In this study, a hybrid learning algorithm for training the Dynamic Synapse Neural Network (DSNN) to high accurate prediction of congestion in TCP computer networks is introduced. The idea behind this technique is to inform the TCP transmitters of congestion before it happens and to make transmitters decrease their data sending rate to a level which does not overflow the routers buffer. Traffic rate data are available in the format of time series and these real data are used to train and predict the future traffic rate condition. Hybrid learning algorithm aims to solve main problems of the Gradient Descent (GD) based method for the optimization of the DSNN, which are instability, local minima and the problem of generalization of trained network to the test data. In this method, Adaptable Weighted Particle Swarm Optimization (AWPSO) as a global optimizer is used to optimize the parameters of synaptic plasticity and the GD algorithm is used to optimize the weighted parameters of DSNN. As AWPSO is a derivative free optimization technique, a simpler method for the train of DSNN is achieved. Also the results are compared to GD algorithm.
M. Shakiba, M. Teshnehlab, S. Zokaie and M. Zakermoshfegh, 2008. Short-Term Prediction of Traffic Rate Interval Router Using Hybrid Training of Dynamic Synapse Neural Network Structure. Journal of Applied Sciences, 8: 1534-1540.