IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-7, NO.

10, OCTOBER 1977 707

[9] --, "Computational techniques for analysis of system dynamics [17] F. Harary, R. Norman, and D. Cartwright, Structural Models: An
models of social systems," J. Socio-Economic Planning Sci., vol. 4, Introduction to the Theory of Directed Graphs. New York: Wiley,
Oct. 1974. 1965.
[10] J. R. Burns, "Error analysis of nonlinear simulations: Applications to [18] M. R. Goodman, Study Notes in System Dynamics. Cambridge, MA:
world dynamics," IEEE Trans. Syst., Man, Cybern., vol. SMC-5, pp. Wright-Allen Press, 1974.
331-340, May 1975. [19] J. R. Burns, "Toward a mathematically rigorous methodology for
[11] , "A preliminary approach to automating the process of simula- simulation of social processes," Proc. Summer Computer Simulation
tion model synthesis," Proc. Seventh Pittsburgh Conf Modeling and Conf, pp. 1139-1147, July 1975.
Simulation, Apr. 1976. [20] J. Kane, "A primer for a new cross-impact language KSIM," Techno-
[12] J. N. Warfield, "Binary matrices in system modeling," IEEE Trans. logy Forecasting and Social Change, vol. 4, pp. 129-142, 1972.
Syst., Man, Cybern., vol. SMC-3, pp. 441-449, Sept. 1973. [21] J. Kane et al., "A methodology for interactive resource policy simula-
[13] "'Developing subsystem matrices in structural modeling," IEEE tion," J. Water Resources Res., vol. 9, no. 1, pp. 65-79, Feb. 1973.
Trans. Syst., Man, Cybern., vol. SMC-4, pp. 74-80, Jan. 1974. [22] J. R. Burns et al., "Causality: Its characterization by methodologies
[14] "Developing interconnection matrices in structural modeling," for modeling socioeconomic systems," to be published.
IEEE Trans. Syst., Man, Cybern., vol. SMC-4, pp. 81-87, Jan. 1974. [23] G. J. Klir, An Approach to General Systems Theory. New York: Van
[15] , "Toward interpretation of complex structural models," IEEE Nostrand Reinhold, 1969.
Trans. Syst., Man, Cybern., vol. SMC-4, pp. 405-417, Sept. 1974. [24] R. Fitz and D. Hornbach, "A participative methodology for design-
[16] M. D. Mesarovic, D. Macko, and Y. Takahara, Theory of Hierarchi- ing dynamic models through structural models," Proc. Seventh
cal, Multilevel, Systems. New York: Academic, 1970. Annual Pittsburgh Conf Modeling and Simulation, Apr. 1976.

A Fuzzy Logic Controller for a Trafc Junction

C. P. PAPPIS AND EBRAHIM H. MAMDANI, MEMBER, IEEE

Abstract Work done on the implementation of a fuzzy logic analysis of such processes, to which the human approach is
controller in a single intersection of two one-way streets is presented. characterized by rough approximations. Note that,
The model of the intersection is described and validated, and the use although stochastic and fuzzy logics can both be regarded as
of the theory of fuzzy sets in constructing a controller based on
linguistic control instructions is introduced. The results obtained derived from a probability logic [1], a stochastic approach
from the implementation of the fuzzy logic controller are tabulated would be methodologically different from the fuzzy disci-
against those corresponding to a conventional effective vehicle- pline which has been used here. It seems, therefore, that the
actuated controller. With the performance criterion being the aver- fuzzy rather than the stochastic approach should be used as
age delay of vehicles, it is shown that the use of a fuzzy logic the domain for the implementation of heuristics.
controller results in a better performance.
Previous work reported in the literature (e.g. [2]-[5]) has
INTRODUCTION shown the merits of the theory of fuzzy sets when applied to
A CONSIDERABLE amount of work has been done on the design of controllers for real dynamic plants, industrial
the problem of modeling and controlling traffic junc- processes, etc. In this study, the system is a traffic junction
tions. Although the major problem in cities concerns sets of and the problem of its control is considered as a classical
intersections (not individual ones), any approach to this example of nonprogrammed decisionmaking, i.e., decision-
problem should also include a sufficient description of the making characterized by the lack of well-specified analytical
events occurring in any individual intersection in the linked means for coping with a particular problem. Thus a lingu-
or disjoint system under study. istic control algorithm is synthesized, capable of dealing
Zadeh's pioneering work on fuzzy sets, by which a with a continuously reproduced decisionmaking situation.
conceptual framework is provided for dealing with prob- The starting point is an adequate (though qualitative)
lems of vagueness in the representation of complex knowledge of the system and a protocol of control instruc-
processes, can be of great help to the task of constructing a tions used by a human operator. A fuzzy set theoretic
controller for such an individual traffic intersection. Indeed, representation of these instructions (which we call "a fuzzy
the strength of the theory of fuzzy sets lies in its capability of logic controller") was tried as an answer to the control
rendering a powerful conceptual basis for the modeling and modeling problem, which gave very satisfactory results.
The work done on the construction of the model ot the
Manuscript received February 10, 1976; revised June 21, 1976, Novem- system and the implementation of the fuzzy logic controller
ber 15, 1976, and April 21, 1977. is presented below. In order to validate the model a fixed-
C. P. Pappis is with the Agricultural Bank of Greece, Athens, Greece.
E. H. Mamdani is with the Department of Electrical and Electronic cycle controller was also simulated. The average delays of
Engineering, Queen Mary College, University of London, London El the vehicles resulted from the implementation of the fuzzy
4NS, England. logic controller were compared to those caused by an
708X IEEE TRANSACTIONS ON SYSTEMS. MAN, AND CY13ERNETICS, V(o. SM(c-7 No. 10. (W 10111FR 1977

