Fuzzy Time Series Forecasting

- Developing a new forecasting model based on high order fuzzy time series
AAUE
November 2009
Semester: CIS 4
Autor: !ens "#ni $ou%sen
Fuzzy Time Series Forecasting
& 'eve%o(ing a ne) *orecasting mo+e% base+ on ig or+er *uzzy time series
Autor: Jens Rni Poulsen
E+ucationa% Institute: Aalborg University sb!erg "AAU#
Semester: $%& '( )ov* +,,-
Su(ervisor: Daniel .rtiz Arroyo
%%
Abstract
)umerous /uzzy 0ime &eries "/0&# models have been proposed in scientific literature
during the past decades or so* Among the most accurate /0& models found in literature are the
high order models* 1owever( three fundamental issues need to be resolved with regards to the
high order models* /irst( current prediction methods have not been able to provide satisfactory
accuracy rates for defuzzified outputs "forecasts#* &econd( data becomes underutilized as the
order increases* 0hird( forecast accuracy is sensitive to selected interval partitions*
0o cope with these issues( a new high order /0& model is proposed in this thesis* 0he
proposed model utilizes aggregation and particle swarm optimization "P&.# to reduce the
mismatch between forecasts and actuals* $omparative e2periments confirm the proposed
model3s ability to provide higher accuracy rates than the current results reported in the
literature* 4oreover( the utilization of aggregation and P&.( to individually tune forecast rules(
ensures consistency between defuzzified outputs and actual outputs( regardless of selected
interval partitions* As a conse5uence of employing these techni5ues( data utilization is
improved by6 "7# minimizing the loss of forecast rules8 "+# minimizing the number of pattern
combinations to be matched with future time series data*
/inally( a fuzzification algorithm( based on the trapezoid fuzzification approach( has been
developed as a byproduct* 0he proposed algorithm ob!ectively partitions the universe of
discourse into intervals without re5uiring any user defined parameters*
Jens Rni Poulsen
Table of Contents
, Intro+uction------------------------------------------------------------------------------------------------------------------------------------,
2 Teoretica% Foun+ation------------------------------------------------------------------------------------------------------------------2
+*7 $onventional &ets vs /uzzy &ets********************************************************************************************+
+*+ 0he Universe of Discourse******************************************************************************************************9
+*9 /uzzy &ubsets**************************************************************************************************************************9
+*9*7 Alpha $ut**************************************************************************************************************************:
+*' Representations of /uzzy &ets************************************************************************************************;
+*: .perations on /uzzy &ets********************************************************************************************************;
+*; /uzzy )umbers***********************************************************************************************************************<
+*< Ran=ing***********************************************************************************************************************************-
+*> ?inguistic @ariables****************************************************************************************************************-
+*- Defuzzification**********************************************************************************************************************7,
+*-*7 4a2-membership principle*********************************************************************************************7,
+*-*+ $entroid method**************************************************************************************************************77
+*-*9 Aeighted average method***********************************************************************************************77
+*-*' 4ean-ma2 membership***************************************************************************************************77
+*7, /uzzy Relations*******************************************************************************************************************7+
+*7,*7 /uzzy Relational $ompositions************************************************************************************79
+*77 Aggregation*************************************************************************************************************************7'
+*77*7 Averaging .perators******************************************************************************************************7<
+*77*7*7 Beneralized 4eans************************************************************************************************7>
+*77*7*+ .rdered Aeighted Averaging .perators ".AA#***************************************************7-
+*77*+ 0riangular )orms "0-)orms and 0-$onorms#***************************************************************+7
+*77*+*7 Duality of 0-)orms and 0-$onorms**********************************************************************++
+*77*9 Averaging .perators and 0riangular )orms in $onte2t************************************************+9
+*7+ Particle &warm .ptimization "P&.#***********************************************************************************+'
+*79 /uzzy 0ime &eries and its $oncepts***********************************************************************************+;
+*7' $onclusion**************************************************************************************************************************+>
. "e%ate+ /or0--------------------------------------------------------------------------------------------------------------------------------29
9*7 &ong and $hissom3s Aor=*****************************************************************************************************+-
9*+ $hen3s Aor=**************************************************************************************************************************9,
9*9 .ther Developments*************************************************************************************************************9;
%%%
9*' $onclusion****************************************************************************************************************************9;
4 Intro+ucing a 1o+i*ie+ Fuzzy Time Series 1o+e%--------------------------------------------------------------------.2
'*7 Algorithm .verview*************************************************************************************************************9>
'*+ /uzzifying 1istorical Data****************************************************************************************************9-
'*9 valuating the Proposed /uzzification Algorithm****************************************************************';
'*' Defuzzifying .utput*************************************************************************************************************'-
'*'*7 stablishment /uzzy &et Broups "/&B3s#***********************************************************************:7
'*'*+ $onverting /&B3s into if statements********************************************************************************:+
'*'*9 0raining the if-then rules with P&.********************************************************************************:'
'*: $onclusion****************************************************************************************************************************:>
3 E4(erimenta% "esu%ts-------------------------------------------------------------------------------------------------------------------50
:*7 $omparing different /0& models*****************************************************************************************;,
:*+ $onclusion****************************************************************************************************************************;7
5 Fina% Conc%usion---------------------------------------------------------------------------------------------------------------------------52
6 "e*erences--------------------------------------------------------------------------------------------------------------------------------------5.
2 A((en+i4 I-------------------------------------------------------------------------------------------------------------------------------------55
9 A((en+i4 II------------------------------------------------------------------------------------------------------------------------------------56
%@
1 Introduction
0his research is carried out as part of the $%& ' semester at AAU and is concerned with the
development of a new forecasting model based on high order fuzzy time series "/0&#* )umerous
/0& models have been proposed during the past decades or so C7(+(9('(:(;(<(>(-(7,(77(7+(79(7'(7:(
7;(7<(7>(7-(+,(+7(++(+9(+'(+:(+;(+<(+>(+-(9,(97(9+(99D* 0he high order /0& models C7:(7<(9+(97(
9,D are the most accurate models found in literature vis-E-vis related modalities* Despite this(
current publications have not been able to provide satisfactory results for defuzzified outputs
"forecasts#* Another problem particularly associated to high order models is the underutilization of
data that occurs as a result of increasing the model3s order* ?astly( current prediction models are
sensitive to selected interval partitions* 0o cope with the shortcomings mentioned here( this pro!ect
sets out to develop a forecast model based on /0& which6 "7# provides higher forecasting accuracy
rates than its high order counterparts8 "+# improves data utilization8 and "9# remains unaffected by
selected interval partitions* A secondary ob!ective is to design a fuzzification algorithm based on the
trapezoid fuzzification approachC'D which automatically generates interval partitions based on some
ob!ective measure*
0he thesis is organized as follows* &ection + deals with relevant theoretical aspects such as
fuzzy sets( fuzzy numbers( defuzzification( fuzzy relations( fuzzy aggregation operators( particle
swarm optimization "P&.# and basic concepts of /0&* &ection ' provides an overview of related
wor=* 0he proposed /0& model is presented and comparatively evaluated in section : and ;(
respectively* /inally( concluding remar=s are provided in section <*
7
2 Theoretical Foundation
0his section reviews various theoretical concepts relevant in the conte2t of this study* 0he main
topics covered here are fuzzy sets( fuzzy numbers( defuzzification( fuzzy relations( fuzzy
aggregation operators( P&. and basic /0& concepts*
2.1 Conventional Sets vs Fuzzy Sets
$onventional set theory rests on the notion of a crisp boundary between which elements are
members and non-members of a particular set* 0hus if someone as=s the 5uestion of whether an
element is in a set( the answer is always yes or no for all elements* /or e2ample( if we consider the
set of tall people( all persons are either tall or not* 0here is nothing in between of being tall and not
being tall* Fasically( conventional sets can be described in two ways8 e2plicitly in a list or implicitly
with a predicate* An e2ample of the former could be the finite set A=,( 7( +( 9* An e2ample of the
latter could be the set of all integers larger than 7, which is an infinite set* ither way( we can
always answer yes or no to which elements are in a set or not*
0he drawbac= of conventional sets is that many concepts encountered in the real world cannot
always be described e2clusively by their membership and non-membership in sets* As an e2ample(
let us again consider the set of tall people* %f we as= a group of people when e2actly a person is tall
and when e2actly heGshe is not( we are li=ely to get a set of deviating answers* 0his is because there
is no crisp boundary between being tall and not being tall "at least not one that is intuitively clear#*
&tated otherwise( the property of being tall is inherently undecidable* Benerally all persons taller
than +,, cm satisfy the property of being tall and everyone shorter than 7:, cm do not* Fut what
about the membership status of those with a height that lies in between these two e2tremesH Aell
some of these people may still be considered tall and some not( the point is however that the further
we move down the scale from +,, cm to 7:, cm( the answer will not remain as clear-cut for all
cases* At some point we will reach a state of ambiguity( that is a state where we cannot e2plicitly
say either yes or no*
/uzzy set theory C9'(9:(9;D e2pands the notion of purely crisp sets by assigning membership
degrees to set elements so the transition from membership to non-membership is gradual rather than
abrupt* )ormally( the membership degree of a set element is a real number between , and 7* 0he
closer the membership degree is to 7( the more an element belongs to a given set* A membership
degree of , means that an element is clearly not a member of a particular set* lements with a
membership degree between , and 7 are more or less members of a particular set* An e2ample is the
+
property of being tall( where a person may be more or less tall* %n conventional set theory( more or
less membership is not allowed*
2.2 The Universe of Discourse
All elements in a set are ta=en from a universe of discourse or universe set that contains all the
elements that can be ta=en into consideration when the set is formed* %n reality there is no such
thing as a set or a fuzzy set because all sets are subsets of some universe set( even though the term
3set3 is predominantly used* %n the fuzzy case( each element in the universe set is a member of the
fuzzy set to some degree( even zero* 0he set of elements that have a non-zero membership is
referred to as the support* Ae will use the notation U for the universe set.
2.3 Fuzzy Subsets
A fuzzy subset A in U is characterized by a membership function "characteristic function# that
maps each element in A with a real number in the unit interval* /ormally( this can be e2pressed as

A
: U | ,(7
where the value

A
( x)
is called the degree of membership of the element x in the
fuzzy set A* 0he membership function declares which elements of U are members of A and which
are not* 0he principle of fuzzifying crisp sets in this manner is called the extension principle*
A classical e2ample of a fuzzy set is the subset of person3s heights considered tall( see figure 7* Ae
refer to this set as 0all* 0he figure shows that if a person3s height is less than or e5ual to 7:, cm
"point a#( the degree of membership in the fuzzy set 0all is e5ual to zero* 0his means that all
persons whose height is less than or e5ual to 7:, cm are completely e2cluded from this set* %f a
person3s height is larger than or e5ual to +,, cm "point b#( the property of being tall is fully
satisfied* 1ence the membership degree in 0all is e5ual to 7* Ahen the height is larger than 7:, cm
9
Figure ,- An e4am(%e o* a membersi( *unction *or te *uzzy set Ta%%-
and less than +,, cm( the property of being tall is more or less satisfied* /or e2ample( a person who
is 7>: cm "point u# is tall to a degree of ,*>* 4athematically the above membership function "a=a
characteristic function# can be defined as
j
Tall
( x)=

,( xa
xa
ba
, axb
7( bx*
"7#
%n the case where a fuzzy set A is a conventional "crisp# set( the corresponding membership function
can be reduced to
j
A
( x)=

