PolicyBoost: Functional Policy Gradient with Ranking-based Reward Objective

Learning policies in nonlinear representations is an important step toward real-world applications of reinforcement learning in robotics. While functional representation has been widely applied in state-of-the-art supervised learning techniques (as known as boosting approaches) to adaptively learn nonlinear functions, in reinforcement learning the boosting-style approaches have been little investigated. Only a few pieces of work explored in this direction, which however may suffer from the occurring-probability-pursuing problem. In this paper, to alleviate the problem, we propose to employ a ranking-based objective function to guide the policy search in a function space, resulting in the PolicyBoost approach. Experiment results verify the effectiveness as well as the robustness of the PolicyBoost.