TRN32 performs topology-sensitive vector quantization. Unlike the SOM and the SOMnia program the adjacency structure of the prototype system is not predetermined but will be learned and unlearned during network training. TRN32 is a Visual Pascal stand-alone version of the TRN and has been enriched with various additional features for initializing, computing statistical neighborhood results, or Sammon projections.

The extended beta version 2008 offers new functionality. It processes up to 30.000 cases, 100 variables, and 30 classes (prototypes). The weighted Simple Structure Index (wSSI) assists in determining the number of classes. New bubble charts visualize the results gained with block data (e.g. if each respondent evaluates the same number of product brands or destinations); the coefficient of similarity competition and a crosstabulation of classes (absolute and normalized frequencies) are provided for each pair of brands requested by the user.

TRN32 (Mazanec, 1997) is based on Martinetz, Th, Berkovich, St G and K Schulten (1993)'s "Neural Gas" Network for Vector Quantization and its Application to Time-Series Prediction', IEEE Transactions on Neural Networks, 4(4) 558-569, and

Martinetz, Th and K Schulten (1994) 'Topology Representing Networks', Neural Networks, 7(5) 507-522.

A first sample application in the marketing and consumer research literature is Mazanec, J. A. (1999), Simultaneous Positioning and Segmentation Analysis with Topologically Ordered Feature Maps: A Tour Operator Example, Journal of Retailing and Consumer Services, Special Issue on Marketing Applications of Neural Networks, 6 (4): 219-235. J. Mazanec and his group employ TRN for data reduction in product positioning and market segmentation. You may use the same test data files CITY468.DAT (26 cities * 18 years = 468 cases) and HOTEL152.DAT (4 hotels * 38 respondents = 152 cases) as for the SOMnia program.

Note: You have to read and accept our download conditions before!