The importance of variance reduction in policy gradient method

Reinforcement learning (RL) has been applied to a wide range of motion control problems in robotics. In particular, policy gradient method (PGM) emerges as a powerful subset of RL that can learn effectively from one's experience. However, when the dynamics is stochastic and is short of samples for learning, the performance of PGM becomes inconsistent and heavily relies on the tweaking of the learning rate. In this work, we argue that this degeneration is mainly due to the high variance in the gradient. Through theoretical justifications, simulations and experiments, we verify that by applying a variance suppression, which is called local baseline, on the gradient, PGM can then be applied to some previously untouchable problems.