An Improved Method for Job Shop Scheduling by Means of Machine Learning and Mathematical Optimization

This paper is concerned with solving a job shop scheduling problem (JSSP) optimally in a shorter time by predicting a good solution based on a database of scheduling performed previously and then providing it as an initial solution in the optimization algorithm. Although several methods based on this framework have been developed, they have a problem that a considerable time is required for building learners. For reduction of learning time, an improved method using multiclass classification is provided in this paper. By numerical evaluation, it was shown that the learners in the proposed method can be built in a considerably shorter time than the previous method when the same hyperparameters are used. But, from the point of view of solution time, a clear advantage could not be confirmed. Further investigation considering accuracy of the learners and appropriate settings of hyperparameters will be performed in a future work.