CORL: Offline Reinforcement Learning Library

[casti...@gmail.com]

Happy to announce CORL — a library that provides high-quality single-file implementations of Deep Offline Reinforcement Learning algorithms and uses Weights and Biases to track experiments.

• SOTA algorithms (Decision Transformer, AWAC, BC, CQL, IQL, TD3+BC, SAC-N, EDAC)

• Benchmarked on widely used D4RL datasets (results match performances reported in the original papers, sometimes even with better results)

• Configs with hyperparameters for better reproduction

• Weights&Biases logs for all of the experiments (so that you don’t have to solely rely on final performances from papers)

github: https://github.com/tinkoff-ai/CORL
paper: https://arxiv.org/abs/2210.07105 (accepted at NeurIPS, 3rd Offline RL Workshop)

Thanks,

-----------------------

Vladislav Kurenkov
--
Research Scientist
Tinkoff
https://vkurenkov.me/