A new improved online algorithm for multi-decisional problems based on MLP-networks using a limited amount of information

In this paper the authors extend their previous (1993) algorithm (iterative conjugate gradient singular value decomposition) to the general case of MLP-networks having an arbitrary number of output units. Moreover, it is shown that the use of some suitable thresholds in the matrices of weights allows a further increase of the efficiency of the method. Numerical experiments confirm that the algorithm is particularly effective for the online training of "medium size" MLP-networks using a low number of patterns.