动态鲁棒决策

Our main research target is the robustness in dynamic decision models. We conduct research on how to combine robust models with the framework of Thompson Sampling, and make conservation decisions based on updating data and corresponding forecasting results. Furthermore, we design iterative algorithms which can guarantee fast updating speed, decision robustness and efficiency of the system. We also constructed robust optimization models corresponding to distribution shape uncertainty (for example, log-concave or increasing/decreasing failure rate), and managed to solve the models efficiently. ...The team has strong background and solid research in the corresponding areas. The applicant has published 3 papers in Operations Research and 3 other papers in Mathematics of Operations Research, in the field of robust optimization, supply chain management and algorithm design. The Co-PI Xi Chen, who studied machine learning as a postdoctoral fellow under Prof. Michael I. Jordan, has published several papers in journals like Journal of Machine Learning Research, etc.

我们的主要研究对象是动态管理决策模型中的鲁棒性。针对数据的不断更新，我们研究如何将鲁棒模型整合进Thomson Sampling方法的框架，通过不断调整预测做出相应的稳健决策。进一步的，我们设计可以快速更新的迭代算法，同时保证策略的更新速度、稳健性以及有效性。针对对数凹和单调失败率等分布特征的假设，我们研究如何建立相应的鲁棒优化模型并求解。申请人关于鲁棒决策、供应链管理及算法方面的论文已在《Operations Research》发表3篇，在《Mathematics of Operations Research》发表3篇。参与人陈溪师从机器学习泰斗Miachel I. Jordan,已在《Journal of Machine Learning Research》等杂志发表论文多篇。项目组其他成员在库存管理和机器学习领域也都有丰富的研究经验。

项目聚焦动态鲁棒决策理论，从模型角度聚焦如何引入高效信息以提升决策精度，从理论角度研究最优分布及最优解的结构性质并设计高效算法，从应用实践层面在收益管理、库存管理等供应链管理领域利用鲁棒决策方法协助企业提升管理效率，从应用理论层面在管理实践中定位并提炼关键管理问题展开理论研究。 研究成果拓展了分布式鲁棒优化决策模型中可利用分布信息的类型，提升了鲁棒决策模型的决策精度，衍生出了新的研究方向。在库存管理及机体失效率分析等具体应用研究领域，大幅提升了独立随机分布和分布的鲁棒分析精度，改进了多个现有经典结论。在应用实践中也取得了突出成效。.项目整体执行顺利，在主要完成指标上大幅超出预期。其中，依托项目发表代表性成果2篇OR, 1篇MS, 2篇POMS, 2篇EJOR及2篇NRL；团队成员何斯迈与杨超林后续分别获得杰出青年基金及优秀青年基金支持，培养研究生及博士生20余人；产学研结合取得突出的应用实践成效，所参与的京东无人仓项目入围INFORMS学会应用大奖Franz Edelman决赛圈（Finalist Award）。