A neural network trained by a genetic algorithm as applied to job-shop scheduling

In three-layered neural networks, a new combined method of weight learning based on overall evaluation without teacher data and structural determination to minimize the necessary number of hidden units was proposed using a genetic algorithm, where the length of chromosomes was permitted to change. The purpose of this study was to discuss how to apply the neural network to job-shop scheduling using a bench mark test. It was clarified that concurrent optimizing of the above-mentioned weight learning and structural determination could be achieved when the whole number of individuals, the mutation rate, the initial number of hidden units, and the number of pickup individuals were adequately determined. One of the Muth and Thompson's JSSP bench mark problems (the 6x6x6 test problem) was discussed to be solved minimizing the total operating time. Plural optimal networks were obtained, and among them the network with the smaller number of hidden units had the more excellent adaptability (generality) to the change of problems.