Some important topics for fast search.

The scanning of the density codebooks is done by using the partial distance computation (PDC). This is an elementary but efficient method for finding the BMU from a codebook. In the PDC the computation of the distance between two vectors aborts when the sum of the added component-wise differences exceeds a reference distance. Therefore the fastest execution time can be expected, if the rejections of the non-K-best matches can be determined as soon as possible. This can be assured on the average, if the candidates are ordered so that the most likely winners are checked first. For the same purpose the components of the feature vectors should be processed in the order of decreasing significance. With no special knowledge about the rank of the candidates except the continuous character of the signal, a good average performance can be expected, for example, if the candidates are scanned according to the distance in SOM topology from the expected winner. The better the topology of the input space is preserved in the SOM the better results this search order will give. Similarly, with no special knowledge about the rank of the components, it is best to organize them according to the decreasing variance, in general.

The frequency of the complete search affects the ability to react for fast changes in the signal characteristics and is, along with the number of the K best matches and the size of the basic search neighborhood, a controllable variable which can be used to increase either the accuracy or the speed of the search. It might become necessary to increase the search area and the frequency of the complete search, when the neighborhood used in the codebook training has decreased to zero for better accuracy and the ordering starts to disappear. However, the success of this approximate search is clearly dependent on the way the codebook is organized so that poor performance may occur, for example, if the best matches lie on the different edges of the grid. One way to reduce the risk of bad approximations might be to monitor the spread of the K best matches after the complete scan and start other local searches around the next best clusters, if necessary.