Abstract
Recently, reinforcement learning with deep neural networks has achieved great success in challenging continuous control problems such as 3D locomotion and robotic manipulation. However, in real-world control problems, the actions one can take are bounded by physical constraints, which introduces a bias when the standard Gaussian distribution is used as the stochastic polic In this work, we propose to use the Beta distribution as an alternative and analyze the bias and variance of the policy gradients of both poli cies. We show that the Beta policy is bias-free and provides significantly faster convergence and higher scores over the Gaussian policy when both are used with trust region policy optimization (TRPO) and actor critic with experience replay (ACER), the state-of-the-art onand offpolicy stochastic methods respectively, on OpenAI Gym’s and MuJoCo’s continuous control environments.