7( xA
,( xA.
"+#
0he above function has only two outputs( , or 7* Ahenever

A
( x)=7(
x is a member of A, if

A
( x)=,(
x is declared a non-member of A* .ther e2amples of membership functions commonly
used in literature are depicted in figure +*
%t needs to be noted that fuzzy membership functions are not necessarily symmetric in nature even
though this is not indicated in the figure* Depending on the application( the shape of a membership
function may or may not be symmetrical* 0here is not yet any universal rule or criterion for
selecting a membership function for a particular type of fuzzy subset* Rather the choice depends on
several factors( for e2ample( the users scientific e2perience and =nowledge or actual needs for the
application in 5uestion* Ahatever membership function is chosen for the problem at hand( will
'
Figure 2- 7arious sa(es o* common%y use+ membersi( *unctions-
more or less be based on the users sub!ective measures* 1owever( !ust as in probability theory and
statics( for e2ample( one may assume that a particular function describes some property( li=e Iit is
assumed that the membership function in figure 7 describes the property of being tallI*
2.3.1 Alha Cut
An important property of fuzzy sets is the alpha cut "J-cut#* Biven a fuzzy set A defined on the
universal set U and any number in the unit interval( o| ,(7 ( the "wea=# J-cut(
A
o
(
and the strong
J-cut(
A
o
(
are the crisp sets which satisfy

A
o
=xAj
A
( x)o
A
o
=xAj
A
( x)>o*
"9#
?ess formally the J-cut of a fuzzy set A is the crisp set
A
o
that contains all the elements of the
universal set U whose membership grades in A are grater than or e5ual to a given J "or strictly
grater than J( if we refer to the strong variant
A
o
#*
Recall that the support of a fuzzy set A within the
universal set U is the crisp set that contains all the elements of U that have non-zero membership
grades in A. 1ence the support of A is e2actly the same as the strong J-cut of A for =,*
A fuzzy subset A in is conve2 iff the sets defined by
A
o
=xAj
A
( x)o "'#
are conve2 for all J-level sets in the interval C,(7D* Another more direct definition of conve2ity is the
following6 /or all pairs
x
7
( x
+
A
and any \| ,(7 ( A is conve2 iff
j
A
|\x
7
+(7\) x
+
min| j
A
( x
7
), j
A
( x
+
)( ":#
where min denotes the minimum operator*
:
Figure .- /ea0 an+ strong a%(a&cut8 res(ective%y-
2.! "eresentations of Fuzzy Sets
Fasically a fuzzy set can be viewed as a collection of ordered pairs
A=( x
7
, j( x
7
)) ,( x
+
, j( x
+
)) ,.,( x
n
,j( x
n
))( ";#
where element x is a member of A and ( x) denotes its degree of membership in A* A single pair
( x , ( x)) is called a fuzzy singleton* 1ence a complete fuzzy set can be viewed as the union of its
constituent singletons* %f a fuzzy set is finite and discrete an often used notation is
A=u
7
/ x
7
+u
+
/ x
+
+.+u
n
/ x
n
* "<#
%t is important to note that neither the slash or the plus sign represent any =ind of algebraic
operation* 0he slash lin=s the elements of the support with their grades of membership in A,
whereas the plus sign indicates that the listed pairs of elements and membership grades collectively
form the definition of the set A*
2.# $erations on Fuzzy Sets
%n classical set theory there are three basic operations that can be performed on crisp sets6 the
complement( intersection and union* 0hese operators also e2ists in fuzzy set theory in addition to a
range of other operators* 0he standard fuzzy set operators complement( intersection and union are
defined by the e5uations
j
A
( x)=7j
A
( x)
(j
A
j
B
)( x)=min| j
A
( x) , j
B
( x)
(j
A
j
B
)( x)=ma2 |j
A
( x) , j
B
( x) *
">#
where A and B are fuzzy subsets of the universal interval U* 0he two operators( min and ma2,
respectively denote the minimum and ma2imum operators* As can be seen from the respective
e5uation( the min and the ma2 operations of two fuzzy sets
A
and
B
is an element-by-element
;
Figure 4- A conve4 an+ a non&conve4 *uzzy subset-
comparison between corresponding elements in the respective sets* %n the complement case( each
membership value of
A
is substracted from 7*
As mentioned( there are also a range of other operators in addition to the standard fuzzy set
operators* 0hese operators can be categorized as follows6 t-norms( averaging operators and t-
conorms* An in-debt discussion about these operators can be found in section +*77*
2.% Fuzzy &u'bers
%n this section we will briefly review some fre5uently used classes of fuzzy numbers*
'e*inition ,: Fuzzy Number
A fuzzy number A is described as any fuzzy subset of the real line with membership function
j
A
which possesses the following propertiesC9<D6
"a#
j
A
is a continuous mapping from to the closed interval | ,( o( ,<o<78
"b#

A
( x)=,(
for all x|( a 8
"c#
j
A
is strictly increasing on | a( b 8
"d#
j
A
( x)=o(
for all x| b( c( where o is a constant and ,w78
<
Figure 3- Te basic set o(erations-
"e#
j
A
is strictly decreasing on | c( d 8
"f#
j
A
( x)=,(
for all x| d ,8
where a ,b , c and d are real numbers* Unless elsewhere specified( it is assumed that A is conve2 and
bounded8 i*e* a ( d. %f o=7 in "d#( A is a normal fuzzy number( and if ,w7 in "d#( A is
a non-normal fuzzy number* /or convenience( the fuzzy number in definition 7 can be denoted by
A=(a ( b( c ( d 8 o)* 0he image "opposite# of A can be given by A=(d (c(b(a 8 w) *
Property "a# can also be written as A6 -| ,( 7 0he membership function of
j
A
can be e2pressed
as
j
A
( x)=

j
A

( x) , axb
o, bxc
j
A
!
( x) , cxd
,( otherwise ,
"-#
where j
A

:| a , b -| ,(o and j
A
!
:| c , d -|,( o*
'e*inition 2: Triangu%ar Fuzzy Number
A triangular fuzzy number A is a fuzzy number with a piecewise linear membership function
j
A
defined by
j
A
=

xa
7
a
+
a
7
, a
7
xa
+(
a
9
x
a
9
a
+
, a
+
xa
9(
,( otherwise ,
"7,#
which can be denoted as a triplet
(a
7(
a
+(
a
9
) *
'e*inition .: Tra(ezoi+a% Fuzzy Number
A trapezoidal fuzzy number A is a fuzzy number with a membership function
j
A
denoted by
j
A
=

xa
7
a
+
a
7
, a
7
xa
+
7( a
+
xa
9
a
'
x
a
'
a
9
, a
9
xa
'
,( otherwise ,
"77#
>
which can be denoted as a 5uartet
(a
7(
a
+(
a
9(
a
'
)*
2.( "an)in*
Ran=ingC9>(9-D is the tas= of comparing fuzzy subsets "i*e* numbers# and arranging them in a
certain order* specially in decision ma=ing situations( appropriate methods are needed to compare
and evaluate different alternatives( i*e* which fuzzy number precedes the other in a given situation*
Ahen dealing with strictly numerical values( this process is 5uite simple( since the order can be
naturally determined* /uzzy numbers( on the other hand( cannot be ordered in the same manner
because the same natural order does not e2ist in fuzzy numbers* 0hat is( we cannot e2plicitly say
that a fuzzy number A is larger than another fuzzy number B as in the numerical case* Ahether A is
larger( smaller or e5ual to B is a matter interpretation* A simple method for ordering fuzzy subsets
consists in the definition of a ran=ing function "( mapping each fuzzy set to the real numbers (
where a natural order e2ists* &uppose
#=j
A$
,j
A%
,., j
An

is a set of n conve2 fuzzy numbers( and

the ran=ing function " is a mapping from # to the real numbers ( i*e* " 6 # -* 0hen for any
distinct pair of fuzzy numbers
j
Ai
, j
A&
# (
the ran=ing function can be defined as
if " (j
Ai
) " (j
A&
) 8 then j
Ai
j
A&
if " (j
Ai
) =" (j
A&
) 8 then j
Ai
=j
A&
if " (j
Ai
) >" (j
A&
) 8 then j
Ai
>j
A&
*
"7+#
0his implies( for e2ample( that if
" (j
Ai
) > " (j
A&
)(
the fuzzy number
j
Ai
is numerically greater than
the fuzzy number
j
A&
*
0he higher
j
Ai
is( the larger
" (j
A&
)
is* A useful techni5ue for ran=ing normal
fuzzy numbers that are conve2( such as triangular- and trapezoid fuzzy numbers( is the centroid(
defined by
" (j
A
)=

xj
A
( x) dx

j
A
( x) dx
( "79#
where
" (j
A
)
represents the centroid of the fuzzy set
j
A
*
2.+ ,in*uistic -ariables
/uzzy numbers are fre5uently used to represent 5uantitative variables( normally referred to as
lin'uistic variables C9:(9;D* ?inguistic variables ta=e words or sentences as values( as opposed to an
algebraic variable which ta=es numbers as values* All values are ta=en from a term set that contains
the set of acceptable valuesGconcepts* ach valueGconcept in the term set is represented by a fuzzy
number which is defined over some universe interval( also called a base variable* %n short this
relationship can be e2pressed as follows6 linguistic variable K fuzzy variable K base variable* /or
-
e2ample( let v be a linguistic variable denoting a person3s height* 0he values of v( which are fuzzy
variables( could be defined by the term set T L Mver( short, short, medium tall, tall, ver( tallN and
the associated base variable could span the interval from 7,, to ++, cm*
An e2ample of a linguistic variable is shown in figure ;* %ts name is 1eight and it e2presses the
height of a person in a given conte2t by five linguistic terms - very short( short( medium tall( tall
and very tall* ach of the basic linguistic terms is assigned one of five trapezoidal fuzzy numbers
which define the range of the base variable*
2.. Defuzzification
0wo central concepts in fuzzy set theory are fuzzification and its counterpart defuzzification*
/uzzification is the process of converting crisp values into fuzzy values by identifying possible
uncertainties or variations in the crisp values* 0his conversion is represented by membership
functions* 0here are various ways this fuzzification process can be carried out( li=e intuition(
genetic algorithms C9'D or neural networ=s C9'D* Defuzzification is the process of converting fuzzy
values into crisp ones* %n the following we will describe some defuzzification methods found in
literature C9'(',D*
2...1 /a01'e'bershi rincile
%n this method( the defuzzified value( )( e5uals the x-value with the highest membership
degree* %t is given by the e2pression
j
A
( x
O
)j
A
( x) for all xA( "7'#
where
x
O
is the value with the highest degree of membership in the fuzzy set A* %f we consider the
following set A=,*9/ 7,+,*':/7++,*;/ 7:+,*-/ 7<* 0hen ma2 ( A)=7<*
7,
Figure 5- A %inguistic re(resentation o* te *uzzy set 9eigt-
2...2 Centroid 'ethod
0his is the most widely used method* %t is also referred to as the center of gravity or center of
area method* %t can be defined by e5uation 79 when
j
A
is continuous* /or the discrete case in which
j
A
is defined on a finite universal set
x
7(
x
+
,., x
n
(
the e5uation is
)=
_
i =7
n
j
A
( x
i
) x
i
_
i=7
n
j
A
( x
i
)
* "7:#
Using the e2ample from before( we get
)=
(,*97,)+(,*':7+)+(,*;7:)+( ,*-7<)
,*9+,*':+,*;+,*-
=7'*:9
2...3 2ei*hted avera*e 'ethod
0his method is only applicable for symmetrical membership functions* %t bears some
resemblance to the centroid method( e2cept it only includes the ma2imum membership value of
membership functions* 0he e2pression is given as
)=
_
ma2(j
A
( x)) x
_
ma2 (j
A
( x))
( "7;#
where
ma2 (
A
( x))
is the ma2imum membership degree of membership function
j
A
and x is the
corresponding value* Assume we have two functions(
j
A
and
j
B
(
where
ma2 (
A
)=,*>
and
ma2 (
B
)=,*<:*
0he corresponding points on the 2-a2is are a and b( respectively* 0hen the
defuzzified value ) can be obtained as
)=
(a,*>)+(b,*<:)
,*>+,*<:
2...! /ean1'a0 'e'bershi
0his method is similar to the ma2-membership principle( e2cept the ma2imum does not
necessarily have to be uni5ue* 1ence the ma2imum membership degree may include more than a
single point( it may include a range of points* 0he e2pression is given as
)=
a+b
+
( "7<#
where a and b are the end points of the ma2imum membership range*
77
2.13 Fuzzy "elations
A relation signifies a relationship between set elements of two or more sets* $risp relations can
be defined by a characteristic function which assigns a value from the binary pair M,(7N to each
subset of the universe set( where , implies no association and 7 implies an association* 0he
$artesian product of two sets A and B( denoted AB( is the set of all possible combinations of the
elements in A and B. All relations are subsets of the $artesian product which therefore represents
the universe set* A fuzzy relation C9'(',D is a fuzzy set defined on the $artesian product of crisp
sets* ach element within the relation may then be associated to a degree between , and 7( in the
same manner as set membership is represented in fuzzy sets* 0he grade indicates the strength of the
relation present between the elements*
0o e2press this more formally( we consider a fuzzy relation between two crisp sets * and +.
0hen a fuzzy relation R is a mapping from * -+ from the $artesian space( * + ( to the unit
interval* 0he strength of the mapping is e2pressed by the membership function(
j
!
( x , () (
of the
relation for all ordered pairs ( x , ()* + * 0his function can be e2pressed as
j
!
( x , () L j
AB
( x , () L min(j
A
( x) , j
B
( ())*

