An Actor-Critic reinforcement learning algorithm based on adaptive RBF network

We introduce an algorithm of actor-critic reinforcement learning methods in continuous state space. In order to cope with large-scale or continuous state spaces, the algorithm utilizes applied radial basis function (RBF) neural network to approximate the state value function. By training self-adapted non-linear processing unit, realizing online adaptive reconstructing of state space, the approximation is improved. In order to improve the efficient of exploration, a hybrid exploration strategy is proposed. Experimental studies concerning a mountain-car control task illustrate the performance and applicability of the proposed algorithm.