/LAP-PAL

Author's PyTorch implementation of LAP and PAL with TD3 and DDQN

Primary LanguagePythonMIT LicenseMIT

An Equivalence between Loss Functions and Non-Uniform Sampling in Experience Replay

PyTorch implementation of Loss-Adjusted Prioritized (LAP) experience replay and Prioritized Approximation Loss (PAL). LAP is an improvement to prioritized experience replay which eliminates the importance sampling weights in a principled manner, by considering the relationship to the loss function. PAL is a uniformly sampled loss function with the same expected gradient as LAP.

The paper will be presented at NeurIPS 2020. Code is provided for both continuous (with TD3) and discrete (with DDQN) domains.

Bibtex

@article{fujimoto2020equivalence,
  title={An Equivalence between Loss Functions and Non-Uniform Sampling in Experience Replay},
  author={Fujimoto, Scott and Meger, David and Precup, Doina},
  journal={arXiv preprint arXiv:2007.06049},
  year={2020}
}