Kuang-Huei Lee, Ian Fischer, Anthony Liu, Yijie Guo, Honglak Lee, John Canny, Sergio Guadarrama

NeurIPS 2020

This repository hosts the open source implementation of PI-SAC, the reinforcement learning agent introduced in Predictive Information Accelerates Learning in RL. PI-SAC combines the Soft Actor-Critic Agent with an additional objective that learns compressive representations of predictive information. PI-SAC agents can substantially improve sample efficiency and returns over challenging baselines on tasks from the DeepMind Control Suite of vision-based continuous control environments, where observations are pixels.

If you find this useful for your research, please use the following to reference:

@article{lee2020predictive,
  title={Predictive Information Accelerates Learning in RL},
  author={Lee, Kuang-Huei and Fischer, Ian and Liu, Anthony and Guo, Yijie and Lee, Honglak and Canny, John and Guadarrama, Sergio},
  journal={arXiv preprint arXiv:2007.12401},
  year={2020}
}

PI-SAC learns compact representations of the predictive information I(X_past;Y_future) that captures the environment transition dynamics, in addition to actor and critic learning. We capture the predictive information in a representation Z by maximizing I(Y_future;Z) and minimizing I(X_past;Z|Y_future) to compress out the non-predicitve part for better generalization, which reflects in better sampled efficiency, returns, and transferability. When interacting with the environment, it simply executes the actor model.

Find out more:

• PDF paper

To train the model(s) in the paper with periodic evaluation, run this command:

python -m pisac.run --root_dir=/tmp/pisac_cartpole_swingup \
--gin_file=pisac/config/pisac.gin \
--gin_bindings=train_pisac.train_eval.domain_name=\'cartpole\' \
--gin_bindings=train_pisac.train_eval.task_name=\'swingup\' \
--gin_bindings=train_pisac.train_eval.action_repeat=4 \
--gin_bindings=train_pisac.train_eval.initial_collect_steps=1000 \
--gin_bindings=train_pisac.train_eval.initial_feature_step=5000

We use gin to config hyperparameters. The default configs are specificed in pisac/config/pisac.gin. To reproduce the main DM-Control experiments, you need to specify different domain_name, task_name, action_repeat, initial_collect_steps, initial_feature_step for each environment.

To use multiple gradient steps per environment step, change train_pisac.train_eval.collect_every to a number larger than 1.

*gs: number of gradient steps per environment step

The PI-SAC code uses Python 3 and these packages:

• tensorflow-gpu==2.3.0
• tf_agents==0.6.0
• tensorflow_probability
• dm_control (egl rendering option recommended)
• gym
• imageio
• matplotlib
• scikit-image
• scipy
• gin
• pstar
• qj

If you ever see that dm_control complains about some threading issues, please try adding --gin_bindings=train_pisac.train_eval.drivers_in_graph=False to put dm_control environment outside of the TensorFlow graph.

Disclaimer: This is not an official Google product.