An Equivalence between Loss Functions and Non-Uniform Sampling in Experience Replay
This work provides a theoretical insight for reinforcement learning researchers, potentially enabling new improvements to PER, though it appears incremental as it builds on existing PER techniques.
The paper tackled the problem of understanding the relationship between loss functions and non-uniform sampling in Prioritized Experience Replay (PER) for deep reinforcement learning, showing that PER can be replaced by an equivalent uniformly sampled loss function without performance loss in some environments and proposing effective modifications.
Prioritized Experience Replay (PER) is a deep reinforcement learning technique in which agents learn from transitions sampled with non-uniform probability proportionate to their temporal-difference error. We show that any loss function evaluated with non-uniformly sampled data can be transformed into another uniformly sampled loss function with the same expected gradient. Surprisingly, we find in some environments PER can be replaced entirely by this new loss function without impact to empirical performance. Furthermore, this relationship suggests a new branch of improvements to PER by correcting its uniformly sampled loss function equivalent. We demonstrate the effectiveness of our proposed modifications to PER and the equivalent loss function in several MuJoCo and Atari environments.