A collection of trained Reinforcement Learning (RL) agents, with tuned hyperparameters, using Stable Baselines.

We are looking for contributors to complete the collection!

Goals of this repository:

1. Provide a simple interface to train and enjoy RL agents
2. Benchmark the different Reinforcement Learning algorithms
3. Provide tuned hyperparameters for each environment and RL algorithm
4. Have fun with the trained agents!

If the trained agent exists, then you can see it in action using:

python enjoy.py --algo algo_name --env env_id

For example, enjoy A2C on Breakout during 5000 timesteps:

python enjoy.py --algo a2c --env BreakoutNoFrameskip-v4 --folder trained_agents/ -n 5000

If you have trained an agent yourself, you need to do:

# exp-id 0 corresponds to the last experiment, otherwise, you can specify another ID
python enjoy.py --algo algo_name --env env_id -f logs/ --exp-id 0

To load the best model (when using evaluation environment):

python enjoy.py --algo algo_name --env env_id -f logs/ --exp-id 1 --load-best

The hyperparameters for each environment are defined in hyperparameters/algo_name.yml.

If the environment exists in this file, then you can train an agent using:

python train.py --algo algo_name --env env_id

For example (with tensorboard support):

python train.py --algo ppo2 --env CartPole-v1 --tensorboard-log /tmp/stable-baselines/

Evaluate the agent every 10000 steps using 10 episodes for evaluation:

python train.py --algo sac --env HalfCheetahBulletEnv-v0 --eval-freq 10000 --eval-episodes 10

Save a checkpoint of the agent every 100000 steps:

python train.py --algo td3 --env HalfCheetahBulletEnv-v0 --save-freq 100000

Continue training (here, load pretrained agent for Breakout and continue training for 5000 steps):

python train.py --algo a2c --env BreakoutNoFrameskip-v4 -i trained_agents/a2c/BreakoutNoFrameskip-v4.pkl -n 5000

Note: when training TRPO, you have to use mpirun to enable multiprocessing:

mpirun -n 16 python train.py --algo trpo --env BreakoutNoFrameskip-v4

We use Optuna for optimizing the hyperparameters.

Note: hyperparameters search is not implemented for ACER and DQN for now. when using SuccessiveHalvingPruner ("halving"), you must specify --n-jobs > 1

Budget of 1000 trials with a maximum of 50000 steps:

python train.py --algo ppo2 --env MountainCar-v0 -n 50000 -optimize --n-trials 1000 --n-jobs 2 \
  --sampler tpe --pruner median

You can specify in the hyperparameter config one or more wrapper to use around the environment:

for one wrapper:

env_wrapper: gym_minigrid.wrappers.FlatObsWrapper

for multiple, specify a list:

env_wrapper:
    - utils.wrappers.DoneOnSuccessWrapper:
        reward_offset: 1.0
    - utils.wrappers.TimeFeatureWrapper

Note that you can easily specify parameters too.

You can specify keyword arguments to pass to the env constructor in the command line, using --env-kwargs:

python enjoy.py --algo ppo2 --env MountainCar-v0 --env-kwargs goal_velocity:10

You can easily overwrite hyperparameters in the command line, using --hyperparams:

python train.py --algo a2c --env MountainCarContinuous-v0 --hyperparams learning_rate:0.001 policy_kwargs:"dict(net_arch=[64, 64])"

Record 1000 steps:

python -m utils.record_video --algo ppo2 --env BipedalWalkerHardcore-v2 -n 1000

Scores can be found in benchmark.md. To compute them, simply run python -m utils.benchmark.

7 atari games from OpenAI benchmark (NoFrameskip-v4 versions).

Additional Atari Games (to be completed):

See https://github.com/bulletphysics/bullet3/tree/master/examples/pybullet/gym/pybullet_envs. Similar to MuJoCo Envs but with a free simulator: pybullet. We are using BulletEnv-v0 version.

Note: those environments are derived from Roboschool and are much harder than the Mujoco version (see Pybullet issue)

PyBullet Envs (Continued)

See https://github.com/maximecb/gym-minigrid A simple, lightweight and fast Gym environments implementation of the famous gridworld.

There are 19 environment groups (variations for each) in total.

Note that you need to specify --gym-packages gym_minigrid with enjoy.py and train.py as it is not a standard Gym environment, as well as installing the custom Gym package module or putting it in python path.

pip install gym-minigrid
python train.py --algo ppo2 --env MiniGrid-DoorKey-5x5-v0 --gym-packages gym_minigrid

This does the same thing as:

import gym_minigrid

Also, you may need to specify a Gym environment wrapper in hyperparameters, as MiniGrid environments have Dict observation space, which is not supported by StableBaselines for now.

MiniGrid-DoorKey-5x5-v0:
  env_wrapper: gym_minigrid.wrappers.FlatObsWrapper

You can train agents online using colab notebook.

Min version: stable-baselines[mpi] >= 2.10.0

apt-get install swig cmake libopenmpi-dev zlib1g-dev ffmpeg
pip install -r requirements.txt

Please see Stable Baselines README for alternatives.

Build docker image (CPU):

./scripts/build_docker.sh

GPU:

USE_GPU=True ./scripts/build_docker.sh

Pull built docker image (CPU):

docker pull stablebaselines/rl-baselines-zoo-cpu

GPU image:

docker pull stablebaselines/rl-baselines-zoo

Run script in the docker image:

./scripts/run_docker_cpu.sh python train.py --algo ppo2 --env CartPole-v1

To run tests, first install pytest, then:

python -m pytest -v tests/

To cite this repository in publications:

@misc{rl-zoo,
  author = {Raffin, Antonin},
  title = {RL Baselines Zoo},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/araffin/rl-baselines-zoo}},
}

If you trained an agent that is not present in the rl zoo, please submit a Pull Request (containing the hyperparameters and the score too).

We would like to thanks our contributors: @iandanforth, @tatsubori @Shade5