This competition, "Optimal Tax Policy with Heterogeneous Agent Models," is one of the challenges hosted by AAMAS 2025. It focuses on designing optimal tax policies in an economy with heterogeneous agents. In real life, taxation and public spending are two key government policies that affect economic growth and wealth distribution. A good government needs to adjust tax and spending policies based on the needs of society to support economic development, social fairness, and long-term stability. However, people’s economic behaviors are different (heterogeneous), making tax policy design a complex challenge.
Traditional economic models assume that people are rational decision-makers who maximize their own benefits. However, in reality, people are different in many ways, including:
These different behaviors together shape the overall economy, affecting wealth distribution, economic growth, employment, and social welfare.
For a government, designing an optimal tax policy in a society with different types of people is a difficult task. The government must balance several goals, such as:
To study these problems, we introduce the TaxAI Simulator, a simulated economy where different economic actors make decisions, interact, and adjust to changing conditions. This simulator allows us to test how AI can make smart decisions in a complex economic system.
The TaxAI Simulator includes four key roles: government, households, companies, and financial institutions. The government and households make decisions that change over time, while companies and financial institutions follow fixed rules as part of the environment.
For a detailed description of the relationship, see GitHub Link: github link 。
The challenge consists of two tracks, focusing on decision making issues in the economy and society from different roles.
Track 1 focuses on optimizing fiscal policy in a simulated macroeconomic environment. Participants will control a government agent responsible for adjusting income tax, asset tax, and government spending to drive economic growth while ensuring social fairness and stability. The government’s primary objective is to maximize GDP while maintaining a balanced wealth distribution. To simulate a complex economy, the environment includes heterogeneous households following three predefined strategies in varying proportions. The government agent must dynamically adjust tax and spending policies in response to economic conditions to achieve optimal outcomes. The performance of each government agent will be evaluated based on economic effectiveness within the simulation, with rankings determined by overall economic performance.
Track webpage: http://www.jidiai.cn/compete_detail?compete=55
Track 2 focuses on optimizing individual decision-making within an economic system. Participants will design household agents that adjust their working hours and wealth investment based on observed economic and social information—such as average wealth, income, and education levels across different social classes (e.g., wealthy, middle class, and low-income groups)—to maximize personal benefits. The environment consists of 100 independent household agents, each making decisions autonomously. The objective is to maximize personal well-being, which depends on both consumption levels and working hours—higher consumption and shorter working hours lead to greater individual benefits. An optimal household agent must strategically balance work and consumption by leveraging observed information, personal assets, and productivity.
During the competition, participant-designed household agents will interact with three predefined household policies in a heterogeneous economic environment. Performance will be evaluated based on overall economic benefits within the simulation, with rankings determined accordingly.
Track webpage: http://www.jidiai.cn/compete_detail?compete=56
The competition will be held online on the Jidi multi-agent open-source platform (Jidi Platform). The Jidi Platform provides participating competitors with fair agent evaluation and real-time ranking functions online. Users only need to register an account on the Jidi Platform and sign up to participate in the challenge.
Each participating team consists of 1-3 members, including a team leader and two team members. The steps to participate are:
According to the competition rankings, we will award certificates to the top 6 teams and invite them to present their work on-site at the AAMAS 2025 conference.
Institute of Automation, Chinese Academy of Sciences
Nanjing Artificial Intelligence Research of IA
Email: haifeng.zhang@ia.ac.cn
Institute of Automation, Chinese Academy of Sciences
Email: miqirui2021@ia.ac.cn
Nanjing Artificial Intelligence Research of IA
Email:fanwentian23@mails.ucas.ac.cn
Nanjing Artificial Intelligence Research of IA
Email:wushuai@airia.cn
Institute of Automation, Chinese Academy of Sciences
Email: zhaozhiyu2022@ia.ac.cn