Fuzzy fundlmentals d

This orderly desgn subject to interpretation by the system

designer.
model of the process does not exist, or ex-
ists but is too difficult to encode, or is too
procedure can save time Thus the variable “temperature’ ’ might complex to be evaluated fast enough for real-
have a range of states: cold, cool, moder- time operation, or involves too much mem-
and help prevent problems ate, warm, hot, very hot. Yet the change ory on the designated chip architecture;
in the development of from one state to another is not precisely de-
fined. At no point can an increase of a tenth
when high ambient noise levels must be
dealt with or it is important to use inexpen-
fuzzy logic systems of a degree be said to change “this is warm” sive sensors and/or low-precisionmicrocon-
into “this is hot.” Consequently, the idea trollers (since fuzzy logic’s interpolation
uzzy logic has rapidly be- of what is cold, what is warm, and what is capabilities work well with 4-and 8-bit con-
come one of the most suc- hot is subject to different interpretations by trollers); and perhaps above all, when an ex-
cessful of today’s technol- different experts at different points in the pert is available who can specify the rules
ogies for developing variable’s domain. underlying the system behavior and the
sophisticated control sys- This subjectivityhas profound implications fuzzy sets that represent the characteristics
tems. With its aid, complex for continuous system modeling, and is at the of each variable.
requirements may be im- heart of the power and flexibility of fuzzy FUZZY MODEL. The components of conven-
plemented in amazingly logic. With fuzzy logic, control statements tional and fuzzy systems are quite alike,
simple, easily maintained, and inexpensive are written in terms of these imprecise ideas differing mainly in that fuzzy systems con-
controllers. The same fuzzy technology, in of what constitutes the states of the varia- tain “fuzzifiers,” which convert inputs into
the form of approximate reasoning, is also ble. As an example, a fuzzy rule in an anti- their fuzzy representations, and ‘ ‘defuzzi-
resurfacing in information technology, where lock braking system might be: fiers, ” which convert the output of the fuzzy
it provides decision-support and expert sys- ,If brake temperature is Warm i w D process logic into “crisp” (numerically pre-
tems with powerful reasoning capabilities speed is Not very fast, then brake pressure cise) solution variables [Fig. 11.
bound by a minimum of rules. is Slightly decreased, In a fuzzy system, the values of a fuzzi-
Of course, fuzzy logic is not the best ap- while a rule in a conventional proportional- fied input execute all the rules in the knowl-
proach for every control problem. As design- integral-derivative (PID) controller would edge repository that have the fuzzified input
ers look at its power and expressiveness, need to be very specific: as part of their premise. This process gener-
they must decide where to apply it, and also ,Ifbrake temperature is greater than 280 ates a new fuzzy set representing each out-
how best to manage software projects based AND speed is less than 45, then brake pres- put or solution variable. Defuzzification cre-
on this new technology. Increasingly, proj- sure is BO. ates a value for the output variable from that
ect managers, system architects, and design Not only is the fuzzy logic rule more nat- new fuzzy set,
engineers are asking the tough questions: ural, and technically more expressive, but For physical systems, the output value is
what types of projects can benefit from the it will fire over a wider range of brake tem- often used to adjust the setting of an actua-
use of fuzzy logic? How do I control projects peratures and car speeds. The degree to tor that in turn adjusts the state of the phys-
with this technology? What is the best way which the variables are considered Warm ical system. The change is picked up by the
to design, develop, deliver, and test systems or Not very fast is the degree to which the sensors, and the entire process starts again.
using fuzzy logic and its related tech- In contrast, a PID controller is based
nologies? on a rigorous mathematical model of
FIT TO BE FUZZY. Fuzzy logic is a meth- some linear process [Fig. 1, bottom].
od of easily representing analog pro- One fuzzy rule These models develop, by means of
cesses on a digital computer. These root locus or some other method, a set
processes are concerned with continu- can replace many of equations that describe the stable
ous phenomena that are not easily bro-
ken down into discrete segments, and -often very many- equilibriumstate of the control surface,
with coefficients being assigned to the
the concepts involved are difficult to
model-sometimes extraordinarily conventional rules proportional, integral, and derivative
aspects of the system. A PID controller
so-along mathematical or rule-based reads a precise sensor value, applies
lines. the mathematical model, and produces
An example might be an anti-lock braking brake pressure is relaxed. Thus one fuzzy a specific output from the mathematical al-
system for an automobile. The control rules rule can replace many-often very many- gorithm.
for a system of this nature might include such conventional rules. And since fuzzy logic While the PID model may seem the sim-
variables as the car’s speed, the brake pres- creates a control surface by combining rules pler and hence, the more economical
sure, the brake temperature, the interval be- and fuzzy sets, it allows designers to build representation, the contrary is more often
tween applications of the brakes, and the controllers even when their understanding true. Fuzzy controllers are in fact easier to
angle of the car’s lateral motion relative to of the mathematical behavior of the system prototype and implement, simpler to de-
its forward motion. These variables are all is incomplete. scribe and venfy, and can be maintained and
continuous, and the range of their values So when is it appropriate to use fuzzy extended with greater accuracy in less time.
~

