Multiple neural networks for selecting a problem solving technique

Summary form only given. A description is given of how to build a modular neural network system for selecting a category that fits an instance of a complex problem. The problems to which the method applies defy attempts at accurate categorization via multilayer single neural networks, yet have structures that yield to the sequential application of three distinct types of neural networks. The first two operate together as an internal classifier, producing information that is used to specify which of several generalization networks to apply to the problem. The technique was successfully applied to the selection of vehicle routing models, a complex problem of considerable importance in operations research. In these experiments, the neural network system was trained with a collection of 250 problems and achieved 100% accuracy in identifying the best model for the problems in the training set. The trained system was then exposed to 100 test problems not used in the training, and achieved 93% accuracy. In parallel experiments with single neural networks having differing numbers of layers and topologies, accuracy rates were in the 70-80% range.<>