"7>#
0his means that each fuzzy set can be regarded as a vector of membership values where each value
is associated with a particular element in each set* %f we consider two fuzzy sets A, containing four
elements( and B( containing five elements* 0hen A is e2pressed as column vector of size 97 and B
a column vector of size 7+* 0he corresponding relation will be a 9+ matri2* 0hat is( a matri2
with four rows and five columns "note6 the resulting matri2 has the same number of rows as A and
the same number of columns as B#* ?ets illustrate this by an e2ample where A is defined on the
universe set
* L x
7
, x
+
, x
9
(
and B is defined on the universe set
+ =(
7
, (
+
*
Ae then have the
two following vectors
A L ,*'/ x
7
+,*</ x
+
+,*7/ x
9
and B L ,*:/ (
7
+,*>/ (
+
*
0he resulting matri2 is then obtained by ta=ing the minimum of each pair of ( x , ()AB* /or
e2ample( the entries
( x
7
, (
+
)
and
( x
+
, (
7
)
of the matri2 are derived by ta=ing the minimum of the
pairs (,*'( ,*>) and (,*<( ,*:) * 1ence the relational matri2 of A and B loo=s as
(
7
(
+

AB L ! L
x
7
x
+
x
9
|
,*' ,*'
,*: ,*<
,*7 ,*7

*
7+
2.13.1 Fuzzy "elational Co'ositions
Relations can be combined in various ways by using the union or the intersection operator* A
generic way to compose fuzzy relations is to pic= the minimum value in a series connection and the
ma2imum value in a parallel connection* Fecause a relation is itself a set( the basic set operations
such as union( intersection( and complement can be applied without modifications* 0he standard
composition ! of two fuzzy relations , and -( normally written as ! L ,-( is formally defined
by
!( x , .)=| ,-( x , .)=ma2
(+
min| j
,
( x , () , j
-
( ( , .)(
"7-#
for all x* and all . ) * ?ess formally this means that the i&-entry of the matri2 ! is derived by
combining the ith row of P with the &th column of -* Using matri2 notation( the same e2pression
can can be written as
| r
i&
L | p
i/
| 0
/&
L ma2 min( p
i/
, 0
/&
)
An illustrative e2ample of a ma2-min composition of two fuzzy sets is shown below
,- L
|
,*< ,*;
,*> ,*9

|
,*> ,*: ,*'
,*7 ,*; ,*<

L
|
,*< ,*; ,*;
,*> ,*: ,*'

*
/or e2ample(
r
7(+
L ma2| min( p
77
, 0
7+
) , min( p
7+
, 0
++
)
L ma2| min(,*<( ,*:) , min(,*;( ,*;)
L ,*;(
r
+(+
L ma2| min( p
+7
, 0
7+
) , min( p
++
, 0
++
)
L ma2| min(,*>( ,*:) , min(,*9( ,*;)
L ,*:*
Another related operation is the min-ma2 operations which is derived in an analogous manner to the
ma2-min operation*
A second e2ample of a relational composition is the ma2-product which is defined by
!( x , .)=| ,-( x , .)=max
(+
| j
,
( x , ()j
-
( ( , . )
"+,#
for all x* and .) * 1ere the min-operator has been replaced by the multiplication operator but
the entries are combined between matrices in the same manner* Fy reusing the previous e2ample(
we get
,- L
|
,*< ,*;
,*> ,*9

|
,*> ,*: ,*'
,*7 ,*; ,*<

L
|
,*:; ,*9; ,*'+
,*;' ,*' ,*9+

*
/or e2ample(
79
r
77
L max |( p
77
0
77
) , ( p
7+
0
+7
)
L max |(,*<,*>) ,(,*;,*7)
L ,*:;(
r
+9
L max| ( p
+7
0
79
) ,( p
++
0
+9
)
L max| (,*>,*') ,(,*9,*<)
L max| (,*9+) ,(,*+7)
L ,*9+*
A third e2ample of a compositional operation is the ma2-average composition( defined by
!( x , .)=| ,-( x , .)=
7
+
max
(+
| j
,
( x , ()+j
-
( ( , .) "+7#
for all x* and .) * $ompared to the previous e2pression( it can be seen that the multiplicative
operator has been replaced by an additive operation such that we now obtain the ma2imum of the
averages between corresponding pairs* Again by using the same e2ample( we get
,- L
|
,*< ,*;
,*> ,*9

|
,*> ,*: ,*'
,*7 ,*; ,*<

L
|
,*<: ,*; ,*;:
,*> ,*;: ,*;

*
/or e2ample(
r
77
L 7/ + ma2|(j
p$$
+j
0$$
) , (j
p$%
+j
0%$
)
L 7/ + ma2|(,*<+,*>) ,(,*;+,*7)
L 7/ + ma2|(7*:) ,(,*<)
L ma2 |(,*<:) ,(,*9:)
L ,*<:(
r
+9
L 7/ +ma2| (j
p%$
+j
0$1
) ,(j
p%%
+j
p%1
)
L 7/ +ma2| (,*>+,*') ,(,*9+,*<)
L 7/ +ma2| (7*+) ,(7*,)
L ma2| (,*;) ,(,*:)
L ,*;*
2.11 A**re*ation
0he purpose of aggregation is to aggregate pieces of data in a desirable way in order to reach a
conclusion or final decision* 0ypically this data is represented by numerical values which ma=e
some =ind of sense in regard to the application* 1ence the aggregation problem is generally
regarded as the problem of reducing a series of numerical values into a single representative*
/ormally an aggregation operator can be defined as function h which assigns a real number ( to any
n-tuple
( x
7(
x
+
****** , x
n
)
of real numbers C'7D6
(=h( x
7(
x
+
****** , x
n
)* "++#
%n literature( aggregation operations are generally defined over the unit interval( meaning that it
is assumed that both the input and output is restricted to the unit interval* An aggregation operation
7'
of dimension n2 7 can therefore be formally described as mapping over the unit interval
h6 | ,(7
n
-| ,(7 *
"+9#
0he case n=7 is represented by the negation operator defined by 3h( x)=7h( x)* /or n2 +( two
classes of operators are of particular interest in fuzzy theory8 triangular norms and the averaging
operators*
ven though the inputGoutput of aggregation operations often times is restricted to the unit
interval( this is not a mandatory characteristic of aggregation operators* 1ence the definition above
can be e2tended to arbitrary intervals as well* %n this conte2t however( we assume that the
inputsGoutputs of the aggregation operators discussed in this thesis are from the unit interval( unless
otherwise stated*
?ogically( certain conditions are e2pected to be imposed on the function h in order for it to
5ualify as an aggregation operator( although there are different views on which basic properties
should be fulfilled( since aggregation operators fre5uently are designed for certain applications* 0he
most important thing in this case is not whether a given aggregation operator satisfies all of the
basic properties associated with aggregation operators( but whether the aggregation operator in
5uestion produces a meaningful outcome from an applicative conte2t( which in turn necessitates the
presence of certain constraints* &ome of the fundamental properties fre5uently associated with
aggregations operators are enlisted below C'7D6
7# h( x)=x "i+entity )en unary#8
+# h(,(., ,)=, and h(7(.,7)=7 "boun+ary con+itions#8
9#
h( x
7
,., x
n
) 4 h( (
7
,., (
n
)
"monotonic increasing#
if
x
i
4 (
i
for all i N8
'# h is continuous with respect to each of its arguments8
:#
h( x
7
,., x
n
)=h( x
p( i)
,., x
p( n)
)
for all permutations p "symmetry#8
;# h( x , x ,., x)=x "i+em(otency#8
<#
h( x
7
, x
+
, x
9
)=h(h( x
7
, x
+
) , x
9
)=h( x
7
, h( x
+
, x
9
))
"associativity#8
># h
| n
( x
7
,., e ,., x
n
)=h
| n7
( x
7
, ., x
n
) "neutra% e%ement#8
-# h
| n
( x
7
,*** , a , *** , x
n
)=a "absorbent e%ement#8
7,#
min( x
7
,****** , x
n
)4 h( x
7
, ****** , x
n
)4 ma2 ( x
7
, ****** , x
n
)
"com(ensation#*
7:
Property 7 only applies to unary operators( i*e* operators ta=ing one argument only* According to
this property( the aggregated result e5uals x if h is unary* Property + defines the worstGbest case
behaviour of aggregation operators* %f the argument x
i
is either completely false "x
i
L ,# or
completely true "x
i
L 7# for all i N( then the aggregated result should reflect the same behaviour*
0he properties of boundary conditions can easily be e2tended to inputGoutputs outside the unit
interval* &ometimes the boundary conditions are e2tended as follows C'+D6
+*7# x [,(7] h (x(,) = h (7(,) x
+*+# x [,(7] h (x(7) = (7 h (7(,)) x + h (7(,).
0hese e2tensions add more constrains to the basic re5uirements of aggregation operators since they
e2clude every aggregation operator which is not an averaging operator* Property +*7 presumes the
value of h"x(,# to be the weighed arithmetic mean of x and ,( and property +*+ presumes the value
of h"x(7# to be the weighted arithmetic mean of x and 7* Actually the re5uirements of property + are
special cases of +*7 and +*+ when x L , and x L 7( respectively* Property 9 states that the aggregated
result as minimum does not decrease if the argument increases* Property ' "continuity# ensures that
a changes in arguments will not result in discontinuous change in the aggregate value* specially
the properties of +-' are considered fundamental to aggregation operators in general C9'D* Property
: "symmetry# is related to the order of arguments in the sense that the order should not have any
influence on the aggregated result* 0his is particularly relevant when all arguments are assumed to
be e5ually important* Another related property associated to n-ary operators with n
+
arguments is
bis(mmetr( C'7D* 0his property simply states that it does not matter whether the aggregation is done
vertically or horizontally( if h is an n-ary operator* Ae can write this as
h(h( x
77
, x
7+
,., x
7n
) , h( x
+7
, x
++
,., x
+n
) ,., h( x
n$
, x
n%
,., x
nn
)) L
h(h( x
77
, x
+7
,., x
n$
) , h( x
7+
, x
++
,., x
n%
) ,., h( x
7n
, x
+n
,., x
nn
))*
"+'#
0his implies that you can either aggregate the column vectors first and then the outputs thereof or
the row vectors first and then the outputs thereof* %t should be noted that symmetry and associativity
implies bisymmetry( but neither symmetry or associativity is implied by bisymmetry* Property ;
"idempotency# states that if x is aggregated n times( the final outcome will be x as well* 0his
condition may be warranted in cases where x is a fuzzy set( because aggregating e5ual sets logically
implies the same set* Property <"associativity# reflects the notion that aggregation is done in
pac=ages but the order of pac=ages has no influence on the aggregated result* Property > "neutral
element# assumes the e2istence of a neutral element e which has no influence on the result when
applied* Property - "absorbent element# assumes the e2istence of an absorbent element a which acts
7;
as an annihilator* Property 7, "compensation# relies on the assumption that the aggregated result is
somewhere between the lowest argument "min# and the highest argument "ma2#* 0his condition is
only valid for averaging operators*
2.11.1 Avera*in* $erators
A type of operator widely studied in literature are averaging operators* Averaging operators of
dimension n2 + can be described by the mapping h: |,(7
n
| ,(7 which meets the following
a2iomatic re5uirements C'9(''D6 "7# h is symmetric8 "+# h is monotonic increasing8 "9# h is
continuous8 "'# h is idempotent* %t has to be noted that the assumption of symmetry may not always
be warranted in every application conte2t* %n that case( the assumption of symmetry has to be
dropped* A good e2ample of an operator where this property is not meet is the weighted average
mean* &ome commonly used averaging operators are listed in table 7*
:(erator E;uation
0he arithmetic mean
7
n
_
i =7
n
x
i
Aeighted arithmetic mean
_
i =7
n
w
i
x
i
where w
i
| ,(7 and
_
i=7
n
w
i
=7
Beometric mean
(
]
i =7
n
x
i
)
7
n
1armonic mean n
_
i =7
n
7
x
i
Puadratic mean
.
7
n
_
i =7
n
x
i
+
4edian &ort the arguments in ascending
order* %f the number of arguments
n is odd( then the middle value is
selected* %f n is even( then ta=e the
mean of the middle pair*
4in and ma2 min ( x
7
, ., x
n
)
ma2 ( x
7
, ., x
n
)
Tab%e ,- E4am(%es o* common%y use+ averaging o(erators-
A common characteristic of aggregation operators is that they cover the entire interval between
min and ma2* 0hat is( any aggregation operator( h"x
7
(***( x
n
#( satisfies the following ine5ualities "a=a
compensation property# C9'D 6
\( x
7
, .x
n
)|,(7
n
6 min( x
7
, ., x
n
) <h( x
7
,., x
n
) <ma2( x
7
,., x
n
)* "+:#
7<
0o show this( let x
min

