science-review.ru
Scientific journal
Scientific Review. Technical science
ISSN 2500-0799
ПИ №ФС77-57440
ISSN 2500-0799
ПИ №ФС77-57440
NEURO-EXPERT ALGORITHM FOR EXCLUSION STATISTICAL FLUCTUATIONS FOR DESIGN OF NEURAL NETWORKS
208 KB
Summed up the experience of the Perm school of artificial intelligence are necessary conditions for the development of adequate neural network model. One of these conditions is the lack of statistical information to unauthorized outliers - observations that do not satisfy the laws that govern the behavior of the vast majority of the examples studied the subject area. The causes of unauthorized outliers can be: not clean enough of the experiment, the measurement errors, failures of devices and equipment, noise and the impact of factors that are not included in the statement of the problem. The idea of the proposed algorithm for detecting outliers based on the fact that if the outliers in the training set are relatively few, and if the neural network has a relatively small number of synaptic weights, then after the application of the procedure of training a neural network on the examples which are outliers, generally, shows higher error learning than the examples that are not outliers. In contrast to other known algorithms for the detection and exclusion of the outliers, the proposed algorithm involves interactive participation of the expert, making the opinion on the legality of removing the detected outlier, and the removal of outliers in each iteration is strictly one.