Skip to content
/ panacea Public

Security Analysis of Safe and Seldonian Reinforcement Learning Algorithms

Notifications You must be signed in to change notification settings

pinaro/panacea

Repository files navigation

Security Analysis of Safe & Seldonian Reinforcement Learning Algorithms

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.

Requirements

  • >= Python 3.6 with pip
  • R

Data Collection

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.

Setup

  1. Download the requirements: pip install -r requirements.txt

  2. To replicate the results in the paper, from the root directory of this project, run:python run.py 0

  3. To run the same experiment with a random behavior and evaluation policy, from the root directory of this project, run:python run.py 1

Results

The plot that will be generated by the code will be placed in "resuts/final_results.pdf"

Configs

constants.py specifies all the hyperparameters used in the experiment and can be changed.

TODO

Implemantation of softmax action selection to run experiments with a random behavior and evaluation policy

About

Security Analysis of Safe and Seldonian Reinforcement Learning Algorithms

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published