stjordanis / spibb Goto Github PK

Implementation of Safe Policy Improvement with Baseline Bootstrapping

This project can be used to reproduce the finite MDPs experiments presented in the ICML2019 paper: Safe Policy Improvement with Baseline Bootstrapping, by Romain Laroche, Paul Trichelair, and Rémi Tachet des Combes. For the DQN implementation, please refer to git repository at this address.

Prerequisites

The project is implemented in Python 3.5 and requires numpy and scipy.

Usage

We include the following:

• Libraries of the following algorithms:
  • basic RL,
  • SPIBB:
    • Pi_b-SPIBB,
    • Pi_{\leq b}-SPIBB,
  • HCPI:
    • doubly-robust,
    • importance_sampling,
    • weighted_importance_sampling,
    • weighted_per_decision_IS,
    • per_decision_IS,
  • Robust MDP,
  • and Reward-adjusted MDP.
• Environments:
  • Gridworld environment,
  • Random MDPs environment.
• Gridworld experiment of Section 3.1. Run: python gridworld_main.py #name_of_experiment# #random_seed#
• Gridworld experiment with random behavioural policy of Section 3.2. Run: python gridworld_random_behavioural_main.py #name_of_experiment# #random_seed#
• Random MDPs experiment of Section 3.3. Run: python randomMDPs_main.py #name_of_experiment# #random_seed#

We DO NOT include the following:

• The hyper-parameter search (Appendix C.2): it should be easy to re-implement.
• The figure generator: it has too many specificities to be made understandable for a user at the moment. Also, it is not hard to re-implement one's own visualization tools.
• The multi-CPU implementation: its structure is too much dependent on the cluster tools.

License

This project is BSD-licensed.

Reference

Please use the following bibtex entry if you use this code:

@inproceedings{Laroche2019,
    title={Safe Policy Improvement with Baseline Bootstrapping},
    author={Laroche, Romain and Trichelair, Paul and Tachet des Combes, R\'emi},
    booktitle={Proceedings of the 36th International Conference on Machine Learning (ICML)},
    year={2019}
}