L min and

x
ma2
L ma2* &ince every averaging operator satisfies the properties of
monotonicity and idempotency( it also satisfies the ine5ualities6
x
min
=h( x
min
,., x
min
) 4 h( x
7
,., x
n
) 4 h( x
max
,., x
max
) = x
max
* "+;#
0his means that all averaging operators are bounded by the standard fuzzy union and the standard
fuzzy intersection operations* $onversely( it follows that all operators bounded by the standard
fuzzy union and standard fuzzy intersection are idempotent since
x=h( x ,., x) 4 h( x ,., x) 4 h( x ,., x)=x
\ x|,(7 *
"+<#
2.11.1.1 4eneralized /eans
4any of the common means belong to the family of the 5uasi-arithmetic means C'7(''D(
defined as
h( x
7
,., x
n
)= f
7
(
7
n
_
i=7
n
x
i
)
( "+>#
where f is a continuous strictly monotonic function( and f
- 7
is its inverse* %t can be noted that the
geometric mean and the harmonic mean are particular cases of "+># with f"x# L log x and f"x# L 7Gx(
respectively* A particularly noticeable case of 5uasi-arithmetic means can be obtained by applying
the function6 f : x-x
o
* Ae can then obtain a 5uasi-arithmetic mean of the form6
h
o
( x
7
,., x
n
)=
|
7
n
_
i=7
n
x
i
o

7
o
*

"+-#
0his class of means is often referred to as power means or 'enerali.ed means because a common
group of well-=nown means can be generalized by changing the J parameter6
e /or J L 7 we obtain the arithmetic mean*
e /or J L + we obtain the s5uare mean "a=a euclidean mean#
e /or J L -7 we obtain the harmonic mean*
e Ahen J converges towards -Q( h
J
converges towards minimum*
e Ahen J converges towards Q( h
J
converges towards ma2imum*
e Ahen J converges towards ,( h
J
converges towards the geometric mean*
0he class of power means can be e2tended with weights as well such that we get weighted
7>
power means( defined by the e5uation6
h
o
(( w
7
, x
7
) ,.,( w
n
, x
n
))=
|
7
n
_
i=7
n
w
i
x
i
o

7
o
where w
i
| ,(7 \i and
_
i=7
n
w
i
=7*
"9,#
.ther well-=nown means can be generalized as well using weighted power means by changing
the J parameter6
e /or J L 7( h
J
e5uals the weighted arithmetic mean*
e /or J L +( h
J
e5uals the weighted s5uare mean*
e /or J L -7( h
J
e5uals the weighted harmonic mean*
e Ahen J converges towards -Q( h
J
converges towards minimum*
e Ahen J converges towards Q( h
J
converges towards ma2imum*
e Ahen J converges towards ,( h
J
converges towards the weighted geometric mean*
2.11.1.2 $rdered 2ei*hted Avera*in* $erators 5$2A6
.ne of the most widely studied operator in fuzzy theory is the .AA operator C'9(''D* 0his
operator is mainly used for aggregating scores associated with the satisfaction of multiple criteria*
An .AA operator of dimension n R + can be described as the function6
56A( x
7(
x
+(
., x
n
)=
_
& L7
n
w
&
x
p( &)

where w
i
| ,(7 and
_
i L7
n
w
i
L7
"97#
and p is a permutation that orders the arguments in a non-increasing order6
x
p( 7)
2 x
p( +)
2*** 2 x
p(n)
*
&ome special cases of .AA( when choosing particular weights( are displayed in table +*
7-
:/A )eigts
4a2imum

w
7
=7
w
i
=, if i =7
4inimum

w
n
L 7
w
i
=, if i =n
Arithmetic mean w
i
L
7
n
\i
4edian

w
n+7
+
L7 if n is odd
w
n
+
=
7
+
and w
n
+
+7
L
7
+
if n is even
w
i
=, else*
Tab%e 2- S(ecia% cases o* :/A C'7D-
An important aspect with respect to .AA is the derivation of appropriate weights which should
represent the problem at hand as closely as possible* 0wo measures of importance in this regard are
orness and dispersion( defined by6
orness ( w)=
7
n7

_
i L7
n
(ni )w
i

"9+#
and
dispersion( w)=
_
i L7
n
w
i
ln (w
i
)| ,( ln( n)
"99#
0he dual measure of orness is referred to as andness( and is defined by 7orness( w)* 0he
dispersion measure reflects the degree of utilization of the information in the argument vector* 0he
more evenly distributed the weights are( the higher dispersion* A normalized dispersion measure to
the unit interval can obtained by dividing by ln7n8 such that the e5uation becomes6
ndispersion ( w)=
7
ln(n)
_
i L7
n
w
i
ln( w
i
)| ,(7
"9'#
.rness and andness can be interpreted as the degree to which an .AA operator represents pure .R
"i*e* ma2# or pure A)D "i*e* min#( respectively* 0he degree of orness of the arithmetic mean is ,*:(
as can be seen from the following e2ample* $onsider the vector of .AA weights
w=(,*+( ,*+(,*+( ,*+( ,*+)* 0he orness can then be calculated as6
orness ( w) =
7
'
(,*>+,*;+,*'+,*+)
=
7
+
*
&o if orness e5uals ,*:( we obtain neutrality with the arithmetic mean* %f orness is strictly less than
+,
,*:( we move towards min( and thereby also the pure A)D* %f orness is strictly larger ,*:( we move
towards ma2( and thereby the pure .R*
$learly( it is possible to obtain the same degree of orness for different weight vectors* Fy using
the dispersion measure( we are able to further distinguish between the .AA weights* 0o illustrate
this( consider the two .AA vectors w
7
=(,(,(7(,(,)
and w
+
=(,*+(,*+(,*+(,*+(,*+)*
0hese two
vectors respectively correspond to the median and the arithmetic mean* 0he orness and the
dispersion for these vectors are calculated as6
orness ( w
7
) =
7
:7
((:7),+(:+),+(:9)7+(:'),+(::),)
=,*:
ndispersion( w
7
) =
(,ln (,))+(,ln(,))+(7ln(7))+(,ln(,))+(,ln( ,))
ln(:)
=,
orness ( w
+
) =,*:(as already shown)
ndispersion( w
+
) =
(,*+ln (,*+))+(,*+ln (,*+))+(,*+ln(,*+))+(,*+ln(,*+))+(,*+ln (,*+))
ln(:)
=7
As can be seen from this e2ample( different results for the normalized dispersion can be obtained(
despite the same degree of orness for the two vectors*
2.11.2 Trian*ular &or's 5T1&or's and T1Conor's6
Another class of operators which have been e2tensively studied in literature( are the so-called
triangular norms C'7(':D which can be divided into two basic operations( namely the t-norm and its
dual the t-conorm* %n fuzzy set theory( the t-norm defines the union and the t-conorm defines
intersection of fuzzy sets* 0his ma=es it possible to use these to characterize the logical connectives
of A)D and .R( respectively*
A t-norm is a function T 6| ,(7
+
-|,(7 which satisfies the following a2ioms6
e T ( x , () =T ( ( , x) "07# commutativity
e T ( x , () <T ( u , v) , if x <u ( <v "0+# monotonicity "increasing#
e T ( x ,(T ( ( , .)) =T (T ( x , () , .)) "09# associativity
e T ( x(7) = x "0'# , as neutra% e%ement
0he result of applying the t-norm operator will never be larger than the minimum of arguments*
/ormally this can be written as6
+7
\t-norms T 6 T ( x , () <min( x , ()* "9:#
Ae can prove this as follows6
7* /rom 0+ and 0' we get T ( x , () <T ( x(7) = x*
+* /rom 07( 0+ and 0' we get T ( x , () <T (7( () = ( *
0hat is T ( x , () < x and T ( x , () < (( hence T ( x , () <min( x , () *
A t-conorm is a function # 6| ,(7
+
-|,(7 which satisfies the following a2ioms6
e #( x , () =#( ( , x) "&7# commutativity
e #( x , () <#( u , v) , if x <u ( <v "&+# monotonicity "increasing#
e #( x ,( #( ( , .)) =# ( #( x , () , .)) "&9# associativity
e #( x (,) = x "&'# 0 as neutra% e%ement
/rom an a2iomatic point of view( t-norms and t-conorms only differ with the respect to their
boundary conditions or neutral element which is 7 and ,( respectively* 0he result of applying the t-
conorm operator is never less than the ma2imum of arguments* 0he formal notation is6
\t-conorms # 6 # ( x , () ma2( x , ()* "9;#
0he proof is trivial and analogous to the one shown previously*
)orm operations are always defined as binary operations( but due to their associative
properties( they can be generalized for n arguments* /or e2ample( the multi-argument forms for the
min and ma2 operators are6
T
i=7
n
( x
i
)
min
=min
i=7
n
( x
i
) and #
i=7
n
( x
i
)
ma2
=ma2
i=7
n
( x
i
)( respectively* "9<#
/or the algebraic product and the algebraic sum( the multi-argument forms are6
T
i=7
n
( x
i
)
ap
=
]
i =7
n
( x
i
) and #
i=7
n
( x
i
)
as
=7
]
i=7
n
(7x
i
)( respectively*
"9>#
Benerally multi-argument forms are trivial with the algebraic sum as an e2ception* 0herefore( the
derivation of the multi-argument form of the algebraic sum is shown as a proof of induction in
Appendi2 %*
2.11.2.1 Duality of T1&or's and T1Conor's
Any t-norm is associated with a dual t-conorm and vice versa C'7(':D* A t-norm and a t-conorm
are said to be dual if the law of De 4organ is satisfied6
++
T ( x , ()=#( x , () ( "9-#
where x denotes the standard negation( defined by x=7 x* &ome common t-norms and their dual
t-conorms are listed in table 9*
t&norm t&conorm
min an+ ma4 min"x( (# ma2"x( (#
a%gebraic (ro+uct an+ sum x 9 ( x S ( - x 9 (
<u0asie)icz t&norm an+ t&conorm ma2"x S ( - 7(,# min"x S ( (7#
+rastic (ro+uct an+ sum

