Skip to main content
Dryad logo

The AI Economist: Taxation policy design via two-level deep reinforcement learning

Citation

Zheng, Stephan et al. (2021), The AI Economist: Taxation policy design via two-level deep reinforcement learning, Dryad, Dataset, https://doi.org/10.5061/dryad.bnzs7h4c4

Abstract

This dataset contains all raw experimental data for the paper "The AI Economist: Taxation Policy Design via Two-level Deep Multi-Agent Reinforcement Learning". 

The accompanying simulation, reinforcement learning, and data visualization code can be found at https://github.com/salesforce/ai-economist.

For the one-step economy experiments, we provide:

  • training histories,

  • configuration files (these experiments do not use phases), and

  • final agent and planner models.

For the Gather-Trade-Build scenario, the data covers 6 spatial layouts: two Open-Quadrant (with 4 and 10 agents), and four Split-World maps with different configurations of the high-skilled and low-skilled agents. It also covers 4 tax policies (the AI Economist, Saez, free-market, and US federal). In addition, the AI Economist has been optimized for two social welfare functions: the product of equality and productivity, and inverse-income weighted utility. The Saez tax policy also uses estimated elasticities. 

Each experiment was repeated with different random seeds: 10 seeds for the Open-Quadrant scenarios, and 5 seeds for the Split-World scenarios. For each individual experiment, we provide: 

  • Training histories (e.g. equality and productivity throughout training)

  • the phase 1 and phase 2 configuration files, 

  • 40 episode dense logs (the final 10 simulation logs across 4 environment replicas),

  • phase 1 final agent models, and

  • phase 2 final agent and planner models.

Finally, we include all data and results used to calibrate the Saez elasticity estimates and to estimate elasticity directly from a sweep over flat-rate tax policies:

  • training histories,

  • the phase 1 and phase 2 configuration files, 

  • phase 1 final agent models, and

  • phase 2 final agent and planner models.

Methods

This data has been generated by applying multi-agent deep reinforcement learning in economic simulations. It contains all key reinforcement learning and economic metrics that support the results in the paper.

We have included tutorials and configuration files (with the raw data itself) on 2-level RL to generate this data. These complement the code available in the following locations: 

The Github repo https://github.com/salesforce/ai-economist has all the simulation and reinforcement learning code that produced this data. We also provide all the visualization and analysis code on Github. All figures in the paper can be generated by using the visualization code on Github with this experimental data. 

In addition, all code is also available on Zenodo \url{https://doi.org/10.5281/zenodo.5644182}.

To reproduce this data, we have also provided instructions on Github to independently run reinforcement learning and the economic simulations. 

This data has only been organized for clarity and not otherwise modified. 

Usage Notes

Note: the 3 GB zip file inflates to ~20 GB. 

For questions: stephan.zheng@salesforce.com or aieconomist@salesforce.com.