This code implements the experiments described in the paper "Security Analysis of Safe & Seldonian Reinforcement Learning Algorithms" by A. Pinar Ozisik, and Philip S. Thomas.
- >= Python 3.6 with pip
- R
This code collects data using a slightly modified version of Jinyu Xie's Simglucose v0.2.1. Xie's code, found in the contents of folder SimGlucose, is incorporated into this repository for convenience.
-
Download the requirements:
pip install -r requirements.txt -
To replicate the results in the paper, from the root directory of this project, run:
python run.py 0 -
To run the same experiment with a random behavior and evaluation policy, from the root directory of this project, run:
python run.py 1
The plot that will be generated by the code will be placed in "resuts/final_results.pdf"
constants.py specifies all the hyperparameters used in the experiment and can be changed.
Implemantation of softmax action selection to run experiments with a random behavior and evaluation policy