TABLE I
AVFRAGE DEIAYS WITH FIXED-CYCLE CONTROLLER ANI) OPTIMI M SETTINGS

N - S traffic E - W traffic Delay Error

(secs/veh) %
(veh/hr) (veh/hr) Model Formula
(dim) (d )

360 360 7.2 7.4 2

360 720 7.4 7.9 - 6
360 1080 7.9 8.4 - 5
360 1440 8.4 9.0 - 7
360 1800 9.3 10.2 - 9
360 2160 12.3 12.9 - 5
360 2520 15.8 18.9 -17
720 720 9.7 10.0 - 3
720 1080 10.8 11.6 - 7
720 1440 12.7 13.8 - 8
720 1800 15.9 17.3 - 8
720 2160 21.8 24.9 -12
1080 1080 13.6 14.9 - 9
1080 1440 17.9 19.7 - 9
1080 1800 25.8 29.2 -11
1440 1440 27.3 30.7 -11

efficient vehicle-actuated one. The results obtained show and the total waiting time of the vehicles in the queue would
that the performance of the system is better under the fuzzy be
logic controller. *a t12

THE MODEL Dn R = n2X 1 QG + n E 1 qnl

I

The major assumption concerning the model is that the Let s be the saturation flow, i.e., the rate at which vehicles
arrival of vehicles at the intersection is considered as being are cleared during the effective green period. At the 11th time
random. More explicitly, arrival times are considered to be unit after the beginning of the effective green, the number of
uniformly distributed. Note that this assumption affects not vehicles not yet cleared would be
only the truthfulness of the model but the selection of the n
control policies as well. The cycle is divided into two periods Sn z Q R + n , qGnl - S *}
of "effective red" and "effective green" for each phase; the =1 1
first corresponding to the halted traffic and the second to where QR is the queue which was built up during the
the traffic having the right of way. A total lost time of 10 previous effective red period of the phase, and is equal to
s/cycle is assumed. Vehicles leave the queue at a constant one when multiplied by a nonnegative quantity and zero
rate equal to the saturation flow during the effective green otherwise.
(see [6], [7] for definitions). The saturation flow equals 3600 These vehicles have been subjected to a delay
vehicles/h at both arms. There is no turning traffic.
For each successive time unit (1 s) a pseudorandom
number is generated and compared to some fixed quantity, Dn,G 11=
'7
ZQR
1
+
ZZ~?2= 1
qn2 S I/I!
which is equal to the mean rate of arrival (in vehicles per Thus during a cycle, the total delay experienced by
second). Thus the arrival of a vehicle is decided. Let vehicles traveling along one arm of the intersection would be
11 if a vehicle arrived during the nth unit interval D = DR,R + D;,o;
qn =
where D RERDG G are the delays during R and G, i.e. the
01 otherwise.
whole effective red and green periods, respectively. Finally
If QG denotes the number of vehicles not cleared during the average delay per vehicle would be
the previous effective green period of a phase, then the queue D
dm =
Q, at the nth time unit after the beginning of the effective red Z q,l
of that phase would be
n= I
n
Qn QG + qn1, The model just described is quite simple in comparison to
n= 1 some more sophisticated ones (see, for example, [8], [9]), yet
PAPPIS AND MAMDANI: LOGIC CONTROLLER FOR TRAFFIC JUNCTION 709

it suffices for the purpose of this work. A measure of its

reliability was obtained by using a fixed-cycle controller
which was implemented to the system. The system was
subjected to a wide range of averages of random vehicle
arrivals. Each time it was run for 7200 simulated seconds h

and the corresponding average delay per vehicle was cal-

culated. Results of the calculations, together with the ex- F2

pected average delays, are given in Table I. The expected

ones have been obtained from the following formula (see [6])
d = 0.65 2
2(1 - AX) 2(1 X)
-
q C2

