APPLYING NEURO-FUZZY DYNAMIC BUFFER TUNING TO

MAKE WEB-BASED TELEMEDICINE SUCCESSFUL

Jackei H. K. Wong, Chen Ye Zhu, Wilfred W. K. Lin and Allan K. Y. Wong

Department of Computing, The Hong Kong Polytechnic University Hung Hom, Kowloon, Hong Kong

Keywords: Neuro-Fuzzy Logic Controller, web-based telemedicine, mobile Internet, fast system response, TCM

(Traditional Chinese Medicine), dynamic buffer tuning.

Abstract: We propose to make web-based medical consultation successful by applying the Neuro-fuzzy Dynamic

Logic Controller (NFLC). This is achieved for the NFLC shortens the service response time for the

physician, who answers the patient requests pervasively, by dynamic buffer tuning. The physician carries a

SFF (small form factor) mobile device (e.g. PDA) that provides the interface for interacting wirelessly with

rest of the web-based telemedicine system (WTS) on the mobile Internet. The WTS in this paper caters to

Traditional Chinese Medicine (TCM) and therefore called TCM-WTS or simply T-WTS. The T-WTS

usability relies on various factors such as correct information exchange, and fast system response. This

paper focuses on the second factor by exploiting real-time dynamic buffer tuning as a solution.

1 INTRODUCTION

We propose to apply the NFLC (Neuro-Fuzzy Logic

Controller), which is a dynamic buffer tuner to

quicken the response of the extant T-WTS (TCM

(Traditional Chinese Medicine) Web-based

Telemedicine System), which was operated by the

Purapharm Group of the Hong Kong SAR. The T-

WTS response is slow and unreliable under heavy

traffic conditions. Our preliminary study showed

that slow response and channel unreliability could be

caused by overflows at the server side. Overflows

cause widespread retransmission and long service

roundtrip time (RTT) (i.e. slow service response).

TCM is enshrined in the Hong Kong local law. Its

popularity invigorates the local drive to make TCM

reach every SAR corner and eventually the rest of

the world. The NFLC can contribute to make this

goal a success.

The T-WTS is distributed on the mobile Internet

and allows pervasive interaction between a mobile

TCM physician and the dedicated surrogate

node/server assigned to the smart space (Patterson et

al., 2003). Before interaction takes place the

physician must move into a smart space, which is a

communication cell that seamlessly supports various

wireless technologies. The interaction relationship is

one-surrogate-to-many-clients (i.e. physicians) or

asymmetric rendezvous.

A physician provides T-WTS based medical

consultations anytime and over any geographical

location via a portable SFF (small form factor)

mobile device (e.g. mobile phone). The device hosts

a logical agent to provide the interface for remote

interaction with the rest of the distributed T-WTS.

With the SFF patient records can be created, stored

and retrieved remotely. The mechanism to support

all these activities in the background is the dedicated

surrogate server. A physician can dispense

prescriptions in a remote fashion. Figure 1 shows the

T-WTS infrastructure as follows: a) it is operating

pervasively over the mobile Internet that supports

both wireless and wireline communications; b) end-

to-end client/server interaction can be wireless and

wireline (server is surrogate (Patterson et al., 2003));

c) T-WTS has many surrogates that collaborate over

a wireline high-speed network. Every surrogate

server is assigned to serve at least one smart space

and those physicians (i.e. clients) within; and d) the

physicians interact with their surrogate via mobile

SFF devices in a wireless manner, e) if a surrogate

cannot serve a request it seeks help from others, in

the cyber foraging mode.

Cyber foraging under Markovian conditions is

the M/M/n (M for Markov) model; n is the number

nodes/surrogates/information-stations) in collaboration.

(1 / )

(1 )

−

151

H. K. Wong J., Ye Zhu C., W. K. Lin W. and K. Y. Wong A. (2008).

APPLYING NEURO-FUZZY DYNAMIC BUFFER TUNING TO MAKE WEB-BASED TELEMEDICINE SUCCESSFUL.

In Proceedings of the International Conference on Wireless Information Networks and Systems, pages 151-156

DOI: 10.5220/0002024101510156

 SciTePress

The speedup produced by the distributed

parallelism for n is , where

is the surrogate

utilization. The traffic stream between a physician

and the surrogate may have a distinctive character

(e.g. self-similar) at a specific period and change

suddenly. Since the all the traffic streams from

different physician merge at the SAP (service access

point – Figure 1; “+” symbol means merging), the

resultant traffic pattern into the surrogate’s queue

can be undefined. It is not uncommon for such

merged traffic to surge the surrogate’s request

reception queue to overflow its buffer easily,

causing widespread retransmissions and thus long

service RTT (i.e. slow response) (Lin et al., 2006).

