mamba413/room

Robust Offline Reinforcement Learning with Heavy-Tailed Rewards

Robust Offline Reinforcement Learning with Heavy-Tailed Rewards

Reproducible code for the paper: Robust Offline Reinforcement Learning with Heavy-Tailed Rewards

Summary of the paper

This paper endeavors to augment the robustness of offline reinforcement learning (RL) in scenarios laden with heavy-tailed rewards, a prevalent circumstance in real-world applications. We propose two algorithmic frameworks, ROAM and ROOM, for robust off-policy evaluation (OPE) and offline policy optimization (OPO), respectively. Central to our frameworks is the median-of-means (MM) method. Our key insight is that employing MoM to offline RL does more than just tackle heavy-tailed rewards—it offers valid uncertainty quantification to address insufficient coverage issue in offline RL as well.

Below it is the numerical performance of our proposal (ROOM-VM & P-ROOM-VM) on the d4rl benchmarked dataset:

File structure

1. requirement.txt: prerequisite python libraries

2. Cartpole directory: code for reproducing results in Figures 3, 4, 6

   • _density directory: functions for estimating the density ratio in marginalize importance sampling based methods
   • _RL directory: employ MM in the TD update in fitted Q-iteration/evaluation based algorithms (Algorithms 4-5)
   • _MM_OPE.py: Algorithm 1 and its variant (ROAM-variant)
   • _MM_OPE.py: Algorithm 2 and its pessimistic variant (P-ROOM)
   • _PB_OPO.py: Bootstrap based variant for OPE.
   • eval_cartpole.py: reproduce Figures 3(a), 4, 6
   • optimize_cartpole.py: reproduce Figures 3(b)

3. SQL:

   • src directory: implement the sparse Q-learning (SQL) for
   • main_SQL.py: the main file for conducting numerical studies for SQL. (reproduce Figure 5)

4. SAC-N:

   • SACN.py directory: implement the soft-actor critic (SAC) of $N$ ensemble.
   • main_SACN.py: the main file for conducting numerical studies for SACN. (reproduce Figure A3)

Citation

@article{zhu2023robust,
  title={Robust Offline Policy Evaluation and Optimization with Heavy-Tailed Rewards},
  author={Zhu, Jin and Wan, Runzhe and Qi, Zhengling and Luo, Shikai and Shi, Chengchun},
  journal={arXiv preprint arXiv:2310.18715},
  year={2023}
}

Reference

• Offline RL with No OOD Actions: In-Sample Learning via Implicit Value Regularization, ICLR (2023)

• Uncertainty-Based Offline Reinforcement Learning with Diversified Q-Ensemble, NeurIPS (2021)