logic? When one or more of the control vari- Moreover, because of their reliance on rules
Earl Cox Metus Systems Group ables are continuous: when a mathematical and knowledge, they give their environ-

58 ooi8-9235/92/$3.0001992 IEEE IEEE SPECTRUM OCTOBER 1992

 [l]In a &hcalfuzzysys-
tem [top lg81,the input is
’
read from an extern1 k
source andtdfuzziFed be-
fore being processed by E
the fuzzy logic. The out-
put of theprocess lo& is g
defuzztjW before being 0”
sent to the extemalphys- 11o 220 330
ical system under con- Temperature, “C
trol. A ~ o t t hcon-
trolsystem[boftomle$]
l
’
has a very similar over- f
all shucture, but without
thefuzzyelements. E
c

[Z] Both the rnputs (tem-

perature and pressure) 8
and the output (throttle Oleo * * 1200 2300
action) are defined as Pressure Pa
fuzzy regions in a fuzzy NL NM NS ZR PS PM NL
logic system. For the a 1
throttle action variable, $
the regionsare: negative 4
large (NL), negative E
medium(NM),neg;ative
small WS), zero (ZR), a,
positive small ( P S ) , g
positive medium (PM), 2
and positive large (PL). c600
Throttle action rnmls
LOW

ments what Lbtfi Zadeh (the father of fuzzy

logic) calls a higher “machine intelligence Q

quotient. ’’
STEAM TURBINE. The architecture of a sim-
-
m
rn
L

a,
Q

a,
ple fuzzy controller for a steam turbine E I ,
shows how this process works. The design 8 0 0
100
of the turbine’s throttle action is based on p (‘0)
Pressure, Pa
two control variables: temperam and pres-
sure. In order to build a controller that
represents the relationship between the in-
puts and the output, each control variable
must f i t be decomposed into a set of con- 400 * o +Eo0
Throttle action. mm/s
trol regions Fig. 2, top and middle] and the E
0
a,
a,
Deflning terms [3] Premises connected by an AND are
combined by taking the degree of mem-
C ~ i I mrrrlrllr:
l a variable that appears in the T(t0) bership of the lesser of the tw as the
premise of a rule and therefore controls the state Temperature, “C value of the combination [top]. Ib com-
of the solution variable. bine rules, OR them together by taking
DefnzzRcrtlo~:the process of converting an out- the larger of the two values as the value
put fuzzy set for a solution miable into a single value of the combination at each point on the
that can be used as an output. horiwntal azis [bottom].
Fuzzy set a set that allows partial membership
states. Ordinary, or crisp, sets have only two mem.
bership states: inclusion and exclusion; fuzzy states
allow degrees of membership as well.
Ororlam: the degree to which the domain of one
fuzzy set overlaps with that of another. As an exam- Pressure, Pa
ple, the right-handedge of a Cool fuzzy set may over-
lap the left-hand edge of a fuzzy set called Warm,
so that some temperatures belong in both sets.
StlnIlo~frnl sol: a temporary fuzzy set created 0
by the fuzzy model to resolve the value of a COP hronle action mrn
respondingsolution variable. When all the rules have
been fired, the solution fuzzy set is defuuified into
the actual solution variable.
Salrtlrn rrrlrllr: the variable whose value the
fuzzy logic Wtem is meant to find.

