Welcome to the weather2alert package

This package provides a Gymnasium environment (formerly OpenAI Gym) for training reinforcement learning agents to control and optimize when to send alerts based on the observed heat index during summer and alerting history to minimize the number of hospitalizations due to heat-related illnesses.

The environment was calibrated from a comprehensive data set between 2006--2016 of heat index values, hospitalizations, and alerts from various sources, including the National Weather Service, the Centers for Disease Control and Prevention, the Census Bureau, Centers for Medicare and Medicaid Services, and the National Oceanic and Atmospheric Administration.

The mathematical model and a detailed description of the data sources and modeling approach is available in the following publication:

@article{considine2023optimizing,
  title={Optimizing Heat Alert Issuance with Reinforcement Learning},
  author={Considine, Ellen M and Nethery, Rachel C and Wellenius, Gregory A and Dominici, Francesca and Tec, Mauricio},
  journal={arXiv preprint arXiv:2312.14196},
  year={2023}
}

Getting Started

The package is still under development and is not yet available on PyPI. To install the package, run the following command:

pip install git+https://github.com/NSAPH-Projects/weather2alert

To create an environment, use the following code:

import weather2alert
env = weather2alert.env.HeatAlertEnv(seed=1234)
obs, info = env.reset(location='06037', similar_climate_counties=False)

The location parameter is a string that represents the FIPS code of the county where the environment is located. The default value is '06037', which corresponds to Los Angeles County, California. When the location is not provided, it will be chosen randomly from the available locations. The keyword argument similar_climate_counties is a boolean that determines whether the environment will use climate data from similar counties to the reference to augment the available episodes. The default value is False. See the paper for details on the data augmentation process for climate transitions.

Use the environment with the standard Gymnasium API. For example:

done = False
while not done:
    action = env.action_space.sample()
    obs, reward, done, info = env.step(action)

The available actions are always 0 (do not send an alert) and 1 (send an alert).

The simulations contain real climate data from 2006 to 2016, and the environment is designed to be episodic. The episode ends when the simulation reaches the end of the data set. A year is chosen randomly from the data set every time the environment is reset.

The rewards represent the (scaled) negative rate of hospitalizations due to heat-related illnesses. The goal is to minimize the number of hospitalizations by sending alerts when the heat index is above a certain threshold.