CFRL documentation

Welcome to CFRL, a Python library for counterfactually fair reinforcement learning! The acronym “CFRL” stands for “Counterfactual Fairness in Reinforcement Learning”. CFRL provides algorithms that ensure counterfactual fairness in reinforcement learning and builds tools for evaluating the value and counterfactual fairness of reinforcement learning policies.

To install CFRL, run

$ pip install cfrl

This project is still being perfected. We will continue adding new functionalities and expanding the use cases of CFRL. We appreciate your patience and support!

[CFRL Github repository]

[CFRL software paper]

Introduction

• Getting Started
  • Installation
  • A Brief Introduction to Counterfactual Fairness
  • Key Functionalities
  • High-level Design
• FAQ

Inputs and Outputs

• Data Requirements
• Trajectory Arrays
  • Sensitive Attributes Format
  • Single-time States Format
  • Full-trajectory States Format
  • Single-time Actions Format
  • Full-trajectory Actions Format
  • Single-time Rewards Format
  • Full-trajectory Rewards Format
• Tabular Trajectory Data

Tutorials

• Example Workflows
  • Preprocessing Only
  • Policy Learning Only
  • Assessing Preprocessors Using Synthetic Data
  • Assessing Policies Using Real Data

Interface

• Interface
  • Reader
  • Preprocessor
  • Environment
  • Agents
  • FQE
  • Evaluation

Customizations

• Custom Preprocessors
• Custom Agents

About CFRL

• The CFRL Team
  • Maintainer
  • Contributors