Connection admission control of ATM network using integrated MLP and fuzzy controllers

Connection admission control of ATM network usingintegrated MLP and fuzzy controllers

Nelson O.L. Ng *, C.K. Tham

Department of Electrical Engineering, National University of Singapore, Singapore 119260, Singapore

Abstract

This paper presents a new approach to the problem of call admission control (CAC) of variable bit rate (VBR) tra�c

in an asynchronous transfer mode (ATM) network. Our approach employs an integrated neural network and fuzzy

controller to implement the CAC controller. This scheme capitalizes on the learning ability of a neural network and the

robustness of a fuzzy controller. Experiments show that this scheme is able to achieve high throughput and low cell loss

while achieving fairness among di�erent classes of VBR tra�c. For comparison, we have also implemented four other

CAC schemes: (1) peak bandwidth method, (2) equivalent bandwidth method, (3) average bandwidth method and

(4) neural network quality of service (QoS) predictor. Results of these experiments are presented in this paper. Ó 2000

Elsevier Science B.V. All rights reserved.

Keywords: ATM; Call admission control; Neural network; Fuzzy control

1. Introduction

Asynchronous transfer mode (ATM) is ahigh-speed packet switching technology for thebroadband integrated services digital network(B-ISDN), in which various kinds of communica-tion services, such as voice, video and data aretransferred over high-speed links [1]. This tech-nology has gained acceptance as the backbonehigh-speed network of the future, as well as thelower bandwidth link to homes and desktops forinteractive multimedia services. ATM supportsdi�erent service classes, such as constant bit rate

(CBR), variable bit rate (VBR), unspeci®ed bitrate (UBR) and available bit rate (ABR).

Due to the diverse mix of tra�c types and ser-vice requirements, resource allocation in B-ISDNmust be related both to the tra�c parameters andthe quality of service (QoS) negotiated between theuser and the network at the establishment of eachcall. Most of the published connection admission(CAC) criteria allocate resources based on theavailability of bandwidth necessary to guaranteethe negotiated QoS. These approaches use eitherthe peak or average cell rate as CAC criteria. Moreelaborate methods introduced the concept ofequivalent capacity, which gives a value betweenthe peak and average, depending on the type oftra�c and desirable maximum cell loss rate.Equivalent bandwidth computation focuses on thebandwidth requirement of the bit rate generatedby sources, and not on the di�erent interactionsthat take place within the network. In addition,

Computer Networks 31 (2000) 61±79

www.elsevier.com/locate/comnet

* Corresponding author. Tel.: +65-373-2826; fax: +65-273-

5452.

E-mail addresses: [email protected] (N.O.L. Ng),

[email protected] (C.K. Tham).

1389-1286/00/$ - see front matter Ó 2000 Elsevier Science B.V. All rights reserved.

PII: S 1 3 8 9 - 1 2 8 6 ( 9 9 ) 0 0 1 2 4 - 3

these methods are all based on some probabilisticrepresentation of the sources of tra�c, and on amethod for determining from the model whetherthe QoS is satis®ed or not for that set of sources.For example, in [2], Gu�erin et al. determines thee�ective bandwidth requirement of a single con-nection and the aggregate bandwidth usage ofmultiplexed connections for on±o� ¯uid-¯owsources. In [3], Elwalid et al. determines the e�ec-tive bandwidth for multiple Markov modulated¯uid-¯ow sources sharing a statistical multiplexerwith a large bu�er. The problem with equivalentbandwidth methods is that the computed band-width is often highly conservative when bu�ers aresmall or of moderate size. The causes of this con-servatism are that it is derived under the asymp-totic region where the product of bu�er size andcell loss probability tends to zero, and that thesemethods use the bu�er over¯ow probability as theQoS requirement. This bu�er over¯ow probabilityis normally larger than the cell loss probabilitywhich results in the reduction of the useful ad-mission region.

Because of the learning abilities of neural net-works, Hiramatsu [4] has proposed the use ofneural networks to learn the relation between theo�ered tra�c characteristics and the negotiatedQoS to perform CAC. Nordstr�om et al. [5] hasalso proposed a hybrid approach based on neuralnetworks and analytical aproximations to performlink admission control. In this approach, a newcall is accepted only if the QoS of new and alreadyestablished calls on all links along a route wouldnot be compromised. This approach uses mathe-matical analysis to determine the upper bound onthe cell loss probability to perform CAC at lowlink loads, while the neural network is used toperform control at higher loads.

This paper proposes a scheme that is based onthe methods proposed by Moh et al. [6] and Neveset al. [7]. Moh et al. used neural network for tra�cprediction to perform dynamic bandwidth alloca-tion while Neves et al. used a quality of operationfunction to perform CAC. In addition, Neves et al.also explored the issue of fairness when perform-ing CAC. Our scheme integrates both of theseframeworks to perform CAC. Like Moh et al., ourscheme uses multilayer perceptron (MLP) neural

networks to perform bandwidth prediction.However, unlike Neves et al. who used neuralnetworks to model di�erent components of theATM network, the fairness and CAC controls inour scheme are implemented using fuzzy control-lers, which are well known for their robustness. Inour implementation, the MLP network is trainedusing an adaptive learning rate algorithm to im-prove the learning performance. Section 2 intro-duces a modularized control framework andSection 3 describes the proposed control schemebased on this framework. Section 4 describes thesimulations that have been performed using anetwork simulator. Following this, results of thesimulations will be presented and discussed inSection 5.

2. Framework of the proposed CAC scheme

The MLP neural network is the most commonlyused neural network architecture [8]. The networkconsists of a set of sensory nodes that constitutethe input layer, one or more hidden layers ofcomputation nodes and an output layer of com-putation nodes. The strengths of the connectionsare called weights and can be adjusted to producethe required input±output mapping. Duringtraining, input vectors are sequentially applied tothe network, while adjusting network weights ac-cording to a predetermined performance objective.The performance objective of weight adaptation isto reduce the error averaged over the training set.The most common error function is the mean-squared error (MSE), expressed as

e�t� � 1

2

Xj

e2j �t�; �1�

where ej�t� is the error signal at the output ofneuron j for iteration t.

