2025年第54期(总第1095期)

演讲主题：Learning to Bid in Non-Stationary Repeated First-Price Auctions

主讲人：胡子豪 香港科技大学研究助理教授

主持人：关旭 供应链管理与系统工程系教授、主任

活动时间：2025年9月22日（周一）10:00-11:30

活动地点：管院大楼107教室

主讲人简介：

胡子豪，香港科技大学数学系研究助理教授。他于乔治亚理工学院（Georgia Tech）获得机器学习博士学位，师从Jacob Abernethy教授。胡博士的研究领域涵盖运筹学与数学优化，研究重点为在线广告拍卖与黎曼优化。其研究得到香港研究资助局优配研究金（GRF）的支持，相关成果已发表于《Mathematical Programming》、COLT、NeurIPS、CVPR及AISTATS等顶级学术期刊与会议。

活动简介：

First-price auctions have gained significant traction in digital advertising markets, exemplified by Google's transition from second-price to first-price auctions. Unlike second-price auctions where bidding one's private valuation is optimal, determining bidding strategies in first-price auctions is more complex. From a learning perspective, the learner (bidder) can interact sequentially with the environment (opponents) to infer their behaviors. Existing research typically assumes specific environmental conditions and benchmarks against the best fixed policy (static benchmark). While ensuring strong learning guarantees, this static benchmark can deviate significantly from the optimal strategy in mildly non-stationary environments.

To address this, we propose a dynamic benchmark—the sum of highest achievable rewards at each time step—as a more suitable objective. However, achieving no-regret learning against the dynamic benchmark requires additional constraints. We introduce metrics to quantify the non-stationarity of opponents' highest bid sequences and provide minimax-optimal characterization of dynamic regret. Our main technical tool is the Optimistic Mirror Descent (OMD) framework with a novel optimism configuration, well-suited for achieving minimax-optimal dynamic regret rates. Synthetic datasets validate our theoretical guarantees and demonstrate superior performance over existing methods.