Using BK-Products of Fuzzy Relations in Quality of Service

Adaptive Communication
Ahmed S. Moussa and Ladislav J. Kohout
Department of Computer Science
Florida State University
Tallahassee, FL 32306-4530, USA
[email protected], [email protected]
I

Abstract constraints. It can be Link constraints, Path

The term QoS is a fuzzy term. When applied to data constraints, Tree constraints, or any other pre-defmed
communication, the measures of the QoS are also constraints. Dealing with such QoS requirements and
fuzzy linguistic measures. Hence, fuzzy logic ought to the resulting problems is a multi-disciplinary field. A
be exploited as a tool for modeling the communication variety of problems need to be dealt with by scientists
system and solving its problems. In this paper we from diverse areas, such as physics, hardware
provide a foundation and overview of the possible engineering, and computer science. In physics,
utilization of fuzzy logic in the area of QoS data research has been running on discovering new media
communication. We start by an introduction to the through which data can be transrnitted as well as trying
subject, which provides justification of the need for to enhance the use of the currently available media.
fuzzy logic to be utilized in this area. We follow this Engineers have been working on using the currently
by a review of the previous approaches for solving the available physics to design the best possible network
problems of QoS data routing in general and the in terms of speed and reliability. Meanwhile, computer
problems associated with the inaccuracy in the scientists have been developing algorithms and
network state information in particular. We follow this architectures that optimally utilize the currently
by presenting our proposed approach, which provides available technology.
the foundation for building a new paradigm for Speaking of the computer science solutions to the
modeling and solving QoS computation problems. We QoS networking, we noticed the absence of utilizing
conclude by pointing out several new avenues for fuzzy logic and soft computing as tools for developing
research and list some open problems. solutions to QoS problems. We cannot conclusively
1. Introduction say that no work has been done in that field, but upon
The term Quality of Service (QoS) has been the review of many research papers in the area of QoS,
introduced to the field of data routing/communication none in this direction has been found.
in computer science only in the past few years. The Yet, the field of QoS is a natural area for applying
concept of QoS has been proposed to capture the fuzzy systems and soft computing techniques. The
qualitatively and/or quantitatively defined following points provide the justification for our
performance contract between the service provider and claim:
the user applications [6]. These qualitative and l-The term QoS itself is a fuzzy term. We do not have
quantitative metrics may not be of importance in some a precise generalized specification of the term QoS.
applications but may be critical to others. For example, 2-We usually describe the quality of service as:
when we FTP a file, quality of service data acceptable, not acceptable, good, bad, . . . etc., which
transmission is not so important; the issue is just are obviously all fuzzy linguistic terms, the type of
whether the file transfer is successfully completed or measures fuzzy logic was invented to deal with.
not. On the other hand, transferring or broadcasting a 3-A number of research papers have been published
transaction in a distributed database system may about problems such as dealing with imprecise state
require a minimum speed and a certain order to be information [5] and inaccurate information [9].
considered successful. Furthermore, when streaming Those problems are predominantly subtle and
video or implementing video/audio real-time important problems in large-scale network
communication, not only the speed but also the communications and are inherently fuzzy logic
synchronization of visual and audio data is essential to problems.
the success of data communication. 2. Previous approaches
Therefore, QoS must be considered when building QoS data communication and its metrics have
any communication system in order for that system to always been specified in the literature as a set (or sets)
meet the requirements of an integrated service network of constraints. Even though the QoS metrics have been
that carry heterogeneous data traffic. The QoS described in the fuzzy linguistic terms mentioned
requirement of a connection is usually given as a set of above, they have not been modeled as such. The