During training, reduction of e�t� is achievedthrough the method of steepest descent given by

Dwji�t� � ÿgji�t�oe�t�

owji�t� ; �2�

where wji�t� is the synaptic weight connecting theoutput of neuron i to the input of neuron j at it-eration t. gji�t� is the learning step size for wji�t�.

62 N.O.L. Ng, C.K. Tham / Computer Networks 31 (2000) 61±79

In the standard backpropagation algorithm,gji�t� is usually set to a small constant, e.g. 0.1.However, a learning rate parameter appropriatefor the adjustment of one synaptic weight is notnecessarily appropriate for the adjustment of othersynaptic weights in the network. In addition, thetask of choosing g is often not easy. If g is toosmall, convergence can be very slow; if it is toolarge, instability can result. Therefore, for ourimplementation, each weight is assigned a learningrate, which is updated at every time step t. Formore details on this algorithm, please refer toAppendix A.

Unlike neural network, the control algorithm ina fuzzy controller is a knowledge-based algorithm,described by the methods of fuzzy logic [9]. Thiscontroller bases its decisions on inputs in the formof linguistic variables derived from membershipfunctions (MFs). MFs are formulas used to de-termine the fuzzy set to which a value belongs andthe degree of membership in that set. The variablesare then matched with the antecedents of linguisticIF-THEN rules (fuzzy logic rules), and the re-sponse of each rule is obtained through fuzzy im-plication. To perform compositional rule ofinference, the response of each rule is weightedaccording to the degree of membership of its in-puts, and the centroid of the responses is calcu-lated to generate the appropriate output.

Fig. 1 shows the modularized framework forour proposed CAC scheme. Each framework is

used to perform CAC for one class of tra�c. Thisscheme has the following main modules:1. MLP bandwidth predictor;2. MLP QoS predictor;3. Fuzzy throughput module;4. Fuzzy fairness module;5. Fuzzy CAC module.

The notations used in the framework are givenin Table 1.

At every time interval, the bandwidth predictoraccepts the cell arrival pattern ai�t�; ai�t ÿ 1�;. . . ; ai�t ÿ L� for the ith tra�c class and predictsbwi max�t � T �, the predicted bandwidth over T. Itthen passes bwi max�t � T � to the ATM network.During the same interval, the QoS predictor istrained using the cell arrival pattern and the ob-served QoS values to predict the QoS values forthe next averaging period, T. Both these modulesare implemented using MLP neural networks.

Whenever there is a new setup request from theith class, ATM network will pass the total band-width usage of all the classes including the newrequest to the throughput module. This totalbandwidth usage is expressed asXm

j�1

bwj max�t � T � � bwi max�t � T �=Ni�t�;

where Ni�t� is the number of currently connectedith class sources and the second term in the ex-pression refers to the estimated bandwidth usageof the new setup request.

Fig. 1. Proposed CAC framework.

Table 1

Notations used in the proposed CAC framework

L Lth previous time step

T Tth future time step

i ith tra�c class

m Number of tra�c classes

n Number of QoS parameters

ai�t� Cell arrival rate for ith tra�c class at time t

din�t � T � Predicted value of nth QoS parameter for

ith tra�c class at Tth future time step

bwm

max�t � T �Predicted total maximum bandwidth us-

age for

mth tra�c class at Tth future time step

Gthroughput Defuzzi®ed output from throughput

module

Gfairness Defuzzi®ed output from fairness module

GCAC Defuzzi®ed output from CAC module

N.O.L. Ng, C.K. Tham / Computer Networks 31 (2000) 61±79 63

As the throughput module is a fuzzy controller,the module will pass the input through its infer-ence system to determine Gthroughput. This modulehas only one universe of discourse, which we calltotal_throughput for convenience. This discourseis divided into divisions of 0.1 from 0 to 1.0 toindicate small bandwidth usage to almost fullbandwidth utilization. For ease of implementa-tion, all the inputs and outputs of the fuzzymodules implemented in this framework are nor-malized to be �0; 1�. For each of the 0.1 division, itis represented by a gaussian MF with the mean atthe corresponding division. The rule base for thethroughput module is set up such that when to-tal_throughput of all classes is very much less thanthe required threshold THthroughput, Gthroughput willbe near 1.0. Gthroughput will approach 0 as thethreshold is reached. In order to understand howthe throughput module works, we will describe itsoperations in details. For this module, there are 10rules in the rule base, as shown in Table 2. Inaddition, each of the 0.1 division for the anteced-ents and consequents are represented by member-ship functions shown in Table 3.

For a better understanding of the fuzzy mech-anism, we consider two values of total_through-put: 0:15 and 0:85. The desired throughputthreshold THthroughput for both cases is selected as0:7. Any value of total_throughput below 0:7should have Gthroughput greater than 0:5; otherwise,Gthroughput should be less than 0:5.

We will now consider the case of to-tal_throughput� 0.15. This value is ®rst substi-tuted into the antecedent MF for each rule in therule base to obtain the degree of ®ring sp, where pis the number of rules in the rule base. The results

are shown in the second column of Table 4. Next,each sp is multiplied with the mean of the conse-quent MF for each rule Mp to give spMp. As themeans of gaussian MFs are known, each sp can beeasily multiplied with the respective Mp. The re-sults are shown in the fourth column of Table 4.Then using centre of area method, the defuzzi®edGthroughput is obtained as

Gthroughput �P10

p�1 spMpP10p�1 sp

� 1:459

1:70

� 0:86:

Since 0:15 is much less than the desiredthreshold of 0:7, Gthroughput would de®nitely behigher than 0:5, as expected.

