大规模生产和服务系统智能化决策的理论研究

生产和服务系统的发展水平很大程度体现在稳定，高效的运营管理决策。本项目的考虑结合大流量近似和统计学习方法智能化求解生产和服务系统一类困难但重要的动态决策问题。其主要学术贡献在于创新性的发现了统计学习方法和大流量近似法这两个学术前沿方向在智能化求解动态决策问题中良好的互补性。项目分三个部分。第一，本项目使用强化学习方法订单装配系统最优生产和库存联合策略求解。项目第二部分为在未知客户耐心的条件下，使用线上学习方法对排队系统负载和调度进行优化。第三部分考虑更广泛的格点上局部转移的马氏决策过程，使用线上强化学习方法对最优策略进行近似求解。本项目计划对以上每个动态决策问题提出首个理论上保证收敛和渐近最优，实际运行中高效和可靠的智能化求解算法，进而提出动态决策问题的一般性求解框架。同时，本项目的现有成果对统计学习方法和大流量近似法均有创新和发展，并为更广泛的决策问题提供了理论支持。

This project is concerned to approximately solve a variety of dynamic decision-making problems that arise in the management of production and service systems. Practical but also challenging, dynamic decision-making problems consider a series of actions taken over time to achieve some overall goal, given that later decisions must respond to the change of the environment and the consequence of earlier actions that have been made. Many established problems in operations research fall into the category, where the operations research community has seen that understanding how these problems are solved can significantly benefit the operation of supply chains, healthcare, manufacturing, and cloud computing, all of which can be especially helpful in the covid-relieving year 2022...Our research proposal aims to combine machine learning algorithms with heavy-traffic asymptotic theory to approximately solve various dynamic decision-making problems in operations research. This research proposal currently has three independent research projects on this topic. First, we consider applying an online learning algorithm to solve a sequential capacity-allocation problem when the system manager needs to learn customers’ patience while periodically reallocating the system. Second, we consider establishing interpretable reinforcement learning methods that guarantee a close-to-optimal solution for a dynamic production and inventory control of assemble-to-order production systems. Third, we consider developing a framework of approximately solving Markov decision problems that are defined on a grid and have only local state transitions. For each of the three dynamic decision-making problem, we plan to formulate an asymptotic approximation of the problem and then develop machine learning algorithms that guarantee a close-to-optimal solution.