, if ( x , ( )C ,(7C
+
(
min( x , () otherwise*

7 if ( x , ()D ,(7D
+
ma2( x , () otherwise*
Tab%e .- Common t&norms an+ teir +ua% t&conorms-
0he minimum or min is the largest t-norm* %t is also the only idempotent t-norm and thus the only t-
norm which is an averaging operator as well* %ts dual t-conorm( i*e* the ma2 operator( is the smallest
t-conorm* %t is the only idempotent t-conorm and thus the only t-conorm which is an averaging
operator as well* 1ence the min and ma2 respectively define the lower and upper bounds of
averaging operators* 0he drastic product is interesting from the point of view that it yields the
smallest t-norm and the largest t-conorm*
2.11.3 Avera*in* $erators and Trian*ular &or's in Conte0t
Figure 6- Te re%ationsi( bet)een triangu%ar norms an+ averaging o(erators C''D -
+9
/igure < summarizes the relationship between the different classes of operators discussed in the
previous sections* %t can be seen from the figure that the boundary between averaging operators and
triangular norms is defined by the min and ma2 operators* Recall that the result of a t-norm
operation is always T min( and for t-conorm operations( the result is always R ma2* %n particular(
these operators are important when distinguishing between triangular norms and averaging
operators because the min and ma2 are the only idempotent triangular norms and thereby the only
triangular norms that are averaging operators as well* 4oreover these operators are the only
associative averaging operators*
Averaging operators satisfy the compensation property which implies that an averaging
operator always yield a result between min and ma2* Aeighted averaging operators( li=e .AA( can
be regarded as parametrized ways of moving between min and ma2* 4oving towards min "or ,#
corresponds to moving towards pure A)D* As we move closer towards A)D( the more restrictive
the operator becomes( since pure A)D re5uires all arguments to be satisfied* 0his is e5uivalent to
the universal 5uantifier which states that all arguments must be fulfilled* At the opposite end of the
e2treme( we have ma2( corresponding to the pure .R* 0his is e5uivalent to the e2istential
5uantifier( which states( that there e2ists at least one argument which is fulfilled* &o( the further we
move towards ma2 "or 7#( the less restrictive the operator becomes* Fetween these two e2tremes
different levels of strictness can be specified* /or e2ample( a 5uery may be satisfied if Imost offI
the arguments are fulfilled or Iat least a fewI*
/rom an a2iomatic point of view( triangular norms and averaging operators have the
symmetry( monotonicity and continuity a2ioms in common* 0he a2ioms regarding associativity and
the e2istence of a neutral element only applies to triangular norms* %n fact( triangular norms cover
all aggregating operations which are associative C9'D* %dempotency( on the other hand( only applies
to triangular norms*
2.12 7article S8ar' $ti'ization 57S$6
P&. C';('<D is an optimization techni5ue applicable to continuous non-linear functions* %t was
first introduced by Uennedy and berhart in C';D* 0he algorithm simulates the social behaviours
shown by various =inds of organisms such as bird floc=ing or fish schooling* %magine a group of
birds randomly foraging in an area* 0he group shares the common goal of locating a single piece
food* Ahile foraging( individual birds may learn from the discoveries and past e2periences of other
birds through social interaction* ach bird synchronizes its movements with group while
simultaneously avoiding collisions with other birds* As the search continues( the birds move closer
+'
toward the place where the food is by following the bird which is closest to the food*
%n P&.( bird floc=s are represented as particle swarms searching for the best solution in a
virtual search space* A fitness value is associated to each particle which is evaluated against a
fitness function to be optimized( and the movement of each particle is directed by a velocity
parameter* During each iteration( particles move about randomly within a limited area( but
individual particle movement is directed toward the particle which is closest to the optimal solution*
ach particle remembers its personal best position "the best position found by the particle itself# as
well as the global best position "the best solution found by any particle in the group#* 0he
parameters are updated each time another best position is found* 0his way( the solution evolves as
each particle moves about*
$ompared to other related approaches such as genetic algorithms and neural networ=s( P&. it
is 5uite simple and easy to implement* %t is initialized with a set of randomly generated particles
which in fact are candidate solutions* An iterative search process is then set in motion to improve
the set of current solutions* During each iteration( new solutions are proposed by each particle
which are individually evaluated against6 "7# the particles own personal best solution found in any
proceeding iteration and "+# the global best solution currently found by any particle in the swarm*
Ae refer to each candidate solution as a position* %f a particle finds a position better than its current
personal best position( its personal best position is updated* 4oreover( if the new personal best
position is better than the current global best position( the global best position is updated* After the
evaluation process is completed( each particle updates its velocity and position with the e5uations6
v
i
=ov
i
+c
7
r
7
( x
i
x
&
)+c
+
r
+
( 'x
&
) "',#
x
&
=x
&
+v
i
, "'7#
where
e v
i
is the velocity of particle p
i
and is limited toC-:
ma2
( :
ma2
D where :
ma2
is user-defined
constant*
e V is an inertial weight coefficient*
e ;<i
is the current personal best position*
e x
!
is the present position*
e W is the global best position*
e c
7
and c
+
are user defined constants that say how much the particle is directed towards good
positions* 0hey affect how much the particle3s local best and global best influence its
+:
movement* Benerally c
7
and c
+
are set to +*
e r
7
and r
+
are randomly generated numbers between , and 7*
)ote that the velocity controls the motion of each particle* 0he speed of convergence( can be
ad!usted by the inertial weight coefficient and the constants c
7
( c
+
* Ahenever computed velocity
e2ceeds its user-defined boundaries( the computed results will be replaced by either -:
min
or :
ma2
*
0he running procedure of basic P&. algorithm is summarized in pseudo code below*
Te basic $S: a%goritm
*or all particlesM
initialize velocities and positions
NGGend for
)i%e stopping criteria is unsatisfied=
*or each particleM
7* compute velocities by e5uation "',#
+* increment positions by e5uation "'7#
i* present fitness value is better than current local best value
9* update local best positions
i* present fitness value is better than current global best value
'* update global best positions
NGGend for
NGGend while
2.13 Fuzzy Ti'e Series and its Concets
%n the following section we will briefly review some of the fundamental concepts of /0& as
they originally were conceived in pioneering publications by &ongG$hissom C7(+('>D and $hen C9(
97D*
'e*inition ,: Fuzzy Time Series
?et + (t )(t =*** , ,(7(+(* **)( a subset of real numbers( be the universe of discourse on which fuzzy sets
f
i
(t )(i =7(+, ***)
are defined* %f " (t ) is a collection of
f
i
(t )(i =7(+(***)(
then " (t ) is called a fuzzy
time series on + (t )(t =*** , ,(7(+(* **)*
'e*inition 2: Fuzzy "e%ation
%f there e2ists a fuzzy relationship !(t 7( t ) ( such that " (t )=" (t 7)!(t 7( t )( where
represents an operator( then " (t ) is said to be caused by " (t 7)* 0he relationship between " (t )
and "(t 7) is denoted by
" (t 7) -"(t )* "'+#
+;
2amples of operators from literature are the ma2-min composition "see section +*7,*7# C7D( the
min-ma2 composition C+D and the arithmetic operator C9D* %f
" (t 7)=A
i
and " (t )=A
&
(
the logical
relationship between " (t ) and "(t 7) is denoted by
A
i
-A
&
(
where
A
i
is called the left hand side
and
A
&
the right hand side of the fuzzy relation* 0he variable t denotes the time* /or e2ample( if t L
7-<9( the fuzzy relationship between " (t ) and "(t 7) is given by " (7-<+)-" (7-<9) * )ote the
right hand side of the fuzzy relation represents the future fuzzy set "forecast#* %ts crisp counterpart is
denoted as + (t )*
'e*inition .: N&:r+er Fuzzy "e%ations
?et " (t ) be a fuzzy time series* %f " (t ) is caused by " (t 7) ( " (t +) (.( "(t n)( then this fuzzy
relationship is represented by
" (t n) (.( " (t +)( "(t 7)-" (t )( "'9#
and is called an n=order fuzzy time series* 0he n-order concept was first introduced by $hen in C97D*
>-order based /0& models are referred to as high order models*
'e*inition 4: Time&Invariant Fuzzy Time Series
&uppose " (t ) is caused by "(t 7) only and is denoted by " (t 7)-" (t )( then there is a fuzzy
relationship between " (t ) and "(t 7) which is e2pressed as the e5uation6
"(t )="(t 7 )!( t 7( t )* "''#
0he relation ! is referred to as a first order model of "(t )* %f !(t 7( t ) is independent of time t (
that is( for different times
t
7
and
t
+
( !(t
7
, t
7
7)=!(t
+
,t
+
7)(
then " (t ) is called a time-invariant
fuzzy time series* .therwise it is called a time-variant fuzzy time series*
'e*inition 3: Fuzzy "e%ationsi( >rou( ?F<">@
Relationships with the same fuzzy set on the left hand side can be further grouped into a
relationship group* Relationship groups are also referred to as fuzzy logical relationship groups or
/?RB 3s in short* &uppose there are relationships such that
A
i
-A
&$(
A
i
-A
&%(

