Модель оптимізації багатоетапного процесу нейромережевого навчання з функцією обмеження втрат у задачі розпізнавання зображень

Abstract

On a pattern of neuronet investigation of the colored image recognition problem the problem of bringing the fastest information processing system to the optimal mode of functioning by conditions of preliminarily unknown distribution of priorities of its parameters is considered in the paper. There has been formalized the corresponding decision making problem, in which alternatives are methods of training the neural network, and its parameters are interpreted as aftereffects. Reasoning from absence of probability measures over the set of indeterminate parameters of the information processing system, throughout the given decision making problem there is solved a matrix game, in which the second player of optimal strategy is used for combined selection of methods of training the neural network, not depending on amount of its being estimated parameters. There has been displayed the advantage of a model of optimizing the multistage process of neuronet training with wastage constraint function in comparison to the clas-sic model of optimal strategy realization, and also to the training method, at which the minimal distance to the utopia point is reached.