P. V. Saraev

The article is devoted to development and investigation of effectiveness of feed-forward neural networks teaching based on weights vector decomposition. Vector of nonlinear weights v including hidden neurons weights and vector of linear weights u including output neurons weights are connected by the analytic equation . The equation contains vector of teacher values , pseudo-inverse matrix to matrix F(v) of outputs of last hidden layer neurons. The equation lets to determine optimal vector u in a non-iterative manner. This leads to reduction of adjusting weights space dimension. It is developed the formula for evaluation of optimization direction for nonlinear weights v each iteration. It is developed the software for experiments with new method of neural networks teaching. The software was used for numerical experiments with the developed method to determine its effectiveness. Analysis of experiments results showed effectiveness of developed teaching method. Particularly, it leads to better decision and better optimization direction. The developed method is recommended to using with small teaching sets. It can be used to teach of radial basis function neural networks and to adjusting of hybrid neuro-fuzzy systems.