Baconian: Boosting model-based reinforcement learning

Baconian [beˈkonin] is a toolbox for model-based reinforcement learning with user-friendly experiment setting-up, logging and visualization modules developed by CAP. We aim to develop a flexible, re-usable and modularized framework that can allow the users to easily set-up their model-based RL experiments by reusing modules we provide.

Installation

You can easily install by (with python 3.5/3.6/3.7, Ubuntu 16.04/18.04):

# install tensorflow with/without GPU based on your machine
pip install tensorflow-gpu==1.15.2
# or
pip install tensorflow==1.15.2 

pip install baconian

For more advance usage like using Mujoco environment, please refer to our documentation page.

Release news:

• 2020.04.29 v0.2.2 Fix some memory issues in SampleData module, and simplify some APIs.
• 2020.02.10 We are including external reward & terminal function of Gym/mujoco tasks with well-written documents.
• 2020.01.30 Update some dependent packages versions, and release some preliminary benchmark results with hyper-parameters.

For previous news, please go here

Documentation

We support python 3.5, 3.6, and 3.7 with Ubuntu 16.04 or 18.04. Documentation is available at http://baconian-public.readthedocs.io/

Algorithms Reference:

Model-based:

1. Dyna

Sutton, Richard S. "Dyna, an integrated architecture for learning, planning, and reacting." ACM Sigart Bulletin 2.4 (1991): 160-163.

2. LQR

Abbeel, P. "Optimal Control for Linear Dynamical Systems and Quadratic Cost (‘LQR’)." (2012).

3. iLQR

Abbeel, P. "Optimal Control for Linear Dynamical Systems and Quadratic Cost (‘LQR’)." (2012).

4. MPC

Garcia, Carlos E., David M. Prett, and Manfred Morari. "Model predictive control: theory and practice—a survey." Automatica 25.3 (1989): 335-348.

5. Model-ensemble

Kurutach, Thanard, et al. "Model-ensemble trust-region policy optimization." arXiv preprint arXiv:1802.10592 (2018).

Model-free

1. DQN

Mnih, Volodymyr, et al. "Playing atari with deep reinforcement learning." arXiv preprint arXiv:1312.5602 (2013).

2. PPO

Schulman, John, et al. "Proximal policy optimization algorithms." arXiv preprint arXiv:1707.06347 (2017).

3. DDPG

Lillicrap, Timothy P., et al. "Continuous control with deep reinforcement learning." arXiv preprint arXiv:1509.02971 (2015).

Algorithms in Progress

1. Random Shooting

Rao, Anil V. "A survey of numerical methods for optimal control." Advances in the Astronautical Sciences 135.1 (2009): 497-528.

2. MB-MF

Nagabandi, Anusha, et al. "Neural network dynamics for model-based deep reinforcement learning with model-free fine-tuning." 2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2018.

3. GPS

Levine, Sergey, et al. "End-to-end training of deep visuomotor policies." The Journal of Machine Learning Research 17.1 (2016): 1334-1373.

Acknowledgement

Thanks to the following open-source projects:

• garage: https://github.com/rlworkgroup/garage
• rllab: https://github.com/rll/rllab
• baselines: https://github.com/openai/baselines
• gym: https://github.com/openai/gym
• trpo: https://github.com/pat-coady/trpo

Citing Baconian

If you find Baconian is useful for your research, please consider cite our demo paper here:

@article{
linsen2019baconian, 
title={Baconian: A Unified Opensource Framework for Model-Based Reinforcement Learning}, 
author={Linsen, Dong and Guanyu, Gao and Yuanlong, Li and Yonggang, Wen}, 
journal={arXiv preprint arXiv:1904.10762},
year={2019} 
}

Report an issue

If you find any bugs on issues, please open an issue or send an email to me ([email protected]) with detailed information. I appreciate your help!