Cox-Fuzzy fundamentals BEST COPY AVAILABLE 59

-- ._-
 horizontal axis.) In other words, if the re-
i Define model's Define post-model
functional + normalization gion is not empty, then OR the outputs to-
and data flow - gether.
[4] Most successful operational & (inter-system Returning to the steam turbine example,
fuzzv logic develop- characteristics connects) vis-a-vis rule N0.2 (If temperature is Cool
ment projects follow ANDpressure is Low, then throttle action is
an iterative develop- PM),it may be seen that T has a 0.48 de-

1
ment cycle like this. Define control gree of membership in Cool and P has a 0.57
ThepinkboxeS@t~- surfaces degree of membership in Low. In accor-
sent the four central
1 (fuzzy sets)
dance with the f i t step of the procedure,
I
steps in the design therefore, the lesser of the two figures-
process. namely, 0.484s taken as the truth level of
the premise. Then, in accordance with the
-4 Define behavior of
control surfaces Define
limits
acceptable
for ~
second step, the level of the output of rule
No. 2-namely, PM-is truncated at that
level, and copied to the output variable fuzzy
compatibility set [Fig. 3, top].
index, stability, When rule No. 3 fires, the ZR fuzzy set
Determine method is also truncated at the truth of its premise
of defuuification
for solution (the lesser of 0.25 and 0.48) and then copied
into the output throttle action region. But
since the region is not empty, this modified
fuzzy set is combined with the PM fuzzy set
by taking the maximum of their respective
Run simulation membership grades at each point along the
of system horizontal axis [Fig. 3, bottom].
From this combined region, one of the
several techniques of defuzzificationis ap-

.
plied to produce an expected value for the
throttle action. In this case, using the cen-
validate system
Tuneand '.i
Connect to
production
t -> ter of gravity, or centroid, of the combined
region, a value of +150 mm/s is produced.
This value is used to adjust the throttle.
After that, the pressure and temperature
output or solution variable then redefined following three-step procedure shows how: sensors will make new measurements, start-
into a set of fuzzy regions [Fig. 2, bottom]. For each premise expression connected by ing the cycle over again.
When the fuzzy sets have been defined, an AND, take the minimum of the truth of METHOOOLOEIES. Although the previous ex-
the conceptual model is completed by writ- the expressions as the truth level of the ample illustrated how fuzzy systems work,
ing the rules that describe the action taken premise. it gave few hints on how to specify, develop,
on each combination of control variables. Truncate the output fuzzy set being built and deliver one. Yet it is important to have
Some of these rules might appear as: at the truth level of the premise. a good design methodology for fuzzy sys-
1. Iftemperature is cool ANDHessure is Copy the newly modified fuzzy set into the tems if only because they are new to most
Weak, then throttle action is PL. output variable's fuzzy set. If that region is designers, who therefore have no hoard of
2 . If temperate is c o o l AND pressure is not empty, combine it with the current con- experience on which to rely for guidance.
Low, then throttle action is PM. tents by taking the maximum of the new Fuzzy models, be they employed in pro-
3. If temperature is Cool ANDpressure is fuzzy region and the currently existing fuzzy cess control or information technology, tend
OK, then throttle action is ZR,. region at each point in the domain (along the to follow the same applicationdevelopment
4. If temperature is Cool ANDpressure is
Strong, then throttle action is NM.
What is needed to complete this example Controlisolution Temperature
is a means for converting these input fuzzy
sets and rules into an output fuzzy set, and variable
then into a crisp output for controlling the
throttle. Note that a fuzzy model is, in a
sense, a parallel processor. All the rules that
have any truth in their premises will fire and
contribute to the output fuzzy set-the one
that will represent the throttle action con- Labels
trol variable in the steam turbine example. (fuzzy sets)
Suppose that at given time, t,the sys-
tem sensors determine the turbine pressure
to be P(t.,) and the temperature to be T(tJ [5] The relationship betmn variables, labels,
[Fig. 31. As the figure shows, T(L) falls into and fuzzy sets is simple but important. Obvi-
a single region of the T variable-namely ously, the label names should be unique and
Cool. But P(tJ has degrees of member- semantically related to the concepts that each
ship in two regions of its fuzzyset-Low and represents.
OK. This combination causes rules No. 2 Fuzzy set
and No. 3 to fire. representation
The two rules have somehow to be com-
bined to form a single system output. The Ternperature,"C 1