A
i
-A
&n(
then they can be grouped into a relationship group as follows6
A
i
-A
&$ (
A
&% ( .(
A
&n
*
"':#
+<
0he same fuzzy set cannot appear more than once on the right hand side or the relationship group*
0he term relationship group was first introduced by $hen in C9D*
2.1! Conclusion
@arious theoretical concepts have been reviewed in this section such as fuzzy sets( fuzzy
numbers( defuzzification( fuzzy relations( fuzzy aggregation( P&. and /0&* 0he main purpose of
this discussion has been to provide self-contained study of the underlying theoretical concepts of the
forecasting model presented in later sections*
+>
3 "elated 2or)
0his section provides an overview of current research* /0& has been sub!ected to e2tensive
research since first introduced almost + decades ago* 1owever( the intention here is not to provide
an e2haustive study of every wor= published* Rather the intention is to provide a general overview
of /0& as an independent research field* /irst we will briefly review &ong and $hissom3s wor= C7(+(
'>D which is the earliest wor= published on /0&* )e2t( a more detailed study is provided of $hen3s
wor= published in C9(97D( as the wor= presented in the respective papers are among the most
important milestones in this particular field of research* /inally( a brief review of more recent
developments is provided*
3.1 Son* and Chisso'9s 2or)
/0& was originally proposed by &ong and $hissomC7(+('>D in a series of papers to forecast
student enrollments at the University of Alabama* 0he motivation for introducing a new forecasting
framewor=( based on fuzzy set theory( was the need to model time series problems when historical
data are defined as linguistic values* 0he first model published was the so-called time-invariant
model which comprises the followings steps6 "7# define the universe of discourse8 "+# partition the
universe of discourse into e5ually lengthy intervals8 "9# define fuzzy sets of the universe of
discourse8 "'# fuzzification of historical data8 ":# establish fuzzy relations8 ";# forecast by
A
i
=A
i7
!(
XYZ[Z ]^_ `YZabcda^efgZ[b`f[h_ZZ_Zi`^fe +*7,*7#8 and "<# defuzzify forecasted
results*
%n step ":#( the fuzzy relation ! is defined by
!
i
=A
s
T
A
0
( for all / relations A
s
-A
0
( !=
i=7
/
!
i
(
"';#
where is the min operator( T is the transpose operator( and is the union operator*
&ubse5uently &ong and $hissom proposed the time-variant model( which basically comprises
the same steps as its time-invariant counterpart* 0he most notable difference is the notion of fuzzy
relationship in step :( denoted as !
w
(t ,t 7)( and defined by
!
i
= f
T
(t i ) f (t i +7) for all / fuzzy relations A
s
-A
0
( !
w
(t ,t 7)=
i=+
w
!
i
(
"'<#
where w is the window base( T is the transpose operator( is the $artesian product( and is the
union*
+-
3.2 Chen9s 2or)
A significant drawbac= of the /0& models developed by &ong and $hissom is that they are
associated with unnecessary high computational overheads due to comple2 matri2 operations in step
: and ;* %n order to reduce the computation overhead of the time-variant and time-invariant models(
$hen C9D proposed a simplified model including only simple arithmetic operations* 0he step-by-step
procedure proposed by $hen is listed as6
7* Partition the universe of discourse into e5ually lengthy intervals*
+* Define fuzzy sets on the universe of discourse*
9* /uzzify historical data*
'* %dentify fuzzy relationships "/?R3s#*
:* stablish fuzzy relationship groups "/?RB3s#*
;* Defuzzify the forecasted output*
%n the following it will be demonstrated how the model is used to forecast student enrollments
at the University of Alabama* Actual enrollment data for the period 7-<7 - 7--+ are shown in table
'*
Aear Stu+ent enro%%ments Aear Stu+ent enro%%ments
7-<7 79,:: 7->+ 7:'99
7-<+ 79:;9 7->9 7:'-<
7-<9 79>;< 7->' 7:7':
7-<' 7';-; 7->: 7:7;9
7-<: 7:';, 7->; 7:->'
7-<; 7:977 7->< 7;>:-
7-<< 7:;,9 7->> 7>7:,
7-<> 7:>;7 7->- 7>-<,
7-<- 7;>,< 7--, 7-9+>
7->, 7;-7- 7--7 7-99<
7->7 7;9>> 7--+ 7>><;
Tab%e 4- 9istorica% stu+ent enro%%ments ,96, & ,9928 at A%abama University-
Ste( ,: 'e*ine te universe o* +iscourse an+ (artition it into e;ua%%y %engty interva%s
0he universe of discourse U is defined as
| ?
min
?
7(
?
max
?
+

where ?
min
and ?
ma2
are the
minimum and ma2imum historical enrollment( respectively* /rom table '( we get
?
min
=79,::
and
?
max
=7-99<*
0he variables ?
7
and ?
+
are !ust two positive numbers( properly chosen by the user* %f
we let ?
7
L :: and ?
+
L ;;9( we get U=|79,,,( +,,,, * $hen used seven intervals which is the
same number used in most cases observed in literature* Dividing U into seven evenly lengthy
9,
intervals u
7
( u
+
( u
9
( u
'
( u
:
( u
;
and u
<
( we get u
7
L C79,,,, 7',,,D( u
+
L C7',,,( 7:,,,D( u
9
L C7:,,,(
7;,,,D( u
'
L C7;,,,( 7<,,,D( u
:
L C7<,,,( 7>,,,D( u
;
L C7>,,,( 7-,,,D and u
<
L C7-,,,( +,,,,D*
Ste( 2: 'e*ine *uzzy sets on te universe o* +iscourse
Assume
A
7
, A
+
,., A
/
to

be fuzzy sets which are linguistic values of the linguistic variable
3enrollments3* 0hen the fuzzy sets
A
7
, A
+
,., A
/
are defined on the universe of discourse as
A
7
=a
77
/ u
7
+a
7+
/ u
+
+.+a
7m
/ u
m
,
A
+
=a
+7
/ u
7
+a
++
/ u
+
+.+a
+m
/ u
m
,

A
/
=a
/$
/ u
7
+a
/%
/ u
+
+.+a
/m
/ u
m
,
where
a
i&
| ,(7 ( 7<i</ (
and 7< &<m* 0he variable a
i&
represents the membership degree of the
crisp interval u
&
in the fuzzy set A
i
* Prior to defining fuzzy sets on the U( linguistic values should be
assigned to each fuzzy set* $hen uses the linguistic values A
7
L "not many#( A
+
L "not too many#( A
9
L
"many#( A
'
L "many many#( A
:
L "very many#( A
;
L "too many# and A
<
L "too many many#* /uzzy sets
can be defined on the universe of discourse as follows6
A
7
=7/ u
7
+,*:/ u
+
+,/ u
9
+,/ u
'
+,/ u
:
+,/ u
;
+,/ u
<
A
+
=,*:/ u
7
+7/ u
+
+,*:/ u
9
+,/ u
'
+,/ u
:
+,/ u
;
+,/ u
<
A
9
=,/ u
7
+,*:/ u
+
+7/ u
9
+,*:/ u
'
+,/ u
:
+,/ u
;
+,/ u
<
A
'
=,/ u
7
+,/ u
+
+,*:/ u
9
+7/ u
'
+,*:/ u
:
+,/ u
;
+,/ u
<
A
:
=,/ u
7
+,/ u
+
+,/ u
9
+,*:/ u
'
+7/ u
:
+,*:/ u
;
+,/ u
<
A
;
=,/ u
7
+,/ u
+
+,/ u
9
+,/ u
'
+,*:/ u
:
+7/ u
;
+,*:/ u
<
A
<
=,/ u
7
+,/ u
+
+,/ u
9
+,/ u
'
+,/ u
:
+,*:/ u
;
+7/ u
<*
Ste( .: Fuzzi*y istorica% +ata
%n this conte2t( fuzzification is the process of identifying associations between the historical values
in the dataset and the fuzzy sets defined in the previous step* ach historical value is fuzzified
according to its highest degree of membership* %f the highest degree of belongingness of a certain
historical time variable( say " (t 7) ( occurs at fuzzy set A
/
( then " (t 7) is fuzzified as A
/
* 0o
e2emplify this( let us fuzzify year 7-<7* According to table '( the enrollment in 7-<7 was 79,::
which lies within the boundaries of interval u
7
* &ince the highest membership degree of u
7
occurs at
A
7
(

the historical time variable " (7-<7) is fuzzified as A
7
* Actual enrollment of 7-<' is 7';-; which
lies within the boundaries of interval u
+
* 1ence " (7-<') is fuzzified as A
+
* A complete overview of
fuzzified enrollments is shown in the table :*
97
Aear Actua% enro%%ment Interva% Fuzzi*ie+ enro%%ment
7-<7 79,:: C79,,,( 7',,,D A
7
7-<+ 79:;9 C79,,,( 7',,,D A
7
7-<9 79>;< C79,,,( 7',,,D A
7
7-<' 7';-; C7',,,( 7:,,,D A
+
7-<: 7:';, C7:,,,( 7;,,,D A
9
7-<; 7:977 C7:,,,( 7;,,,D A
9
7-<< 7:;,9 C7:,,,( 7;,,,D A
9
7-<> 7:>;7 C7:,,,( 7;,,,D A
9
7-<- 7;>,< C7;,,,( 7<,,,D A
'
7->, 7;-7- C7;,,,( 7<,,,D A
'
7->7 7;9>> C7;,,,( 7<,,,D A
'
7->+ 7:'99 C7:,,,( 7;,,,D A
9
7->9 7:'-< C7:,,,( 7;,,,D A
9
7->' 7:7': C7:,,,( 7;,,,D A
9
7->: 7:7;9 C7:,,,( 7;,,,D A
9
7->; 7:->' C7:,,,( 7;,,,D A
9
7->< 7;>:- C7;,,,( 7<,,,D A
'
7->> 7>7:, C7>,,,( 7-,,,D A
;
7->- 7>-<, C7>,,,( 7-,,,D A
;
7--, 7-9+> C7-,,,( +,,,,D A
<
7--7 7-99< C7-,,,( +,,,,D A
<
7--+ 7>><; C7>,,,( 7-,,,D A
;
Tab%e 3- Fuzzi*ie+ istorica% enro%%ments-
Ste( 4: I+enti*y *uzzy re%ationsi(s
Relationships are identified from the fuzzified historical data* %f the time series variable " (t 7) is
fuzzified as A
/
and " (t ) as A
m
( then A
/
is related to A
m
* Ae

denote this relationship as
A
/
-A
m
(
where A
/
is the current state of enrollment and

A
m
is the ne2t state of enrollment* /rom table :( we
can see that year 7-<7 and 7-<+ both are fuzzified as A
7
(

which provides the following relationship6
A
7
-A
7
*
0he complete set of relationships identified from table : are presented in the table ;*
A
7
K A
7
A
7
K A
+
A
+
K A
9
A
9
K A
9
A
9
K A
'
A
'
K A
'
A
'
K A
9
A
'
K A
;
A
;
K A
;
A
;
K A
<
A
<
K A
<
A
<
K A
;
Tab%e 5- Fuzzy set re%ationsi(s-
)ote that even though the same relationships may appear more than once( these are ignored since
there can only be one relationship of the same =ind*
9+
Ste( 3: Estab%is *uzzy re%ationsi( grou(s ?F<">Bs@
%f the same fuzzy set is related to more than one set( the right hand sides are merged* Ae refer to
this process as the establishment of /?RB3s* /or e2ample( we see from table ; that A
7
is related to
itself and to A
+
*

0his provides the

following /?RB6
A
7
-A
7
( A
+
*
A complete overview of the
relationship groups obtained from table ; is shown table <*
Broup 76 A
7
K A
7
A
7
K A
+
Broup +6 A
+
K A
9

Broup 96 A
9
K A
9
A
9
K A
'
Broup '6 A
'
K A
'
A
'
K A
9
A
'
K A
;
Broup :6 A
;
K A
;
A
;
K A
<
Broup ;6 A
<
K A
<
A
<
K A
;
Tab%e 6- F<">Bs-
Ste( 5: 'e*uzzi*y te *orecaste+ out(ut
Assume the fuzzified enrollment of " (t 7) is A
&
( then forecasted output of " (t ) is determined
according to the following principles6
7* %f there e2ists a one-to-one relationship in the relationship group of A
&
( say
A
&
-A
/
(
and the
highest degree of belongingness of A
/
occurs at interval u
/
( then the forecasted output of
" (t ) e5uals the midpoint of u
/
*
+* %f A
&
is empty( i*e*
A
&
-Y,
and the interval where A
&
has the highest degree of belongingness
is u
&
( then the forecasted output e5uals the midpoint of u
&
*
9* %f there e2ists a one-to-many relationship in the relationship group of A
&
( say
A
&
-A
7
( A
+
( .( A
n
(
and the highest degrees of belongingness occurs at set
u
7
( u
+
(.( u
n
(
then the forecasted output is computed as the average of the midpoints
m
7
( m
+
(.( m
n
of
u
7
( u
+
( .( u
n
*
0his e5uation can be e2pressed as6
m
7
+m
+
+.+m
n
n
*
/or e2ample( year 7-<+ is forecasted using the fuzzified enrollments of 7-<7* According to
table :( the fuzzified enrollments of year 7-<7 is A
7
*

