Status: Archive (code is provided as-is, no updates expected)

Leveraging Procedural Generation to Benchmark Reinforcement Learning

[Blog Post] [Paper]

This is code for training agents for some of the experiments in Leveraging Procedural Generation to Benchmark Reinforcement Learning (citation). The code for the environments is in the Procgen Benchmark repo.

We're currently running a competition which uses these environments to measure sample efficiency and generalization in RL. You can learn more and register here.

Supported platforms:

• macOS 10.14 (Mojave)
• Ubuntu 16.04

Supported Pythons:

• 3.7 64-bit

Install

You can get miniconda from https://docs.conda.io/en/latest/miniconda.html if you don't have it, or install the dependencies from environment.yml manually.

git clone https://github.com/openai/train-procgen.git
conda env update --name train-procgen --file train-procgen/environment.yml
conda activate train-procgen
pip install https://github.com/openai/baselines/archive/9ee399f5b20cd70ac0a871927a6cf043b478193f.zip
pip install -e train-procgen

Try it out

Train an agent using PPO on the environment StarPilot:

python -m train_procgen.train --env_name starpilot

Train an agent using PPO on the environment StarPilot using the easy difficulty:

python -m train_procgen.train --env_name starpilot --distribution_mode easy

Run parallel training using MPI:

mpiexec -np 8 python -m train_procgen.train --env_name starpilot

Train an agent on a fixed set of N levels:

python -m train_procgen.train --env_name starpilot --num_levels N

Train an agent on the same 500 levels used in the paper:

python -m train_procgen.train --env_name starpilot --num_levels 500

Train an agent on a different set of 500 levels:

python -m train_procgen.train --env_name starpilot --num_levels 500 --start_level 1000

Run simultaneous training and testing using MPI. 1 in every 4 workers will be test workers, and the rest will be training workers.

mpiexec -np 8 python -m train_procgen.train --env_name starpilot --num_levels 500 --test_worker_interval 4

Train an agent using PPO on a level in Jumper that requires hard exploration

python -m train_procgen.train --env_name jumper --distribution_mode exploration

Train an agent using PPO on a variant of CaveFlyer that requires memory

python -m train_procgen.train --env_name caveflyer --distribution_mode memory

View training options:

python -m train_procgen.train --help

Reproduce and Visualize Results

Sample efficiency on hard environments (results/hard-all-runN):

mpiexec -np 4 python -m train_procgen.train --env_name ENV_NAME --distribution_mode hard
python -m train_procgen.graph --distribution_mode hard

Sample efficiency on easy environments (results/easy-all-runN):

python -m train_procgen.train --env_name ENV_NAME --distribution_mode easy
python -m train_procgen.graph --distribution_mode easy

Generalization on hard environments using 500 training levels (results/hard-500-runN):

mpiexec -np 8 python -m train_procgen.train --env_name ENV_NAME --num_levels 500 --distribution_mode hard --test_worker_interval 2
python -m train_procgen.graph --distribution_mode hard --restrict_training_set

Generalization on easy environments using 200 training levels (results/easy-200-runN):

mpiexec -np 2 python -m train_procgen.train --env_name ENV_NAME --num_levels 200 --distribution_mode easy --test_worker_interval 2
python -m train_procgen.graph --distribution_mode easy --restrict_training_set

Pass --normalize_and_reduce to compute and visualize the mean normalized return with train_procgen.graph.

Citation

Please cite using the following bibtex entry:

@article{cobbe2019procgen,
  title={Leveraging Procedural Generation to Benchmark Reinforcement Learning},
  author={Cobbe, Karl and Hesse, Christopher and Hilton, Jacob and Schulman, John},
  journal={arXiv preprint arXiv:1912.01588},
  year={2019}
}