0-7803-7078-3/01/$10.00 (C)2001 IEEE. Page: 681

 quality of service and its metrics have only been l A name to capture the meaning of the base variable
modeled as constraints that can be satisfied or involved.
dissatisfied. However, some attempts have been done l A base variable with its range of values (an interval
to deal with the imprecise and inaccurate state of real numbers).
information when taking the routing decisions. Guerin l A set of linguistic terms that refer to values (or
and Orda [9] published a paper on QoS-based routing states) of the base variable.
in networks with inaccurate information. Their l A semantic rule, which assigns to each linguistic
approach was based on the probability distribution term its meaning.
functions. To deal with the inaccuracy, for each link, In this context, our model is exemplified in Figure
the source node only knows the quantities pZ(x), where (1). The name of the linguistic variable is “QoS”. The
pZ(x) is the probability that link I can accommodate a base variable whose meaning is captured by QoS is
connection which requires x units of bandwidth. ranged from 0 to 100 as a percentage of being
Similarly, the source node maintains, for every link Z, satisfactory. There may be a value, possibly below
the quantities pZ(d) where pZ(d) is the probability that 100, in which no communication requirement may not
link Z have a delay of d units, where d ranges from be satisfied. This value would obviously be system and
zero to maximum possible value. The goal is to problem dependent. In our example, the value is 80.
maximize the probability of finding a feasible path. The linguistic values (states) of the linguistic variable
However, they did not address the problem of how to are very poor, poor, acceptable, good, and very good.
maintain the probability distribution. Each of these linguistic terms is assigned one of the
Another research by Chen and Nahrstedt [5] was trapezoidal-shape fuzzy numbers by specified
published on distributed QoS routing with imprecise semantic rules.
state information. They developed a multi-path
distributed routing scheme, called ticket-based
probing. They claim that their algorithm tolerates
information imprecision by searching multiple paths in
parallel, increasing the chance of finding a feasible
path. We believe that they developed a good scheme,
which considers the optimality of the selected path as
well as the QoS requirement. However, we do not
believe that searching multiple paths in parallel is a
proper solution to the problem of imprecise state
information. The imprecision will still be there,
affecting all the paths being searched in parallel.
3. Our Approach
As explained above, the term QoS is a fuzzy term 10 20 30 40 50 60 70 80 90 100
and its metrics are determined in terms of linguistic FIGURE (1): An exampleof our systemmodel
fuzzy measures and are often subjective. Hence, we
believe that fuzzy set theory, and its consequent fuzzy In this regard, it is important to note that the fuzzy
relations and BK-products, is an ideal tool for dealing number functions do not have to be trapezoidal. In
with this type of computation. As stated by Klir in fact, we believe that this could be one of the open
[lo], “An important feature of fuzzy set theory is its research areas in the utilization of fuzzy logic in QoS
capability to capture the vagueness of linguistic terms networking. It could also be application dependent.
in statements of natural language”. Consequently, we The main system modeled above is an aggregation
propose a new paradigm for modeling and solving the of several subsystems that can be modeled the same
QoS computation problems by utilizing fuzzy logic. way. For example, the delay requirement, which is part
In our proposed model, a QoS data communication of the QoS requirements, can be specified as in Figure
system is modeled as a system whose variables range (2). The name of the linguistic variable is “The
over states. Those states are represented with fuzzy Delay”. The linguistic values (states) of the linguistic
numbers. These fuzzy numbers may represent the variable are very good, good, acceptable, bad, and not
linguistic variables that provide the QoS metrics, such acceptable. The fuzzy number represents the degree of
as very poor, poor, acceptable, good, very good, truth in the statement: The delay is variable (where
excellent, and so on. Each of those linguistic variables variable in this example can be very good, good,
is defined in terms of a base variable whose values are acceptable, bad, and not acceptable, respectively).
real numbers within a specific range. As specified by Therefore, in the example of figure (2), a delay of 8
Klir in [lo], each linguistic variable consists of: nanoseconds is considered to be 75% very good, 25%

