Authors: V.S. Doroganov, A.G. Pimonov

Title of the article: Modified network of Word and hybrid training method for the prediction of indicators of quality of metallurgical coke

Year: 2015, Issue: 3, Pages: 141-148

Branch of knowledge: Informatics, computing and control

Index UDK: 004.032.24:004.032.26

DOI: -

Abstract: The article presents a modification of the model of artificial neural network of Word. The essence of modification is variability parameters of the transfer functions. The transfer functions are the Fermi function, identity, hyperbolic functions, the dome of Gauss and sine wave. Back propagation algorithm is not suitable for training, as they do not change the parameters of the transfer functions. For learning to use a hybrid method. It based on genetic algorithm and stochastic methods. To select characteristics of neural network (the number of layers and neurons in each layer) experiment has conducted: each combination studied during 2000 iterations 40 times. The result of study presented in tabular form and graphically. The article describes the technology of parallel computations for the neural network. We present the results of the investigation of the influence-network features on the training in tabular and graphical form. The results of experimental evaluation velocity calculations on two multi-core processors.

Key words: neural networks genetic algorithms metallurgical coke forecasting multi-threaded computing multicore processors

CITE DOWNLOAD