In April, Salesforce detailed the AI Economist, a research environment for understanding how AI could improve economic design. The company pledged to eventually make the codebase available in open source, and today marks the release of the initial version.
Studies have shown that income inequalities can negatively impact economic growth, opportunity, and health, and tax policy has similar effects. Over-taxation can discourage people from working, for example, leading to lower productivity. But it’s difficult to experiment with policies in the real world because economic theory relies on tough-to-validate assumptions (e.g., people’s sensitivity to taxes).
With the AI Economist, Salesforce wants to spearhead the development of a tool to guide tax policy. The company is calling on AI researchers, the economics community, and policymakers to contribute code and collaborate on research; volunteer their expertise and build rich simulations; and indicate which social issues could be addressed with the framework.
“The moonshot goal of this project is to build a reinforcement learning framework that will recommend economic policies that drive social outcomes in the real world, such as improving sustainability, productivity, and equality,” Salesforce machine learning research scientist Stephan Zheng wrote in a blog post. “To achieve this, we’ll need to advance AI, challenge conventional economic thinking, and create AI that can ground and guide policymaking. While none of these tasks are easy, together they make for a true moonshot.”
Above: The game environment in the AI Economist.
Image Credit: Salesforce
The AI Economist is a two-level, deep reinforcement learning framework that uses a system of rewards to spur software agents to identify tax policies. Agents simulate how people might 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, and they trade with other agents to exchange resources for coins or move around the environment to gather resources from tiles.
While each agent in the simulation earns money, an AI planner module (“the economist”) learns to enact taxes and subsidies to promote certain global objectives. Concretely, the planner learns a tax schedule analogous to U.S. federal income taxes. It also incorporates a social welfare function that considers the trade-off between income equality and productivity, where “equality” is defined as the complement of an index on the distribution of wealth (in other words, the cumulative number of coins owned by an agent after taxation and distribution). As it does all this, the agents learn to “game” the function and tax schedule to lower their effective tax rate, in part by exploiting loopholes like alternating between tax periods with high and low incomes.
The AI planner and agents engage in this fiscal tug-of-war until a semblance of stability is achieved. Millions of years’ worth of economies are simulated in the course of a single run. During experiments, Salesforce says the AI Economist arrived at a more equitable tax policy than a free-market baseline, the U.S. federal single-filer 2018 tax schedule, and a prominent tax framework called the Saez tax formula.
Salesforce cautions against applying the AI Economist’s policies to real economies. But the company asserts that as a theoretical tool used ethically with sound scientific judgment, the framework could give economists and governments unprecedented modeling capabilities to augment research into improving sustainability, productivity, and equality — particularly in the economic aftermath of COVID-19.
The research team behind the AI Economist plans to host a live Reddit Q&A on Friday at 2 p.m. Eastern (11 a.m. Pacific) on r/Futurology.