where Pi processing of vehicles within the junction along route i,

Ri regulating function (signal setting),
Hi data processing,
d, average delay per vehicle on the particular arm, K optimizing function (control algorithm),
C cycle time, Ui vehicles on route i entering the junction,
A proportion of the cycle which is effective green for the C vehicles on route i leaving the junction,
phase under consideration (g/C), Yij subprocess's Pi interaction with subprocess Pi
ti premeasured data concerning ci (saturation flow),
q flow, Ui information input concerning oncoming vehicles.
s saturation flow (i.e., the flow which would be ob- Fig. 1. System control process.
tained if there were an endless queue of vehicles and
they were given a continuous green),
X degree of saturation (qlXs). extention operation would result in delays as those of Table
It can be verified that dm and d, yield comparable delays. I for the respective flow rates. These delays were the basis for
Indeed, Table I shows a fair agreement between the cal- the comparison between vehicle-actuated controllers and
culated delays and those obtained from the above formula, the fuzzy logic one (a fixed-cycle controller for a single
thus providing a validation of the model. intersection is scarcely worthy of comparison).
The results of Table I actually correspond to optimum The system control process is shown in Fig. 1. The
settings, i.e., optimum cycle and effective green times for the intervention of the controller takes place every 10 s during
respective flow rates according to each phase's effective green period; the first intervention
1.5L + 5 taking place just after the first 7 s of the period. At each
c = intervention the length of the extension of the effective green
1°
- =t y~
'-
time for the phase having the right of way is decided.
Information concerning the flow pattern is collected by
gi-y (Co -L) i= 1,2, detection pads, which, it is assumed, have been installed
before the traffic lights in both arms of the intersection. The
where role of the detection pads is very important, as will be made
Co optimum cycle time. clear in the sequel. It was assumed for the calculations that
effective green time, i 1,2, =
the flow pattern, as detected at the pads, is preserved
- qi s, where i = 1,2, throughout the period after each intervention for the phase
Yi
y = y1 + Y2, having the right of way. The distance between the pads and
L total lost time per cycle (10 s). the stop lines is sufficient for the controller to be informed
about the arrivals of vehicles in both arms ofthe intersection
It should be noted that the delays of Table I correspond to during the next 11 1 s, assuming that the effective green ends
random arrivals having fixed averages. In other words, at the middle of the 3-s amber period.
should the average rates of arrival be different from those for Thus vehicle i passing over the detectors is registered in
which optimum settings have been found and used for the the following way. Its speed vi is calculated. Assuming that
control of the intersection, the resulting delays would also be its speed is preserved constant during its trip from the
different from those of Table I. The use of timers, for tuning detectors to the junction, vehicle i will be at the "critical
the controller in order to adjust its settings to the daily flow point" in (I/vi - 1.5 s) time (see Fig. 2). The "critical point" is
pattern, would not be an easy task. Consequently, the actual the point where, should the lights turn to amber, it would be
delays occurring in any intersection controlled by fixed- possible for the vehicle just to pass. Let
cycle controllers would be by far in excess of those shown in
Table I, especially in cases of heavy traffic. Ni= lv- 1.5 s
In the case of vehicle-actuated controllers, the results of be the number of seconds required for the vehicle to arrive at
Table I correspond to those that should be expected with an the critical point. Ni indicates the position of vehicle i in the
efficient vehicle-actuated installation [7]. That is, a vehicle- flow pattern array for the next 10-s interval. The control
actuated controller with speed timing or with a low fixed input parameters are two continuously updated arrays
710 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL SMC- No. 10, ()(oCit
1 i1 1977

Iix statements. The above basic ideas of the theory of fuzzy sets
were used for the quantitative interpretation of these in-
IL
1 / //V/4v '
I I
structions as well as the decisionmaking process.
The fuzzy control instructions (see Appendix for a com-
plete set) are of the form
L
5 critical
if T = medium
points and A = mt(medium)
and Q = lt(small)
then E = medium
. . . I . . I I -

7-
detecting pads
t else
Fig. 2. Time-space diagram. if T= long
and A = mt(many)
corresponding to the halted traffic and the traffic having the and Q = lt(medium)
right of way. then E = long
else, etc. where
THE Fuzzy LOGIC CONTROLLER
In order to make our exposition self-contained, some of T the fuzzy variable "time," which is assigned values like
the basic definitions of the theory of fuzzy sets ([10], [11]) "very short," "short," "medium," etc.,
that were used to model the control algorithm are given A the fuzzy variable "arrivals," i.e., the number of
below. vehicles arriving at the arm having the right of way,
A fuzzy set F of a universe of discourse U = {x} is defined which may be assigned values like "many," "more
as a mapping IIF(x): U -+ [0,1] by which each x is assigned a than a few," etc.,
number in [0,1] indicating the extent to which x has the Q the fuzzy variable "queue," which is assigned values
attribute F. Thus if x is the number of vehicles in a queue, like "any," "less than small," etc.,
then "small" may be considered as a particular value of the E the fuzzy variable "extension," which has values
fuzzy variable "queue" and each x is assigned a number identical to "time."
Psmall(x) E [0,1] which indicates the extent to which that x is The terms "medium," "more than medium," "less than
considered to be small. small," etc., are labels of fuzzy sets defined on the relevant
Given the fuzzy sets A, B, or U, the basic operations on universes of discourse T,A,Q,E. Tables II-IV show the fuzzy
A,B are sets used in this application. Further to the above basic
i) the complement A of A, defined by operations, in this application we have introduced the
operators mt and It, standing for "more than" and "less
HA(X) = 1 HA(X), -
than," respectively. These are defined as follows. If A is a
ii) the union A u B of A and B, defined by fuzzy set defined on the real line R' = {xit, klA(xi) is its grade
of membership function and x0 is the element of R' for
PA B(x) = max {IA(X)4,B(X)}, which PA(Xi) is maximum, then lt(A) and mt(A) are fuzzy sets
iii) the intersection A n B of A and B, defined by defined on U such that
PA B(x) = min {tA(X),1B(X)}. 0. forx; >x I - -
-
r,
Plt(A) X A Xi)
for xi < x0
A fuzzy relation R from U = {x} to V = ty} is a fuzzy set
on the Cartesian product U x V, characterized by a func-
tion IR(X,Y), by which each pair (x,y) is assigned a number in PmtA)xj
Hmt()(i )1
=0 PAfor
(X i )
x(i < XO
for xi > x0. -

