A Linear Online Guided Policy Search Algorithm

摘要

In reinforcement learning (RL), the guided policy search (GPS), a variant of policy search method, can encode the policy directly as well as search for optimal solutions in the policy space. Even though this algorithm is provided with asymptotic local convergence guarantees, it can not work in a online way for conducting tasks in complex environments since it is trained with a batch manner which requires that all of the training samples should be given at the same time. In this paper, we propose an online version for GPS algorithm, which can learn policies incrementally without complete knowledge of initial positions for training. The experiments witness its efficacy on handling sequentially arriving training samples in a peg insertion task.