COVID-19 & Economy Simulation

Jointly model pandemic health outcomes and economic activity across 51 US state-level agents and a federal government planner.

The COVID-19 & Economy simulation jointly models infection dynamics and economic activity. It is built on empirical data and fitted epidemiological and economic models drawn from the real-world trajectory of the COVID-19 pandemic in the United States.

This simulation requires an activation code. When you instantiate CovidAndEconomyEnvironment, verify_activation_code() is called automatically. See the ethical review below for context on why this requirement exists.

What the simulation models

The environment comprises 51 agents — each representing a US state or Washington D.C. — and one federal government planner. At each timestep (one day):

• State agents set the stringency level of their pandemic policy response (e.g. lockdown intensity).
• The federal government provides direct payment subsidies to eligible individuals.
• Infection dynamics follow a fitted SIR (Susceptible-Infected-Recovered) model.
• Economic productivity is driven by the fraction of the working-age population that is employed and not too sick to work.

The simulation can operate in three modes:

Mode	use_real_world_data	use_real_world_policies	Description
Fully fitted	False	False	Fitted models drive both health and economic dynamics
Real-world policies	False	True	Real policy stringency levels are replayed; dynamics from fitted models
Full replay	True	True	Both dynamics and policies match historical data exactly

Scenario registration

The scenario is registered under the name:

CovidAndEconomySimulation

The Python class is CovidAndEconomyEnvironment, defined in:

ai_economist/foundation/scenarios/covid19/covid19_env.py

Key parameters

Parameter	Type	Default	Description
use_real_world_data	bool	False	Replay real-world SIR, unemployment, policy, and vaccination data
use_real_world_policies	bool	False	Replay historical policy stringency levels
path_to_data_and_fitted_params	str	""	Path to the directory containing fitted parameters and real-world data. Defaults to the bundled dataset.
start_date	str	"2020-03-22"	Simulation start date in YYYY-MM-DD format
pop_between_age_18_65	float	0.6	Fraction of the population in the working-age bracket
infection_too_sick_to_work_rate	float	0.1	Fraction of infected individuals unable to work
risk_free_interest_rate	float	0.03	Interest rate at which the federal government borrows for direct payments
economic_reward_crra_eta	float	2	CRRA eta parameter for economic reward non-linearity
health_priority_scaling_agents	float	1	How much state agents weight health over the economy relative to the real world (1 = real-world calibration)
health_priority_scaling_planner	float	1	Same as above for the federal government
reward_normalization_factor	float	1	Scalar applied to all rewards before returning them

Real-world data and fitted parameters

The bundled dataset is located at:

ai_economist/datasets/covid19_datasets/data_and_fitted_params/

It includes:

• real_world_data.npz: Historical SIR, unemployment, government policy stringency, and vaccination time series for all 51 US regions.
• Fitted epidemiological and economic model parameters, loaded at init via load_model_constants() and load_fitted_params().

To re-derive the fitted parameters from raw data, run the notebook:

ai_economist/datasets/covid19_datasets/gather_real_world_data_and_fit_parameters.ipynb

GPU acceleration with WarpDrive

The COVID-19 simulation supports CUDA C acceleration via WarpDrive. When a GPU is available, the simulation automatically imports WarpDrive constants and data feeds:

from warp_drive.utils.constants import Constants
from warp_drive.utils.data_feed import DataFeed

If no GPU is detected, the simulation falls back to CPU execution automatically. Install WarpDrive with:

pip install rl-warp-drive

Instantiating the environment

import ai_economist.foundation as foundation

env_config = {
    "scenario_name": "CovidAndEconomySimulation",
    "episode_length": 540,          # ~18 months of daily timesteps
    "n_agents": 51,                 # 50 US states + Washington D.C.

    # Data and policy mode
    "use_real_world_data": False,
    "use_real_world_policies": False,
    "start_date": "2020-03-22",

    # Population parameters
    "pop_between_age_18_65": 0.6,
    "infection_too_sick_to_work_rate": 0.1,

    # Economic parameters
    "risk_free_interest_rate": 0.03,
    "economic_reward_crra_eta": 2,

    # Policy priority weights (1.0 = real-world calibration)
    "health_priority_scaling_agents": 1.0,
    "health_priority_scaling_planner": 1.0,
}

env = foundation.make_env_instance(**env_config)
obs = env.reset()

When use_real_world_data=True, you must also set use_real_world_policies=True. If use_real_world_policies=True, the episode_length cannot exceed the number of available real-world policy timesteps from start_date.

Observation and reward structure

State agents observe:

• Current SIR state (susceptible, infected, recovered fractions)
• Current unemployment rate
• Current and recent policy stringency levels
• Subsidy amounts received
• Time index

The federal government observes the aggregate state across all 51 regions.

Rewards balance health outcomes and economic productivity. The relative weight between health and economy is controlled by health_priority_scaling_agents and health_priority_scaling_planner.

Ethical review

This simulation was designed with care. Before using it in research or policy work, please review:

• Foundation Simulation Card — intended use and ethical review of the broader framework.
• COVID-19 Simulation Card — ethical aspects specific to the pandemic simulation.
• AI Economist web demo and ethical review

Tutorials

COVID-19 and economic simulation

Introduction to the simulation and how to study health and economic policy trade-offs.

Multi-agent GPU training with WarpDrive

Use WarpDrive for high-throughput GPU-accelerated training on this scenario.