The above steps are repeated for the case oftotal throughput � 0:85 and we obtain the corre-

Table 2

Rule base used for throughput module

Rule 1: IF total_throughput � 0.1 THEN Gthroughput � 0:9



Rule 4: IF total_throughput � 0.4 THEN Gthroughput � 0:7Rule 5: IF total_throughput � 0.5 THEN Gthroughput � 0:7



Rule 8: IF total_throughput � 0.8 THEN Gthroughput � 0:1Rule 9: IF total_throughput � 0.9 THEN Gthroughput � 0:1


Table 3

Membership functions used for throughput module

Division Membership function (MF)

0:1 1ÿ exp�ÿ�0:1=�j0:1ÿ total throughputj��4�0:2 1ÿ exp�ÿ�0:1=�j0:2ÿ total throughputj��4�0:3 1ÿ exp�ÿ�0:1=�j0:3ÿ total throughputj��4�0:4 1ÿ exp�ÿ�0:1=�j0:4ÿ total throughputj��4�0:5 1ÿ exp�ÿ�0:1=�j0:5ÿ total throughputj��4�0:6 1ÿ exp�ÿ�0:1=�j0:6ÿ total throughputj��4�0:7 1ÿ exp�ÿ�0:1=�j0:7ÿ total throughputj��4�0:8 1ÿ exp�ÿ�0:1=�j0:8ÿ total throughputj��4�0:9 1ÿ exp�ÿ�0:1=�j0:9ÿ total throughputj��4�1:0 1ÿ exp�ÿ�0:1=�j1:0ÿ total throughputj��4�

Table 4

Sample computations for total_throughput � 0.15

Rule sp Mp spMp

�a� �b� �a� � �b�1 1:00 0:9 0:900

2 0:63 0:8 0:504

3 0:06 0:8 0:048

4 0:01 0:7 0:007

5 0:00 0:7 0:000

6 0:00 0:6 0:000

7 0:00 0:5 0:000

8 0:00 0:1 0:000

9 0:00 0:1 0:000

10 0:00 0:1 0:000

Total 1:70 1:459


sponding results given in Table 5. The defuzzi®edGthroughput is now 0:340=2:41 � 0:14, which is muchsmaller than 0:5. Since the total_throughput isnow greater than 0:7, Gthroughput should be near 0.Although the values of Gthroughput are given for twocases only, the above steps are applicable to allvalues of total_throughput between 0 and 1, thusgiving a smooth continuous range of values forGthroughput.

The next module, fairness module is also a fuzzycontroller. Whenever there is a new setup request,this module takes the di�erent bandwidth usage ofall the m classes from the ATM network and de-termines Gfairness. This bandwidth usage also in-cludes the setup request. The fairness module hasone universe of discourse for each class, which isdivided into divisions of 0:1 from 0 to 1:0 to in-dicate the bandwidth usage amount for each class.The purpose of the module is to make the band-width usage for each class to be almost equal. Therule set implemented in this module is based on thereasoning that if the current class is using lessbandwidth compared to the other tra�c classes,Gfairness should be high; if the current class is usinga lot of bandwidth, the new request should be re-jected. This is because when one class already has ahigh bandwidth usage, accepting a new requestfrom this class will merely accentuate the di�er-ences, while rejecting new request from a lowbandwidth usage class will also have a similar ef-fect. Therefore, a request from a low usage class

should be accepted, while a request from a highusage class should be rejected. To better under-stand the operations of this module, an examplewhich involved three classes will be considered. Inthis example, a simple rule base with 10 rules is setup for a Class 1 controller (see Table 6), whereBW1, BW2 and BW3 are the antecedents repre-senting the bandwidth usages of the three classeswhile Gfairness is the consequent. The membershipfunctions used in this example are similar to thatshown in Table 3.

As in the throughput module, two cases will beconsidered:· Case 1: BW1� 0.15, BW2� 0.35 and BW3�

0.35.· Case 2: BW1� 0.45, BW2� 0.15 and BW3�

0.15.The ®rst case relates to a low bandwidth usage

for Class 1 while the second case relates to a highbandwidth usage. In both cases, the values of BW1include the estimated bandwidth usage of the newsetup request.

The ®rst case will now be discussed. Table 7shows the results for the whole fuzzy mechanism.Unlike the throughput module where there is onlyone universe of discourse, there are now threeuniverses of discourse on the antecedent side.Therefore, the degree of ®ring for each antecedentsp

j , 8jj16 j6 3 is ®rst multiplied together to give aresultant

Q3j�1 sp

j in the ®fth column. Next, theresultant degree of ®ring is multipled with Mp togive

Q3j�1 sp

j Mp. Using the centre of area method,the defuzzi®ed Gfairness is obtained as

Table 5

Sample computations for total_throughput� 0.85

Rule sp Mp spMp

�a� �b� �a� � �b�1 0:00 0:9 0:000

2 0:00 0:8 0:000

3 0:00 0:8 0:000

4 0:01 0:7 0:007

5 0:01 0:7 0:007

6 0:03 0:6 0:018

7 0:18 0:5 0:090

8 1:00 0:1 0:100

9 1:00 0:1 0:100

10 0:18 0:1 0:018

Total 2:41 0:340

Table 6

Sample rule base used for fairness module in Class 1 controller

Rule BW1 BW2 BW3 Gfairness

1 0:1 0:3 0:4 0:92 0:1 0:2 0:3 0:8

3 0:2 0:3 0:4 0:7

4 0:2 0:2 0:3 0:6

5 0:3 0:2 0:4 0:56 0:3 0:3 0:1 0:2

7 0:4 0:1 0:3 0:1

8 0:4 0:4 0:1 0:1

9 0:5 0:2 0:1 0:110 0:5 0:1 0:3 0:1


Gfairness �P10

p�1

Q3j�1 sp

j MpP10p�1

Q3j�1 sp

j

� 1:687

2:280

� 0:74:

Since Class 1 is using less bandwidth as comparedto the other two classes, Gfairness should be greaterthan 0:5, as expected.

The above steps are repeated for the secondcase where Class 1 has a high bandwidth usage andthe corresponding results are given in Table 8. Thedefuzzi®ed Gfairness is now 0:418=3:228 � 0:13,which is much smaller than 0:5. Since Class 1 isalready consuming a lot of bandwidth, accepting a

new request from Class 1 will merely accentuategreater unfairness. Therefore, the new requestshould be rejected.