[0,1] indicating the extent to which the relation R is true for

result of these operations on the fuzzy sets of Tables III
(x,y). There are several ways of constructing P R(X,Y). The one The
and IV above is shown in Tables V and VI. Obviously
used here will be seen later.
Finally given a fuzzy relation R from U to V and a fuzzy lt(A) or mt(A) = not (A)
set on U, fuzzy
A a set B on V is induced, given by the
lt(A) and mt(A) =0
compositional rule of inference
B=A R -
or

or
max {[Llt(A)(XiA),Jmt(A)(Xi)} = 1 -IA(Xi)

UB(y) = max {min tPIR(x,Y),A(x)}}. {,jt(A)(Xj),Pmt(A)(Xi)} = 0. min

x
Note that if a fuzzy assignment like "A = small" is
A heuristic approach to the control problem was characterized by the poor content of the information
employed, which resulted in a set of linguistic control conveyed, a fuzzy assignment like "A = less than small" is
PAPPIS AND MAMDANI: LOGIC CONTROLLER FOR TRAFFIC JUNCTION
7111

TABLE II
Fuzzy SETs DEFINED ON TIME (OR EXTENSION)
time
(secs)
Fuzzyse 1 2 J 4 5 b 7 8 9 10

very short 1 .5 0 0 0 0 0 0 0 0

short 0 .5 1 .5 0 0 0 0 0 0

medLium 0 0 0 .5 1 .5 0 0 0 0

long 0 0 0 u 0 .5 1 .5 0 0

very long 0 0 0 0 0 0 0 .5 1 1

TABLE III
Fuzzy SETs DEFINED ON ARRIVALS
rIvaIs
(veh) 1 2 3 4 -5 6 7 8 9 10
F. sets

none .5 .2 .1 0 0 0 0 0 0 0

a few 1 .5 .2 .1 0 0 0 0 0 0

f ew .5 1 .5 .2 .1 0 0 0 0 0

medium .2 .5 1 5 .2 .1 0 0 0 0

many .1 .2 .5 1 .5 .2 .1 0 0 0

too many 0 .1 .2 .5 1 .5 .2 .1 0 0

TABLE IV
Fuzzy SETs DEFINED ON QUEUES
queue
(veh) 4 5 6 7 8 9 10 1112 1314 1516 17 18192021 2223 2425 2627 2829 3031 32
f.sets

very small 0 .5 .7 .9 .1 .9 .7 .5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

small 0 0 0 0 0 .5 .7 .9 1 .9 .7 .5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

small plus 0 0 0 0 0 0 0 0 0 .5 .7 .9 1 .9 .7 .5 0 0 0 0 0 0 0 0 0 0 0 0 0

medium 0 0 0 0 0 0 0 0 0 0 0 0 0 .5 .7 .9 1 .9 .7 .5 0 0 0 0 0 0 0 0 0

long 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 .5 .7 .9 1 .9 .7 .5 0 0 0 0 0

