In machine learning with large or continuous state space, it is a hot topic to combine the function approximation and reinforcement learning. The study also faces a very difficult problem of how to balance the exploration and exploitation in reinforcement learning. In allusion to the exploration and exploitation dilemma in the large or continuous state space, this paper presents a novel policy iteration algorithm based on Gaussian process in deterministic environment. The algorithm uses Gaussian process to model the action-value function, and in conjunction with generative model, obtains the posteriori distribution of the parameter vector of the action-value function by Bayesian inference. During the learning process, it computes the value of perfect information according to the posteriori distribution, and then selects the appropriate action with respect to the expected value of the action-value function. The algorithm achieves the balance between exploration and exploitation to certain extent, and therefore accelerates the convergence. The experimental results on the Mountain Car problem show that the algorithm has faster convergence rate and better convergence performance.