WhyNot is a Python sandbox for developing, testing, and benchmarking causal inference and sequential decision-making tools within dynamic environments. It targets researchers, practitioners, and students seeking a flexible platform to explore and evaluate methods in complex, simulated scenarios. The primary benefit is providing a unified interface to diverse simulators and experimental designs, facilitating robust evaluation and pedagogical use.

How It Works

WhyNot integrates causal inference and reinforcement learning techniques with dynamic simulators. It offers a structured approach to generating datasets for causal analysis, including randomized control trials, confounding, and mediation scenarios. For sequential decision-making, it leverages the OpenAI Gym interface, enabling experimentation with reinforcement learning agents in simulated environments. This dual focus allows for comprehensive evaluation of methods across different problem types.

Quick Start & Requirements

• Install: pip install whynot
• Prerequisites: Python 3.x. For advanced estimators, whynot_estimators requires R.
• Resources: Basic installation is straightforward. Complex simulations may require significant computational resources.
• Docs: https://github.com/socialfoundations/whynot
• Walkthrough Video: https://www.youtube.com/watch?v=91_2030_f1Q

Highlighted Details

• Supports causal inference experiments for average treatment effects, heterogeneous treatment effects, time-varying treatments, and causal structure discovery.
• Integrates with OpenAI Gym for sequential decision-making and reinforcement learning tasks.
• Includes a "Strategic Classification" module for evaluating classifiers in adversarial, dynamic settings.
• Offers a wide array of simulators from economics, epidemiology, and social sciences, including World3, HIV, and DICE.

Maintenance & Community

The project is under active development, with contributions welcomed via GitHub issues for bugs and feature requests. The README does not specify community channels like Discord or Slack.

Licensing & Compatibility

MIT License. This permissive license allows for commercial use and integration into closed-source projects.

Limitations & Caveats

The simulators are designed for technical challenges in causal inference and dynamic decision-making, not as faithful models of the real world, and should not be used for direct policy design. The project is still under active development, implying potential for breaking changes or incomplete features.