MODIFICATION OF NEURAL NETWORK MODEL RKELM
WITH ADDITIONAL TRAINING
Y.A. Asanov, S.Y. Beletckaya
The aim of this work is developing of an artificial neural network model (ANN) capable of working in dynamically changing conditions. Despite a large number of research and development in this sphere, there are still no models that satisfy the limited resources of mobile systems (primarily – performance). This article proposes a developed modification of the Huang Extreme Learning Model, which differs from the original approach in the training process – training on common conditions, without increasing the weight matrix and the training sample, followed by further training for specific conditions. As a test sample of data, a dataset from the open source machine-learning repository UCI was used. Vast experiments were performed, the purpose of which was to identify the most suitable model, the choice was made from RKELM, SVM and ELM. The selection criteria for the model were performance and classification accuracy. The model with extreme training of Huang turned out to be the most suitable, it was used as the basis of the developed modification. The results of comparing the original and modified models are presented. The proposed approach surpassed the competition in speed and performance, while only slightly inferior in accuracy of data classification in the initial conditions, but turned out to be much more accurate in the new conditions in which the model was not trained.
Keywords: artificial neural network, RKELM modification, model with an additional training.