HYBRID TRAINING METHOD OF ARTIFICIAL NEURAL NETWORK ON THE BASIS OF THE MODIFIED ANT ALGORITHM

Abstract

Back propagation of error algorithm is one of the training methods of artificial neural network. Its creation gave an additional incitement to the development of artificial neural network theory. An increase and appearance of the new tasks had led to the appearance of hybrid algorithms based on classical and stochastic ones. They emphasize the advantages and decrease the drawbacks of each other. The article suggests the hybrid training algorithm of neural network based on the ant colony and back propagation of error algorithms. The ant colony algorithm is used to choose the priority ways of moving along the neural network and to change the network balance with the help of pheromones. In comparison with the basic algorithm the ant colony one is changed. Instead of distance records between each pair of connected nodes and total network error records, they use the record of error value on the each exit node at the end of each ant travel. In the modified algorithm, each network weight changes the coefficient with the glance of the value of pheromone matrix elements, the error on the each node of exit layer and suggested function for pheromone accounting. The algorithm performance was checked on the famous databases [http://archive.ics.uci.edu/ml/]. The results of the experiments show that in comparison with another algorithm the modified one accelerates the training of neural network. The algorithm implementation was approved on the Microsoft SQL Server 2005 in Analysis Services. It will help to apply suggested algorithm to solve another problems of classification.