Experiments for the RATS algorithm

This repository contains an implementation of the RATS algorithm presented in the paper "Non-Stationary Markov Decision Processes a Worst-Case Approach using Model-Based Reinforcement Learning" along with the reported experiments.

Installation and use

First clone the repository:

git clone git@github.com:SuReLI/rats-experiments.git
cd rats-paper-experiments

Ensure all the dependencies described in the next section are installed within your python environment. Now you can do your stuff.

python example.py

List of dependencies

Standard dependencies include the following: numpy, csv, multiprocessing, time, itertools, scipy, math.

Non-standard dependencies include gym, available at this location.

Run the experiments of the paper

To reproduce the experiment of the paper "Non-Stationary Markov Decision Processes a Worst-Case Approach using Model-Based Reinforcement Learning", run the following script sequentially:

python nsbridge_experiment.py
python results_exploitation.py

Cite

If you use this code in your work or build on the paper, please cite the latter. Here is an example of bibtex entry:

@inproceedings{lecarpentier2019non,
    title={{Non-Stationary Markov Decision Processes a Worst-Case Approach using Model-Based Reinforcement Learning}},
    author = {Lecarpentier, Erwan and Rachelson, Emmanuel},
    booktitle = {Advances in Neural Information Processing Systems 32},
    editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett},
    pages = {7214--7223},
    year = {2019},
    publisher = {Curran Associates, Inc.}
}