Job-Shop Scheduling Using a Neural Network to Estimate Margins to Due-Dates

Abstract

Job-shop scheduling is not easy to be solved analytically, therefore, it is ordinarily solved by computer simulation using heuristic dispatching rules. The SLACK rule to give the priority to the job having the shortest margin to its due-date is effective to keep the due-dates. The margin is simply calculated by subtraction of remained operation-time from the time to the due-date. However, actual margins become shorter than the calculated ones due to the confliction between jobs.
In the preceding paper of authors^{(16, 17)}, a method to estimate rather accurate margins using two neural networks was proposed in which a three-layer neural network (numbers of neurons; the first layer-the second-the third: 10-20-1) estimated the margins and the other three-layer neural network having the composition of 11-20-1 judged the accuracy of the estimation because the reliability of the estimation was low. The estimated margins were used for scheduling using the SLACK rule instead of ones calculated by the ordinary method when its reliability was judged high. In this paper, a method using only a neural network for the estimation of which the organization is more adequately adjusted is proposed. In the proposed method, the neural network having four layers with fewer neurons as 8-8-8-1 learns the margins to due-dates from schedules made by using the ordinary SLACK rule, as human schedulers do. It is verified by scheduling simulation that the proposed method is effective to improve the maximum lateness to due-dates on job-shop schedules. Furthermore, re-learning of the neural network based on schedules made by the above method is revealed to be effective for further improvement of the schedules.