An Analysis of Measure-Valued Derivatives for Policy Gradients

Joao Carvalho, Jan Peters

Reinforcement learning methods for robotics are increasingly successful due to the constant development of better policy gradient techniques. A precise (low variance) and accurate…