Advanced Policy Gradient Methods
Michael Y. Hu · Apress eBooks · 2023
One of the primary challenges associated with policy gradient methods is their instability and sensitivity to hyperparameters, such as the learning rate. This can lead to oscillations in the agent’s performance, resulting in slow convergence or even divergence. Furthermore, these methods often suffer from high variance in gradient estimates, which hampers convergence speed. Moreover, standard policy gradient methods exhibit poor sample efficiency, as they utilize training samples only once for learning purposes.