FIELD: machine training.
SUBSTANCE: group of inventions relates to machine training and can be used to evaluate training objects. Method comprises obtaining a first set of training samples, comprising many features, iterative training of the first predictive model based on a plurality of features and generating of a corresponding first prediction error indicator; analysis of the corresponding first prediction error index for each iteration to determine an “over-fit” point, and determining at least one initial evaluation point; receiving data from a new set of training objects and iterative re-learning using at least one training object of the first predictive model, starting from at least one initial evaluation point, to obtain a plurality of retrained first predictive models, and the formation of an appropriate prediction error index after retraining. Based on a plurality of prediction error indicators after retraining and a plurality of corresponding first prediction error indicators one of the first sets of training samples and at least one trained object is selected.
EFFECT: technical result is higher efficiency of a machine training algorithm while saving computing resources.
16 cl, 6 dwg
