GOALI: Planning Horizons for Optimal Decision Making Over Time with Applications to Production Systems Optimization

在当今迅速变化的世界背景下，企业每天都面临着做出决策的挑战，这些决策不仅在现在具有经济意义，而且还使企业能够应对技术和市场迅速发展的未来。做出这些决定的关键是所采用的规划视野。从历史上看，规划视野的选择一直是基于传统或工程判断。如果这一视野是短视的，那么由此产生的决定可能是短视的，无法预见可能使今天的决定变得不明智的事件。如果视野太长，可能会花费相当多的资源来收集不相关的数据，由此产生的问题可能会带来巨大的计算需求。本研究的总体目标是提供一个合理的框架，导致规划视野的选择，是有效的，但有远见的，导致的决定是不受意外的研究结束的影响。 这类决策问题的例子包括产能扩张的规模和时机、生产调度和维护任务的规划，以及新设备的更换和购置。我们建议验证的模型和方法开发的问题，共同优化制造和维修计划的背景下，汽车生产沿着在通用汽车公司的生产线的集合。这项研究将与密歇根大学的教师和学生以及通用汽车研发实验室的研究工程人员合作进行。这项工作的智力价值包括建立的方法和条件下，一个可以计算一个最佳的第一个政策的问题与无限的数据，从而扩展解决方案的程序，以涵盖非齐次马尔可夫决策过程。 这项工作的广泛影响在于开发的模型，算法和经验法则的潜力，以帮助决策者决定他们需要在多远的未来做出明智的决定。这项研究将扎根于通用汽车公司的真实的应用竞技场，旨在提高国家制造系统的生产率和可靠性。学生将在通用汽车公司实习，从而直接协助在研究和教育领域的技术转让。

Within the context of today's rapidly changing world, firms are confronted daily with the challenge of making decisions that not only make economic sense now but also position the firm for grappling with a future often characterized by rapidly evolving technology and markets. Key to making these decisions are the planning horizons employed. Historically, selection of planning horizons has been based upon tradition or engineering judgment. If that horizon is myopic, the resulting decision can be short-sighted and fail to anticipate events that can render today's decision unwise. If the horizon is too long, considerable resources may be expended to collect irrelevant data and the resulting problem can present formidable computational demands. The overall goal of this research is to provide a rational framework that leads to a planning horizon choice that is efficient and yet far-sighted, leading to decisions which are undistorted by unanticipated end-of-study effects. Examples of such decision-making problems include the sizing and timing of capacity expansions, planning for production scheduling and maintenance tasking, and the replacement and acquisition of new equipment. We propose to validate the models and methods developed on the problem of jointly optimizing manufacture and maintenance schedules in the context of vehicle production along a collection of production lines at General Motors. The research will be collaboratively pursued with faculty and students at the University of Michigan and research engineering staff at General Motors R & D Laboratories. The intellectual merit of this work includes establishing methods and conditions under which one may finitely compute an optimal first policy to a problem with infinite data, thereby extending solution procedures to cover nonhomogeneous Markov decision processes. The broad impact of this work lies in the potential of the models, algorithms, and rules-of-thumb developed to assist decision makers in deciding how far into the future they need to look to make a wise decision today. This research will be grounded in a real application arena at General Motors with the intent of increasing the productivity and reliability of national manufacturing systems. Students will serve as interns at GM thus directly assisting in the transfer of technology both in the research and educational domains.