On the Input Parameter Representation for Input-Output Mappings Using Vector Quantization or Neural Networks

In many applications, a mapping between two sets of parameters is desired. Quite often, however, the relationship between input and output is known only from experimental data. One way to form an input-output mapping in such situations is to cluster the input using either a classical clustering algorithm or an unsupervised neural network. Each resulting centroid is then labelled with an appropriate output, and the mapping becomes essentially a table look-up procedure. If sufficient centroids are used, such systems work well, but generally a relatively low mapping accuracy is obtained for a given number of centroids.