Skip to content

A Python sandbox for decision making in dynamics

License

Notifications You must be signed in to change notification settings

wearepal/whynot

 
 

Repository files navigation

WhyNot Logo

Build Status Documentation Status Code style: black

WhyNot is a Python package that provides an experimental sandbox for decisions in dynamics, connecting tools from causal inference and reinforcement learning with challenging dynamic environments. The package facilitates developing, testing, benchmarking, and teaching causal inference and sequential decision making tools.

For an introduction to WhyNot and a brief tutorial, see our walkthrough video. For more detailed information, check out the documentation.

Table of Contents

  1. Basic installation instructions
  2. Quick start examples
  3. Simulators in WhyNot
  4. Using estimators in R
  5. Frequently asked questions
  6. Citing WhyNot

WhyNot is still under active development! If you find bugs or have feature requests, please file a Github issue. We welcome all kinds of issues, especially those related to correctness, documentation, performance, and new features.

Basic installation instructions

  1. (Optionally) create a virtual environment
python3 -m venv whynot-env
source whynot-env/bin/activate
  1. Install via pip
pip install whynot

You can also install WhyNot directly from source.

git clone https://github.com/zykls/whynot.git
cd whynot
pip install -r requirements.txt

Quick start examples

Causal inference

Every simulator in WhyNot comes equipped with a set of experiments probing different aspects of causal inference. In this section, we show how to run experiments probing average treatment effect estimation on the World3 simulator. World3 is a dynamical systems model that studies the interplay between natural resource constraints, population growth, and industrial development.

First, we examine all of the experiments available for World3.

import whynot as wn
experiments = wn.world3.get_experiments()
print([experiment.name for experiment in experiments])
#['PollutionRCT', 'PollutionConfounding', 'PollutionUnobservedConfounding', 'PollutionMediation']

These experiments generate datasets both in the setting of a pure randomized control trial (PollutionRCT), as well as with (unobserved) confounding and mediation. We will run a randomized control experiment. The description property offers specific details about the experiment.

rct = wn.world3.PollutionRCT
rct.description
#'Study effect of intervening in 1975 to decrease pollution generation on total population in 2050.'

We can run the experiment using the experiment run function and specifying a desired sample size num_samples. The experiment then returns a causal Dataset consisting of the covariates for each unit, the treatment assignment, the outcome, and the ground truth causal effect for each unit. All of this data is contained in NumPy arrays, which makes it easy to connect to causal estimators.

import numpy as np

dataset = rct.run(num_samples=200, seed=1111, show_progress=True)
(X, W, Y) = dataset.covariates, dataset.treatments, dataset.outcomes
treatment_effect = np.mean(dataset.true_effects)

# Plug-in your favorite causal estimator
estimated_ate = np.mean(Y[W == 1.]) -  np.mean(Y[W  == 0.])

WhyNot also enables you to run a large collection of causal estimators on the data for benchmarking and comparison. The main function to do this is the causal_suite which, given the causal dataset, runs all of the estimators on the dataset and returns an InferenceResult for each estimator containing its estimated treatment effects and uncertainty estimates like confidence intervals.

# Run the suite of estimates
estimated_effects = wn.causal_suite(
    dataset.covariates, dataset.treatments, dataset.outcomes)

# Evaluate the relative error of the estimates
true_sate = dataset.sate
for estimator, estimate in estimated_effects.items():
    relative_error = np.abs((estimate.ate - true_sate) / true_sate)
    print("{}: {:.2f}".format(estimator, relative_error))
# ols: 1.06
# propensity_score_matching: 1.38
# propensity_weighted_ols: 1.37

In addition to experiments studying average treatment effect, WhyNot also supports causal inference experiments studying

  1. Heterogeneous treatment effects,
  2. Time-varying treatment policies
  3. Causal structure discovery

Sequential decision making

WhyNot supports experimentation with sequential decision making and reinforcement learning via unified interface with the OpenAI gym. In this section, we give a simple example showing how to use the HIV simulator for sequential decision making experiments.

