Üretim sistemlerinin işarete bağlı eşik kuralı ile veriye dayalı kontrolü

With shop-floor data becoming more available, production systems can be controlled more effectively by using data-driven control methods. This thesis focuses on the following research question: how can the decision to produce or not to produce at any time be given depending on the real-time information about a production system?; how can the collected data be used directly in optimizing the policy parameters?; what is the effect of using different information sources on the performance of the system? and how can this choice be made? In order to answer these questions, a production/inventory system that consists of a production stage that produces to stock to meet random demand is considered. The system is not fully observable but partial production and demand information, referred to as markings is available. We propose using the marking-dependent threshold policy to decide whether to produce or not based on the observed markings in addition to the inventory and production status at any given time. An analytical method that uses a matrix geometric approach is developed to analyze a production system controlled with the marking-dependent threshold policy when the production, demand, and information arrivals are modeled as Marked Markovian Arrival Processes. A mixed integer programming formulation is presented to determine the optimal thresholds. Then a mathematical programming formulation that uses the real-time shop-floor data for joint simulation and optimization (JSO) of the system is presented. Using numerical experiments, we compare the performance of the JSO approach to the analytical solutions. The results indicate that the marking-dependent policy introduced here is able to make use of the available partial information effectively and the data-driven joint simulation and optimization is an efficient way for setting the parameters of the policy. Next, we investigate how machine learning methods can be employed to facilitate the optimization of systems controlled with the marking-dependent policies and compare the performance of a number of well-known machine learning methods on this task. Through numerical experiments we show that using machine learning and active learning can improve the performance of complex production systems by speeding up time-consuming optimization problems. Finally, we implement the concepts of this thesis using a Lego production system and discuss the usage of Lego production systems in educational and research related to data-driven control.