A Hybrid Transfer Algorithm for Reinforcement Learning Based on Spectral Method

Abstract

For scaling up state space transfer underlying the proto-value function framework, only some basis functions corresponding to smaller eigenvalues are transferred effectively, which will result in wrong approximation of value function in the target task. In order to solve the problem, according to the fact that Laplacian eigenmap can preserve the local topology structure of state space, an improved hierarchical decomposition algorithm based on the spectral graph theory is proposed and a hybrid transfer method integrating basis function transfer with subtask optimal polices transfer is designed. At first, the basis functions of the source task are constructed using spectral method. The basis functions of target task are produced through linearly interpolating basis functions of the source task. Secondly, the produced second basis function of the target task (approximating Fiedler eigenvector) is used to decompose the target task. Then the optimal polices of subtasks are obtained using the improved hierarchical decomposition algorithm. At last, the obtained basis functions and optimal subtask polices are transferred to the target task. The proposed hybrid transfer method can directly get optimal policies of some states, reduce the number of iterations and the minimum number of basis functions needed to approximate the value function. The method is suitable for scaling up state space transfer task with hierarchical control structure. Simulation results of grid world have verified the validity of the proposed hybrid transfer method.