The last fuzzy controller is the CAC module.This module takes both Gfairness and the predictedQoS values over T, di

q�t � T �, 8qj16 q6 n as in-puts. Each universe of discourse is de®ned from 0to 1:0 with divisions of 0:1. The defuzzi®ed outputGCAC will have a value greater than 0:5 only ifGfairness is greater than 0:5. In addition, the pre-dicted QoS values must not violate any QoSthresholds. This GCAC is then passed to thethreshold comparator which accepts the new setuprequest only if both the values of Gthroughput andGCAC are greater than 0:5; otherwise the request isrejected.

Table 8

Sample computations for BW1� 0.45, BW2� 0.15 and BW3� 0.15

Rule sp1 sp

2 sp3

Q3j�1 sp

j MpQ3

j�1 spj Mp

�a� �b� �c� �a� � �b� � �c� �d�1 0:01 0:18 0:18 0:000 0:9 0:000

2 0:01 1:00 1:00 0:010 0:8 0:008

3 0:03 0:18 0:18 0:001 0:7 0:001

4 0:03 1:00 1:00 0:030 0:6 0:018

5 0:18 1:00 1:00 0:180 0:5 0:090

6 0:18 0:18 0:18 0:006 0:2 0:001

7 1:00 1:00 1:00 1:000 0:1 0:100

8 1:00 0:03 0:03 0:001 0:1 0:000

9 1:00 1:00 1:00 1:000 0:1 0:100

10 1:00 1:00 1:00 1:000 0:1 0:100

Total 3:228 0:418

Table 7

Sample computations for BW1� 0.15, BW2� 0.35 and BW3� 0.35

Rule sp1 sp

2 sp3

Q3j�1 sp

j MpQ3

j�1 spj Mp

�a� �b� �c� �a� � �b� � �c� �d�1 1:00 1:00 1:00 1:000 0:9 0:900

2 1:00 0:18 0:18 0:032 0:8 0:026

3 1:00 1:00 1:00 1:000 0:7 0:700

4 1:00 0:18 0:18 0:032 0:6 0:019

5 0:18 0:18 0:18 0:006 0:5 0:003

6 0:18 1:00 1:00 0:180 0:2 0:036

7 0:03 0:03 0:03 0:000 0:1 0:000

8 0:03 1:00 1:00 0:030 0:1 0:003

9 0:01 0:18 0:18 0:000 0:1 0:000

10 0:01 0:03 0:03 0:000 0:1 0:000

Total 2:280 1:687


3. Proposed connection admission control scheme

Fig. 2 shows the pseudocode for the proposedCAC scheme. Each of the modules described inSection 2 applies this algorithm for each class oftra�c. At regular intervals Dt, network informa-tion is passed to the controller which is used topredict the maximum amount of bandwidth re-quired for the class of tra�c over the next aver-aging period, T. This prediction, represented asbwi max�t � T �, is passed to the ATM network.During the same time interval, training of bothMLP bandwidth predictor and QoS predictor arealso carried out.

As proposed by Hiramatsu [4], pattern tableswhich store previously observed training patternsare used to improve the performance of the MLPneural networks. For the bandwidth predictor, onlyone table is used to store previously observedtraining patterns. However, for the QoS predictor,two pattern tables are used, one for low-cell-loss-rate events where CLR is less than target CLR,and another for high-cell-loss-rate events. Patternsobserved at each Dt are inserted into the appro-priate table. When either table is full, the oldestpattern is replaced with the current observed pat-tern. During training, the high-cell-loss-rate pat-tern table is selected with ®xed probability Pt, andthe low-cell-loss-rate table with probability�1ÿ Pt�. A training pattern is then chosen ran-domly from the selected table and used for trainingthe neural network. This permits the neural net-work to be trained with di�erent input patternsselected from the input domain in a random orderand can help it to achieve faster and more accuratelearning.

In this scheme, the neural networks are alwaystrained with the current information. After that,during the same interval, training is repeated usingthe chosen pattern from the selected pattern table.This ensures that the neural networks utilize newinformation as soon as it becomes available, whilenot ``forgetting'' older patterns. Consequently, thecontroller is able to react accordingly to a suddenchange in the tra�c situation. In addition, acompanding scheme involving logarithmic con-version is used to transform the average cell lossrate value to the target value used for training, as

the range of the cell loss probability is exponen-tially wide.

Whenever there is a setup request, the ATMnetwork passes the bandwidth usages of all dif-ferent classes, including the new request to theCAC controller. The aggregation of all thesebandwidths is sent to the fuzzy throughput modulewhich determines Gthroughput. The bandwidth usagesof all classes are also passed to the fairness modulewhich determines Gfairness. The next module, CACmodule will then accept the predicted CLR andGfairness to determine GCAC. Both GCAC and Gthroughput

are ®nally passed to the threshold comparator whichaccepts the new request only if there is no violation ofany criteria.

Fig. 3 shows the pseudocode of the proposedmultiplexer. At regular intervals, the multiplexerwill pass the network information to the CACcontroller, which returns the bandwidth require-ment for a particular class of tra�c. The multi-plexer then compares the predicted value againstthe weighted values of average bandwidth andequivalent bandwidth. This is necessary becauseneural network may predict the extreme lowbandwidth requirement, which might result in cellloss if a new setup request is accepted. A mini-mum bandwidth requirement has to be speci®ed.However, using average bandwidth would de®-nitely result in cell loss, as simulations in Section4 will show. On the other hand, using equivalentbandwidth would be rather conservative. There-fore, weighted values of these bandwidths areused.

4. Simulations

4.1. ATM network simulator

The NIST ATM network simulator [10] pro-vides an interactive ATM network modeling en-vironment with a graphical user interface. Itenables the user to: (1) create di�erent networktopologies, components and applications; (2)measure simulated network activity at di�erentpoints in the network; and (3) modify componentand application parameters to evaluate their ef-fects. The core of the simulator is a discrete eventsimulation kernel. Additional modules have been


Fig. 2. Pseudocode for proposed CAC scheme.


added to enable it to exchange data with di�erentnetwork processes using UNIX shared memory, aswell as to implement the tra�c source and CACcontroller described in this paper.

