Model Based Reinforcement Learning Benchmarking Library (MBBL)
Introduction
Arxiv Link Project Page Abstract: Model-based reinforcement learning (MBRL) is widely seen as having the potential to be significantly more sample efficient than model-free RL. However, research in model-based RL has not been very standardized. It is fairly common for authors to experiment with self-designed environments, and there are several separate lines of research, which are sometimes closed-sourced or not reproducible. Accordingly, it is an open question how these various existing MBRL algorithms perform relative to each other. To facilitate research in MBRL, in this paper we gather a wide collection of MBRL algorithms and propose over 18 benchmarking environments specially designed for MBRL. We benchmark these MBRL algorithms with unified problem settings, including noisy environments. Beyond cataloguing performance, we explore and unify the underlying algorithmic differences across MBRL algorithms. We characterize three key research challenges for future MBRL research: the dynamics coupling effect, the planning horizon dilemma, and the early-termination dilemma.
Installation
Install the project with pip from the top-level directory:
pip install --user -e .
For sub-packages of algorithms not integrated here, please refer to the respective readmes.
Algorithms
Some of the algorithms are not yet merged into this repo. We use the following colors to represent their status.
indicates
Merged into this repo
.
indicates
In a separate repo.
Shooting Algorithms
Rao, Anil V. "A survey of numerical methods for optimal control." Advances in the Astronautical Sciences 135.1 (2009): 497-528. Link
python main/rs_main.py --exp_id rs_gym_cheetah_seed_1234 \
--task gym_cheetah \
--num_planning_traj 1000 --planning_depth 10 --random_timesteps 10000 \
--timesteps_per_batch 3000 --num_workers 20 --max_timesteps 200000 --seed 1234
The following script will test the performance when using ground-truth dynamics:
python main/rs_main.py --exp_id rs_${env_type}\
--task gym_cheetah \
--num_planning_traj 1000 --planning_depth 10 --random_timesteps 0 \
--timesteps_per_batch 1 --num_workers 20 --max_timesteps 20000 \
--gt_dynamics 1
Also, set --check_done 1
for agents to detect if the episode is terminated (needed for gym_fant, gym_fhopper).
Nagabandi, Anusha, et al. "Neural network dynamics for model-based deep reinforcement learning with model-free fine-tuning." arXiv preprint arXiv:1708.02596 (2017). Link
python main/mbmf_main.py --exp_id mbmf_gym_cheetah_ppo_seed_1234 \
--task gym_cheetah --trust_region_method ppo \
--num_planning_traj 5000 --planning_depth 20 --random_timesteps 1000 \
--timesteps_per_batch 1000 --dynamics_epochs 30 \
--num_workers 20 --mb_timesteps 7000 --dagger_epoch 300 \
--dagger_timesteps_per_iter 1750 --max_timesteps 200000 \
--seed 1234 --dynamics_batch_size 500
Chua, K., Calandra, R., McAllister, R., & Levine, S. (2018). Deep reinforcement learning in a handful of trials using probabilistic dynamics models. In Advances in Neural Information Processing Systems (pp. 4754-4765). Link
See the codebase for POPLIN, where you can benchmark PETS-RS and PETS-CEM following the readme. PETS-RS with ground-truth is essentially RS with ground-truth, and to run the PETS-CEM with ground-truth dynamics:
python main/pets_main.py --exp_id pets-gt-gym_cheetah \
--task gym_cheetah \
--num_planning_traj 500 --planning_depth 30 --random_timesteps 0 \
--timesteps_per_batch 1 --num_workers 10 --max_timesteps 20000 \
--gt_dynamics 1
Policy Search with Backpropagation through Time
Deisenroth, M., & Rasmussen, C. E. (2011). PILCO: A model-based and data-efficient approach to policy search. In Proceedings of the 28th International Conference on machine learning (ICML-11) (pp. 465-472). Link
We implemented and benchmarked the environments in this repo PILCO (will be set public soon).
