https://github.com/vermouth1992/bracp

Improved Behavior Regularized Offline Reinforcement Learning
https://github.com/vermouth1992/bracp

Last synced: about 2 months ago
JSON representation

Improved Behavior Regularized Offline Reinforcement Learning

• Host: GitHub
• URL: https://github.com/vermouth1992/bracp
• Owner: vermouth1992
• License: apache-2.0
• Created: 2020-10-03T02:45:11.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2021-12-08T10:02:43.000Z (about 3 years ago)
• Last Synced: 2023-03-03T03:41:40.574Z (almost 2 years ago)
• Language: Jupyter Notebook
• Homepage:
• Size: 1.03 MB
• Stars: 4
• Watchers: 1
• Forks: 1
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# BRAC+: Improved Behavior Regularized Offline Reinforcement Learning

This repository is the official implementation of [BRAC+: Improved Behavior Regularized Actor Critic for Offline Reinforcement Learning](https://arxiv.org/abs/2110.00894).

## Requirements

We high recommend that you create a new Python environment to test our code

#### Conda Environment

```shell

conda create -n bracp python=3.8

```

To install requirements:

#### Python package

```shell

pip install -r requirements.txt

```

#### D4RL library

```shell

pip install git+https://github.com/rail-berkeley/d4rl@master#egg=d4rl

```

#### rlutils library

```shell

pip install rlutils-python==0.0.3

```

## Training

```shell

python d4rl_bracp.py train --env_name halfcheetah-medium-v0 --seed 110

```

The script will first pretrain the behavior policy and the initial policy that minimize the KL divergence.

## Logging

The logs will be placed at data/d4rl_results/