In light of telemedicine (Kaar, 1999) for which T-

WTS is an example, such overflows are not

acceptable at all. A logical solution to prevent such

mishaps is to make buffer always covers queue. This

is exactly the principle for dynamic buffer tuning

paradigm (Lin et al., 2006).

Activities inside a channel for end-to-end

communication are considered at the system level.

Normally a request sent from a client would be

routed through many routers, which have their own

local reception queues, before reaching the

destination. To prevent local routing congestion a

router may throttle any sender that send too much

and too fast by choke packets. This throttling

process is called active queue management (AQM).

It does not, however, reduce overflow due to merged

traffic at the user level (i.e. surrogates and clients)

(Lin et al., 2006).

Figures 2 and 3 are screen captures for the

following wireless operations via the T-WTS

respectively: a) login request by a TCM physician;

and b) request to the surrogate for retrieving a

patient record. The control bar shows some of the T-

WTS icons, namely, Login, Patient Record,

Prescription, and Dispensing. Each icon can be

exploded for more detailed operations. Field tests of

the basic

T-WTS prototype with no dynamic tuning

support indicated that its response time could vary

significantly over 24 hours. Our analysis indicates

that one cause of the variations was the transient

mass transit population through a smart space

(Jamioom 2004). This concurs with the findings by

others (e.g. (Kaar, 1999)). The mass transit can

seriously increase the traffic volume between SFF

mobile clients and the surrogate at peak hours. One

solution to lessen the congestion is setting a

maximum number of SFF-surrogate connections in a

smart space. This solution, however, cannot prevent

surrogate buffer overflow caused by traffic ill

effects. This paper only focuses on how to apply the

NFLC to deal with traffic volume.

Figure 1: Pervasive T-WTS infrastructure.

Figure 2: Patient record retrieval.

2 RELATED WORK

We observed that slow T-WTS system response

could be caused by frequent buffer overflow at the

reception side. This cannot be resolved by AQM

(active queue management) alone (Braden et al.,

1998). It is naturally to augment the AQM with user-

level reception buffer overflow by dynamic buffer

WINSYS 2008 - International Conference on Wireless Information Networks and Systems

152

Figure 3: Prescription preparation.

tuning ((Lin et al., 2006), (Lin et al., 2007)). For this

reason propose to apply the novel NFLC (Neuro-

Fuzzy Logic Controller) dynamic buffer tuner,

which we developed for telemedicine systems. Our

simulations showed that the NFLC had

outperformed other extant dynamic buffer tuners

(e.g. FLC (Fuzzy Logic Controller) (Lin et al.,

2006)). The NFLC proposal here is partially based

on a theoretical controller (Wang et al., 2001), which

could not be realized due to absence of details such

as: how to train the neural network part; how to

specify the fault tolerance; and how to avoid diving

by zero. The NFLC carefully addresses these

shortcomings in light of usability.

3 THE NEURO-FUZZY LOGIC

CONTROLLER

The novel Neuro-Fuzzy Logic Controller (NFLC)

shortens the T-WTS response time by reducing the

channel error probability

that encapsulates various

hardware and software errors. One of the

contributors to

is buffer overflow at the reception

side (e.g. the surrogate server in the T-WTS setup).

The

value affects the average number of trials

(ANT) to get a successful transmission, for:

[(1)]1/(1)

ANT j

ρρ

→∞

−

−≈ −

∑

Dynamic buffer tuning shrinks

because it

eliminates reception buffer overflow at the user

level (i.e. servers outside the channel domain – e.g.

surrogates). The theoretical foundation of the NFLC

can be summarized by the equations (3.1), (3.2) and

(3.3), for which the parameters include: i)

for the

current control cycle; ii) CA for Convergence

Algorithm which is equation (3.3); the basis of

integral (I) control of

k cycle that involves the

current sample of size

f as well as the last

predicted mean (i.e.

1−k

M ) as feedback;

||/

IC n

→∞

∑

iii)

RIC which is the ratio using the typical/mode

value of the

f data points (i.e.

s ) in cycle k as

the reference, can be positive or negative; and iv)

QL is queue length. The actual NFLC output is

delB ; n for the current

n control cycle since

the buffer tuning process had started. The factor

shows the integral nature of the

delB calculation.

()/

kk k

k CA typical typical

RIC QL QL QL=−

(3.1)

(3.2)

()

(1)

CA k

QL M

−

∑

(3.3)

The NFLC has two main control parts: fuzzy logic

(FL), and artificial neural network (ANN). The FL

leverages two parameters, namely, the QOB (queue

length over buffer length) ratio and the rate of

changes in the queue length -

dtdQ . The FL

output in the

i cycle is a sign, ?},,{)(

−

+=•

(i.e. add, subtract, or uncertain) for buffer

adjustment, as depicted by equation (3.4). The

)(

•

is the input to the ANN part that ascertains if

“

? ” should be plus (

) or minus (-) so that the

buffer adjustment size

delB can be properly

computed as shown by equation (3.5).