0-7803-7078-3/01/$10.00 (C)2001 IEEE. Page: 682

 good, and 0% otherwise. A delay of 20 nanoseconds is For example, if certain application is only bandwidth
100% good and 0% otherwise. requirement dependent, then we can only model the
bandwidth and compute the QoS as the outcome of the
IDelay bandwidth computation without further aggregation of
other factors. However, if a path change needs to be
computed then the triangle products (R D R) and
(R a R) of the two paths must be computed in order
to guarantee the inclusion of the parameters of the
replaced path in the newly selected one. Finally, when
certain numerical guarantees are required, the
computation can be done on the a-cut at the required
value.
4. Justification of validity
10 20 30 40 50 60 70 80 90 100 We believe that adapting this paradigm will be an
FIGURE (I): An example of our systemmodel ideal approach for the solution of the problem of QoS
computation for the following reasons:
The same technique (possibly with different fuzzy 1. Representing the states of the variables by fuzzy
functions) can be used to model other components of numbers provides good and realistic quantization.
the system such as bandwidth, cost, . . . etc. When the
This quantization is capable of capturing the limited
various components have been modeled and the
resolution of measuring the system features. In
corresponding computations have been done, the
addition, it also captures the error and imprecision in
system can be aggregated using the BK-products [ 141,
the network state information by avoiding the
[ 151 to compute the QoS available. Let us now
computation on single values or intervals with sharp
describe a concrete example of the computation of a
divisions between them, which are known to be
routing decision using BK-products to aggregate the
unrealistic.
system. For simplicity, and without loss of generality,
2. It is very difficult, if not impossible, to maintain and
we assume the above scheme has been used to model update a precise global network state like most
delay, bandwidth, and cost. Furthermore, we assume
routing algorithms require [4, 16, 18, 191. The
that each parameter has been computed for an entire global state will always be imprecise in a dynamic
path and only three paths are available. Let the three
network, and will even be worse in large wide-area
paths be denoted I, J, and K. The fuzzy sets specifying
networks [5]. The sources of imprecision can be one
the delay (D), bandwidth (B), and cost (C) will be
or more of the following: propagation delay of state
defined by the system designer. The level of the
messages, periodic updates (more updates means
service is defined by the contract, for this example we
more overhead), hierarchical state aggregation,
assume the service is required to be good. The
reliability of nodes, and imprecision in the links.
membership grades of each path in the three sets
This shows that the imprecision will be always
express the degree of the path being able to provide
present and our best bet is to live with it. If the
the service with the level good. We encode all the imprecision will always be there, why then do the
available network information in the following relation
computation assuming precise state information?
3. Other than the forced quantization due to the lack of
accurate information, and since the level of accuracy
‘4 will always be limited, we can utilize the opted
quantization, which is desirable when lower
precision is sufficient for a given task [ 11, 121.
To compute the routing decision in this case, we Therefore, performing the computation on the
aggregate the system by computing the quantized variable states significantly reduces the
relation (R 0 R), = v y (& A R, ) . The fuzzy sets computation demands, which in turn frees the
intersection is needed to guarantee the minimum system for more optimized solutions. The resulting
satisfactory system parameter. Then, the path with drop in computational cost is badly needed in the
highest degree of overall performance is the one to be field of QoS networking. For example, finding a
selected, in this case path J. In a 2-way synchronized feasible path with two independent path constraints
traffic that requires two paths with equivalent states, is NP-complete [8]. Other examples are the problem
of finding the least-cost path with bounded delay
the square product ( R tl R ) would be needed.
and the problem of finding a path with both bounded
In some cases, aggregation may not be even needed,
delay and bounded delay jitter. These are examples
if only one QoS requirement needed to be satisfied.