Tassa, Y., Erez, T., & Todorov, E. (2012, October). Synthesis and stabilization of complex behaviors through online trajectory optimization. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems (pp. 4906-4913). IEEE. Link
python main/ilqr_main.py --exp_id ilqr-gym_cheetah \
--max_timesteps 2000 --task gym_cheetah \
--timesteps_per_batch 1 --ilqr_iteration 10 --ilqr_depth 30 \
--max_ilqr_linesearch_backtrack 10 --num_workers 2 \
--gt_dynamics 1
Levine, Sergey, and Vladlen Koltun. "Guided policy search." International Conference on Machine Learning. 2013 Link
We implemented and benchmarked the environments in this repo GPS.
Heess, N., Wayne, G., Silver, D., Lillicrap, T., Erez, T., & Tassa, Y. (2015). Learning continuous control policies by stochastic value gradients. In Advances in Neural Information Processing Systems (pp. 2944-2952). Link
We implemented and benchmarked the environments in this repo SVG (will be set public soon).
Dyna-Style Algorithms
Kurutach, Thanard, et al. "Model-Ensemble Trust-Region Policy Optimization." arXiv preprint arXiv:1802.10592 (2018). Link
We implemented and benchmarked the environments in this repo ME-TRPO.
Luo, Y., Xu, H., Li, Y., Tian, Y., Darrell, T., & Ma, T. (2018). Algorithmic Framework for Model-based Deep Reinforcement Learning with Theoretical Guarantees. Link
We implemented and benchmarked the environments in this repo SLBO
Clavera, I., Rothfuss, J., Schulman, J., Fujita, Y., Asfour, T., & Abbeel, P. (2018). Model-based reinforcement learning via meta-policy optimization. arXiv preprint arXiv:1809.05214. Link We implemented and benchmarked the environments in this repo MB-MPO (will be set public soon).
Model-free Baselines
Schulman, John, et al. "Trust region policy optimization." International Conference on Machine Learning. 2015. Link
python main/mf_main.py --exp_id trpo_gym_cheetah_seed1234 \
--timesteps_per_batch 2000 --task gym_cheetah \
--num_workers 5 --trust_region_method trpo --max_timesteps 200000
Schulman, John, et al. "Proximal policy optimization algorithms." arXiv preprint arXiv:1707.06347 (2017). Link
python main/mf_main.py --exp_id ppo_gym_cheetah_seed1234 \
--timesteps_per_batch 2000 --task gym_cheetah \
--num_workers 5 --trust_region_method ppo --max_timesteps 200000
Fujimoto, S., van Hoof, H., & Meger, D. (2018). Addressing function approximation error in actor-critic methods. arXiv preprint arXiv:1802.09477. Link
We implemented and benchmarked the environments in this repo TD3.
Haarnoja, T., Zhou, A., Abbeel, P., & Levine, S. (2018). Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. arXiv preprint arXiv:1801.01290. Link
We implemented and benchmarked the environments in this repo SAC.
Disclaimer
As mentioned in the project webpage, it is a developing (unfinished) project. We are working towards a unified package for MBRL algorithms. but it might take a while given that we lack the manpower.
Engineering Stats and 1 Million Performance
Env
Here is available environments and their mappings to the name used in the paper.
Mapping Table | ||||||
---|---|---|---|---|---|---|
Env | Pendulum | InvertedPendulum | Acrobot | CartPole | Mountain Car | Reacher |
Repo-Name | gym_pendulum | gym_invertedPendulum | gym_acrobot | gym_cartPole | gym_mountain | gym_reacher |
Env | HalfCheetah | Swimmer-v0 | Swimmer | Ant | Ant-ET | Walker2D |
Repo-Name | gym_cheetah | gym_swimmer | gym_fswimmer | gym_ant | gym_fant | gym_walker2d |
Env | Walker2D-ET | Hopper | Hopper-ET | SlimHumanoid | SlimHumanoid-ET | Humanoid-ET |
Repo-Name | gym_fwalker2d | gym_hopper | gym_fhopper | gym_nostopslimhumanoid | gym_slimhumanoid | gym_humanoid |