Learning Algorithms for Dynamic Inventory and Pricing Optimization Problems

学习算法旨在解决决策者对系统结构的一部分或整个结构的先验信息有限或没有先验信息的动态优化问题。事实上，在许多应用中，系统是如此复杂，以至于可能不可能预先给出一个包含所有已知系统参数的精确理论模型。在这些设置中，决策者需要在决策过程中学习这些信息，例如，通过从收集的数据中提取信息来设计算法以提高系统性能。这种将优化视为动态学习过程的观点近年来变得非常突出，并产生了一些有希望的结果。本研究将为供应链管理中的动态操作优化问题开发有效的数据驱动学习算法。它将通过整合和扩展机器学习和随机优化的思想和技术来完成，算法的有效性将通过后悔来衡量，后悔被定义为单位时间内利润（成本）的平均损失（增量），与拥有完整底层系统结构信息的千里眼相比。该项目涉及多个学科，如制造、计算、运筹学和商业分析，多学科方法将鼓励代表性不足的群体参与，并对研究生和本科教育产生积极影响。将针对几种动态操作优化问题开发高效的数据驱动算法，包括带缺货替代的多产品动态库存控制、客户选择模型下的多产品定价与库存控制、变化与季节环境下的库存与定价优化、竞争环境下的动态优化、参考点效应下的动态库存控制与定价。以及动态的联合运营和营销决策。该研究整合了统计学、博弈论、机器学习、运筹学和行为科学等各个领域的前沿知识和思想，将产生在理论和经验上都表现良好的高效学习算法。随着公司中数据可用性的增加，该项目的研究将帮助他们更好地利用数据进行智能定价和库存决策，从而增加收入并最大限度地降低成本。

Learning algorithms aim to solve dynamic optimization problems in which the decision maker has limited or no prior information about either a part of or the entire system structure. Indeed, in many applications, the system is so complex that it may not be possible to lay out an exact theoretical model with all system parameters known in advance. In these settings, the decision maker needs to learn such information during the decision making process, e.g., by extracting information from the collected data, to design algorithms for improved system performance. This view of optimization as a dynamic learning process has become prominent in recent years and has led to some promising results. This research will develop efficient data-driven learning algorithms for dynamic operations optimization problems in supply chain management. It will be accomplished by incorporating and extending ideas and techniques from machine learning and stochastic optimization, and the effectiveness of the algorithms will be measured by regret, defined as its average loss (increment) in profit (cost) per unit time compared with a clairvoyant who has complete information about the underlying system structure. This project involves several disciplines such as manufacturing, computing, operations research, and business analytics, and the multidisciplinary approach will encourage participation from under-represented groups and positively impact graduate and undergraduate education. Efficient data-driven algorithms will be developed for several classes of dynamic operations optimization problems, including multi-product dynamic inventory control with stockout substitutions, multi-product pricing and inventory control under customer choice models, inventory and pricing optimization under changing and seasonal environments, dynamic optimization in competitive environments, dynamic inventory control and pricing with reference point effect, and dynamic joint operations and marketing decision making. The research integrates cutting-edge knowledge and ideas from various areas, such as statistics, game theory, machine learning, operations research, and behavioral sciences, and it will lead to efficient learning algorithms that perform well both theoretically and empirically. With the increasing availability of data in companies, the research from this project will help them better utilize data for intelligent pricing and inventory decisions, and increase revenue and minimize cost.