4.2. Tra�c characteristics

The network topology used in the simulations isa local area network (LAN) as shown in Fig. 4.Each of the link in this con®guration is of length 1km. As can be seen, there are six hosts in thiscon®guration. Tra�c is one directional fromHost1A to Host2A, Host1B to Host2B andHost1C to Host2C. Host1A, Host1B and Host1Care each connected to large number of variable bitrate (VBR) applications, where each host repre-sents a multiplexed tra�c of VBR sources with thesame characteristics. Each VBR application isrepresented as an ON±OFF model whereby boththe burst period (ON) and the silence period(OFF) are drawn from an exponential distribu-tion. During the ON mode, cells are transmitted atpeak cell rate (PCR) (see Fig. 5). Parameters suchas the mean burst length (tON), the mean OFFperiod (tOFF), and the PCR, are to be speci®ed.

4.3. Experiments

In this paper, two sets of simulations are per-formed, each trying to show di�erent aspects of theproposed scheme. The ®rst set aims to ®nd out ifthe proposed scheme is able to achieve the targetQoS. It also compares the performance of thisscheme against the other CAC schemes. The sec-ond set shows how fairness can be improved withthe proposed scheme. The equivalent bandwidth

Fig. 4. LAN peer-to-peer con®guration.

Fig. 3. Pseudocode for multiplexer behavior using proposed CAC scheme.

Fig. 5. A simple ON±OFF VBR model.


indicated in Table 9 is computed using the equa-tions derived in [2]. Let K be the bu�er size ex-pressed in number of cells, � the target CLR(unitless) and r the fraction of time the source isactive. A simpli®ed solution to the equivalentbandwidth c for a single source is given as

c �aÿ K �

��aÿ K�2 � 4Kar

q2a

PCR; �3�

where

a � PCR�1ÿ r�tON ln�1=��; �4�

r � tON

tON � tOFF

: �5�

Detailed tra�c characteristics of the three classesof VBR sources are shown in Table 9. The di�erenttra�c parameters for each source are related bythe following expression:

tcall � StONPCR

tON � StONPCR

� �tOFF: �6�

Each set of simulation is performed for each CACscheme. For comparison purposes, four other

Table 9

Tra�c characteristics of VBR sources used in CAC simulations

Set VBR tra�c characteristics Source

Class 1 Class 2 Class 3

1 Call interarrival time (s) 0.01 0.01 0.01

Peak cell rate, PCR (Mbytes/s) 3.0 5.0 10.0

Equivalent bandwidth (Mbytes/s) 2.61 4.58 9.56

Average bandwidth (Mbytes/s) 1.5 2.5 5.0

Mean ON period, tON (ms) 10 10 10

Mean OFF period, tOFF (ms) 10 10 10

Mean call holding time, tcall (ms) 677 677 677

Mean transmission size, S (Mbits) 1.011 1.685 3.370

2 Call interarrival time (s) 0.005 0.05 0.05

Peak cell rate, PCR (Mbytes/s) 5.0 1.5 1.0

Equivalent bandwidth (Mbytes/s) 4.58 1.17 0.72

Average bandwidth (Mbytes/s) 2.5 0.75 0.5

Mean ON period, tON (ms) 10 10 10

Mean OFF period, tOFF (ms) 10 10 10

Mean call holding time, tcall (ms) 2632 2632 2632

Mean transmission size, S (Mbits) 6.560 1.968 1.312

Table 10

Parameters used in CAC simulations

Parameter Value

Parameters common to all schemes

Sampling interval, Dt (ms) 1

Bu�er size of multiplexer, K (cells) 100

Link speed, Lc (Mbytes/s) 150

Link distances, x (km) 1

Target CLR, � 10ÿ4

Companded Target CLR 0.5

Parameters pertaining to PureNN

Size of pattern table 100

Pattern table selection probability, Pt 0.5

Size of input pattern to MLP 3

Number of output from MLP 1

Parameters pertaining to Fuzzy/NN

Size of pattern table 100

Pattern table selection probability, Pt 0.5

Size of input pattern to bandwidth predictor 5

Number of output from bandwidth predictor 1

Size of input pattern to QoS predictor 5

Number of output from QoS predictor 1

Size of input pattern to throughput module 1

Threshold for throughput module,

THthroughput

0.9

Size of input pattern to fairness module 3

Size of input pattern to CAC module 2

Weighted a of average bandwidth and equiv-

alent bandwidth

0.8


CAC schemes are also implemented. All the im-plemented CAC schemes are listed below:1. Fuzzy/NN: This is the proposed scheme which

uses integrated MLP neural networks and fuzzycontrollers.

2. PureNN: This scheme uses MLP neural net-work to predict the CLR from the number ofconnections of di�erent classes. A new requestis accepted only when the predicted CLR doesnot violate the target CLR.

3. Peak: This scheme utilizes the peak bandwidthrequirement of each request to perform CAC,i.e. a new request is accepted only when

X3

i�1

XNi�t�

j�1

PCRij6 1;

where PCRij is the PCR of jth source from ith

class and Ni�t� is the number of currently con-nected ith class sources.

4. Equivalent: This scheme utilizes the equivalentbandwidth requirement of each request to per-form CAC, i.e. a new request is accepted onlywhen

X3

i�1

XNi�t�

j�1

cij6 1;

where cij is the equivalent bandwidth of jth

source from ith class, computed using Eq. (3).5. Average: This scheme utilizes the average band-

width requirement of each request to performCAC, i.e. a new request is accepted only when

X3

i�1

XNi�t�

j�1

tiON;j

tiON;j � ti

OFF;j

PCRij6 1;

where tiON;j and ti

OFF;j are the mean ON andOFF periods of jth source from ith class.

Fig. 6. Results for Set 1 at the end of 20 s simulation time: Cell loss rate of Classes 1±3 tra�c, normalized total link utilization and

total declared peak and mean bit rates vs time for Average.


