【主讲人简介】：彭一杰，南京大学至诚特聘教授，工程管理学院、计算机学院博导，复杂系统管理研究院院长、智能系统管理系主任。北京大学信息技术高等研究院多智能体与工业智能实验室主任。本科毕业于武汉大学数学与统计学院，从复旦大学管理学院获博士学位。曾任美国马里兰大学史密斯商学院博士后，美国乔治梅森大学工学院、北京大学工学院助理教授，北京大学光华管理学院副教授，北京大学人工智能研究院多智能体与社会智能中心执行主任。主要研究方向包括仿真建模与优化、金融工程与风险管理、人工智能、健康医疗等。主持优秀青年科学基金、原创探索计划、杰出青年科学基金等。在《Journal of Machine Learning Research》，《Operations Research》，《INFORMS Journal on Computing》和《IEEE Transactions on Automatic Control》等高质量期刊与人工智能顶会上发表学术论文，曾获INFORMS Outstanding Simulation Publication Award、教育部第九届高等学校科学研究优秀成果二等奖。担任或曾担任Journal of System Science and Engineering、《系统管理学报》领域主编，Asia-Pacific Journal of Operational Research、Journal of Systems Science and Information、《运筹学报》副主编，北京运筹学会副理事长、管理科学与工程协会人工智能技术与管理应用分会副理事长、管理科学与工程协会理事。

【内容简介】：Modern retail and platform operations couple four levers—production, inventory, pricing, and recommendation—through shared inventory dynamics and a common profit objective, yet most models optimize subsets of these decisions in isolation. We formulate the integrated problem as a finite-horizon Markov decision process and develop the Pretrained Coordination Model (PCM), a management world model built on a single sequence-model backbone: it distills logged trajectories of incumbent heuristic and reinforcement-learning policies into one coordinated policy whose per-product tokenization keeps its parameter count invariant to the number of products; its shared dynamics and outcome heads answer counterfactual queries about joint decisions; and it improves decisions by planning over short imagined rollouts behind an error-calibrated guard. Our analysis provides a separability benchmark identifying when coordination pays, approximation and return-conditioned stitching guarantees under which the distilled policy can exceed every data-generating policy, an information-structure bound quantifying the cost of siloed control, and a guarded-improvement guarantee—exact in the deterministic limit—converting model accuracy into a no-regress planning property. Computational experiments show that (i) coordinated control raises expected profit and reduces dispersion relative to heuristic and deep reinforcement-learning baselines, most strongly under tight cross-process coupling and nonstationary demand; (ii) a single fixed-capacity policy trained from logs alone matches a full-information reference at small scale and retains roughly half to two-thirds of its performance as the assortment grows from 2 to 100 products, where a flat deep reinforcement-learning controller collapses; and (iii) operated in planning mode, PCM exceeds the full-information reference at small scale and matches it at larger scale—with higher service levels and fewer stockouts throughout—a gain traced through counterfactual evaluation to dynamics models trained with action coverage rather than expert logs alone.

【讲座时间】：2026年7月16日（星期四）下午14:00

【讲座地点】：东南大学九龙湖校区人文社科科研楼18层1801室