Death, Taxes, and the ‘AI Economist’
Economists, fond of their models, may have another AI-based tool for designing new tax policies that address growing economic inequality while attempting to boost the productivity that would give new meaning to the aphorism, “A rising tide lifts all boats.”
In a paper published this week by the research arm of enterprise software giant Salesforce (NYSE: CRM) and Harvard University, researchers used reinforcement learning (RL) techniques to design a tax policy that addresses income inequality and its relationship to productivity. RL varies from supervised machine learning, in which algorithms are retrained to maintain accuracy, by instead employing a feedback loop of learning “agents” built directly into the process.
The “AI Economist” is based on a RL framework that combines an agent with tax policy to learn using “observable data alone” rather than modeling assumptions. The platform is touted as able to learn “dynamic tax policies” that boost equality without sacrificing productivity in a simulated economy.
The researchers said their AI approach improved the tradeoffs between equality and productivity by 16 percent, outperforming alternative tax frameworks, including a proposal called “progressive wealth taxation.”
“Our vision for the AI Economist is to enable an objective study of policy impact on real-world economies, at a level of complexity that traditional economic research cannot easily address,” the researchers said in a blog post describing their findings. “We believe the intersection of machine learning and economics presents a wide range of exciting research directions, and gives ample opportunity for machine learning to have positive social impact.”
Reinforcement learning frameworks use two or more learning agents, creating what the researchers note is a challenge since agents must learn collectively. Mistakes in exploring random environments can be introduced, thereby slowing learning. Those error can hinder the framework’s simulations in which a “social planner” learns a tax policy in parallel with agents seeking to optimize their behavior.
Hence, the researchers specified that each learning agent use information only it could observe. This “hybrid approach” boosted learning efficiency “without assuming information” or sharing among learning agents, the researcher said.
The bottom line? The researchers said AI Economist yielded a 47 improvement in equality compared to a “free market” approach with no taxation or wealth redistribution. Productivity declined, but only by 11 percent, they reported.
Specifially, the RL framework set a higher top tax rate for annual income above $510,000, lower rates for income between $150,000 and $510,000 annually and a combination of tax rates for incomes below $160,000 a year.