Efficient Continuous Control with Double Actors and Regularized Critics

This repo contains our work in AAAI 2022, Efficient Continuous Control with Double Actors and Regularized Critics.

Overview

DDPG is the fine-tuned version of vanilla DDPG which could achieve much better performance than DDPG at various environments. TD3 use fine-tuned DDPG as the baselines, and so we do in our work. The implementation of DARC is based on the open-source TD3 codebase.

We use main.py to run results where DDPG.py along with TD3.py are served as baselines and DARC.py is the core file for our work. We use seeds 1-5 for all algorithms during training and different seeds (the current seed + 100) during evaluation (see run.sh for more details).

Evaluate True Value

One need to set the sampled state as the initial state in MuJoCo to evaluate the true value. Please refer to openai/gym#1617 for details.

Requirements

• python: 3.7.9
• mujoco_py: 2.0.2.13
• torch: 1.8.0
• gym: 0.18.0
• box2d-py
• pybulletgym

Install PybulletGym

Please refer to the open-source implementation of pybulletgym here

Before installing pybullet, make sure that you have gym installed. Then run the following commands to install pybulletgym.

  git clone https://github.com/benelot/pybullet-gym.git
  cd pybullet-gym
  pip install -e .

Use PybulletGym

import pybulletgym

For detailed environments in pybulletgym, please refer here.

Usage

Utilize GPUs to accelerate training if available

export CUDA_VISIBLE_DEVICES=1

Run the following commands to reproduce results in the submission

Reproduce results in the submission

./run.sh

Run DARC

python main.py --env <environment_name> --save-model --policy DARC --dir ./logs/DARC/r1 --seed 1 --qweight 0.1 --reg 0.005

Run DDPG/TD3/DADDPG/DATD3

python main.py --env <environment_name> --seed 1 --policy <algorithm_name> --dir './logs/' --save-model

Citation

If you find our work helpful, please consider cite our work.

@inproceedings{Efficient2022Lyu,
  title={Efficient Continuous Control with Double Actors and Regularized Critics},
  author={Jiafei Lyu and Xiaoteng Ma and Jiangpeng Yan and Xiu Li},
  booktitle={Thirty-sixth AAAI Conference on Artificial Intelligence},
  year={2022},
}