Machine learning for scheduling problems

Abstract

In the modern competitive world, it is crucial for manufacturers to be able to quickly and accurately predict the best possible manufacturing schedule for their machines. In this thesis we try to create a machine learning algorithm that is capable of creating a near optimal schedule, based on prior knowledge i.e. previously obtained optimal solutions, to quickly predict the daily schedule of a flow shop. As described, solving an integer programming problem to obtain the optimal schedule is not viable. Though other techniques like heuristics exist, we propose a different way, not only to obtain near optimum schedules but also to provide lower and upper bounds to new problems so that when a new problem arises it will be easier for mixed integer programs to solve them. Heuristics have the disadvantage that it is not clear how close the solution that is proposed is to the optimal one. With machine learning it is clear the linkage between the schedule that is proposed and the optimum one due to the metrics that are in place.