Philip Mars

The limitations of artificial neural networks for traffic prediction

B-ISDN is expected to support a variety of services, each with their own traffic characteristics and quality of service requirements. Such diversity however, has created new congestion control problems, some of which could be alleviated by a traffic prediction scheme. This paper investigates the applicability of artificial neural networks for traffic prediction in broadband networks. Previous work has indicated that such prediction is indeed possible, as the neural networks are able to learn a complex mapping between past and future arrivals. Such work however has been based on the use of artificially generated traffic, and by definition the past and future arrivals are related. We consider real traffic and show that prediction is possible for certain traffic types but not for others. We also demonstrate that simple linear regression prediction techniques perform equally as well as neural networks.

Application of recurrent neural networks to communication channel equalization

The paper examines the mechanism by which recurrent neural networks (RNNs) achieve equalization whilst operating on simple digital communication channels. The mode of operation is seen to be essentially similar to the conventional decision feedback equalizer (DFE) and the RNN node nonlinearity is identified as a limiting factor. Two versions of an alternative RNN structure are formulated for channels with longer impulse responses based on soft-decision feedback. Simulations demonstrate the improved BER performance compared with the DFE.

Satisfying QoS with a learning based scheduling algorithm

Considers the problem of scheduling packets in a multiplexer, the aim being to provide sufficient service to each traffic stream such that their objectives are just satisfied, thus maximising the resources available for other streams. We propose a novel scheduling scheme based on stochastic learning automata which is capable of satisfying a variety of delay requirements in a dynamic traffic environment, and we show via simulation that the scheme outperforms several existing scheduling algorithms.

Self-similar traffic generation and parameter estimation using wavelet transform

Previous studies of real network traffic have shown that packet traffic exhibits fractal properties such as self-similarity which are fundamentally different from features of traditional Poisson-based traffic models. Fractional Brownian motion (FBM) is a widely-used self-similar process which can be characterized by the Hurst parameter. To gain a better understanding of queueing and network-related performance issues with regard to FBM models, it is essential to be able to accurately and quickly generate traffic from FBM processes. In the mean time, it is of crucial importance to accurately estimate the Hurst parameter. In this paper we attack the above two problems by using the wavelet transform. Our analysis indicates that the wavelet approach is as fast or faster than existing methods and appears to generate a closer approximation to true self-similar sample paths than the other known fast method (random midpoint displacement). Moreover, the Hurst parameter can also be accurately estimated from the power-law behaviour of the wavelet coefficients variance.

ATM traffic prediction using FIR neural networks

ATM networks support a wide range of multimedia traffic. Various BISDN VBR sources generate traffic at significantly different rates. The traffic can often have time-varying characteristics which are not well understood currently. However, traffic management techniques require traffic parameters that can capture the various traffic characteristics and adapt to the changing network environment. In this paper, we present a novel neural network approach to characterize and predict the complex arrival process. The FIR multilayer perceptron model and its training algorithm are discussed in this paper. It is shown that the FIR neural network can adaptively predict the traffic by learning the relationship between the past and the future traffic variations. Based on the experimental results, we conclude that the FIR neural network is an attractive tool for traffic prediction and hence has an excellent potential for use in some congestion control schemes.

Application of artificial neural networks to effective bandwidth estimation in ATM networks

A prime instrument for controlling congestion in ATM networks is admission control, which limits calls and guarantees a grade of service (GOS) determined by delay and loss probability in the multiplexer. It is essential for an admission control scheme to characterize, for a given GOS, the effective bandwidth requirement of the aggregate bandwidth usage of multiplexed connections. In this paper, an accurate and computationally efficient approach is proposed to estimate the effective bandwidth of multiplexed connections. In this method, a feedforward neural network is employed to model the complex relationship between the effective bandwidth and the traffic situations and a GOS measure. It is trained and tested via a large number of patterns generated by the accurate fluid flow model. Due to the neural networks adaptive learning, high computation rate and generalization features, this method can increase the link utilization and is suitable for real-time network traffic control applications.

An NN-based dynamic time-slice scheme for bandwidth allocation in ATM networks

In this paper, we propose a neural network (NN) approach for adaptive bandwidth allocation in ATM networks. This method is essentially based on the dynamic time-slice (DTS) scheme proposed by K. Sriram (1993) which guarantees a required bandwidth for each traffic class and/or virtual circuit (VC). Instead of using analytical static traffic tables to allocate bandwidth, we use NNs to adaptively estimate the effective bandwidths of different call types to reflect the time-varying nature of traffic conditions. Simulation results show that the neural estimation is more accurate and hence leads to higher resource utilization. The NN approach also provides faster response in reallocation of bandwidth to meet the stringent delay requirements.

Multiplexing Gains in ATM Networks

The focus of this paper is to investigate the effects of various traffic and switch characteristics on multiplexing gains and their implications for different bandwidth allocation and admission control algorithms. We show that the total multiplexing gain due to the independent combination of identical sources can be resolved into two factors, one expressing the advantage gained (by means of buffering) from the statistical rate variations within a source and the other expressing the efficiency of statistical multiplexing of i.i.d. streams (gain across sources). Both simulation and theoretical analysis illustrate that although bursty sources require more bandwidth, multiplexing gains are increasing with burstiness. The effective bandwidth approach is found to work well in the region with high buffer size to burst length ratio, high source utilization and small number of sources, whereas the Gaussian approximation performs well in the region with small buffer size to burst length ratio, high source utilization and large number of sources. Finally, we also give some quantitative information related to self-similar traffic, i.e., FBM. It is shown that even for LRD traffic with high-H values, it is possible to obtain higher multiplexing gains when a large number of independent sources with the same Hurst parameter are multiplexed for combined transmission.

A hybrid approach for generating fractional Brownian motion

Self-similar traffic models have been found to be more appropriate for the representation of bursty telecommunication traffic. The fractional Brownian motion (FBM) processes is one of the two most commonly used traffic models to interpret self-similarity. The discrete fractional Gaussian noise (dFGN) and random midpoint displacement (RMD) algorithms have been used to synthesize self-similar sample paths. However, the dFGN is very inefficient and the inaccuracy of the RMD is usually unacceptable. In this paper we use the dFGN with interpolated RMD subtraces to get a faster and more accurate algorithm in which the dFGN generates the end points and the RMD produces a subtrace with a level of depth that determines the number of samples for each subtrace. The hybrid algorithm improves the computational time significantly, and still keeps the accuracy of the expected Hurst value.

Global search of adaptive IIR filter error surfaces using stochastic learning automata

The multimodal nature of adaptive IIR filter error surfaces limits the use of gradient search adaptive algorithms. Recently an intelligent learning approach was suggested by the authors to tackle the problem. This paper describes further research results and shows that stochastic learning automata are capable of locating the global minimum under different conditions. Computer simulation results for a system identification application are presented to illustrate that it is possible to achieve global convergence irrespective of insufficient filter order or input colouration or both. Stability during adaptation is also maintained.© (1990) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.