No label defined (Q4955682)

The authors describe an architecture of a neural net with two hidden layers which use fuzzy expert rules in its basis functions. Problems of semantic preserving learning algorithm construction are discussed. The authors propose to formulate the constraints which should be preserved in terms of order of parameters of basis functions. The learning is based on a gradient descent technique. Variation of active set technique is used to improve possible incorrect order after each learning step. The problem of elimination of neurons or parameters which are only approximating the noise in the data is considered. Results of the German index DAX forecasting are discussed. For the entire collection see [Zbl 0929.00046].