In April, the site detailed a research environment tool designed to understand how AI can improve economic design, AI Economist, according tomedia Salesforce. The site promises to eventually open the source code library, and today, the first version of the code base will be released. Studies have shown that income inequality has a negative impact on economic growth, opportunities and health, and that tax policies can have a similar impact. For example, excessive taxation can discourage people from working and lead to lower productivity.
But experimenting with policy in the real world is difficult because economic theory relies on hard-to-verify assumptions.
Salesforce hopes to use AI economists to lead the way in developing a tool that guides tax policy, calling on AI researchers, economists and policymakers to contribute code and collaborate on research, voluntarily provide their expertise and build rich simulations, and suggest which social issues can be addressed through the framework.
“The goal of this project is to create an enhanced learning framework that will recommend economic policies that drive real-world social outcomes such as sustainability, productivity and equity,” Stephan Zheng, a scientist at Salesforce’s machine learning research division, wrote in a blog post. To achieve this, we need to advance AI technology, challenge traditional economic thinking, and create AI that can lay and guide policy development. Although all these tasks are not easy, they together make a real plan for the moon landing. “
AI Economist is a two-tiered, deep-reinforcement learning framework that uses a reward system that encourages software agencies to identify tax policies. Agents simulate how people react to taxes in a two-dimensional grid world called Gather-and-Build. These agents collect resources and earn coins by building houses of stone and wood, exchanging resources with other agents in exchange for coins or moving in the environment to collect resources from tiles.
While each agent in the simulation was making money, an AI planning module (The Economist) learned to set taxes and subsidies to promote certain global goals. Specifically, planners learn tax returns similar to u.S. federal income taxes. It also includes social welfare functions that take into account the trade-off between income equality and productivity. In doing all this, agents learn to “play around” with features and tax schedules to lower their effective tax rates, in part by exploiting loopholes such as alternating between high-income and low-income tax periods.
AI planners and agents have been engaged in a fiscal tug-of-war until they appear to be stable. A single run simulates an economy worth millions of years. In the experiment, Salesforce said, the AI economist came up with a fairer tax policy than the free-market baseline, the Federal 2018 single tax return, and a prominent tax framework known as the Saez tax formula.
Although Salesforce cautions against applying the AI economist’s policies to the real economy. But the company says that as a theoretical tool based on ethical and scientific judgment, the framework could provide economists and governments with unprecedented modeling capabilities to enhance research on sustainability, productivity and equality — especially in the economic aftermath of the new crown pneumonia pandemic.