报告题目：Conditional Generative Learning for Joint Pricing and Inventory Control
报告人：叶志盛教授新加坡国立大学
报告时间：2026年10月15日10:00-11:00
报告地点：腾讯会议ID：531-321-691 密码1015
校内联系人：赵世舜 [email protected]
报告摘要：
We consider data-driven joint pricing and inventory control with contextual information, in which the retailer makes pricing and inventory decision based solely on historical data consisting of demand, price and features. Learning the optimal policy as a function of the features is challenging, as it requires knowledge of the conditional demand distribution, which is difficult to estimate and varies with the pricing decision. We propose a provably efficient deep conditional generative framework to learn the demand distribution conditional on selling price and covariates from historical data. The conditional demand estimator is a deep generator driven by price, features and an easy-to-sample noise variable independent of both. This generator explicitly encodes the indirect effect of the pricing decision on demand, and it allows the data-driven control to be reformulated as a stochastic program and solved via stochastic gradient descent. We establish finite-sample error bounds for the excess cost or risk of the data-driven policy relative to the ground-truth optimal policy, as well as for the estimation error of the cost or risk estimator. Although the policy is learnt from observational data, our sensitivity analysis suggests that confounding biases may have a limited impact on its performance. Comprehensive simulations show that the proposed approach is markedly faster than the state-of-the-art prescriptive benchmarks and achieves a significantly lower expected cost. A case study on real meal delivery data further confirms that our method is at least four times faster than the machine learning benchmarks with at least a 10% reduction in costs under a semi-synthetic evaluation.
报告人简介：
叶志盛教授本科毕业于清华大学，获得材料科学与工程及经济学双学士学位，博士毕业于新加坡国大工业与系统工程系。目前担任新加坡国立大学工业系统工程与管理系教授，主要研究方向包括韧性与可靠性工程、工业统计以及数据驱动的运营管理，在可靠性、统计学和运营管理领域的顶尖期刊上发表了100多篇论文。