very long 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 .5 .7 .9 1 .9 .7 .5 0
712 '112
~~~~~~~~~~~IEEE
TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. smc-7, NO. 10, OCTOBER 1977

TABLE V
'MORE THAN"-OPERATION ON Fuzzy SETS OF TABLE III

Arrivals
eicls) 1 2 3 4 5 6 7 8 9 10
fuzzy set

mt(none) .5 .8 .9 1 1 1 1 1 1 1

mt(a few) 0 .5 .8 .9 1 1 1 1 1 1

mt(few) 0 0 .5 .8 .9 1 1 1 1 1

mt(medium) 0 0 0 .5 .8 .9 1 1 1 1

mt(many) 0 0 0 0 .5 .8 .9 1 1 1

mt(too many) 0 0 0 0 0 .5 .8 .9 1 1

TABLE VI
"LEss THAN"-OPERATION ON Fuzzy SETS OF TABLE IV
queue
(veh) 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
f. sets

lt(very small) 1 .5 .3 .1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

lt(small) 1 11 11 .5 .3 . 1 00 00 00 00 00 00 0 000 0000 0

lt(small plus) 1 1 1 111 1 1 1.5 .3 . 100 00 00 00 00 00 0000 0

lt(medium) 1 1 1 1 1 1 1 1 1 1 1 1 1 .5 .3 .1 0 0 0 0 0 0 0 0 0 0 0 0 0

lt(long) 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 .5 .3 .1 0 0 0 0 0 0 0 0 0

lt(very long) 1 1 11 1 1 1 111 1 11 11 111 11 1. .5.3 .100 00 0

conveying even less information. In other words, fuzzy The fuzzy implication "if P then E = very short" is also a
assignments like "A less than small" are used whenever fuzzy phrase R defined on T x A x Q x E with grades of
the grade of fuzziness is high. membership function
"~Any" is considered as a fuzzy set with all the elements of
its universe of discourse assigned a grade of membership
/IR(t,a,q,e) = min {Iv.short(t),/Lmt( none)(a )/a.y(q) short(e)}. v.

equal to one.
Finally two or more fuzzy implications R,S, ' , connected -

A total of 25 rules were used (5 for each intervention). by ".else" form a fuzzy clause C defined on T x A x Q x E
Every rule is a fuzzy relation between the inputs T, A, Q, and
with grades of membership function
the output E. The connectives "and" and "else" are in-

terpreted as the operators "min") and "max" respectively. /IC(t,a,q,e) = max {JiR(t,a,q,e),us(t,a,q,e),.
Thus
T =
very short In this application, since each fuzzy rule is represented by
and A =
mt(none) a four-dimensional array, the fuzzy algorithm employed at
and Q =any each intervention for deciding the control action is repre-
(see [12]) the universe of
sented by the union of five such arrays, as five rules operate
at each intervention. 25(5 x 5) rules given in the Appendix
is a fuzzy phrase P defined on

thus provide for a maximum of five interventions (each

discourse T x A x Q with grades of membership function

jip(t,a,q) min tpvshort(t),ILmt(none)(a)jitany(q)}. consisting of five rules) taking place at 7th, 17th, 27th, 37th,
PAPPIS AND MAMDANI: LOGIC CONTROLLER FOR TRAFFIC JlJNCTION 713

and 47th s. Thus the maximum possible effective green time THE PROCEDURE FOR DECIDING THE
is 57 s. At each intervention the five rules are invoked in the CONTROL ACTION
manner described below ten times (i.e., for each of the next Having determined the entries ofthe array corresponding
10 s). Note that, as the detecting pads are sufficiently far to the algorithm for each intervention, the process of
away from the junction, at each intervention, data is avail- inferring the control action is carried out as follows.
able for each of the next 10 s. Consider now the (ti,aJ,ak,e i) For each successive time unit (1 s) for the next 10 s, data
entry of the array C 2, corresponding to the algorithm used at concerning vehicles crossing the critical point and vehicles
the second intervention of the controller, where added to the queue are used as input to the algorithm array
tj 8 i.e., we consider the next 8 s, in use. The corresponding entry of the array is thus deter-
aj=4 i.e., 4 vehicles will cross the critical point if no mined. This entry is a measure of the confidence with which
change of the current state of the system occurs the algorithm may be applied, for the corresponding data.
during the next 8 s, Obviously, that extension will be selected which corre-
qk= 5 i.e., five vehicle queue will build up if no change of sponds to the maximum degree of confidence. In other
the current state of the system occurs during the words, fuzzy predictive decisionmaking implies that, that
next 8 s, action is selected that minimizesfuzziness. Thus given a set of
e, = 8 i.e., the extension given to the present state of the fuzzy rules, choose the one which is provided for coping with
system is 8 s. conditions as similar to the actual ones as possible. And,
The first control statement R 1 for the second intervention
given a set of alternative actual conditions, consider those
which are as similar to the conditions, for which the
(see Appendix) is algorithm provides, as possible.
if T = very short The explicit description of the procedure for deciding the
and A = mt(none) control action is given below, by means of an example. Thus
and Q= any we consider the controller's second intervention. Arm N-S
then E = very short. has the right of way. There are five vehicles queued at E- W
arm. Data, concerning number of vehicles crossing the
From Tables II, V, VI we have critical point (N-S traffic) and queued (E- W traffic) at each
[v.short(8) = 0.0 successive time unit during the next 10 s, is summarized in
arrays a and a' respectively
Jimt(none)(4) = 1.0
x=(O 1 0 1 1 1 1 0 0 1)
Many(5) = 1.0.
Thus x'=(O 1 0 0 1 0 0 1 0 0).
From af' arrays Il,/3' are constructed
PR1(8,4,5,8) = min {uv.short(8),/lmt(none)(4),uany(5),/v.short(8)}
= min {o,1.0,1.0,0} = 0. fl=(O 1 1 2 3 4 5 5 5 6)
Similarly we find ,B'=(5 6 6 6 7 7 7 8 8 8)
as follows. If the ith elements of x and fl are a i,bi, respectively,
PR2(8,4,5,8) = min {I1short(8),Pmt(a few)(4), it is
btlt(v. short) (5 )itshort(8 )}
= min {0,0.9,0.5,0} = 0 bi= E aj,
j=1
JR3(8,4,5,8) = min {Jmedium(8 ),Jlmt((few) (4), and if the ith elements of a' and ,B' are a',b', respectively, it is
