Wavenet Based Link State Predictor Computer Science Essay

Published Date: 02 Nov 2017

Abstract In this paper, a wavelet neural network (WNN) (or wavenet) predictor is used to predict link status (congestion and loading state) of each link in the computer network. The proposed WBLSP predictor generates two indicators for congestion and loading for each link based on the square of utilization values of each link measured in the previous time intervals. WNNs possess the learning and generalization capabilities of the traditional neural networks together with the local characteristics of wavelet functions that enhance network ability to deal with sudden changes and burst network load in efficient manner . The square of utilization distribute the values in more efficient manner. The proposed predictor can be used in the context of active congestion control and link load balancing techniques to provide the status of all links in the computer network.

Keywords— Computer Networks; Wavelet-Neural Networks; Prediction; WBLSP; Congestion control, Link utilization, Link Loads .

INTRODUCTION

Network traffic prediction takes significant interest in many domains such as adaptive applications, congestion control, and network management [1]. Traffic prediction enables proactive network management which improves the global performance of the network through congestion control and prevention [2].

Data traffic in computer network can be considered as non linear and non stationary time signal that exhibits a high degree of long range dependence characteristics in additional to short range characteristics so that classical prediction systems cannot model network traffic well. Heuristic systems such as artificial neural networks can approximate non-linear systems more accurately because of their characteristics and approximation capabilities [3].

The prediction of link state (congested, uncongested, loaded or unloaded ) help the designers to develop routing protocols that are capable to detect congestion in a network before it begins effecting on a network performance. If congestion state is predicted in early stages, the effect of congestion may be avoided using an efficient congestion control technique that uses the predicted information to control the amount of traffic and manages network resources to prevent congestion problem. Such congestion control techniques are called active congestion control [2].

In the present Internet, congestion control mechanisms rely on queue management algorithms (dropping packets randomly or based on their priority) or TCP (Transmission Control Protocol) congestion avoidance (reducing the sending rate). From the end-user perspective, these solutions are not optimal because they mean lost packets or a reduced bitrate, both affecting the quality of transmission. Therefore some of new researches propose techniques for solving congestion by changing the paths of traffic to avoid congested links or adding a new paths .

Employing a prediction-based approach helps to match network resources to the traffic demand and a prediction-based approach will be faster, in terms of congestion detection and elimination, than reactive methods which detect congestion only after it significantly influenced the operation of the network, as demonstrated in [4].

This work develop a simple topology independent predictor for congestion and loading states in next time interval based on wavenet neural network and the square utilization values of the previous time intervals.

The rest of this paper is organized as follows . Section 2 briefly presents previous work regarding prediction using artificial intelligence techniques. In section 3 , wavenet fundamentals are given. Section 4 presents the structure of the proposed predictor (WBLSP ). In Section 5 testing the proposed predictor with one of the recent predictors with similar tasks and comparing their performance.

Related works

Using artificial intelligence techniques for traffic prediction and congestion control are not strange. There many recent research that use neural network for predicting traffic patterns, the author of [6] propose an a simple Feedforward neural network to predict severe congestion in a network and he used neural networks to predict the source or sources responsible for the congestion, and he designed a simple control method for limiting the rate of the offending sources so that congestion can be avoided. in [7] a network traffic prediction hybrid model based on αTrous wavelet analysis and Hopfield neural network is proposed , which can be used to predict the network traffic flow. in [8] two different artificial neural network (ANN) architectures, multilayer perceptron (MLP) and fuzzy neural network (FNN) are used to predict one-step ahead value of the MPEG and JPEG video, Ethernet and Internet traffic data. the output of the individual ANN predictors are combined using different combination schemes.

The use of Wavelet neural network are proposed in [3] for congestion prediction and in [9] in congestion and load state prediction using the previous link utilization values.