60 IEEE SPECTRUM OCTOBER 1992

cycle Fig. 41. The methodology attempts Finally, the density of the fuzzy sets maximum, on the other hand, produces a re-
to formalize and structure a procedure in should be highest around the optimal con- sult that is sensitive to the truth produced
which the conceptual design is done on trol point of the system and should thin out by the single rule that has the highest predi-
paper, and the later steps are an iterative as the distance from that point increases. cate truth. By and large, process control ap-
cycle of modeling and simulation, carried out This rule was clearly applied in the throttle plications use centroid, while information-
on a computer using fuzzy logic development action variable of the steam turbine exam- based applications like risk evaluation and
tools, and continued until the model behaves ple, where the labels are bunched together terrain analysis use the composite
as desired. in the middle of the domain but are more maximum.
It cannot be overemphasized that the widely spaced toward the edges. Once the fuzzy model has been construct-
paper portion of the process is critical: un- The third step, “define the behavior of the ed, the process of simulation and protocy-
derstandingthe mechanicsbehind a system’s control surfaces,’’ involves writing the rules cling begins. In this adjunct to the method-
behavior, and identifymg the system dynam- that tie the input values to the output model ology, the model is compared against known
ics in terms of the conventional input-pro- properties. These rules are expressed in an test cases to validate its results. When the
cess-output model is an absolutely essen- English-like language with a syntax like: results are not as desired, changes are made
tial part of fuzzy system design. either to the fuzzy set descriptions or
fOUR STEPS. Four of the steps in the to the mappings encoded in the rules.
cycle illustrated in Fig. 4 are central to Tools are available to help project
this process. In the first, “Define the
model functional and operational
Fuzzy logic fits best managers and system designers evalu-
ate fuzzy models and isolate problems
characteristics,” the goal is to estab-
lish the architectwal characteristics of
when variables are at the fuzzy set or rule level. These
tools measure such factors as the statis-
the system, and also to define the
specific opexatmg properties of the pro-
continuous andlor tical compatibility between the model
and test-bed data, and the stability of
posed fuzzy model. The system is al-
ways described in terms of an
mathematical models the model based on a loss of informa-
tion. In general, though, a methodolo-
input-process-output model. The f i t do not exist I gy will only be as good as the design-
step in designing a fuzzy system follows er’s understanding of the problem.
the analysis techniques used by com- TO PROBE FURTHER. For an excellent in-
mercial application systems analysts and FZf +zzy proposition >, then <firzzy troduction to fuzzy logic, readers may want
intelligent-systems knowledge engineers. proposition >, to consider a US $195 education kit from
The fuzzy system designer’s task lies in where the fuzzy propositions are of the Motorola Inc., which contains a PC-based
defining what information (data points) flows form, “xis Y” or “xis not Y,”x being a introductory course (on both 5.25- and 3.5-
into the system, what basic transformations scalar variable and Y being a fuzzy set as- inch disks), a demonstration version of Ap-
are performed on the data, and what data sociated with that variable. tronix Inc.’s fuzzy inference development
elements are output from the system. Even Such a group of rules forms a fuzzy as- environment (FIDE, pronounced feeday),
if the designer lacks a mathematical model sociative memory. When a set of input and a collection of fuzzy logic freeware with
of the system process, it is essential that she values are read, each of the rules that has documentation. Further information on the
have a deep understanding of these three any truth in its premise will be executed. kit may be obtained by writing Fuzzy logic,
