Codebase for PERRY: Policy Evaluation with Confidence Intervals using Auxiliary Data
This repository accompanies the paper PERRY: Policy Evaluation with Confidence Intervals using Auxiliary Data, which will appear at TMLR.
├── README.md
├── halfcheetah
│ ├── VLBM.py
│ ├── d4rl.ipynb # OPE
│ ├── d4rl_policies.json
│ ├── dg_b.py # data generation
│ ├── learned_log_w_IS.pkl
│ ├── learned_log_w_PDIS.pkl
│ ├── log_w_CP.pkl
│ ├── log_w_IS.pkl
│ ├── log_w_PDIS.pkl
│ ├── saved_model # VAE checkpoint
│ └── utils.py
├── inventory
│ └── inventory.ipynb # Simulator & OPE
├── mimic-iv
│ ├── VLBM.py
│ ├── dg_b.py # data generation
│ ├── mimic.ipynb # OPE
│ ├── mimic_iv_behavior_trajectories.pkl
│ ├── mimic_iv_target_trajectories.pkl
│ ├── mimic_iv_trajectories_behavior.pkl
│ ├── mimic_iv_trajectories_target.pkl
│ ├── saved_model # VAE checkpoint and behavior cloning policies
│ ├── skip_list.pkl
│ └── utils.py
└── sepsis
├── mdptoolboxSrc
├── sepsis.ipynb # OPE
└── sepsisSimDiabetes # Simulator
If you use this code, please cite:
@misc{mandyam2025perrypolicyevaluationconfidence,
title={PERRY: Policy Evaluation with Confidence Intervals using Auxiliary Data},
author={Aishwarya Mandyam and Jason Meng and Ge Gao and Jiankai Sun and Mac Schwager and Barbara E. Engelhardt and Emma Brunskill},
year={2025},
eprint={2507.20068},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2507.20068},
}