/rom table < it can be seen that A
7
is related to
A
7
and A
+
* 0he highest degrees of belongingness of A
7
and A
+
are the sets of u
7
and u
+
(

where

u
7
L
C79,,,( 7',,,D and u
+
L C7',,,( 7:,,,D* 0he midpoints of the intervals( u
7
and u
+
(

are 79:,, and
7':,,( respectively* Using rule 9( the forecasted enrollment of 7-<+ is computed as
"79:,,S7':,,#G+ L 7',,,* jear 7->, is forecasted using the fuzzified enrollments of 7-<- as basis*
99
Fecause the fuzzified enrollment of 7-<- is A
'
( we have the following /?RB6
A
'
-A
9
( A
'
( A
;
*
0he
highest degrees of belongingness for the fuzzy sets A
9(
A
'
and A
;
are at intervals u
9
L C7:,,,( 7;,,,D(
u
'
L C7;,,,( 7<,,,D and u
;
L C7>,,,( 7-,,,D( respectively( and the midpoints of u
9
( u
'
and u
;
are
7::,,( 7;:,, and 7>:,,( respectively* 0herefore the forecasted output is calculated as
"7::,,S7;:,,S7>:,,#G9 L 7;>99* &ince there are no empty relationship groups( rule + is never
applied in this e2ample*
Aear Actua%
enro%%ment
Forecaste+
enro%%ment
F<">Bs Interva% mi+(oints
7-<7 79,:: A
7
K A
7
( A
+
79:,,8 7':,,
7-<+ 79:;9 7',,, A
7
K A
7
( A
+
79:,,8 7':,,
7-<9 79>;< 7',,, A
7
K A
7
( A
+
79:,,8 7':,,
7-<' 7';-; 7',,, A
+
K A
9
7::,,
7-<: 7:';, 7::,, A
9
K A
9
( A
'
7::,,8 7;:,,
7-<; 7:977 7;,,, A
9
K A
9
( A
'
7::,,8 7;:,,
7-<< 7:;,9 7;,,, A
9
K A
9
( A
'
7::,,8 7;:,,
7-<> 7:>;7 7;,,, A
9
K A
9
( A
'
7::,,8 7;:,,
7-<- 7;>,< 7;,,, A
'
K A
9
( A
'
(

A
;
7::,,8 7;:,,8 7>:,,
7->, 7;-7- 7;>99 A
'
K A
9
( A
'
(

A
;
7::,,8 7;:,,8 7>:,,
7->7 7;9>> 7;>99 A
'
K A
9
( A
'
(

A
;
7::,,8 7;:,,8 7>:,,
7->+ 7:'99 7;>99 A
9
K A
9
( A
'
7::,,8 7;:,,
7->9 7:'-< 7;,,, A
9
K A
9
( A
'
7::,,8 7;:,,
7->' 7:7': 7;,,, A
9
K A
9
( A
'
7::,,8 7;:,,
7->: 7:7;9 7;,,, A
9
K A
9
( A
'
7::,,8 7;:,,
7->; 7:->' 7;,,, A
9
K A
9
( A
'
7::,,8 7;:,,
7->< 7;>:- 7;,,, A
'
K A
9
( A
'
(

A
;
7::,,8 7;:,,8 7>:,,
7->> 7>7:, 7;>99 A
;
K A
;
( A
<
7>:,,8 7-:,,
7->- 7>-<, 7-,,, A
;
K A
;
( A
<
7>:,,8 7-:,,
7--, 7-9+> 7-,,, A
<
K A
;
( A
<
7>:,,8 7-:,,
7--7 7-99< 7-,,, A
<
K A
;
( A
<
7>:,,8 7-:,,
7--+ 7>><; 7-,,,
Tab%e 2- Forecaste+ enro%%ments *or te (erio+ ,962 & ,992-
0he model e2plored in so far is referred to as a first order /0& model* $hen later introduced its
high order counterpart which incorporates n-order relationships in C97D* %n the high order variant(
relationships of order n R + can be e2pressed as
A
i ,7
( A
i , +
(.( A
i ,n
-A
i ,n+7
*
/or e2ample( a second
order relationship is denoted by
A
i ,7
( A
i , +
-A
i ,9
*
A third order relationship is denoted by
A
i ,7
( A
i , +
( A
i ,9
-A
i , '
*
All second order relationships identified from table : are listed in table -*
9'
A
7
( A
7
K A
7
A
9
( A
9
K A
'
A
9
( A
'
K A
;
A
7
( A
7
K A
+
A
9
( A
'
K A
'
A
'
( A
;
K A
;
A
7
( A
+
K A
9
A
'
( A
'
K A
'
A
;
( A
;
K A
<
A
+
( A
9
K A
9
A
'
( A
'
K A
9
A
;
( A
<
K A
<
A
9
( A
9
K A
9
A
'
( A
9
K A
9
A
<
( A
<
K A
;
Tab%e 9- Secon+ or+er re%ationsi(s-
0he grouping of relationships is somewhat different compared to the first order variant* %n high
order models( relations with identical left hand sides are not merged into a single entity in the same
manner as in the first order case* 0o illustrate this( consider the following relationship from table -(
for t L 7-<96
" (7-<7) ( " (7-<+)-" (7-<9)=A
7
( A
7
-A
7
*
An ambiguity occurs in this case because another relation is found with an identical left hand side(
namely
A
7
( A
7
-A
+
*
0o deal with this ambiguity( the current relationship is e2tended to a third order
relation as follows6
" (7-<,)( "(7-<7)( " (7-<+)-" (7-<9)=k( A
7
( A
7
-A
7
*
0he k symbol indicates null set since " (7-<,) does not e2ist* %f no other third order relation e2ists
with identical left hand sides( " (7-<9) can be defuzzified by the same principles as in the first
order case( otherwise another n S 7 order e2tension has to be made*
0able 7, shows the performance of different n-order models in form of 4& "see e5uation :'#*
:r+er 1SE
7 ',<:,<
+ >-,-9
9 >;;-'
' >-9<;
: -':9-
; ->+7:
< 7,',:;
> 7,+7<-
- 7,+<>-
Tab%e ,0- Forecast accuracy *or +i**erent or+ers-
%n some cases( higher forecasting accuracies can be accomplished with higher model orders( as is
the case with the enrollment data* Fut increasing the order from n to n S 7( does not necessarily
result in higher accuracy rates for all cases*
9:
3.3 $ther Develo'ents
Benerally( the focus of current /0& research has been on the establishment of fuzzy
relationships and interval partitions* arly studies( in particular( were entirely devoted to the former
issue such as &ongG$hissom C7(+('>D( $hen C9D( 1wang et al C7;D( &ullivanGAoodall C+,D* .ther
more recent studies dealing with the relationship aspect can also be found( li=e 0saurGAoodall C7'D
and &ing C:D*
4ore recently( interval partitions have received a considerable amount attention in current
studies* A ma!or reason for this paradigm shift is the need for formalized approaches to interval
partitioning* %n early studies( intervals were assumed to be sub!ectively defined by the user( in the
same manner as shown in the e2ample provided of $hen3s model C9D( in section 9*+* 1uarng C7-D(
was probably the first researcher to focus on the interval partition aspect* %n C7-D( 1uarng proposed
the distribution- and average-based length approaches to determine the lengths of intervals*
/urthermore( the study conducted by 1uarng in C7-D( was the first to investigate the influence of
interval lengths on forecast results* .ther e2amples of formalized approaches to interval
partitioning can be found in C'(7,(+-D*
A common factor shared by the models published in C'(7,(7-(+-D( is that interval lengths are
determined independently of forecast accuracy* %n contrast( $henG$hung C7:D apply a somewhat
different strategy where they e2ploit genetic algorithm "BA# to tune interval lengths in order to
improve forecast accuracy* A similar study is published by Uuo et al in C9,D( where particle swarm
optimization "P&.# is e2ploited in an analogous manner* 0he studies published in C7:D and C9,D are
highlighted in this pro!ect because they have presented the best results currently published in
literature* Foth of these models will be used as targets for comparison with the respect to the high
order model presented in later sections*
.ther studies can also be found which both deal with the fuzzy relationship aspect and the
interval partition aspect( such as the ones presented in C>(7>(99D* 0he current study belongs to this
latter category of pro!ects( as it sets out to develop new and hopefully better ways of creating
interval partitions and relational computations*
3.! Conclusion
specially two issues seem to be the primary focus of current research* 0he first is the
selection of interval partitions "i*e* the length and number of intervals#* 0he second is the
formulation of fuzzy relationships* Foth of these factors highly influence forecast accuracy and thus
are considered central to /0&* %t has been found that current research efforts are united by one
9;
common goal6 to enhance consistency between forecast rules and the data they derive from* 0his
implies that performance of different /0& models by tradition is evaluated under =nown conditions*
%n plain words( it means that forecast rules are validated using the same old data they originate
from( rather than validating them on future datasets "see computations of forecasts in section 9*+#*
0o ma=e this pro!ect comparable with those of others( we will follow the same principle in the
evaluation phase in section :*
1igh order models are highlighted in this conte2t( because they are among most accurate
models found in literature( and thus are selected as targets for comparison* 0he findings of this
related wor= study has lead to the identification of the following =ey problems with regards to high
order models6 "7# there is a lac= of consistency between forecast rules and the data they represent8
"+# forecast accuracy is sensitive to selected interval partitions8 "9# data becomes underutilized as
the model3s order increases* %n "9#( the underutilization of data manifests itself in two ways* /irst the
number forecast rules "fuzzy relationships# reduces as the order increases* &econd( the combination
of patterns "fuzzy sets# to be matched with future patterns increases with order increments* 0his( in
turn( reduces the probability of finding e5uivalent pattern combinations in future time series data*
&olving the problems "7#-"9# is the primary ob!ective of this pro!ect*
A secondary ob!ective is to further improve the trapezoid fuzzification algorithm proposed by
$heng et al in C'D* 0his ob!ective is motivated by the need for an algorithm capable of generating
trapezoidal fuzzy numbers "or intervals# automatically( based on the characteristics in data*
9<
! Introducin* a /odified Fuzzy Ti'e Series /odel
%n the following sections a modified /0& model is presented* /irst we will discuss the data
fuzzification part* A novel fuzzification algorithm based on the trapezoid fuzzification approach C'D
is be presented and evaluated* )e2t a novel approach to defuzzify forecasted output is presented
based on P&. and aggregation* /inally( in section :( the proposed /0& model will be evaluated by
comparing it to other related developments*
!.1 Al*orith' $vervie8
Fefore elaborating on the details of the forecasting model presented in here( we will initially
provide an overview of the algorithmic process* 0he overall structure of the algorithm discussed in
the following sections is depicted in figure >*
0he proposed model proposed is divided into two main components( fuzzification and
defuzzification* Foth of these main components are decoupled which implies that they can be
integrated independently with other alternatives* /or e2ample( the fuzzification module can be
integrated with $hen3s first order model C9D or $hen3s high order model C97D* 0his will be
demonstrated in section '*9* 0he defuzzification module can be integrated with other fuzzification
algorithms( such as the ones published in C7,(7:(7>(7-D*
9>
Figure 2- :vera%% a%goritm structure-
0he fuzzification module can be further decomposed into a si2-step process where the first
four steps are data preprocessing functions* 0he fuzzification tas= itself comprises the last two
steps* Ahen data has been fuzzified( it is further processed by the defuzzification module* During
the defuzzification phase( data is grouped into patterns which are converted into corresponding if-
statements* 0he if-rules are trained individually to match the data they represent* Ahen training is
completed( data is defuzzified by matching the if-then rules with e5uivalent patterns in the dataset*
!.2 Fuzzifyin* :istorical Data
0he fuzzification algorithm "/A# proposed here generates a series of trapezoidal fuzzy sets
from a given dataset and establishes associations between the values in the dataset and the fuzzy
sets generated* %t is inspired by the trapezoid fuzzification approach proposed by $heng et al in C'D*
0hey introduced an approach where the crisp intervals( generally defined by the user at the initial
step of /0&( are replaced with trapezoidal fuzzy sets with overlapping boundaries* 0his overlap
implies that a value may belong to more than one set* %f a value belongs to more than one set( it is
associated to the set where its degree of membership is highest* 0he /A introduced here follows the
same principles but differs from the approach described by $heng et al C'D by performing
automatically the calculation of the fuzzy intervalsGsets* 0he fuzzification approach published in C'D(
re5uires the user to specify the number of sets* 0his is an undesirable re5uirement in situations
where multiple forecasting problems need to be solved* /or e2ample( a grocery store may need
forecast information related to thousands of products* 0he proposed algorithm aims to solve this
problem by determining the number of sets on basis of the variations in data*
Another aspect this algorithm attempts to capture( is the notion of a non-static universe set*
Ahenever values are encountered which fall outside the boundaries of the current universe set( the
universe set has to augment accordingly* 0his aspect( in particular( has not received much attention
in current publications* 0he most li=ely reason for this is that current modalities rely on the
assumption of predetermined outcomes "see section 9*'#( and therefore( no revisions of the universe
set are re5uired* %n real life situations though( future outcomes are rarely =nown* 0he basic idea of
the algorithm described in the following paragraphs( is to repeat the fuzzification procedure when
the dataset is updated* 0he proposed procedure can be described as a si2-step process6
Ste( ,: &ort the values in the current dataset in ascending order*
Ste( 2: $ompute the average distance between any two consecutive values in the
sorted dataset and the corresponding standard deviation*
Ste( .: liminate outliers from the sorted dataset*
9-
Ste( 4: $ompute the revised average distance between any two remaining consecutive
values in the sorted dataset*
Ste( 3: Define the universe of discourse*
Ste( 5: /uzzify the dataset using the trapezoid fuzzification approach C'D*
/irst the values in the historical dataset are sorted in ascending order* 0hen the average
distance between any two consecutive values in the sorted dataset is computed and the
corresponding standard deviation* 0he average distance is given by the e5uation6
A?( x
i
.x
n
)=
7
n7
_
i=7
n7
x
p( i)
x
p(i+7)
( "'>#
where p is permutation that orders the values ascendantly6
x
p( i)
<x
p(i +7)
*
0he standard deviation is
computed as
c
A?
=
.
7
n
_
i=7
n
( x
i
A?)
+
"'-#
Foth the average distance and standard deviation are used in step 9 to define outliers in the
sorted dataset* .utliers are values which are either abnormally high or abnormally low* 0hese are
eliminated from the sorted dataset( because the intention here is to obtain an average distance value
free of distortions* An outlier( in this conte2t( is defined as a value less than or larger than one
standard deviation from average* After the elimination process is completed( a revised average
distance value is computed for the remaining values in the sorted dataset( as in step +* 0he revised
average distance( obtained in step '( is used in step : and ; to partition the universe of discourse
into a series of trapezoidal fuzzy sets* Fasically( the intention is to create a series of trapezoidal
appro2imations which capture the generic nature of data as closely as possible( in the sense that we
neither want the spread of individual functions to be to narrow or to wide*
%n step :( the universe of discourse is determined* %ts lower and upper bound is determined by
locating the largest and lowest values in the dataset and augment these by6 "7# subtracting the
revised average distance from the lowest value and "+# adding the revised average distance to the
highest value* 4ore formally( if ?
ma2
and ?
min
are the highest and lowest values in the dataset(
respectively( and A?
!
is the revised average distance( the universe of discourse U can be defined as
U L C?
min
- A?
!
( ?
ma2
S A?
!
D*
Ahen the U has been determined( fuzzy subsets can be defined on U* &ince the subsets are
represented by trapezoidal functions( the membership degree( for a given function
A
and a given
value x( is obtained by e5uation 77*
',
j
A
=

