400 KB17 KB197720262001Novak, Bojanštevilka:1letnik:25str. 83-88ISSN:0350-5596COBISSID:6268182URN:URN:NBN:SI:doc-RYFWT60NenSlovensko društvo InformatikaInformatica (Ljubljana)mehka logikanevronske mrežestrojno učenjeSoft computing on small data sets|The fusion of artificial neural networks (ANN) with soft computing enables to construct learning machines that are superior compared to classical ANN because knowledge can be extracted and explained in the form of simple rules. If the data sets are small it is hard to find the optimal structure of ANN because classical statistical laws do not apply. One possible remedy is the structural risk minimization method applied together with a VC dimension estimation technique. The construction of the optimal ANN structure is done in higher dimensional space. The distortion of an image in this transformationcan happen and the widely used expression for VC estimations based on minimal input data enclosing hypersphere and margin is not precise. An improvement ov VC dimension estimation is presented. It enables better actual error estimation and is particularly suitable for the small data sets. Tests on some real life data sets have confirmed the theoretical expectationsTEXTznanstveno časopisjejournalsSlovenian National E-content AggregatorNational and University Library of SloveniaSlovensko društvo Informatika