This repository contains an implementation of "Importance Weighted Actor-Learner Architectures", along with a dynamic batching module. This is not an officially supported Google product.

For a detailed description of the architecture please read our paper. Please cite the paper if you use the code from this repository in your work.

@inproceedings{impala2018,
  title={IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures},
  author={Espeholt, Lasse and Soyer, Hubert and Munos, Remi and Simonyan, Karen and Mnih, Volodymir and Ward, Tom and Doron, Yotam and Firoiu, Vlad and Harley, Tim and Dunning, Iain and others},
  booktitle={Proceedings of the International Conference on Machine Learning (ICML)},
  year={2018}
}

TensorFlow >=1.9.0-dev20180530, the environment DeepMind Lab and the neural network library DeepMind Sonnet. Although we use DeepMind Lab in this release, the agent has been successfully applied to other domains such as Atari, Street View and has been modified to generate images.

We include a Dockerfile that serves as a reference for the prerequisites and commands needed to run the code.

Training on explore_goal_locations_small. Most runs should end up with average episode returns around 200 or around 250 after 1B frames.

python experiment.py --num_actors=48 --batch_size=32

Adjust the number of actors (i.e. number of environments) and batch size to match the size of the machine it runs on. A single actor, including DeepMind Lab, requires a few hundred MB of RAM.

Training on the full DMLab-30. Across 10 runs with different seeds but identical hyperparameters, we observed between 45 and 50 capped human normalized training score with different seeds (--seed=[seed]). Test scores are usually an absolute of ~2% lower.

python experiment.py --job_name=learner --task=0 --num_actors=150 \
    --level_name=dmlab30 --batch_size=32 --entropy_cost=0.0033391318945337044 \
    --learning_rate=0.00031866995608948655 \
    --total_environment_frames=10000000000 --reward_clipping=soft_asymmetric

for i in $(seq 0 149); do
  python experiment.py --job_name=actor --task=$i \
      --num_actors=150 --level_name=dmlab30 --dataset_path=[...] &
done;
wait

python experiment.py --mode=test --level_name=dmlab30 --dataset_path=[...] \
    --test_num_episodes=10