Numerical investigation of the multilayer perceptron retraining (Q1281242)

Neural networks with an architecture of multilayer of perceptrons (MLP) are widely used in mathematical test data handling. Especially they are successfully used now in plasma diagnostics of tokamak systems where it is necessary to handle voluminous database in a real time. The MLP's have the important property that they can be trained and retrained. Two methods of perceptrons retraining -- an iteration method and a direct one -- are proposed in the article. The influence of the number of latent layers in the MLP on its stability with respect to the input signal noises after retraining is investigated.