Neural Network to Select Dynamic Scheduling Heuristics

ABSTRACT Frequently, several heuristic strategies are relevant to a given production scheduling problem. A choice must be made among them whenever these heuristics have different performances, and when none of them is globally better than the other ones. A neural network approach for selecting the most suited heuristic is discussed. The configuration of the shop floor, the characteristics of the manufacturing program to be carried out, and the performance criteria to optimize are presented as inputs to the first layer. The most suitable heuristic is given as output. Such a neural network is trained using a large sample of simulation results as training examples. This technique is illustrated through the dynamic scheduling problem of a simplified flow shop. A back propagation neural network finds the most appropriate dispatching rules in ninety four percent of the cases. The benefits of the neural network approach over other possible methods are discussed.