For all simulations, the MLP networks used inboth Fuzzy/NN and PureNN have 10 neurons inthe hidden layer. However, the number of neu-rons in the input and output layers for bothschemes are di�erent and they are indicated inTable 10.

5. Results and discussion

5.1. Set 1

5.1.1. AverageFigs. 6±9 show plots of CLR for the three VBR

tra�c classes, total link utilization as well as totaldeclared peak and mean bit rate for the di�erentCAC schemes in simulation Set 1. Tables 11 and12 give the throughput and cell loss results, re-spectively, while Table 13 shows the call blockingprobabilities for each tra�c class. This set of

simulations aims to ®nd out the performance ofthe di�erent CAC schemes in terms of throughput,CLR and call blocking capability. From Table 11,it can be seen that Average achieves the highestthroughput. However, this high throughput isobtained at the expense of high CLR. This can beseen in Fig. 6, where cell loss occurs for the threetra�c classes throughout the entire simulationperiod. This is a direct consequence of its optimisticbandwidth allocation strategy, which only focuseson long-term bandwidth requirement and disre-gards any short-term bit rate ¯uctuations whichmay lead to temporary bu�er over¯ow. From Fig.6, it is also observed that the total declared mean bitrate is maintained at approximately 150 Mbytes/sthroughout the simulation, which agrees with in-tuition since bandwidth is allocated according tothe declared mean bit rates. The wide gap betweenthe total declared peak and mean bit rates indicatethat Average is not very biased against bursty


total declared peak and mean bit rates vs time for (a) Peak and (b) Equivalent.


applications, thus it is able to bene®t from statisti-cal multiplexing gain. This observation is alsoshown in Table 13 where Class 3 sources whichhave the highest PCR, exhibit only slightly highercall blocking probability than Class 1 tra�c.

5.1.2. PeakUnlike Average, the bandwidth allocation

strategy used by Peak is over-conservative. FromFig. 7(a), it is observed that no cell loss occursthroughout the entire simulation period. This is tobe expected because su�cient bandwidth has beenallocated to each VBR session to accomodate theirPCR. However, as the VBR sources do nottransmit at PCR most of the time, the overallbandwidth utilization is rather low. This can beseen in the link utilization plot in Fig. 7(a), wherethere is a lot of wasted bandwidth. Table 11 alsoshows that Peak has the lowest throughput amongall CAC schemes.

From Fig. 7(a), it is observed that the totaldeclared peak bit rate is maintained below 150Mbytes/s throughout the simulation. This agreeswith intuition that any new connection requestwhich may cause the total peak bit rate to exceedthe link capacity of 150 Mbytes/s will be rejectedby the Peak CAC scheme. The narrow gap be-tween the peak and mean bit rate plots indicatethat Peak does not bene®t much from statisticalmultiplexing gain. In fact, Table 13 shows thatPeak has the highest call blocking probabilities foreach tra�c class among all the CAC schemes. Thisis again due to Peak being over-conservative in itsbandwidth allocation.

5.1.3. EquivalentLike Peak, Equivalent also exhibits no cell loss

because it is over-conservative in its bandwidthallocation. As a result, it also has a low through-put, albeit slightly higher than Peak. From


total declared peak and mean bit rates vs time for PureNN.


Fig. 7(b), it is observed that the total declared peakbit is maintained at approximately 150 Mbytes/sthroughout the simulation. This is because thecomputed equivalent bandwidths shown in Table 9are only slightly lower than the PCRs. This also

accounts for the slightly higher throughput whencompared to Peak. As shown in Fig. 7(b), the gapbetween the peak and mean bit rate is only slightlylarger than Peak. This is explained by the slightlylower call blocking probabilities in Table 13.

Fig. 9. Results for Set 1 at the end of 20 s simulation time: Cell loss rate of Classes 1± 3 tra�c, normalized total link utilization and

total declared peak and mean bit rates vs time for Fuzzy/NN.

Table 11

Throughput results for Set 1 at the end of 20 s simulation time

CAC scheme Cells received (% link utilization)

Class 1 Class 2 Class 3 Total

Fuzzy/NN 1,236,351 1,966,972 1,969,857 5,173,180

(17.47%) (27.80%) (27.84%) (73.11%)

PureNN 1,345,827 1,700,183 3,695,509 6,741,519

(19.02%) (24.03%) (52.23%) (95.28%)

Peak 751,132 1,080,140 1,641,064 3,472,336

(10.62%) (15.27%) (23.19%) (49.08%)

Equivalent 870,646 1,175,634 1,696,700 3,742,980

(12.31%) (16.62%) (23.98%) (52.90%)

Average 1,394,688 2,047,785 3,344,003 6,786,476

(19.71%) (28.94%) (47.26%) (95.92%)


5.1.4. PureNNFrom Table 11, it is observed that PureNN is

able to achieve a high throughput. This is fur-ther shown in Fig. 8 where PureNN is able toreach full link utilization most of the time.However, the cell loss plots also indicate theextremely high CLR experienced by all the threetra�c classes. This is due to the learning mech-anism of the MLP predictor. Although it is ableto learn the input±output mapping, the burstynature of the tra�c sources presents di�culty forthe MLP to learn accurately. This is because forthe same set of connection pattern of di�erentclasses, it is possible to have either cell loss orno cell loss. Therefore, the MLP will learn onlythe average CLR for a particular connection

pattern. The large gap between the peak andmean bit rate indicates that PureNN is able tobene®t greatly from statistical multiplexing. Ta-ble 13 also shows that this scheme has the lowestcall blocking probabilities for all the three tra�cclasses.