phenomena. This step is also the t h e to de- Since these rules are declarative xather than Box 600, Mail Drop MO, Austin, Texas
fine exactly where the fuzzy subsystem fits procedural, their order in the knowledge 78762, or calling Norma Williams at
into the total system architecture, which pro- repository is unimportant. Nevertheless, in 512-505-8101.
vides a clear picture of how inputs and out- the interests of maintainability, it is recom- Fuzzy Sets, Uncertain& and Information
puts flow to and from the subsystem. mended that the rules be grouped by their by George J. Klir and Tina A. Folger (Pren-
Locating the fuzzy subsystem within the premise control variables. tice Hall, Englewood Cliffs, N.J., 1988) is an
overall system helps the designer estimate Generally, the number of rules a system especially readable and well thought-out
the numbers and ranges of inputs and out- requires is simply related to the number of presentation of fuzzy logic as it relates to in-
puts that will be required. It also reinforces controlvariables. The steam turbine system, formation modeling. Unlike the majority of
the input-process-output design step. And for example, has two control variables, tem- recent books on the subject, it does not
it is usually best done before tackling the de- perature and pressure, each of which is delve deeply into complex mathematics (al-
tails of input and output since the latter may divided into five fuzzy regions. Since there though it certainly covers the mathematics
well be affected by the localization. are a total of 25 possible input combinations, of fuzzy logic) and it is not concerned exclu-
In the second step, ‘ ‘define the control the system requires 25 rules. sively with process control.
surfaces,” each control and solution varia- In some cases, it is possible to use fewer Another excellent text is Neural N e t w k s
ble in the fuzzy model is decomposed into rules, but there are dangers in so doing. The and Fuzzy Systems-A Dynamical Systems
a set of fuzzy regions. These regions are rules represent knowledge, so if any are Approach to Machine Intelligence by B.
given unique names, called labels, within the deleted, knowledge is removed from the Kosko (Prentice Hall, Englewood Cliffs,
domain of the variable. Finally, a fuzzy set system-knowledge that may become im- N.J., 1992), which is especially good on the
that semanticallyrepresents the concept as- portant if the system is later modified. overlapping of the labels of a variable. Kosko
sociated with the label is created [Fig. 51. The fourth of the central steps, “select shows that Aristotelian and Boolean logic are
Some rules of thumb help in defining fuzzy a method of defuzzification,” is the finalpart special cases of fuzzy logic. He also does a
sets. First, the number of labels associated of creating the basic fuzzy model. There are good job of presenting the differences be-
with a variable should generally be an odd several ways to convert an output fuzzy set tween probability and possibility theory.
number between five and nine. Second, into a crisp solution variable, but the two ABOUT THE AUTHOR. Earl Cox is the founder
each label should overlap somewhat with its most common are the composite maximum and chief executive officer of the Metus Sys-
neighbors. This overlap, in fact, is what gives and the composite moment, or centroid. tems Group, an advanced systems science
a fuzzy controller its smooth, stable surface. The centroid method takes the center of consulting and software services organi-
The overlap should be between 10 and 50 gravity of the final fuzzy space and produces zation in Chappaqua, N.Y. Prior to that, he
percent of the neighboring space, and the a result that is sensitive to all the rules-in founded Knowledge Based Technologies,
sum of the vertical points of the overlap particular, the results tend to move smoothly where he developed Telus, a mainframe-
should always be less than one. across the control surface. The composite based fuzzy logic expert system. +
>x-Fuzzy fundamentals 61