First, we initialize the environment and set the random seed.

import whynot.gym as gym

env = gym.make('HIV-v0')
env.seed(1)

Observations in the simulator are a set of 6 states, capturing infected and uninfected T-lymphocytes, macrophages, immune response, and copies of free virus. Actions correspond to choosing between different drugs and dosages for treatment.

For illustration, we repeatedly chose actions, which correspond to treatment policy decisions, in the environment and measure both the next state and the reward. In this case, the reward weighs the strength of the immune response, the virus count, and the cost of the chosen treatment.

observation = env.reset()
for _ in range(100):
    action = env.action_space.sample()  # Replace with your treatment policy
    observation, reward, done, info = env.step(action)
    if done:
        observation = env.reset()

For more details on the simulation, as well as a fully worked out policy gradient example, see this notebook.

Strategic classification

Beyond settings typically studied in sequential decision making, WhyNot also supports experiments with standard supervised learning algorithms in dynamic settings. In this section, we show how to use WhyNot to study the performance of classifiers when individuals being classified behave strategically to improve their outcomes, a problem sometimes called strategic classification.

First, we set up the credit environment.

import whynot.gym as gym

env = gym.make('Credit-v0')
env.seed(1)

Observations in this environment correspond to a dataset of features for each individual and a label indicating whether they experience financial distress from the Kaggle GiveMeSomeCredit dataset.

dataset = env.reset()

Actions in the environment correspond to choosing a classifier to predict default. In response, individuals then strategically adapt their features in order to obtain a more favorable credit score. The subsequent observation is the adapted features, and the reward is the classifier's loss on this distribution

theta = env.action_space.sample() # Your classifier
dataset, loss, done, info = env.step(theta)

We can then experiment with the long-term equilibrium arising from repeatedly updating the classifier to cope with strategic response.

def learn_classifier(features, labels):
    # Replace with your learning algorithm
    return env.action_space.sample()

dataset = env.reset()
for _ in range(100):
    theta = learn_classifier(dataset["features"], dataset["labels"])
    dataset, loss, _, _ = env.step(theta)

For more details on the simulation and a complete example showing the standard retraining procedures perform in a strategic setting, see this notebook.

Beyond strategic classification, WhyNot also supports simulators and experiments evaluating other aspects of machine learning, e.g. fairness criteria, in dynamic settings.

For more examples and demonstrations of how to design and conduct experiments in each of these settings, check out usage and our collection of examples.

Simulators in WhyNot

WhyNot provides a large number of simulated environments from fields ranging from economics to epidemiology. Each simulator comes equipped with a representative set of causal inference experiments and exports a uniform Python interface that makes it easy to construct new causal inference experiments in these environments, as well as an OpenAI gym interface to perform reinforcement learning experiments in new environments.

The simulators in WhyNot currently include:

For a detailed overview of these simulators, please see the simulator documentation.

Using causal estimators in R

WhyNot ships with a small set of causal estimators written in pure Python. To access other estimators, please install the companion library whynot_estimators, which includes a host of state-of-the-art causal inference methods implemented in R.

To get the basic framework, run

pip install whynot_estimators

If you have R installed, you can install the causal_forest estimator by using

python -m  whynot_estimators install causal_forest

To see all of the available estimators, run

python -m  whynot_estimators show_all

See whynot_estimators for instructions on installing specific estimators, especially if you do not have an existing R build.

Frequently asked questions

1. Why is it called WhyNot?

Why not?

2. What are the intended use cases?

WhyNot supports multiple use cases, some technical, some pedagogical, each suited for a different group of users. We envision at least five primary use cases:

  • Developing: Researchers can use WhyNot in the process of developing new methods for causal inference and decision making in dynamic settings. WhyNot can serve as a substitute for ad-hoc synthetic data where needed, providing a greater set of challenging test cases.

