[论文解读] The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies

Stephan Zheng, Alexander Trott+5 more|arXiv (Cornell University)|Apr 28, 2020

Energy, Environment, and Transportation Policies被引用 51

一句话总结

AI Economist 训练一个两层深度强化学习社会规划者，在经济仿真器中学习动态税收政策，实现在平等-生产力权衡方面比 Saez 框架高出 16%，并对策略性行为具有鲁棒性，且在人类参与实验方面结果令人乐观。

ABSTRACT

Tackling real-world socio-economic challenges requires designing and testing economic policies. However, this is hard in practice, due to a lack of appropriate (micro-level) economic data and limited opportunity to experiment. In this work, we train social planners that discover tax policies in dynamic economies that can effectively trade-off economic equality and productivity. We propose a two-level deep reinforcement learning approach to learn dynamic tax policies, based on economic simulations in which both agents and a government learn and adapt. Our data-driven approach does not make use of economic modeling assumptions, and learns from observational data alone. We make four main contributions. First, we present an economic simulation environment that features competitive pressures and market dynamics. We validate the simulation by showing that baseline tax systems perform in a way that is consistent with economic theory, including in regard to learned agent behaviors and specializations. Second, we show that AI-driven tax policies improve the trade-off between equality and productivity by 16% over baseline policies, including the prominent Saez tax framework. Third, we showcase several emergent features: AI-driven tax policies are qualitatively different from baselines, setting a higher top tax rate and higher net subsidies for low incomes. Moreover, AI-driven tax policies perform strongly in the face of emergent tax-gaming strategies learned by AI agents. Lastly, AI-driven tax policies are also effective when used in experiments with human participants. In experiments conducted on MTurk, an AI tax policy provides an equality-productivity trade-off that is similar to that provided by the Saez framework along with higher inverse-income weighted social welfare.

研究动机与目标

• Motivate the design of tax policies to balance equality and productivity in dynamic economies.
• Develop a data-driven, model-free framework that learns tax schedules from observable data without relying on strong economic assumptions.
• Demonstrate emergent agent behavior and policy robustness in simulated and human-subject experiments.
• Evaluate how AI-driven taxes perform against traditional benchmarks and under potential tax-avoidance behaviors.

提出的方法

• Introduce a two-level reinforcement learning framework with a government planner and adaptive agents in a Gather-and-Build economy.
• Model the tax policy as a neural network-based schedule over income brackets.
• Define agent rewards via a concave utility over coins and linear labor costs, enabling RL to optimize total discounted utility.
• Use a dynamic economic simulator to validate learned policies against economic theory (e.g., agent specialization).
• Stabilize two-level RL training with curricula and entropy regularization while sharing policy parameters across agents.

实验结果

研究问题

• RQ1 Can an AI planner learn tax policies that improve the trade-off between equality and productivity in a dynamic economy?
• RQ2 Do AI-driven tax policies exhibit qualitative differences (e.g., top tax rates, subsidies for low incomes) compared to baseline schemes like Saez?
• RQ3 Are AI-driven policies robust to emergent tax-avoidance or gaming behaviors by agents?
• RQ4 Do AI tax policies transfer effectiveness to human experiments (e.g., MTurk) and compare favorably on social welfare metrics?

主要发现

• AI-driven tax policies improve the equality-productivity trade-off by 16% over baseline policies including the Saez framework.
• Learned tax schedules differ qualitatively from baselines, e.g., higher top tax rates and higher net subsidies for low incomes.
• AI-driven policies remain effective under emergent tax-avoidance strategies by AI agents.
• Experiments with human participants (MTurk) show the AI policy achieves a competitive equality-productivity trade-off with higher inverse-income weighted social welfare.
• The methodology shows promise for applying learned tax policies in real-world-like settings without explicit economic modeling assumptions.