Purdue University

A researcher from the University of Nebraska-Omaha used Purdue’s Anvil supercomputer to develop a new artificial intelligence (AI) technique that can derive optimal taxation strategies for governments. This new method leveraged Anvil’s advanced GPUs to factor in household differences across families within a population in order to determine how taxes should be applied for the best possible outcome.

Dr. Zhigang Feng is a professor in the Department of Economics at the University of Nebraska-Omaha. He, along with his collaborators hailing from multiple institutions, combined machine learning techniques with economic theory to tackle everyone’s favorite economic subject—taxes.

Taxation is an oft-debated subject for governments worldwide, with different opinions and theories as to what works best for individual countries or locales. How differing taxation strategies affect the economic choices of households in a population is an extraordinarily complex problem to solve. Many models have been developed to try and understand and predict the effects of taxes on the economy, with varying levels of success. Most models fail to account for household heterogeneity in the context of dynamic economic fluctuations. This shortcoming is precisely what Feng and his colleagues set out to remedy.

Research has long shown that household heterogeneity needs to be factored in to accurately model economic behavior and therefore design optimized fiscal policies. However, heterogeneity takes an already complex mathematical problem and adds in an infinite-dimensional object. Feng’s goal was to develop a novel machine learning-based approach that successfully factored in household differences. To do this required a massive amount of computing power due to the curse of dimensionality problem, which is why he and his collaborators turned to Anvil.

"This problem isn’t something traditional numerical methods in the standard economist's toolbox can handle—even with a handful of CPUs using MPI, let alone an average computer," says Feng. "We needed multiple GPUs running in parallel to harness the optimization power of modern AI techniques, and we needed them on demand. We also required a machine with massive memory to store the state of every simulated individual. Thankfully, Anvil was able to provide us with both."

The group utilized both CPUs and the advanced GPUs on Anvil to create a Markov decision process in Wasserstein space. They combined deep neural networks for equilibrium function approximation, a histogram-based distribution approximation, an analytically derived distribution transition kernel, and a modified value and policy iteration with an augmented Lagrangian method, all of which together allowed them to address the problem of infinite dimensions. After developing the new approach, the group also needed to run the model simulations for multiple scenarios, showing the cause-and-effect of different taxation strategies.

Overall the group was very happy with Anvil’s performance. The queue for the GPUs was short, allowing them the access they needed to quickly conduct their research. Feng also noted that anytime the team hit any snags or had issues, they reached out to the Anvil support team and received help promptly. All of this combined enabled the group to efficiently proceed with a project that otherwise would not have been possible.

“To solve these models, we needed Anvil,” says Feng. “There’s no question—without it, this is not something we would have been able to achieve.”

Though the research publication is in its preliminary stages, it shows promising results and will have important implications for policymakers and researchers wanting to design effective fiscal policies. The novel machine learning method developed by Feng and his colleagues is also scalable and can be applied to a wide range of other economic models.

For more information about this project, as well as other research conducted by Dr. Feng, please visit his Research Page.

To learn more about High-Performance Computing and how it can help you, please visit our “Why HPC?” page.

Anvil is one of Purdue University’s most powerful supercomputers, providing researchers from diverse backgrounds with advanced computing capabilities. Built through a $10 million system acquisition grant from the National Science Foundation (NSF), Anvil supports scientific discovery by providing resources through the NSF’s Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS), a program that serves tens of thousands of researchers across the United States.

Researchers may request access to Anvil via the ACCESS allocations process. More information about Anvil is available on Purdue’s Anvil website. Anyone with questions should contact anvil@purdue.edu. Anvil is funded under NSF award No. 2005632.

Publications utilizing Anvil

1. Chen C, Feng Z, Gu J. HEALTH, HEALTH INSURANCE, AND INEQUALITY. International Economic Review. Published online July 4, 2024. doi:https://doi.org/10.1111/iere.12722
2. Feng, Zhigang and Han, Jiequn and Zhu, Shenghao, Optimal Taxation with Incomplete Markets–An Exploration Via Reinforcement Learning. Available at SSRN: http://dx.doi.org/10.2139/ssrn.4758552

Written by: Jonathan Poole, poole43@purdue.edu