0-7803-7078-3/01/$10.00 (C)2001 IEEE. Page: 683

 of two classes of problems, namely, Path- ordered set, which provides the scale for
Constrained Path-Optimization routing (PCPO) and measurement of delay, hence computational saving
Multi-Path Constrained routing (MPC), respectively. is achieved. In the examples of Figures (1) and (2),
The entire two classes are NP-complete [6]. This we only have five overlapping equivalence classes
shows how badly reducing the computational cost is to deal with.
needed here. It is hoped that utilizing our proposed 5. Fuzzy control
paradigm would lead eventually to the reduction of Beyond the scope of the classes of problems and
the complexity of some of these problems. However, points of strength presented above, other research has
even if the level of complexity remains the same, the been done on designing new architectures for QoS-
use of quantized states of variables will certainly adaptive communication such as the one presented by
reduce the computation, which spares more Abdelzaher and Shin in [ 11. In their paper, Abdelzaher
computational power for dealing with more complex and Shin present a new communication subsystem
problems. architecture for QoS-adaptive real-time applications.
4. When specifying the QoS requirements and metrics, The main contribution of that architecture is its
users always defme the service in linguistic dynamic QoS-optimization mechanism and support for
measures such as good, bad, . . . etc. Translating flexible QoS contracts. Those QoS contracts between
these linguistic measures into precise numbers the communication client and server specify multiple
without mzzification may unnecessarily reduce the levels of QoS with acceptable degradation mechanism.
system effectiveness and waste considerable Under overload, the communication subsystem
computations. For example, suppose we specify one degrades connections predictably according to their
QoS constraint such as the delay to be, say 9 respective contracts. The contracts also associate
microseconds. If the system performs the necessary rewards with each level. The communication
computations and reaches the conclusion that the subsystem consists of three basic modules: Load
best path available has a delay of 9.0 1 Control, Protocol Processing, and System monitoring.
microseconds, it is likely that the system will reject We propose the use of fuzzy control to enhance this
the connection even though the 9.01 microseconds system. Fuzzy control has been the most successful
could be satisfactory. The problem is that the area of applications of fuzzy set theory by far [ 121 and
constraint had been precisely specified as a smaller our proposed fuzzy controller will have several
value and the model used does not make room for advantages: l- The use of a fuzzy controller should be
the jitter. These jitters can be captured and dealt ideal for dealing with subjective and objective QoS
with by our paradigm. levels and rewards. 2- The multi-level control
5. In addition to the previous point, if a precise, implemented by the load control module requires
nonnegotiable, value is ever required as a constraint, smooth and controlled transitions from one level to
our proposed model can still deal with it by taking another. The capability of fuzzy sets to capture gradual
the a-cut at this value, returning to the crisp transitions from one set to another, and even from a set
relation’s computation. Hence, we propose a more to its own complement, considerably enhances their
flexible paradigm that could deal with different expressive power in comparison with crisp sets.
degrees of satisfaction of requirements, which can Therefore, fuzzy set theory can effectively deal with
still be turned to the known crisp model whenever this class of problems. 3- The dynamic optimization
needed. This makes the proposed paradigm a using the monitoring and feedback module incurs three
superset of the currently known one. types of difficulties: additional complexity and
6. After all, a main contribution of our approach is the computational costs, the previously mentioned smooth
substantial computational savings that are achieved transition between states, and the need for a bridge
by developing equivalence classes of QoS measures between the mathematical model and the associated
[3]. Therefore, the computation can be done on physical reality. These are three areas that Klir
equivalence classes rather than individual values. specifically addressed in [lo] when discussing the
Let us take delay as an example, if the delay is capabilities of fuzzy set theory and fuzzy logic. 4- Our
considered to be good when it is between 5 and 9 proposed fuzzy controller is not designed to only
nanoseconds, then all the values falling between handle the degradation of QoS levels. In stead, it is
these two values can be dealt with as an equivalent designed to also upgrade the levels whenever possible.
class. Changes within this range should not result in Hence, maintaining optimal performance.
the re-computation of information for the whole We propose a fuzzy network controller that does not
system. Hence, computation can be done only when only take the current network state into account, but
one QoS measure changes status from equivalence also generates and maintains a curve representing the
class to another. So, by employing fuzzy history of the difference between the target and current
equivalence classes we reduce the cardinality of the states of the underlying network. Therefore, by

