Dynamic Job Shop Scheduling Using a Neural Network. 1st Report. A Two-Stage Training Method to Minimize Tardiness Cost.

Abstract

We consider dynamic job shop scheduling problems with a tardiness cost criterion. The most practical scheduling method in dynamic scheduling environments is using dispatching rules. However, there is no dispatching rule that performs best under all scheduling conditions. In this paper, we apply a multi-layer feed-forward neural network to the scheduling, and propose a two-stage training method to minimize the average tardiness cost of jobs under various scheduling conditions. In the first stage, each neural network for its specified scheduling condition is trained separately by a simulated annealing method. In the second stage, a final neural network for all scheduling conditions is trained by a back-propagation algorithm. Simulation results show that the trained neural network outperforms the best dispatching rules in any levels of due-date tightness and job arrival rate.