P1t(v.short)(5 ),Jmedium(8)}
b= Q + aa,
= min {0,0.8,0.5,0} = 0 j=1
P R4(8,4,5,8) = min {!long(8),[mt(medium)(4), where Q is the present queue at E-W arm. For example, if no
change of the current state of the system occurs during the
Plt(v.short)(5),Ilong(8 )} next 6 s, it is seen (from /3 and ,B') that four vehicles will cross
=min {0.5,0.5,0.5,0.51 = 0.5 the critical point and there will be a total of seven vehicles in
the queue at E-W arm after 6 s from now.
PR5 (8,4,5,8 )= min {Iv.Iong(8 ),,Lmt(many)(4), The ith elements of arrays /3,B', i = 1, , 10, determine the
Umtmshor(5),v. 1ong(8 0. appropriate entry of the algorithm matrix C2, which indi-
cates the applicability of the algorithm to the situation
min {50.501,0.5} = 0.
= described by these elements of the arrays. Thus for t = 1 s
Thus the (tj,aj,qk,el) entry of matrix C2 is (i.e., considering an extension of 1 s) we have that no vehicle
will cross the critical point (first element of array /3) and that
PC2(t,,aj,qk,el) = max {0,0,0,0.5,0} = 0.5. the queue will remain the same (five vehicles, first element of
714 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-7, NO. 10, OCTOBER 1977

TABLE VII
DECISION TABLE FOR CONTROL ACTION

Fuzzy Control Statement Time (seconds)

time passing queue 1 2 3 4 5 6 7 8 9 10

vehicles

very mt(none) any .5 0 0 0 0 0 0

short | O

short mt(a few) lt(very O O O .3 0 O 0 0 0 0

short)

medium mt(few) lt(very 0 0 0 .1 .1 0 0 0 0

short)

long mt(medium) lt(very | | O

0 .1 .1 0( 0
short)

very mt(many) lt(short) 0 0 0 0 0 0 .5 .5 .8

long |

Fuzzy Algorithm 0 .5 0 .3 .1 .1 .1 .5 .5 .8

array #'). It is easy to show that the rules of the algorithm are trial and error method was used in order to obtain an
assigned the grade 0, and, consequently, the algorithm is effective set of rules (or what is termed "satisficing control"
assigned the grade 0 for t = 1 s. The results have been in management science). In other words, a learning
summarized in Table VII. Obviously, the controller will procedure was employed, by which the human performance
select the extension of 10 s. Thus the state of the system will in a similar real life situation of controlling a traffic junction
remain the same for the next 10 s and the above procedure is derived.
will be repeated (with new ,B and ,B') at the end of the 10-s It is interesting to note that, having defined (in terms of
period. If the extension given to the present state of the fuzzy sets) what was considered to be a "small queue" or
system were less than 10 s, the state of the system would "few arrivals," it was the rules rather than the fuzzy sets
change at the end of the extension period. which were modified. It is quite apparent that if the par-
Note that if all the entries of the last row ofTable VII were ameters describing the membership functions were in-
less than 0.5, no extension would be given and the state of the troduced as additional input to the decision concerning the
system (i.e., the phase) would be immediately changed. control algorithm, the dimensionality of the problem would
Finally, if the maximum of the entries of the last row were radically increase, thus imposing severe difficulties in the
not unique, i.e., if two or more alternative extension periods obtainment of its solution.
were indicated, then the maximum of these alternative The question of stability has been part of the whole
extension periods would be selected. Of course, some other problem of obtaining what is termed "effective set of rules."
rule might be used instead (e.g., one giving the minimum In this particular application, the stability of the system is
extension period or the median or one randomly selected defined as the condition of the system not getting saturated if
among these alternatives). For this, however, another con- subjected to a wide range of flow rates.
trol algorithm would be required in the place of the one used As far as fuzzy set theory is concerned, its basic concept,
in this implication, which was based on the rule giving the "fuzziness," characterizes only a state of knowledge. It exists
maximum extension period. neither in the system nor in the controller but in the human
mind. Although the controller which was actually designed
RESULTS AND FINAL REMARKS is termed "a fuzzy logic controller," it actually acts deter-
Because of the random nature of the arrivals assumed, ministically. That is, the algorithm by means of which the
many runs of the model were needed in order to get reliable decision is taken, although conceived in fuzzy linguistic
results. The simulation work was carried out on the terms, is not fuzzy after the actual design is completed, i.e.,
ICL-1900 general purpose computer. after the fuzzy sets and implications are established [13].
The results have been summarized in Table VIII, whence Some interesting questions arise from the application
it is clear that the system's performance is best under the reported here. These are related to the advantage gained by
fuzzy logic controller for all possible combinations of flow using the theory of fuzzy sets in modeling a controller in
rates. Note that the effectiveness of the controller, as in- general and a traffic one in particular. They are also related
dicated by the percentage improvement in delays, is not to the feasibility of the construction of the fuzzy logic
seriously affected by the total volume of traffic through the controller, as well as cost considerations concerning its
junction. construction.
These results have been obtained after several First, it is clear that the improvement of the particular
modifications of the control instructions initially set. The traffic junction of this study should be attributed to a)
PAPPIS AND MAMDANI: LOGIC CONTROLLER FOR TRAFFIC JUNCTION 715