  • Testing: Researchers can use WhyNot to design robustness checks for methods and gain insight into the failure cases of these methods.

  • Benchmarking: Practitioners can use WhyNot to compare multiple methods on the same set of tasks. WhyNot does not dictate any particular benchmark, but rather supports the community in creating useful benchmarks.

  • Learning: Students of causality and dynamic decision making might find WhyNot to be a helpful training resource. WhyNot is easy-to-use and does not require much prior experience to get started with.

  • Teaching: Instructors can use WhyNot as a tool students engage with to learn and solve problems.

3. What uses are not intended?

  • Basis of real-world policy and interventions: The simulators included in WhyNot were selected because they offer realistic technical challenges for causal inference and dynamic decision making tools, not because they offer faithful models of the real world. In many cases, they have been contested or criticized as representations of the real world. For this reason, the simulators should not directly be used to design real-world interventions or policy.

  • Substitute for healthy debate: Success in simulated environments does not guarantee success in real scenarios, but a failure in simulated environments can nonetheless lead to insight into weaknesses of a particular approach. WhyNot does not obviate the need for debate around common assumptions in causal inference.

  • Substitute for real world experiments and data: WhyNot does not substitute for high-quality empirical work on real data sets. WhyNot is a tool for understanding and evaluating methods for causal inference and decision making in dynamics, not certifying their validity in real-world scenarios.

  • Substitute for theory: WhyNot can help create understanding in contexts where theoretical analysis is challenging, but does not reduce the need for theoretical guarantees and formal analysis.

4. Why start from dynamical systems?

Dynamical systems provide a natural setting to study causal inference. The physical world is a dynamical system, and causal inference inevitably has to grapple with data generated from some dynamical process. Moreover, the temporal structure of the dynamics gives rise to nontrivial problem instances with both confounding and mediation. Dynamics also naturally lead to time-varying causal effects and allow for time-varying treatments and sequential decision making.

5. What what simulators are included and why?

WhyNot contains a range of different simulators, and an overview is provided in the documentation here.

6. What’s the difference between WhyNot and CauseMe?

CauseMe is an online platform for benchmarking causal discovery methods. Users can register and evaluate causal discovery methods on an existing repository of data sets, or contribute their own data sets with known ground truth. CauseMe is an excellent platform that we recommend in addition to WhyNot. We encourage users to export data sets derived from WhyNot and make them accessible through CauseMe. In this case, we ask that you reference WhyNot.

7. What’s the difference between WhyNot and CausalML?

CausalML is a Python package that provides a range of causal inference methods. The estimators provided by CausalML are available in WhyNot via the whynot_estimators package. While WhyNot provides simulators and derived experimental designs on synthetic data, CausalML focuses on providing estimators. We made these estimators available for use on top of WhyNot.

8. What’s the difference between WhyNot and EconML?

EconML is a Python package that provides tools from machine learning and econometrics for causal inference. Like CausalML, EconML focuses on providing estimators, and we made these estimators available for use on top of WhyNot.

9. How can I best contribute to WhyNot?

Thanks so much for considering to contribute to WhyNot. The package is open source and MIT licensed. We invite contributions broadly in a number of areas, including the addition of simulators, causal estimators, sequential decision making algorithms, documentation, performance improvements, code quality and tests.

Citing WhyNot

If you use WhyNot for published work, we encourage you to cite the project. Please use the following BibTeX entry:

@software{miller2020whynot,
  author       = {John Miller and
                  Chloe Hsu and
                  Jordan Troutman and
                  Juan Perdomo and
                  Tijana Zrnic and
                  Lydia Liu and
                  Yu Sun and
                  Ludwig Schmidt and
                  Moritz Hardt},
  title        = {WhyNot},
  year         = 2020,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.3875775},
  url          = {https://doi.org/10.5281/zenodo.3875775}
}

About

A Python sandbox for decision making in dynamics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 80.7%
  • JavaScript 19.3%