Because of link utilization reflect the link congestion in better manner than link loads [3] , it is used in this paper but with another new modification that are using the square function before using it in a wavenet network. The using the square function would improve the distribution of inputs and increase the sensitivity in congestion state prediction so the missing of congestion in the next time interval are very little and the of prediction network become smaller .

Wavenet theoretical background

Wavenet can be considered a particular case of the feed forward basis function neural network model. In ordinary network, several types of basis functions, such as radial basis functions, splines and polynomial functions of synapse neurons are used instead of sigmodial function. The connection weights are taken to represent the corresponding coefficients. The output layer performs the sum of the output of all synapse neurons. Since wavelets have been shown their excellent performance in non stationary signal analysis and nonlinear function modeling, the neural network using wavelet basis function, wavenet, provides higher availability of rate of convergence for the approximation than an ordinary feed forward neural network [10].

The structure of WNN are shown in Fig 1 ,which comprised of an input layer, a wavelet layer, and an output layer. The input data in the input layer of the network is u = [u1 u2 … uN], where N is the number of dimensions. The input data are directly transmitted into the wavelet nodes in the wavelet layer. For the discrete wavelet transform, the mother wavelet φ(x) describes the dilation a and the translation b as follow:

… (1)

Figure 1 WNN structure

In this paper, the Mexican-hat function (shown in Fig 2) are used as activation function , The Mexican-hat function is the second derivate of the Gaussian function and expressed as follow:

Figure 2 The Mexican-hat function with different dilation and translation values

…(2)

In the case of multi-dimensional input the wavelons consist of multidimensional wavelet activation functions. They will produce a non-zero output when the input vector lies within a small area of the multidimensional input space [11]. The output is defined as:

...(3)

This wavelon is in effect equivalent to a multidimensional wavelet[11].

....(4)

Where are j = 1, 2, …, M is used as a nonlinear transformation function of hidden nodes, and wj, j =1, 2, …, m is used as the adjustable weighting parameters to provide the function approximation[12].

The most used training algorithm are the gradient that is commonly used to minimize the error and obtain the suitable network parameters. The process can be summarized as that the variations of error energy function with respect to each network parameters i.e., the gradient factors are calculated. Then these factors are used as incremental factors to update the current network parameters in the direction leads to minimize the error [3].

The proposed WBLSP structure

There are an independent WNN for each unidirectional link in the computer network all operate in parallel to compute the output prediction victors. If there are k unidirectional link (bidirectional link is considered as a 2 unidirectional link ) in the computer network then there are k independent WNN network. Each WNN ,as shown in Fig. 4, consist of 3 nodes in the input layer, 5 nodes (wavelon) in the hidden layer and 2 node in the output layer. Each input is the square of the average link utilization for one minute in the near past. Then the total inputs of the WNN reflect the utilization behavior during the last three minutes.

Figure 3The of WNN for WBLSP

The Mexican hat wavelet is used as activation function in the hidden layer nodes since it is very appropriate for function approximation and prediction because it is continuous, differentiable, provide a softer output, and improve the interpolation capabilities. It is also reduces the number of iterations that results in faster convergence and for escaping from local minima [3].

Each WNN generates two bit output represents both the congestion and load states of the corresponding link. The congestion bit is set when one of the average of utilization values of the last three minutes is greater than 70% or the average of utilization values of next minute may be greater than 70% else the output is cleared. And the loading bit is set when one of the average of utilization values of the last three minutes is greater than 40% or the average of utilization values of next minute may be greater than 40%. The total predicted state is described by two vectors the congestion vector and load vector. For a computer network consists of K links, there are K bits in the each vector.

Simulation and results

For training and testing purpose OPNET 14.5 simulator is used to simulate the same two network topologies used in [3]. Fig 4 and 5 showing the two topologies and how LANs and servers are connected to each one.

Figure 4 Network topology 1

Figure 5 Network topology 2

For the accuracy purpose three network application (HTTP,FTP and EMAIL) are used in the simulation in three different scenarios:

Scenario 1 : some LANs apply heavy loads in some times and light loads in other times while the other LANs apply only light loads.

