Informed Sampling of Prioritized Experience Replay

Experience replay an essential role as an information-generating mechanism plays in reinforcement learning systems that use neural networks as function approximators. It enables the artificial learning agents to store their past experiences in a sliding-window butter, effectively recycling them in the process of a continual re-training of a neural network. The intermediary process of experience caching opens a possibility for an agent to optimize the order in which the experiences are sampled from the butter. This may improve the default standard, i.e., the stochastic prioritization based on Temporal-Difference error (or TD-error), which focuses on experiences that carry more Temporal-Difference surprise for the approximator. A notion of informed prioritization is proposed, a method relying on fast on-line confidence estimates of approximator predictions in order to be able to dynamically exploit the benefits of TD-error prioritization only when its prediction confidence about the selected experiences increases. The presented informed-stochastic prioritization method of replay butter sampling, implemented as a part of standard staple Deep Q-learning algorithm outperformed the vanilla stochastic prioritization based on TD-error in 41 out of 54 trialed Atari games.