Usage

• Causal Inference
  • Treatment Effects
  • Provided Estimators
  • Creating New Experiments
• Causal Graph Discovery
• Sequential Decision Making

Causal Inference

WhyNot provides a powerful framework that generates data to stress-test methods for estimating both Average Treatment Effects and Heterogeneous Treatment Effects in a wide variety of simulated environments. After generating data with WhyNot, the package provides a collection of causal estimators to get started estimating treatment effects, though users are also encouraged to try out their own estimators.

All of the Simulators implemented in WhyNot come equipped with a number of causal inference experiments. Beyond these experiments, WhyNot easily and flexibly supports the creation of new experiment designs to probe other causal questins. Creating New Experiments describes how to use primitives in WhyNot to construct new experiments.

Causal Graph Discovery

WhyNot also provides tools to automatically construct causal graphs associated with runs of the simulators and to generate causal graphs for experiments implemented in WhyNot. Equipped with these graphs, users can go beyond causal inference and study problems of causal structure discovery. See Causal Graph Discovery for more details on how to automatically construct causal graphs in WhyNot and how to use these graphs to probe questions in causal discovery.

Sequential Decision and Reinforcement Learning

WhyNot is also an excellent test bed for sequential decision making and reinforcement learning in diverse dynamic environments. WhyNot offers RL environments compatible with the OpenAI Gym API style, so that existing code for OpenAI Gym can be adapted for WhyNot with minimal changes. See Sequential Decision Making for more details on how to use available environments in WhyNot and how to define new custom environments on top of WhyNot simulators.