5.1.5. Fuzzy/NNFuzzy/NN is the proposed CAC scheme that

utilizes the learning abilities of neural networksand the robustness of fuzzy controllers. Results forPureNN has shown that using MLP neural net-works alone is not su�cient to perform CAC ef-fectively. By adding fuzzy controllers, theperformance could be further improved. For ex-ample, the CLR plots in Fig. 9 show that cell lossoccurs only at the initial stage of the simulationtime. This is because the MLP bandwidth predic-tor has not learnt the tra�c patterns yet, andtherefore it is unable to predict the requiredbandwidth usage accurately. This is of little con-sequence, because the di�erent fuzzy modules willplay their roles in the CAC and reject calls ac-cording to their rule bases, while allowing theMLP networks to learn the tra�c patterns moreaccurately. Table 14 shows the breakdown of the

Table 14

Breakdown of the average call blocking probabilities given in Table 13 for Fuzzy/NN

Rejection due to Class 1 Class 2 Class 3

(i)Pm

j�1 bwj max�t � T �P 1:0 0.002 0.003 0.006

(ii) Throughput module alone 0.722 0.725 0.047

(iii) Fairness module alone 0 0 0.126

(iv) Prediction from QoS predictor 0 0.001 0

(v) Throughput and fairness modules 0 0 0.674

(vi) Prediction from QoS predictor and throughput module 0.006 0.020 0

(vii) Prediction from QoS predictor and fairness module 0 0 0

(viii) Prediction from QoS predictor, throughput and fairness modules 0 0.001 0.017

Table 12

Cell loss results for Set 1 at the end of 20 s simulation time

CAC scheme Cells discarded


Fuzzy/NN 183 2830 2501 5514

PureNN 25,044 353,091 474,345 852,480

Peak 0 0 0 0

Equivalent 0 0 0 0

Average 3745 61,746 65,685 131,176

Table 13

Average call blocking probabilities for Set 1 at the end of 20 s

simulation time

CAC scheme Class 1 Class 2 Class 3

Fuzzy/NN 0.73 0.75 0.87

PureNN 0.70 0.73 0.7

Peak 0.84 0.86 0.89

Equivalent 0.81 0.85 0.89

Average 0.70 0.73 0.78


Fig. 10. Results for Set 2 at the end of 20 s simulation time: Normalized link utilization for the three classes of VBR tra�c for:

(a) Fuzzy/NN; (b) PureNN; (c) Peak; (d) Equivalent; (e) Average.

Table 15

Throughput results for Set 2 at the end of 20 s simulation time

CAC scheme Cells received (% link utilization)


Fuzzy/NN 2,201,944 (31.12%) 1,620,967 (22.91%) 1,108,254 (15.66%) 4,931,165 (69.69%)

PureNN 6,744,616 (95.32%) 128,138 (1.81%) 53,412 (0.75%) 6,926,166 (97.89%)

Peak 3,140,313 (44.38%) 222,027 (3.14%) 148,018 (2.09%) 3,510,358 (49.61%)

Equivalent 3,489,807 (49.32%) 238,362 (3.37%) 162,002 (2.29%) 3,890,171 (54.98%)

Average 6,435,195 (90.95%) 315,514 (4.46%) 208,264 (2.94%) 6,958,973 (98.35%)

Table 16

Cell loss results for Set 2 at the end of 20 s simulation time

CAC scheme Cells discarded


Fuzzy/NN 0 0 0 0

PureNN 588,965 88,529 91,071 768,565

Peak 0 0 0 0

Equivalent 0 0 0 0

Average 8962 4816 5329 19,107


call blocking probabilities given in Table 13 forthe three tra�c classes. As can be seen, most of therejections of the new setup requests are due to thethroughput module for Classes 1 and 2 tra�c,while most of the rejections for Class 3 tra�c aredue to both the throughput and fairness modules.For Class 3 tra�c, it is observed that the fairnessmodule plays a rather important role. This is be-cause each Class 3 VBR source has a higherbandwidth requirement than the other two tra�cclasses. Thus, accepting a new setup request from aClass 3 source would easily cause the bandwidthusage of the Class 3 tra�c class to be greater thanthe other two classes. This results in unfairness,which causes the fairness module in the Class 3controller to reject subsequent new setup requests.From Table 14, it can also be seen that very few ofthe rejections are actually due to the predictionfrom MLP QoS predictor, which further empha-sises the robustness of this scheme.

Table 11 shows that Fuzzy/NN is able toachieve a reasonably high throughput of 73.11%,which is obtained from ��Total cells received�424�=Lc� � 20 s simulation time. The utilizationplot in Fig. 9 shows that Fuzzy/NN is able to

maintain the link utilization below the THthroughput

of 0.9, although there are occurences of link utili-zation reaching the link capacity. The gap betweenthe peak and mean bit rates in Fig. 9 shows thatthis scheme could only bene®t slightly from sta-tistical multiplexing.

5.2. Set 2

Fig. 10 shows the normalized link utilizations ofthe three classes of VBR tra�c (Tables 15 and 16).This set of simulations aims to further highlightthe advantages of the proposed scheme. Class 1tra�c is intentionally setup to have small call in-terarrival times, so that Classes 2 and 3 tra�c areoverwhelmed. For this set of simulations, all thethree classes are assumed to have the same prior-ity. Fig. 10 shows that PureNN, Peak, Equivalentand Average are unable to allocate bandwidthfairly, with PureNN and Average showing themost unfairness. On the other hand, Fuzzy/NN isable to prevent Class 1 tra�c from overwhelmingthe other two classes by rejecting su�cient Class 1setup requests to ensure available bandwidth whenthere are Classes 2 and 3 requests. Table 17 showsthe high call blocking probability for Class 1 tra�cand the low call blocking probabilities for Classes2 and 3 tra�c. A breakdown of the call blockingprobabilities given in Table 17 is shown inTable 18. As can be seen, most of the rejections bythe Class 1 controller are due to both thethroughput and fairness modules, with fairnessmodule playing a greater role. Table 15 also showsthat Fuzzy/NN is able to achieve high throughputwhile maintaining fairness.