(,( ))(){,,?}

ii i i

FL QOB dQ dt

•=+−

(3.4)

() (| |/ )

iiCA k

delB QL RIC n

→∞

=•

∑

(3.5)

] / | [|

n RI

QLdelB

∑

∞ →

APPLYING NEURO-FUZZY DYNAMIC BUFFER TUNING TO MAKE WEB-BASED TELEMEDICINE

SUCCESSFUL

153

Figure 4: NFLC is a swapping twin system.

Table 1: Concise 7-step NFLC pipeline.

Procedure Input Output

Sample queue

length

and

inter-arrival

times (IAT)

for requests

Time series of

and the IAT

among

requests

Expected

computed

for the interval

Normalize

and

dtdQ /

and its rate

of change

dtdQ /

Normalized

and

dtdQ /

Compute

fuzzy set;

membership

functions for

dtdQ /

Normalized

and

dtdQ /

Fuzzy set of

dtdQ /

membership

functions

(

32x

outputs)

Train/learn by

ANN for

},0{ Δ

Fuzzy

and

dtdQ /

values)

Predicted next

normalized

de-normalize

predicted

Predicted next

normalized

Predicted next

de-normalized

Compute

delB

Predicted next

de-normalized

Predicted next

de-normalized

buffer length

(

imun

min

considered)

Tune buffer

size by

delB

and for the

ascertained

)(•

Predicted next

de-normalized

buffer length

(

imun

min

considered)

Fulfilling the

equations (3.4)

and (3.5)

In operation, the three basic NFLC modules (i.e.

Chief NFLC, Learner NFLC and CA) are running in

parallel. The CA execution time has no impact on

the execution time NFLC for timing analyses of by

using the Intel’s VTune Performance Analyzer

showed that on average the NFLC needed 9800

clock cycles to execute and CA needs only 300

clock cycles.

If the Chief needs

M (equation (3.3) for its

computation, it simply fetches the current value

directly from the much faster CA entity. The

importance of the fast CA is to instantaneously

capture system changes. This is essential for

accurate and qualitative control. The Chief and

Learner modules form a twin system (Figure 4).

While the Chief NFLC performs actual control the

Learner acquires new knowledge by supervised

training. The teacher signal is the given safety

margin

of the

},0{ Δ objective function, which

was absent in the concept proposed in (Lin et al.,

2006). Training completes once if the output from

the Learner is within the

Δ± band. The current

Chief swaps position with the Learner that has just

completed its training cycle.

The internal NFLC dynamics goes through a

pipeline of 7 steps: i) sample queue length

and

IAT (inter-arrival times) of requests; ii) normalize

and

dtdQ /

; iii) compute fuzzy set; iv) train/learn

by ANN; v) de-normalize the predicted

; vi)

compute

delB

; and vii) tune buffer size by

delB

with respect to the sign ascertained

for

?},,{)(

−

•

; Table 1.

To recap, NFLC is represented by the following

elements: i) the fuzzy logic (FL) to determine the

sign of dynamic buffer adjustment (i.e. to elongate

or shorten the buffer length); ii) if the FL could not

decide the sign for

)(•

then the ANN

downstream would ascertain it before computing the

dynamic buffer size

delB for the current

adjustment/control cycle; and iii) the above

operations can be summarized as a 7-step pipeline

procedure.

4 EXPERIMENTAL RESULTS

The NFLC power in shortening the T-WTS response

time by eliminating surrogate buffer overflows was

verified by simulations. These simulations are

separated into two categories. In the first category

known discrete waveforms or distributions (e.g.

Poisson and self-similar) were used to mimic the

merge traffic in light of the IAT (inter-arrival times)

WINSYS 2008 - International Conference on Wireless Information Networks and Systems

154

among the requests to the surrogate SAP. These

waveforms verified that the “NFLC + T-WTS”

combination indeed worked stably in known traffic

conditions. In the second category wireless traces

collected in the Hong Kong Polytechnic University

(Lin) were used to verify that the same combination

indeed worked for real situations. In all simulations

the same waveform would simultaneously excite the

two dynamic buffer tuners running in parallel: PIDC

and NFLC. The aim was to compare the results by

NFLC and PIDC under the same traffic conditions.

The comparison should confirm if the NFLC was the

right choice to yield a faster T-WTS response. Many

simulations were conducted: i) T-WTS with no

dynamic buffer tuner support; (ii) T-WTS with

NFLC, and iii) T-WTS with PIDC. The preliminary

results indicate that NFLC converges faster to the

steady state reference than PIDC. There was less

oscillation in the control process as well. When

dynamic buffer tuning was absent the T-WTS

produced frequent surrogate buffer overflows.