TABLE VIII
COMPARISON BETWEEN DELAYS CAUSED BY EFFICIENT VEHICLE-ACTUATED
CONTROLLER AND Fuzzy LOGIC ONE
N - S traffic E - W traffic Average overall delay Improvement
(secs/veh)
(veh/hr) (veh/hr) %
Vehicle-actuated Fuzzy-logic
controller controller

360 360 7.2 5.7 +21

360 720 7.4 6.1 +18
360 1080 7.9 6.6 +17
360 1440 8.4 7.3 +13
360 1800 9.3 8.4 +10
360 2160 12.3 10.0 +19
360 2520 15.8 13.6 +14
720 720 9.7 7.4 +21
720 1080 10.8 8.8 +19
720 1440 12.7 10.9 +14
720 1800 15.9 14.1 +11
720 2160 21.8 18.5 +15
1080 1080 13.6 12.0 +12
1080 1440 17.9 15.4 +14
1080 1800 25.8 21.6 +16
1440 1440 27.3 22.9 +16

..1. I. I

placing the detecting pads at a considerable distance from hand, special problems would arise in this case owing to the
the stop lines, b) detecting the traffic flow at both arms ofthe hierarchical structure of the system and consequently the
intersection simultaneously, and c) implementing the set of control policy itself. It is thought also that in the case of
rules which resulted as the final product of the trial and error integrated traffic control systems the theory of fuzzy sets
procedure used. However, it is very important to note that would show its merits much more so than in the present
fuzzy set theory is credited with the convenience it offers in simple case of an individual intersection.
modeling linguistic rules. Thus in view of the design of a
controller for a complex process, it is easier for the designer APPENDIX
to synthesize a control algorithm in linguistic terms and THE Fuzzy ALGORITHM
subsequently implement it to the process, after interpreting
them in terms of the theory. This advantage of the theory of Intervention: 7th second
fuzzy sets was fully appreciated in this application. Further- if T = very short
more, modification of the initial set of linguistic rules was and A = mt(none)
carried out in terms of computer instructions in a natural and Q = any
manner, by considering labels of fuzzy sets rather than then E = very short
values of the variables. else
Second, hardware implementation of the fuzzy logic if T = short
controller is obviously a straightforward computer exercise and A = mt(a few)
exploiting present day LSI technology, the decision being and Q = lt(very small)
taken by table look-up. This was demonstrated earlier, by then E = short
means of exanples. Furthermore, although cost analysis else
was not a major concern of this study, it seems fair to state if T = medium
that investment for the implementation of the controller and A = mt(few)
would pay for itself. This is justified in view of the fact that and Q = lt(very small)
considerable reduction in delay times, and consequently in then E = medium
the associated cost, is caused by this controller, in respect to else
conventional vehicle-actuated ones. if T = long
It is hoped that further work will be done on the problem and A = mt(medium)
of controlling traffic by use of the theory offuzzy sets. It must and Q = lt(very small)
however be kept in mind that the fuzzy logic controller was then E = long
designed for the purpose of controlling traffic characterized else
by randomness. In a linked system the traffic would be if T = very long
modulated. This should be taken into account when con- and A = mt(many)
sidering a controller for an intersection being part of a whole and Q = lt(very small)
network, forming an integrated control system. On the other then E = very long.
716 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOI- SMC-7, N). 10, oCB)IlBR 1977

Intervention: 17th second else

if T = very short if T = short
and A = mt(none) and A = mt(a few)
and Q = any and Q = lt(small plus)
then E = very short then E = short
else else
if T = short if T = medium
and A = mt(a few) and A = mt(medium)
and Q = lt(very small) and Q = lt(small plus)
then E = medium
then E = short else
else if T = long
if T = medium
and A = mt(few) and A = mt(many)
and Q = lt(very small) and Q= lt(medium)
then E = long
then E = medium else
else if T = very long
if T = long
and A = mt(too many)
and A = mt(medium) and Q = lt(long)
and Q = lt(very small) then E = very long.
then E = long
else
if T = very long Intervention: 47th second
and A = mt(many)
and Q = lt(small) if T = very short
then E = very long. and A = mt(none)
and Q = any
then E = very short
Intervention: 27th second else
if T = very short if T= short
and A = mt(none) and A = mt(a few)
and Q = any and Q = lt(long)
then E = very short then E = short
else else
if T = short if T= medium
and A = mt(a few) and A = mt(medium)
and Q = lt(very small) and Q = lt(long)
then E = short then E = medium
else else
if T = medium if T= long
and A = mt(few) and A = mt(too many)
and Q = lt(very small) and Q = /t(very long)
then E = medium then E = long
else else
if T = long if T = very long
and A = mt(medium) and A = mt(too many)
and Q = lt(very small) and Q = lt(very long)
then E= long then E = very long.
else
if T = very long
and A = mt(many) REFERENCES
and Q = lt(small) [1] B. R. Gaines, "Stochastic and fuzzy logics," Electr. -ertterrs
vol. 11, Pp.
then E = very long. 188-189, 1975.
[2] E. H. Mamdani and S. Assilian, "An experiment in linguistic syn-
thesis with a fuzzy logic controller," Inrtern. J. Mau-Maachine Stiudies,
Intervention: 37th second vol. 7, pp. 1-13, 1975.
if T = very short [3] E. H. Mamdani, "Application of fuzzy algorithms for the control of a
dynamic plant," Proc. IEE, vol. 121, no. 12, pp. 1585 ---1588, 1974.
and A = mt(none) [4] P. J. King and E. H. Mamdani, "The application of fuzzy control
and Q= any systems to industrial processes," Round Table on Fuzzy Automata
then E = very short and Decision Processes, Proc. Sixth TrienIial IFAC World
Congress, Cambridge/Boston, MA, August 24- 30, 1975.