Scenario 2 : all LANs apply heavy and light loads so that the congestion appears on network links in non simultaneous fashion.

Scenario 3 : all LANs apply heavy loading on a network so that the congestion appears simultaneous fashion.

All scenarios run for 6 hour and use OSPF routing protocol in both topologies (there are sample each minute the first 10 minutes are canceled because they may contain transient data so that there are 350 sample per link).

Six links are selected from Scenario 2 of topology 1 to produce 2100 training sample that are used to train the one WNN in WBLSP predictor and for comparison purpose the samples are also used to train the WBCP-LI predictor proposed in [6] that also use WNN for congestion and loading state prediction. the six links are selected in a manner so there are two congested links and 2 loaded uncongested links and two unloaded links.

The data gathered from previous scenarios are used to test the proposed WBLSP and the WBCP-LI. The prediction accuracy is computed for each scenario by dividing the total number of correct prediction to the total number of samples that equal ,for topology 1, to 44*(5*60+50)=15400 sample and for topology 2 are equal to 36*(5*60+50)=12600 sample.

The hit error rate is the ratio of the number of incorrect prediction of congestion or loading while the link is not congested to the total number of samples.

The miss error rate is the ratio of the number of incorrect missing of congestion or loading while the link is congested or loaded in to the total number of samples.

The two predictors are tested and their performance are compared in the next Fig 6 to 17. The difference between WBLSP and WBCP-LI result occured only on congested links for congestion prediction and in loaded links in loading prediction because of the prediction accuracy is 100% in both predictors for uncongested links for congestion prediction and for unloaded links in loading prediction .

Figure 6 Prediction accuracy of congestion state prediction for topology 1

Figure 7 Prediction accuracy of loading state prediction for topology 1

Figure 8 Prediction accuracy of congestion state prediction for topology 2

Figure 9 Prediction accuracy of loading state prediction for topology 2

Figure 10 Miss error for congestion prediction for topology1

Figure 11 Miss error for loading prediction for topology1

Figure 12 Miss error for congestion prediction for topology2

Figure 13 Miss error for loading prediction for topology2

Figure 14 Hit error for congestion prediction for topology1

Figure 15 Hit error for loading prediction for topology1

Figure 16 Hit error for congestion prediction for topology2

Figure 17 Hit error for loading prediction for topology2

From previous figures it can be noticed that both predictors gives high prediction accuracy and the WBLSP is more sensitive in congestion prediction than WBCP-LI because it has low miss error rates than WBCP-LI that have a balancing between the two types of error in congestion prediction that have less hit error rate than WBLSP and higher miss error rate congestion prediction. Although there are a simple bias in congestion state prediction accuracy for WBCP-LI but the cost of miss error is more costly than hit error in congestion because of missing congestion may effect on the performance of computers network .

In loading state prediction the prediction of the WBSLP is more accurate than WBCP-LI and have less sensitivity so that it has less hit errors than WBCP-LI that is more sensitive and this come from the fact that square function (that are used for the utilization values in WBLSP ) enhance the large numbers and reduce the small numbers so that the predictor has more sensitivity for congestion than WBCP-LI and less sensitivity for loading.

Because of WBLSP have less missing congestion errors than WBCP-LI and present more accuracy in loading state prediction and when talking about the cost of WBLSP that have only 3 node in input layers and 5 weavlon in hidden layers therefore it is recommended to use WBLSP in congestion and loading prediction than using WBCP-LI that have 5 nodes in input layer and 7 node in hidden layer.

Conclusion

The wavenet based link status predictor are proposed and tested and compared with one of previous wavenet based predictor and the WBLSP give high prediction accuracy and more efficient prediction results by adding the benefits of square function on the input of the predictor.

From the results, it can be seen that the WBLSP is topology independent and load independent (congestion-load) predictor based on simple Wavenet neural network to produce detailed congestion- load state indicator with a few bytes vectors.