RICE: Breaking Through the Training Bottlenecks of Reinforcement Learning with Explanation

This repo contains the code for the paper "RICE: Breaking Through the Training Bottlenecks of Reinforcement Learning with Explanation"

Paper citation:

@inproceedings{cheng2024rice,
title={RICE: Breaking Through the Training Bottlenecks of Reinforcement Learning with Explanation},
author={Cheng, Zelei and Wu, Xian and Yu, Jiahao and Yang, Sabrina and Wang, Gang and Xing, Xinyu},
booktitle={Proc. of ICML},
year={2024}

Requirement

The codebase is written with python3.7 and Pytorch. We provide a requirements.txt for your reference. If you run errors in some programs, install the missing lib via pip install as the error report.

Introduction

Basics

• We implement our methods in four dense/sparse mujoco games and four real-world applications.

• Note that our implementation of the explanation method is based on StateMask.

• For each game, we provide code for training a mask network and refining the target agent with an explanation.