0-7803-7078-3/01/$10.00 (C)2001 IEEE. Page: 684

 computing the derivative of the curve at some point, represented by the solid line and dE/dt is represented
we can predict system changes, overloads, and under-
utilization before they occur. This results in a
continuously optimized network. Here we present a
simple network fuzzy control system for resource
management. The system uses the network parameters
recorded periodically and on demand to generate a
curve for the corresponding parameter, the bandwidth
for example. The data is subtracted from the target to
determine the error. Then the derivative with respect to -10 -5 -10 5 10 --10 -5 o+l 5 10
time can be computed in order to determine the slope Figure (3): Error Figure (4): Error curve
(rate of change) of the error. The fuzzy controller’ by the gray line.
outputs are used to start the process of initiating Now, let the inputs to the fuzzy controller be as
further resource reservation if the fuzzy engine decides depicted in Figures (3) and (4), E = -1 (this means
that overload is present (or projected) or start the the system is 25% overloaded and 75% optimal) and
process of releasing some resources for other agents of
the network if the fuzzy controller decides that under- dE/dt = +l (this means the system is getting
utilization is present (or projected). The controller may overloaded with a rate of 25% but is still stable with a
also keep the current network state unchanged if the degree of 75%). Then, we apply rules number 4, 5, 7,
system is in an optimum performance. Let E be the and 8. By Rule 4, OUTUT = Rq with degree of
error, T be the target, and C be the current state value. membership =min(O.25,0.75) = 0.25. By Rule 5,
Then, E = T - C. We have the following three cases: OUTPUT = NC with degree of membership
E > 0 z system under-utilization, E < 0 a system =min (0.75,0.75) = 0.75. By Rule 7, OUTPUT =
overload, and E = 0 => optimum performance. with degree of membership =
Rq
Furthermore, we compute dE/dt which also results min(O.25,0.25) = 0.25. By Rule 8, OUTPUT =
in the following three cases: dE/dt > 0 z system is R9 with degree of membership
getting overloaded, dE/dt < 0 a system is getting
=min(0.75,0.25) = 0.25 . The next step is to
perform the defuzzification in order to convert the
under-utilized, and dE/dt = 0 3 system is stable. overall conclusion into a real number. The most
The next step is to define the fuzzy inference rules in common defuzzification methods are the Center of
which the correct action will be based upon. Let Rel = Gravity Defuzzification (CGD) and the Root Sum
Release Resource, Rq = Request Resource, and NC = Square (RSS). According to CGD, given a fuzzy set A
No Change. Then, the rule base is as follows: defined on the interval [al, 4,
Rule 1:If E < 0 and dE/dt < 0 then OUTPUT = Rq.
CGD = I21 xA(x)G!x/ I2 A(X)&. According to
Rule 2:If E = 0 and dE/dt < 0 then OUTPUT = Rel.
RSS, for outputs 01, 02, and 03,
Rule 3:If E > 0 and dE/dt < 0 then OUTPUT = Rel.
RSS=d o12 +022 +032 . Since we have few
Rule 4:If E < 0 and dE/dt = 0 then OUTPUT = Rq. discrete values here, we suggest the use of RSS. Then
Rule 5:If E = 0 and dE/dt = 0 then OUTPUT = NC.
we have: Rq = dO.252 + 0.252 + 0.252 = 0.433
Rule 6: If E > 0 and dE/dt = 0 then OUTPUT = Rel.
(by rules 4, 7, and 8) and NC = J0.7s2 = 0.75 (by
Rule 7:If E < 0 and dE/dt > 0 then OUTPUT = Rq. rule 5). Then the final output is computed as follows:
Rule 8:If E = 0 and dE/dt > 0 then OUTPUT = Rq. -100xRq+OxNC+100xRe~
output =
Rule 9:If E > 0 and dE/dt > 0 then OUTPUT = Rel. Rq+NC+ReZ =
Figure (3) and Figure (4) provide examples of - 100 x 0.433 + 0.0 x 0.75 + 100 x 0.0
possible membership functions for error and error = -36.60/0.
curve, respectively, of the good bandwidth with 0.433 + 0.75 + 0.0
hypothetical numerical values. In Figure (3), E is Then, the conclusion is that the system is in overload
state and needs to request 36.6% of the allowed
represented by the solid line and E is represented by bandwidth.
the gray line. Similarly, in Figure (4), dE/dt is 6. Conclusion and Pointers to Possible Future
Research

