Categorization of Markov Chains M/ M /c/k by Feed-forward Neural Networks

Abstract

The report presents the study of the possibility of applying Artificial Intelligence to analyze traffic data in order to define their category against the observed circuits for different number of server stations in M/M/c/k queuing model, respectively c= 0, c= 5 and c=20. Feed-forward neural models with Scaled Conjugate Gradient training are used at given “sigmoid” activation types in hidden and “softmax” in output network layers. The selection is performed in range of 3 to 25 intermediate neurons in assessing the information content of five input indicators -”client ID”, “service request time”, “service starting time”, “server id” and “release time of requests”. A model is synthesized for four input variables without the “service request time” parameter and accuracy of 87.7% was achieved. For the established model, the experiment was extended to 39 neural units, to reaching a potential accuracy level of over 90.0%. The highest accuracy of 90.7% is achieved at 33 neurons.