A neural-net approach to real time flow-shop sequencing

This paper considers the classic problem of sequencing a set of jobs that arrive in different combinations over time in a manufacturing flow-shop. We focus on the development of a two-level neural network that incrementally learns sequencing knowledge. Based on the knowledge gained from learning using a set of training exemplars, the neural network makes real time sequencing decisions for a set of jobs that arrive in different job combinations. In addition to explaining the details regarding the workings of the neural network, we also evaluate its performance for flow-shop sequencing problems. The practical benefit of the neural-net approach is that the neural network incrementally learns the sequencing knowledge and can apply the knowledge for sequencing a set of jobs on a real time basis. We also show that the neural network can be used to develop hybrid genetic algorithms. The experimental results demonstrate that (1) the neural-net approach produces consistently superior solution quality (i.e., makespan) with significantly less computational time than the traditional heuristic approaches; (2) when compared to genetic algorithms the neural-net approach's performances are within 3.4% of those of genetic algorithms, but using only less than 0.2% of the computational time needed by genetic algorithms; and (3) the neural-net approach further improves solution quality and computational time by combining it with genetic algorithms. These results support the efficacy of using the neural-net approach for real time flow-shop sequencing.