Neural performance criterion

The selection of cluster number, which is generally not known beforehand, represents the primary performance criterion. Optimization of performance therefore requires trial-and-error adjustment of the number of clusters. Once the cluster number is established, the neural network structure is used as a way to determine the linear discriminant for interpretation. In effect, the RBFN makes use of known transformed features space defined in terms of prototypes of similar patterns as a result of applying /c-means clustering. 

The fundamental idea behind training, for all neural network architectures, is this pick a set of weights (often randomly), apply the inputs to the network and see how the network performs with this set of weights. If it doesn t perform well, then modify the weights by some algorithm (specific to each architecture) and repeat the procedure. This iterative process is continued until some pre-specified stopping criterion has been reached. 

As mentioned above, Cho and Wysk (1993) utilized the multilayer perceptron to take the place of the knowledge-based system in selecting candidate scheduling rules. In their proposed framework, the neural network will output a goodness index for each rule based on the system attributes and a performance measure. Sim et al. (1994) used an expert neural network for the job shop scheduling problem. In their approach, an expert system will activate one of 16 subnetworks based on whether the attribute corresponding to the node (scheduling rules, arrival rate factor, and criterion) is applicable to the job under consideration. Then the job with the smallest output value wiU be selected to process. 

Sun and Yih (1996) adopted their idea to develop a neural network-based controller for manufacturing cells. In their approach, a neural network was trained to serve as decisionmaker that will select a proper dispatching rule for its associated machine to process the next job. Based on their results, the controller performs well under multiple criterion environments. In addition, when the production objectives change, the controller can respond to such change in a short time. 

As an additional experiment, we apply the MOEA-based method to a different formulation of the neural network optimal design, one in which the optimization criteria are the minimization of validation error and network computational complexity. The latter is measured here via the Akaike Information Criterion (AIC) [6], defined as —2 In(MSE) - - 2k, where k is the number of weights in the ANNs. This second goal might be important, for instance, in contexts where the classifier ANN must be used in real-time and therefore should be computationally cheap, still guaranteeing robust classification performance. 

---