光华讲坛——社ng28南宫国际app名流与企业家论坛第6983期
主题：Data-Driven Uncertainty Sets数据驱动的不确定集
主讲人：新加坡国立大学商南宫28加拿大软件 Melvyn Sim教授
主持人：工商管理南宫28加拿大软件 章宇教授
时间：7月9日10:00-11:00
地点：柳林校区诚正楼1122ng28南宫国际app议室
主办单位：工商管理南宫28加拿大软件 国际交流与合作处 科研处
主讲人简介：
Professor Melvyn Sim is a Professor and Provost’s Chair Professor in the Department of Analytics and Operations at the National University of Singapore (NUS) Business School. His research covers decision-making and optimization under uncertainty, robust optimization, distributionally robust optimization, stochastic modeling, operations research, and business analytics, with broad applications in finance, supply chain management, healthcare, and engineered systems. He has published more than 30 papers in UTD 24 leading business journals, including about 22 papers in Operations Research, 10 papers in Management Science. His representative work, “The Price of Robustness,” is regarded as one of the foundational papers in robust optimization and currently has approximately 6567 citations on Google Scholar. Professor Sim currently serves as a Department Editor for Manufacturing & Service Operations Management (MSOM). His research has substantial academic influence and practical relevance in robust optimization, supply chain management, and healthcare operations, reflecting broad recognition from both the international academic community and relevant industry practice.
Melvyn Sim教授现任新加坡国立大学商南宫28加拿大软件分析与运营系教授、教务长讲席教授，研究领域涵盖不确定性环境下的决策与优化、鲁棒优化、分布鲁棒优化、随机建模、运筹学与商业分析，并广泛应用于金融、供应链管理、医疗健康和工程系统等场景。他在Management Science、Operations Research、Production and Operations Management等UTD24顶级商南宫28加拿大软件期刊发表论文三十余篇，其中 Operations Research约22篇、Management Science约10篇。其代表作The Price of Robustness是鲁棒优化领域的基础性论文之一，Google Scholar当前显示引用量约6567次。Melvyn Sim教授目前担任Manufacturing & Service Operations Management(MSOM)的 Department Editor，其研究成果在鲁棒优化、供应链与医疗运营等领域具有较高学术影响力和实际应用价值，体现出国际学术界与相关产业实践领域对其研究的广泛认可。
内容提要：
This talk focuses on a central question in data-driven robust optimization: how to construct uncertainty sets directly from historical data in a statistically justified yet non-overly-conservative way. Traditional robust optimization often starts by assuming or estimating a probability distribution and then deriving an uncertainty set. In contrast, this paper develops a deviation-constrained minimum-volume coverage framework that builds uncertainty sets directly from empirical observations by minimizing the volume of the coverage region while controlling expected deviation penalties. This approach provides a refined balance between robustness and conservativeness. The talk will introduce how the framework accommodates symmetric and asymmetric deviation metrics, different norms, and penalty functions, and how it establishes finite-sample feasibility guarantees and uniform convergence bounds that clarify the relationship among sample size, uncertainty-set complexity, and estimation error. The talk will also discuss tractable convex and conic reformulations, as well as the way the framework gives a unified geometric interpretation of classical statistical estimators such as empirical means, covariance matrices, quantiles, and mean absolute deviations. Finally, the talk will present extensions to heterogeneous datasets, directional uncertainty, sub-Gaussian-inspired uncertainty sets, and robust decision-making with predictive information, offering a unified geometric perspective for data-driven decision-making and optimization.
本悟空体育围绕数据驱动鲁棒优化中的一个核心问题展开：如何直接从历史数据中构造既有统计依据、又不过度保守的不确定集。传统鲁棒优化通常先假定或估计概率分布，再据此设计不确定集；而本文提出的偏差约束最小体积覆盖框架，直接基于样本观测构造不确定集，在控制期望偏差惩罚的同时最小化覆盖区域体积，从而在鲁棒性与保守性之间取得更精细的平衡。悟空体育将介绍该框架如何统一处理对称与非对称偏差度量、不同范数和惩罚函数，并给出有限样本可行性保证与一致收敛界，说明样本量、不确定集复杂度和估计误差之间的关系。同时，悟空体育还将讨论其可计算的凸优化与锥优化重构形式，以及该框架如何把经验均值、协方差矩阵、分位数和平均绝对偏差等经典统计估计量解释为最小体积原则下的特殊情形。最后，悟空体育将介绍该方法在异质数据、方向性不确定性、次高斯型不确定集和带预测信息的鲁棒决策中的扩展，为数据驱动决策与优化提供一个统一的几何建模视角。
En
