DonNTU   Masters’ portal

Abstract

Contents

Introduction

One of the main telecommunication market trends is increasing users demand for multimedia services. Therefore the Triple Play concept becomes popular, according to which on the basis of three core services (transmission of audio, video, data) various multimedia services are combined, provided to the subscriber through a single multi‐service converged network infrastructure. Active adoption of multimedia services in the network causes a significant load on the communication channels and equipment, which leads to a deterioration in the quality of service performance.

1 Theme urgency

Master's thesis is devoted to an urgent problem of the load balancing system development, allowing to raise the quality of service and efficiency of using the equipment and resources of a telecommunication network.

Traffic in modern telecommunication networks is a complex stochastic process, which shows signs of self‐similarity [1, 2]. Self‐similar processes are characterized by long‐term dependencies, so based on these dependencies it is possible to build predictive models that can be used for controlling traffic in a telecommunications network. Different mathematical methods are used to build prediction models, one of which is use of feedforward neural network.

2 Goal and tasks of the research

The goal is to improve the quality of service parameters and efficiency of using telecommunication network equipment by implementation of load balancing system.

At this stage of the master's thesis the efficiency of using neural network for building a load balancing system in a converged telecommunication network investigated.

This requires accomplishment of the following tasks:

  1. Consider mathematics which is used in feedforward neural networks.
  2. Evaluate the effectiveness of the neural network traffic forecasting.
  3. Compare the results of traffic forecasting based on the feedforward neural network and the standard autoregressive model.
  4. Compare the efficiency of forecasting for different types of traffic.

In the course of the work it is planned to obtain scientific results in the following areas:

  1. Development of forecasting algorithm for efficient self‐similar traffic prediction in the telecommunication networks.
  2. Development of load balancing system using the methods of dynamic channel capacity allocation based on the predicted network load.

3 Development of telecommunications network traffic forecasting algorithm

3.1 Mathematics of feedforward neural networks

Consider feedforward network with one hidden layer (figure 1). According to Cybenko theorem, feedforward artificial neural network with one hidden layer which activation function is sigmoidal can approximate any continuous function of several variables arbitrarily well [3].

Neural network structure

Figure 1 – Neural network structure

Research was conducted using a neural network with rational sigmoid activation function in the hidden layer (figure 2):

Rational sigmoid
Rational sigmoid

Figure 2 – Rational sigmoid

The reason for introducing non‐linearity is mathematically proved possibility to obtain an arbitrarily accurate approximation of any continuous function of several variables, using the operations of addition and multiplication by a number, a superposition of functions, linear functions, as well as a continuous nonlinear function of one variable [4]. Rational sigmoid is the most efficient sigmoidal activation function, because for it calculation only three mathematical operations are needed.

The neural network operates in compliance with the feedforward algorithm, where xvec — time series, θij(l) — synaptic connections weights, x_vec — neural network output [5]:

Feedforward
Feedforward

Neural network training is based on the backpropagation method (figure 3). The method consists in dynamic change of synaptic connections weights, depending on the forecast error [6, 7].

Backpropagation

Figure 3 – Backpropagation
(animation: 6 frames, 7 cycles of repeating, 148 kilobytes)

3.2 Traffic forecasts using neural network

Consider two types of traffic — streaming video traffic (Hurst parameter H = 0,852) [8] and data traffic (Hurst parameter H = 0,643). There have been chosen parameters of the neural network and calculated the forecast based on learning curves — for video traffic relative error equals 12,2%, for data traffic — 37,6%. Type I and type II errors have the same weight.

3.3 Prediction of traffic based on autoregressive model and comparison of results

To better assess the quality of prediction obtained using a neural network, compare the result with the prediction made on the basis of Box‐Jenkins models [9, 10]. When forecasting video traffic and data traffic based on autoregressive model, prediction error equals 35,3% and 39,7% respectively. A significant amount of errors are type II errors. For streaming video traffic comparative measure of prediction accuracy equals 0,915, for data traffic — 1,024. This suggests that using neural networks for self‐similar traffic forecasting gives a better prognosis than the autoregressive model.

Conclusion

Load forecasting in the telecommunication network can give a way to manage the network bandwidth for each class of traffic, improve QoS in the network. Using neural network for traffic forecasting on a telecommunication network nodes allows getting more accurate forecast values than the standard autoregressive model. Neural network prediction provides smaller type II error rate than the AR model, which is important when implementing forecasting systems in telecommunication networks.

Applying considered forecasting method in the load balancing system will improve quality of service and efficiency of using network equipment.

This master's work is not completed yet. Final completion: December 2012. The full text of the work and materials on the topic can be obtained from the author or his scientific adviser after this date.

References

  1. On the Self‐Similar Nature of Ethernet Traffic / Leland W.E., Taqqu M.S., Willinger W., Wilson D.V. // IEEE/ACM Transactions of Networking. — IEEE Press Piscataway, 1994 — Vol. 2(1). — 15 p.
  2. Петров В.В. Структура телетрафика и алгоритм обеспечения качества обслуживания при влиянии эффекта самоподобия : Диссертация на соискание ученой степени кандидата технических наук / Петров В.В. — М., 2004. — 199 с.
  3. Cybenko G.V. Approximation by Superpositions of a Sigmoidal function / Cybenko G.V. // Mathematics of Control, Signals and Systems, 1989. — Vol. 2. — № 4. — P. 303–314.
  4. Kreinovich V.Y. Arbitrary nonlinearity is sufficient to represent all functions by neural networks: A theorem / Kreinovich V.Y. — University of Texas at El Paso, 1990.
  5. Искусственная нейронная сеть : Википедия — свободная энциклопедия [Electronic resource] / Mode of access to the resource: http://ru.wikipedia.org/wiki/Искусственная_нейронная_сеть
  6. Метод обратного распространения ошибки : Википедия — свободная энциклопедия [Electronic resource] / Mode of access to the resource: http://ru.wikipedia.org/wiki/Метод_обратного_распространения_ошибки
  7. Нейронные сети — математический апарат : BaseGroup Labs [Electronic resource] / Mode of access to the resource: http://www.basegroup.ru/library/analysis/neural/math/
  8. MPEG‐4 and H.263 Video Traces for Network Performance Evaluation / Video Trace Library [Electronic resource] / Mode of access to the resource: http://trace.eas.asu.edu/TRACE/trace.html
  9. Box E.P.G. Time series analysis : forecasting and control / Box E.P.G., Jenkins G.M., Reinsel G.C. — Prentice‐Hall, 1994. — 598 p.
  10. Basic Definitions and Theorems about ARIMA models / Xycoon [Electronic resource] / Mode of access to the resource: http://www.xycoon.com/basics.htm