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-7, NO. 10. OCTOBER 1977 717

[5] E. H. Mamdani and N. Baaklini, "Prescriptive method for denrving [9] W. D. Ashton, The Theory oj Road Traffic Flow. London: Methuen,
control policy in fuzzy-logic controller," Electr. Letters, vol. 11, p. 1966.
625, Dec. 1975. [10] L. A. Zadeh, "Fuzzy sets," Inform. Contr., vol. 8, pp. 338-353, 1965.
[6] F. V. Webster, "Traffic signal settings," Technical Paper 39, Road [11] "Outline of a new approach to the analysis of complex systems
Research Laboratory, 1958. and decision processes," IEEE Trans. Syst., Man, Cybern., vol.
[7] F. V. Webster and B. M. Cobbe, "Traffic signals," Technical Paper SMC-3, pp. 28-44, 1973.
56, Road Research Laboratory, 1966. [12] R. C. T. Lee and C. L. Chang, "Some properties of fuzzy logic,"
[8] H. H. Goode, C. H. Pollmar, and J. B. Wright, "The use of a digital Inform. Contr., vol. 19, pp. 417-431, 1971.
computer to model a signalized intersection," Proc. High. Res. Ed., [13] P. N. Marinos, "Fuzzy logic and its application to switching
vol. 35, pp. 548-577, 1956. systems," IEEE Trans. Comput., vol. C-18, pp. 343-348, 1969.

Machine Recognition of Abnormal Behavior

in Nuclear Reactors
R. C. GONZALEZ, MEMBER, IEEE, AND L. C. HOWINGTON, MEMBER, IEEE

Abstract-A multivariate statistical pattern recognition system Unlike many pattern recognition applications where re-
for reactor noise analysis is presented. The basis of the system is a presentative patterns from each class of interest are avai-
transformation for decoupling correlated variables and algorithms
for inferring probability density functions. The system is adaptable to lable, it is difficult to obtain patterns characterizing
a variety of statistical properties of the data, and it has learning, abnormal behavior in the components of a nuclear plant.
tracking, updating, and dimensionality reduction capabilities. Simulation of abnormal behavior by planned failure of
System design emphasizes control ofthe false-alarm rate. Its abilities components in a plant is usually not a feasible alternative for
to learn normal patterns and to recognize deviations from these generating the desired abnormal patterns because of govern-
patterns were evaluated by experiments at the Oak Ridge National
Laboratory (ORNL) High-Flux Isotope Reactor. Power perturba- mental regulations and implementation costs. The design of
tions of less than 0.1 percent of the mean value in selected frequency a pattern recognition system for surveillance is thus reduced
ranges were readily detected by the pattern recognition system. to characterizing normal behavior and establishing limits
beyond which the operation of a component is classified as
I. INTRODUCTION abnormal. This approach is based on the assumption that
INTEREST in the application of pattern recognition the data which are labeled as normal are derived from plant
techniques to perform automatic monitoring functions in components whose operation is within design specifications.
nuclear power plants stems from a need to provide assist- One of the principal features of a pattern recognition
ance to plant operators in assimilating quickly the large system designed to aid plant operators is that the system
number of interrelated signals that provide insight into the output should contain information beyond a simple
operational status of the plant. "normal" or "'abnormal" message. This information can be
The recognition problem in this field can be broadly in the form of graphs, tables, or numerical data, but, in the
divided into two categories: surveillance and diagnosis. The presence of an abnormality, it must contain reference to the
surveillance problem is one of classifying the status of the characteristics and limits of normal behavior. This informa-
components being monitored into two classes: normal or tion aids the operator in detecting trends and also in
abnormal. The diagnosis problem is concerned with iden- evaluating the relative importance of an alarm.
tifying the source and degree of a detected abnormality. At The system described in this paper is oriented in this
the present time most of the work in this area is being direction. It utilizes the Hotelling transformation to de-
focused on the surveillance problem. couple the multivariate input data and thus allows simpli-
fied interpretation of one-dimensional variables. In addi-
tion, this transformation is used for dimensionality
Manuscript received August 20, 1976; revised May 4, 1977. This work reduction to compress archival storage of a plant's opera-
was supported by the Energy Research and Development Administration tional history. The system also utilizes the Mahalanobis
under contract with the Union Carbide Corporation. A shorter version of
this paper was presented at the Third International Joint Conference on distance as a global measure of performance in the original
Pattern Recognition, Coronado, CA, November 8-11, 1976. pattern space. Following detection of an abnormal pattern,
R. C. Gonzalez is with the Department of Electrical Engineering, the system makes available to the operator a set of options
University of Tennessee, Knoxville, TN 37916, and the Instrumentation
and Controls Division, Oak Ridge National Laboratory, Oak Ridge, TN to help him analyze and interpret the pattern.
37830. The performance of the pattern recognition system was
L. C. Howington was with the Department of Electrical Engineering, evaluated with noise data from the Oak Ridge National
University of Tennessee, Knoxville, TN 37916, and the Oak Ridge Nat-
ional Laboratory, Oak Ridge, TN 37830. He is now with the Develop- Laboratory (ORNL) High-Flux Isotope Reactor (HFIR).
ment Division, Y-12 Plant, Union Carbide Corporation, Oak Ridge, TN. The extraction and use of information from noise signals