README.md

CORMPO: Clinically-aware OOD-regularized Model-based Policy Optimization

Overview

This repository includes an offline RL algorithm, CORMPO, and a digital twin environment for RL evaluation. CORMPO addresses out-of-distribution (OOD) challenges in offline reinforcement learning by incorporating clinical domain knowledge and regularization techniques for safer policy optimization. See the paper: Guardian-regularized Safe Offline Reinforcement Learning for Smart Weaning of Mechanical Circulatory Devices.

Dependencies / Installation

Install all required dependencies:

pip install -r requirements.txt

Usage

MCS Digital Twin and RL Environment

See the README in the abiomed_env folder for environment implementation details and example scripts for using the environment.

CORMPO Training

Train CORMPO with WS penalty on noiseless synthetic dataset:

python cormpo/mbpo_kde/mopo.py --config cormpo/config/noiseless_synthetic/mbpo_kde_ws.yaml

on noiseless synthetic dataset:

python cormpo/mbpo_kde/mopo.py --config cormpo/config/noisy_synthetic/mbpo_kde.yaml

CORMPO Policy Evaluation

Evaluate a saved policy trained on noisy synthetic dataset:

python cormpo/helpers/evaluate.py --config cormpo/config/evaluate/noisy/cormpo.yaml --policy_path "checkpoints/policy/noisy_synthetic/policy_abiomed.pth"

To evaluate the policy trained on noiseless dataset, change policy_path to:

--policy_path "checkpoints/policy/noiseless_synthetic/policy_abiomed.pth"

Reference

• The implementation of MOPO and MBPO-KDE is built largely on this implementation of MOPO algorithm: https://github.com/junming-yang/mopo