guarantee the learning with continuous actions. Then the Q value will be updated by the rules of Dyna algorism, which are not only based on the current neighbor states' practical knowledge but also the priori neighbor states' experience. Computer simulations involving the Maze and Acrobat problems illustrate the validity of the proposed reinforcement learning method and fast convergence in learning an optimal policy.