xuzhiyuan1528 / KTM-DRL

Knowledge Transfer in Multi-Task Deep Reinforcement Learning for Continuous Control

This repository is the official implementation of KTM-DRL.

Dependencies

• Python 3.7.7 (Optional, also works with older versions)
• PyTorch 1.5.1
• MuJoCo 200
• mujoco-py 2.0.2.5
• OpenAI Gym 0.17.2
• gym-extensions (Optional, only for HalfCheetah task group)

Evaluation

To use the pre-trained models of task-specific teachers and multi-task agent on HalfCheetah task group, download from the Dropbox Link and put the half folder into ./model/ or /your_own_path/.

To evaluate the multi-task agent and its corresponding task-specific teachers, run this command:

python3 eval-ktm.py --seed 0 --cfg ./config/HALF/cfg-mt-half-eval.json --dir ./model/half --name EVA

Training

To train the model(s) in the paper, run this command:

python3 train-ktm.py --seed 0 --cfg ./config/HALF/cfg-mt-half-train.json --dir ./model/half --name TRN

Bibtex

@article{xu2020knowledge,
  title={Knowledge Transfer in Multi-Task Deep Reinforcement Learning for Continuous Control},
  author={Xu, Zhiyuan and Wu, Kun and Che, Zhengping and Tang, Jian and Ye, Jieping},
  journal={arXiv preprint arXiv:2010.07494},
  year={2020}
}