Table 17

Average call blocking probabilities for Set 2 at the end of 20 s

simulation time

CAC scheme Class 1 Class 2 Class 3

Fuzzy/NN 0.96 0.07 0.04

PureNN 0.87 0.88 0.88

Peak 0.95 0.87 0.87

Equivalent 0.94 0.86 0.86

Average 0.89 0.82 0.82

Table 18

Breakdown of the average call blocking probabilities given in Table 17 for Fuzzy/NN

Rejection due to: Class 1 Class 2 Class 3

(i)Pm

j�1 bwj max�t � T �P 1:0 0.001 0.005 0.005

(ii) Throughput module alone 0.004 0.065 0.035

(iii) Fairness module alone 0.262 0 0

(iv) Prediction from QoS predictor 0 0 0

(v) Throughput and fairness modules 0.693 0 0

(vi) Prediction from QoS predictor and throughput module 0 0 0

(vii) Prediction from QoS predictor and fairness module 0 0 0

(viii) Prediction from QoS predictor, throughput and fairness modules 0 0 0


6. Conclusion

This paper proposes an integrated neural andfuzzy controller to perform call admission controlin an ATM network. Although neural networksare well-known for their learning capabilities, theirapplications to commercial uses are few. This isbecause of the inherent instabilities of neural net-works where they might produce spurious re-sponse at times. On the other hand, fuzzycontrollers are more frequently in use. However,these controllers usually require careful analysis ofthe control problem to set up the rule-base. Thesetting up of rule base requires ®nding the rela-tionships between important parameters, which isusually di�cult, if not impossible. Therefore, byintegrating the learning ability of neural networkand the robustness of fuzzy controller, we come upwith a control framework that is able to performCAC e�ectively. The advantages of this scheme ishighlighted through di�erent sets of simulationsperformed with the help of a network simulator. Itis shown that this scheme is able to achieve highthroughput and low cell loss while maintainingfairness among di�erent classes of tra�c.

Appendix A. Adaptive learning rate algorithm

This section describes brie¯y the adaptivelearning rate algorithm to train the MLP neuralnetwork. For this algorithm, the error functionused is the MSE function, expressed as

e�t� � 1

2

Xj

e2j �t�; �A:1�

where ej�t� is the error signal at the output ofneuron j for iteration t.

During training, reduction of e�t� is achievedthrough the method of steepest descent given byEq. (A.2). The weights are then updated usingEq. (A.3):

Dwji�t� � ÿgji�t�oe�t�

owji�t� ; �A:2�

where wji�t� is the synaptic weight connecting theoutput of neuron i to the input of neuron j at it-eration t. gji�t� is the learning step size for wji�t�:

wji�t � 1� � wji�t� � Dwji�t�: �A:3�At every time step t, in addition to the updating ofwji�t�, each gji�t� is updated using Eqs. (A.4)±(A.7):

At every time step t, in addition to the updating ofwji�t�, each gji�t� is updated using Eqs. (A.4)±(A.7):

Dgji�t � 1�

�j if Sji�t ÿ 1�Dji�t� > 0;

ÿbgji�t� if Sji�t ÿ 1�Dji�t� < 0;

0 otherwise;

8><>: �A:4�

where j and b 2 �0; 1�,

Dji�t� � oe�t�owji�t� ; �A:5�

Sji�t� � �1ÿ n�Dji�t ÿ 1� � nSji�t ÿ 1�; �A:6�

gji�t � 1� � gji�t� � Dgji�t�: �A:7�

For more information on neural networks andthe adaptive learning rate algorithm, please referto Ref. [8].

References

[1] ATM Forum, Tra�c Management Speci®cation Version

4.0, AF-TM-0056.000, April 1996.

[2] R. Gu�erin, H. Ahmadi, M. Naghshineh, Equivalent

capacity and its application to bandwidth allocation in

high-speed networks, IEEE Journal on Selected Areas in

Communications 9 (7) (1991) 968±981.

[3] A.I. Elwalid, D. Mitra, E�ective bandwidth of general

markovian tra�c sources and admission control of high

speed networks, IEEE/ACM Transactions on Networking

1 (3) (1993) 329±343.

[4] A. Hiramatsu, ATM communications network control by

neural networks, IEEE Transactions on Neural Networks

1 (1) (1990) 122±130.

[5] E. Nordstr�om, J. Carstr�om, O. G�allmo, L. Asplund,

Neural networks for adaptive tra�c control in ATM

networks, IEEE Communications 33 (10) (1995) 43±49.

[6] W. Melody Moh, M.-J. Chen, N.-M. Chu, C.-D. Liao,

Tra�c prediction and dynamic bandwidth allocation over

ATM: a neural network approach, Computer Communi-

cations 18 (8) (1995) 563±571.

[7] J.E. Neves, M.J. Leit~ao, L.B. Almeida, Neural networks in

B-ISDN ¯ow control: ATM tra�c prediction or network

modeling, IEEE Communications 33 (10) (1995) 50±56.


[8] S. Haykin, Neural Networks ± a Comprehensive Founda-

tion, MacMillan, New York, 1994, Ch. 6.

[9] C.-T. Lin, C.S. George Lee, Neural-network-based fuzzy

logic control and decision system, IEEE Transactions on

Computers 40 (12) (1991) 1320±1336.

[10] N. Golmie, A. Koenig, D. Su, The NIST ATM network

simulator, Technical Report NISTIR 5703, National

Institute of Standards and Technology, USA, August 1995.

Chen-Khong Tham is an AssistantProfessor at the Department of Elec-trical Engineering of the NationalUniversity of Singapore. He lecturesand conducts research in the areas ofquality of service in computer net-works, real-time and embedded sys-tems, and network management. Dr.Tham is also the program manager forIT & Networks in the Laboratory forConcurrent Engineering & Logistics atNUS, and was until recently an Ap-plications Manager at SingAREN, theSingapore Advanced Research & Ed-

ucation Network. Dr. Tham obtained his Ph.D. and B.A.(Honours) degrees in Engineering from the University ofCambridge, UK.

Ng Onn Lum Nelson is currently asystems engineer working in the Sin-gapore Ministry of Defence (MIND-EF). He develops secure messagingapplications, as well as explores newtechnologies that could be applied tothe infrastructure of MINDEF. Nel-son obtained his M.Eng. and B.Eng.(Honours) degrees from the NationalUniversity of Singapore.


Connection admission control of ATM network using integrated MLP and fuzzy controllers

Documents