Figure 5 shows how NFLC rectified the PIDC

problem, which was locking too much buffer

memory even when it was no longer needed. This

PIDC problem lowered T-WTS performance

because it deprived other T-WTS tasks of needed

memory. The NFLC achieved this by rigorously

maintaining the given safety margin

between the

buffer length and the queue length on the fly. Surely,

the NFLC is a more accurate, smoother, faster, and

usable dynamic buffer tuner than PIDC. The benefit

of settling quickly to the steady state is less or no

buffer overflow and thus shorter roundtrip time

(RTT) (i.e. quicker system response).

Figure 5: NFLC unlocked unused buffer space.

5 CONCLUSIONS

The NFLC (Neuro-Fuzzy Logic Controller) is

proposed to yield shorter T-WTS response. This is

achieved for NFLC produces more dependable

client/server interaction over an end-to-end channel.

This advantage becomes obvious if NFLC

performance is compared to the PIDC’s. TCM

physicians in a smart space need to hook onto the

dedicated surrogate to conduct pervasive medical

consultations. The response, however, can be

seriously affected by the transient mass through the

smart space. This mass can create unpredictable

traffic volume and pattern for the merged traffic that

enters the queue of the dedicated surrogate. If

dynamic buffer tuning is absent, the merge traffic

could surge the queue to overflow the surrogate

buffer. If this happens, the clients in the smart space

would suffer from long service RTT and their

chance to benefit from the cyber foraging in the

pervasive computing infrastructure. The NFLC

prevents surrogate buffer overflow by ensuring that

the buffer always cover the queue by the given

safety margin. This makes the channel

dependable, as confirmed by the simulation results.

The focus of this research is to explore if the

proposed NFLC can indeed prevent user-level buffer

overflow effectively. But, we analyzed the

simulation results as well for any possible

correlation between traffic patterns and NFLC

accuracy. Our analysis indicated that such a

correlation exists. Internet traffic aggregates (time

series) can be stationary or chaotic (unstable), and

stationary traffic is either SRD (short-range-

dependence) or LRD (long-range dependence). SRD

includes Markovian traffic time series and LRD has

self-similar and heavy-tailed patterns. As observed

from the “T-WTS + NFLC” simulations, each traffic

pattern could produce distinctive ill effect on NFLC

control accuracy and convergence smoothness to the

steady state. Therefore, the next logical step for the

research is to explore and establish the correlation

between traffic patterns and their ill effects (e.g.

oscillatory convergence).

ACKNOWLEDGEMENTS

The authors thank the Hong Kong Polytechnic

University for the A-PA9H research grant.

APPLYING NEURO-FUZZY DYNAMIC BUFFER TUNING TO MAKE WEB-BASED TELEMEDICINE

SUCCESSFUL

155

REFERENCES

H. Jamjoom, P. Pillai and K.G. Shin, Resynchronization

and Controllability of Bursty Service Requests,

IEEE/ACM Transactions on Networking, 14(4),

August 2004, 582- 594

J.F. Kaar, International Legal Issues Confronting

Telehealth Care, Telemedicine Journal, March 1999

Lin, The Wireless LAN Traces, Department of

Computing, Hong Kong Polytechnic University,

http://www4.comp.polyu.edu.hk/~cswklin/research/tra

ces/wireless/

Wilfred W.K. Lin, Allan K.Y. Wong and Tharam S.

Dillon, Application of Soft Computing Techniques to

Adaptive User Buffer Overflow Control on the

Internet, IEEE Transactions on Systems, Man and

Cybernetics, Part C, 36(3), 2006, 397-410

Wilfred W.K. Lin and Allan K.Y. Wong, Tharam S.

Dillon, Elizabeth Chang, Detection of Fractal

Breakdowns by the Novel Real-Time Pattern

Detection Model (Enhanced-RTPD+Holder Exponent)

for Web Applications, Proc. 10

IEEE Int’l

Symposium on Object and Component-Oriented Real-

Time Distributed Computing, May 2007, 79 - 86

C.A. Patterson, R.R. Muntz and C.M. Pancake, Challenges

in Location-aware Computing, IEEE Pervasive

Computing, 2(2), April-June 2003, 80-89

Braden et.al., Recommendations on Queue Management

and Congestion Avoidance in the Internet, RFC 2309,

April 1998

D. Wang, Allan K.Y. Wong, and Tharam S. Dillon,

Heuristic Rule Based Neuro-Fuzzy Approach for

Adaptive Buffer Management for Internet-based

Computing, FUZZ-IEEE, 2001 .

WINSYS 2008 - International Conference on Wireless Information Networks and Systems

156