DescriptionScheduling is a master key to succeed in the manufacturing companies in global competition. Better process scheduling leads to competitive advantage by reducing production cost and increasing productivity. Global competition has obliged the companies to expend their investments in new manufacturing systems. With the arises of these new systems, many of manufacturing problems have appeared such as scheduling problems has gained attention by researchers. Especially, it is important to develop new methodologies in order to improve manufacturing scheduling effectiveness of these new sophisticated systems. Many of researchers have used machine learning for scheduling problems because it is discovering implicit knowledge of expert schedulers that can applied for future schedules generation. In addition, machine learning is able to create flexible schedules depending on the state of the system.
In this dissertation, we present several methodologies for machine learning to scheduling in different manufacturing processes. For the scheduling problem of traditional industries, we first present a machine learning approach for dynamic scheduling of multiple machines. Existing dynamic scheduling algorithms based on classification methods that do not utilize all the available data for the better scheduling problem. A regression-based dynamic scheduling (RDS) algorithm is proposed to improve scheduling performance of classification-based dynamic scheduling. Due to the unknown relationship between predictor variables and output variables, kernel ridge regression is presented to predict the performance of the scheduling based on system status attributes. The scheduling outputs of the proposed RDS algorithm is evaluated with scheduling results of all combinations of dispatching rules from the static job shop scheduling and a classification-based dynamic scheduling.
For the scheduling problem of semiconductor manufacturing system, we present a new machine learning algorithm for complex semiconductor scheduling. An adjustable dispatching rule (ADR) that calculates weighted sum of control factors for determining which job should be processed first. Then, to find near-optimal weight values of the ADR for improving system performance, the real coded genetic algorithm (GA) with fitness approximation is proposed. For the fitness approximation, kernel ridge regression and polynomial regression are applied by using relatively small number of fitness evaluations. The performance of the proposed algorithms is evaluated by using an extensive experiment with existing dispatching rules, fixed weights, and GAs without fitness approximation.
Finally, in order to improve the scheduling performance of semiconductor wafer fabrication, we propose new variable ranking algorithms to identify the contributions of each input variable to output variable. We present a new kernelized general dominance weight (GDW) for ranking of scheduling factors in semiconductor manufacturing system. To build kernel version of GDW, the relevance vector machine regression technique is applied.