2025年第85期(总第1126期)
演讲主题：Data-Driven Robust Inventory Management with Time-Series Demand
主讲人：徐汾 香港中文大学商公司博士后研究员
主持人：杨彦 武信息管理与数据科学系主任、教授
活动时间：2025年11月27日（周四）14:30
活动地址: 管院大楼406教室
主讲人简介：
徐汾，香港中文大学商公司博士后研究员，2021年于中国科公司数学与系统科学研究院获运筹学博士学位。她曾任阿里巴巴达摩院高级算法工程师、清华大学深圳国际研究生院博士后。徐汾的科研兴趣包括分布鲁棒优化、数据驱动的决策制定、随机建模与优化、强化学习、随机动态规划、凸分析、排队网络等，以及他们在供应链与库存管理、动态定价与收益管理等领域的应用。研究成果发表于SIAM Journal on Control and Optimization、Omega等学术期刊。
活动简介：
We study the multi-period stochastic inventory management problem with time-series demand in a data-driven setting. When historical data is limited, the estimate-then-optimize method often suffers from overfitting and poor out-of-sample performance. To address this, we propose a data-driven robust optimization approach that constructs a Wasserstein ambiguity set capturing demand correlation and uncertainty across the entire planning horizon. We identify that this approach enables a recursive solution via robust dynamic programming, and we show that the resulting robust value functions are piecewise-linear and jointly convex, thus a state-dependent base-stock policy is robustly optimal. We also characterize the worst-case distribution as having a threshold form, which reduces the computational complexity of identifying these distributions from exponential to linear in sample size. For the positively correlated first-order autoregressive model, we prove submodularity of the value function, revealing strategic complementarity between inventory levels and realized demand, and establishing monotonicity of both the worst-case distribution’s threshold and the base-stock level. Statistically, we derive finite-sample performance guarantees for the data-driven robust policy relative to the full-information optimal policy, extending existing results by explicitly accounting for demand correlation and distributional uncertainty. Numerical experiments demonstrate the superior out-of-sample performance of our data-driven robust policy, particularly with limited data, and underscore the importance of modeling general time-series demand.
