Data-driven Dynamic Scheduling

在动态调度问题中，目标是在不断变化的环境（例如，机器故障；新的更紧急的病人到达）中，随着时间的推移，将到达的工作（例如，制造业中的产品订单；医疗保健中的病人）最优地分配给资源（例如，制造业中的机器；医疗保健中的护士）。虽然调度软件中已经提出并实现了许多调度优化方法，但在实践中，最终得到的最优调度并不一定能为最终用户所接受。通常，在实现之前需要手动修改进度，并且，如果进度反复不能满足用户的期望，软件本身可能会被丢弃。该建议的主要前提是，系统的真实特征和用户偏好是通过收集有关问题及其过去（实现的）解决方案的数据来捕获的，因此，通过数据驱动的建模，可以减少或消除软件创建的时间表与用户期望的时间表之间的不匹配。现代信息系统能够收集过去作业和资源特征（例如，到达时间、处理时间）、实时系统状态更新（例如，系统中当前作业的数量）和过去的调度决策（例如，过去由管理人员做出的决策）的数据。传统上，这些数据源都是由不同的研究领域单独考虑的，分别是排队理论、经典调度和逆调度。我的研究计划将通过组合调度、排队理论和逆优化技术的整合，利用历史和实时系统和偏好数据的可用性。由此产生的混合调度模型将比传统方法更好地捕获系统特征和用户偏好，产生更容易被用户接受的调度。因此，这项研究有可能在实践中增加调度软件的采用，这将反过来导致这些系统的更大的生产力和效率。

In dynamic scheduling problems, the aim is to optimally assign arriving jobs (e.g., product orders in manufacturing; patients in healthcare) to resources (e.g., machines in manufacturing; nurses in healthcare) over time in a continuously-changing environment (e.g., machines break down; new more urgent patients arrive). While many optimization methods for scheduling have been proposed and implemented in scheduling software, in practice the resulting optimal' schedules are not necessarily acceptable to the end users. Frequently, the schedules are manually modified before implementation, and, if the schedules repeatedly do not meet user expectations, the software itself might be discarded. The main premise of this proposal is that the true characteristics of the system and user preferences are captured by the data collected about the problem and its past (implemented) solutions, and that as a result the mismatch between the schedules created by software and those desired by the user can be reduced or eliminated through data-driven modelling. Modern information systems are capable of collecting data on past job and resource characteristics (e.g., arrival times, processing times), real-time system status updates (e.g., number of jobs currently in the system), and past scheduling decisions (e.g., decisions made by managers in the past). Traditionally, each of these data sources has been considered in isolation by distinct research areas, namely queueing theory, classical scheduling, and inverse scheduling, respectively. My research program will leverage the availability of both historical and real-time system and preference data through the integration of combinatorial scheduling, queueing theory and inverse optimization techniques. The resulting hybrid scheduling models will capture system characteristics and user preferences better than traditional approaches, producing schedules that would be more readily accepted by the users. Thus, this research has the potential to increase the adoption of scheduling software in practice, which will in turn lead to greater productivity and efficiency of those systems.