0-7803-7078-3/01/$10.00 (C)2001 IEEE. Page: 685

 Fuzzy systems and soft computing possess far International Conference on Computer and
greater capabilities than the techniques currently used Communication Networks, October 1998.
to capture the notion of QoS data communications. As [6] S. Chen and K. Nahrstedt. An Overview of
a consequence, their use improves the bridge between Quality-of-Service Routing for the Next
mathematical models and the associated physical Generation High-Speed Networks: Problems and
reality, which would result in more realistic QoS data Solutions. IEEE Network. Special Issue on
routing systems. In addition, they are expected to Transmission and Distribution of Digital Video.
result in more efficient systems by dramatically November/December, 1998.
reducing the computational cost. The resulting system [7] I. Cidon, R. Rom, and Y. Shavitt. Multi-Path
is expected also to be more flexible since it enables Routing combined with Resource Reservation.
flexible measures, and possible negotiations, between IEEE INFOCOM, 1997.
the user application and service provider. [8] M. Garey and D. Johnosn. Computers and
Fuzzy control can also be used in the design of Intractability: A Guide to the Theory of NP-
system controllers for QoS communication. It provides Completeness. Freeman and Co., 1979.
all the features that a system designer in this field [9] R. Guerin and A. Orda. QoS-based Routing in
would be looking for. Besides, it is the most Networks with Inaccurate Information: Theory
successfully established area of application of fuzzy and Algorithm. IEEE INFOCOM, 1997
systems. Hence, no risky exploration is involved here. [lo] G. J. Klir. Fuzzy Logic: Its meaning and
Finally, after starting with this foundational paper, Signzjicance. Department of Systems Science and
we point out several areas of research that can be Industrial Engineering, Watson School of
explored. First, a fuzzy controller can be designed and Engineering and Applied Science, State
used to enhance the architecture proposed in [l] and University of New York, Binghamton, New York,
the results can be compared with the original system. USA.
Second, system requirements such as delay, [ 1l] G. J. Klir and B. Yuan. Fuzzy Sets and Fuzzy
bandwidth, cost, and other requirements can be Logic: Theory and Applications. Prentice Hall,
fuzzified for specitic computation. The computation 1995.
can be done on the fuzzified data and then defuzzify [ 121G. J. Klir, U. H. St.Clair, and Bo Yuan. Fuzzy Set
the result to take the routing decision. The resulting Theory: Foundation and Applications. Prentice
system can be tested for evaluation of several points of Hall, 1997.
criteria such as performance and rate of successful [ 131L. Kohout. A Perspective on Intelligent Systems:
routing. Last but not least is a research on different A Framework for Analysis and Design. Chapman
fuzzy number functions such as trapezoidal, triangle, and Hall Computing, 1990.
bell curve, . . . etc. Some experimentation can be done [ 141L. J.Kohout. Boolean andfuzzy relations. In P.M.
to verify whether one of them is superior to the others Pardalos and C.A. Floudas, editors, The
in all cases or if the choice of the function is Encyclopedia of Optimization. Kluwer, Boston,
application dependent. If it turns out to be application 2001. In press.
dependent, then further work can be done to develop a [ 151L.J. Kohout and W. Bandler. Fuzzy relational
scheme for selecting the fuzzy membership function. products in knowledge engineering. In V. Novak
7. References et al., editor, Fuzzy Approach to Reasoning and
[l] T. Abdelzaher and K Shin. End-host Architecture Decision Making,. Academia and Kluwer, Prague
for QoS-Adaptive Communication. IEEE Real- and Dordrecht, 1992, pp 5 l-66.
time Technology and Applications Symposium, [ 161Q. Ma and P. Steenkiste. Quality of Service
Denver, Colorado, USA, June 1998. Routing with Performance Guarantees.
[2] W. Bandler and L.J. Kohout. Relations, Proceedings of the 4* International IFIP
mathematical. In M.G. Sigh, editor, Systems and Workshop on Quality of Service, May 1997.
Control Encyclopedia,. Pergamon Press, Oxford, [ 171C. Qiao and M. Yoo. Optical Burst Switching
1987, pp 4000-4008. (OBS) - A New Paradigm for an Optical Internet.
[3] W. Bandler and L.J. Kohout. Specialproperties, LANDER, University at Buffalo, Buffalo, New
closures and interiors of crisp andfuzzy relations. York, USA.
Fuzzy sets and Systems, 26(3): 317-332, June [ 181H. F. Salama, D. S. Reeves, and Y. Viniotis. A
1998. Distributed Algorithm for Delay-constrained
[4] S. Chen and K. Nahrstedt. On Finding Multi- Unicast Routing. INFOCOM 97, Japan, April
constrainedpaths. IEEE International Conference 1997.
on Communications, June 1998. [ 191Z. Wang and J. Crowcroft. QoS Routing for
[5] S. Chen and K. Nahrstedt. Distributed QoS Supporting Resource Reservation. JSAC,
Routing with Imprecise State Information. September 1996.

0-7803-7078-3/01/$10.00 (C)2001 IEEE. Page: 686