报告人
米祈睿 中国科学院自动化研究所 时间 2025年11月4日 星期二 下午 14:00-15:00 地点
602会议室
Abstract Economic decision-making problems are ubiquitous, spanning micro-level bargaining, market-level pricing and advertising, and macro-level policymaking. These problems involve dynamic and asymmetric interactions under resource constraints, often with unstructured text data. Macro-level policymaking is particularly challenging due to the participation of large-scale, heterogeneous agents. The high-dimensional and non-stationary nature of such environments further limits the effectiveness of traditional methods. This talk introduces AI-agent-based methods to tackle these challenges, with a focus on economic simulation and policy optimization. For simulation, a task-specific tax simulator, TaxAI, is extended into EconGym—a unified simulation platform for training, evaluating, and benchmarking AI algorithms across 25+ real-world economic tasks. For policy optimization, the Dynamic Stackelberg Mean-Field Game (DSMFG) framework learns macroeconomic policies through multi-agent reinforcement learning based on population-level dynamics. To further align simulations with real-world human behavior, MF-LLM introduces a mean-field framework that uses large language models to simulate population decision-making grounded in real-world data. Together, these results demonstrate the feasibility of AI-agent-based methods for economic simulation and policy optimization, contributing to the advancement of AI for Economics.
Biography 米祈睿，中国科学院自动化研究所 2021 级直博生，师从汪军教授与张海峰副研究员。2024–2025 年期间，于南洋理工大学安波教授团队开展访问研究。主要研究方向为多智能体系统、群体决策智能与 AI for Economics，致力于构建融合人工智能与经济学理论的智能经济仿真系统。代表性工作包括通用经济仿真平台 EconGym。相关成果发表于 NeurIPS、AAMAS、ECAI 等国际顶级会议，并担任 TPAMI、AAAI、AAMAS、DAI 等国际重要期刊 / 会议审稿人。科研方面，作为核心成员参与国家自然科学基金原创探索计划项目及青年科学基金项目等国家级科研任务。
