Dynamic buffer allocation based on traffic prediction with FIR neural network

Analysis of network traffic has shown that traffic data exhibits high degree of self-similarity at different time scale. The feature of self-similarity indicates that there exists a predictive structure of the network traffic. In this paper, an improved finite-impulse-response neural network which utilizes momentum-modified temporal back-propagation algorithm is proposed for traffic prediction. Then the prediction results are used for dynamic buffer allocation to reduce packet loss in queuing systems. Simulation results show that the FIR neural network has good ability for traffic prediction, and the dynamic buffer allocation scheme based on traffic prediction with FIR neural network can effectively reduce packet loss rate.