xa
7
a
+
a
7
, a
7
xa
+
7( a
+
xa
9
a
'
x
a
'
a
9
, a
9
xa
'
,( otherwise*
Prior to the fuzzification of data( we need to =now the number of subsets to be defined on U*
0he number of sets( n( is computed by
n=
!#
+&
( ":,#
where ! denotes the range of the universe set and # denotes the segment length* 5uation :,
originates from the fact that we =now the following about #6
#=
!
(+n+7)
. ":7#
0he range( !( is computed by
!=UBB( ":+#
where UB and B respectively denote the upper bound "?
ma2
S A?
!
# and lower bound "?
min
- A?
!
#
of U* 0he segment length( #( e5uals the average revised distance A?
!
which in turn constitutes the
length of left spread "ls#( core "c# and right spread "rs# of the membership function "see figure -#*
0hat is( ls L A?
!
( c @

A?
!
and rs

L

A?
!
*
%n short( the tas= here is to decide how many fuzzy sets to generate when the length of each
segment( #( and the range( !( are =nown* Ahen the number of sets has been computed( the sets can
be defined on U and data can be fuzzified which completes the final step of the algorithm*
%n the following e2ample( we will fuzzify the first four years of student enrollment in Alabama
University* 0he respective values to be fuzzified are 79,::( 79:;9( 79>;< and 7';-; "see table :#*
'7
Figure 9- Te segments o* a tra(ezoi+a% *uzzy number-
Fecause the se5uence is already in ascending order( the sorting part is omitted* 0he average distance
and the standard deviation are respectively computed as
A?=
79,::79:;9+79:;979>;<+79>;<7';-;
9
=
:,>+9,'+>+-
9
=:'<
and
c
A?
=
.
(:,>:'<)
+
+(9,':'<)
+
+(>+-:'<)
+
9
=+7;*7-+7;*
)e2t( possible outliers are eliminated* Recall that outliers include the values less than or larger than
one standard deviation from A?* 0his means only the values satisfying the condition6
:'<+7;<x<:'<++7;(
are ta=en into consideration when computing the revised average distance* %n this case( only one of
the three values satisfy the above condition( namely :,>* 0hus the revised average distance( A?
!
(
and the segment length( #( e5uals :,>* At this point( step 7 - ' are completed* Prior to defining the
universe set U( we need to determine the lower bound "B# and the upper bound "U,# of U*
/ollowing e5uation :+( B and U, are computed as
B L 79,:: - :,> L 7+:'<
UB L 7';-; S :,> L 7:+,'*
1ence U L C7+:'<( 7:+,'D* 0he range( !( is computed as difference between UB and B* 1ence we
get 7:+,' - 7+:'< L +;:<* /inally the number of sets( n( is computed as
n=
+;:<:,>
+:,>
=+*7+-+*
Unowing the universe of discourse and the parameters of >( ! and #( the fuzzy sets are
generated as shown in figure 7, and table 77*
Fuzzy set Tra(ezoi+a% *uzzy number ?a8 b8 c8 d@ Cris( interva%
A
7
"7+:'<(79,::(79;,+(7'7'-# u
7
L C79,::(79;,+D
A
+
"79;,+(7'7'-(7';-;(7:+,'# u
+
L C7'7'-(7';-;D
Tab%e ,,- Fuzzi*ying te *irst *our years o* enro%%ment-
'+
)ote the difference between the points a( b( c and d( in the fuzzy number A
$
and A
%
( in table 77
and figure 7,( is not e2actly :,>* 0his is because the implemented algorithm adapts the segment
length( such that the lowest value in the dataset always appears as the left bound in the first crisp
interval( and the highest value in the dataset always appears as the right bound in the last crisp
interval* /rom table 77and figure 7, it can be seen that the lowest of the four values "i*e* 79,::#(
appears as the lower bound of the first crisp interval( u
7
( and the highest value "i*e* 7';-;#( appears
as the upper bound in the second crisp interval( u
+
* )ormally these values cannot be matched
precisely without ad!usting the segment length( due to rounding errors occurring as a result of
e5uation '> and :,*
)onetheless( we are now able to fuzzify the first four historical enrollments according to
membership functions A
7
and A
+
( defined by6
A
7
=

,( x7+:'<
x7+:'<
79,::7+:'<
, 7+:'<x79,::
7( 79,::x79;,+
7'7'-x
7'7'-79;,+
, 79;,+x7'7'-
,( x>7'7'-*
and
A
+
=

,( x79;,+
x79;,+
7'7'-79;,+
, 79;,+x7'7'-
7( 7'7'-x7';-;
7:+,'x
7:+,'7';-;
, 7';-;x7:+,'
,( x>7:+,'*
'9
Figure ,0- >enerate+ membersi( *unctions-
)ote the intervals overlap so more than one interval may be met* /or e2ample( the enrollment for
year 7-<9 is 79>;<* 0his value meets both membership functions* 0he membership degree in A
7
is
,*:7:: l ,*:+( and in A
+
( it is ,*'>': l ,*'>* 1ence the enrollment for 7-<9 is fuzzified as A
7
* A
special case occurs when the membership degree is ,*:( as this implies a that value has the same
membership status in two different sets* %n such cases( the respective value is associated to both A
7
and A
+
*
Aear Enro%%ment Fuzzy set 1embersi( +egree
7-<7 79,:: A
7
7
7-<+ 79:;9 A
7
7
7-<9 79>;< A
7
,*:+
7-<' 7';-; A
+
7
Tab%e ,2- Fuzzi*ie+ enro%%ments ,96, & ,964-
Fy processing the entire enrollment dataset from table '( the resultant trapezoidal sets are as
shown in table 79* A complete overview of the fuzzified enrollments is shown in table 7'*
''
Figure ,,- Fuzzi*ying year ,96.-
Fuzzy set Fuzzy number
A7 "7+>;7(79,::(79+':(79'9;#
A+ "79+':(79'9;(79;+;(79>7;#
A9 "79;+;(79>7;(7',,<(7'7-<#
A' "7',,<(7'7-<(7'9>>(7':<>#
A: "7'9>>(7':<>(7'<;>(7'-:-#
A; "7'<;>(7'-:-(7:7'-(7:99-#
A< "7:7'-(7:99-(7::9,(7:<+,#
A> "7::9,(7:<+,(7:-7,(7;7,7#
A- "7:-7,(7;7,7(7;+-7(7;'>+#
A7, "7;+-7(7;'>+(7;;<+(7;>;+#
A77 "7;;<+(7;>;+(7<,:9(7<+'9#
A7+ "7<,:9(7<+'9(7<'99(7<;+'#
A79 "7<'99(7<;+'(7<>7'(7>,,'#
A7' "7<>7'(7>,,'(7>7-:(7>9>:#
A7: "7>7-:(7>9>:(7>:<;(7><;;#
A7; "7>:<;(7><;;(7>-:;(7-7'<#
A7< "7>-:;(7-7'<(7-99<(7-:97#
Tab%e ,.- >enerate+ *uzzy sets by (rocessing te enro%%ment +ata *rom ,96, & ,992-
Aear Enro%%ment Fuzzy Set
7-<7 79,:: A7
7-<+ 79:;9 A+
7-<9 79>;< A9
7-<' 7';-; A:
7-<: 7:';, A<
7-<; 7:977 A<
7-<< 7:;,9 A<
7-<> 7:>;7 A>
7-<- 7;>,< A77
7->, 7;-7- A77
7->7 7;9>> A7,
7->+ 7:'99 A<
7->9 7:'-< A<
7->' 7:7': A;
7->: 7:7;9 A;
7->; 7:->' A>
7->< 7;>:- A77
7->> 7>7:, A7'
7->- 7>-<, A7;
7--, 7-9+> A7<
7--7 7-99< A7<
7--+ 7>><; A7;
Tab%e ,4- Fuzzi*ying annua% enro%%ments-
':
Benerally it is assumed that the fuzzy sets(
A
7
( A
+
(.( A
n
(
individually represent some linguistic
value* Aith 7< intervals( however( linguistic values may not ma=e much sense* %n the model
proposed here( this issue is ignored because linguistic values generally do not serve any purpose in
/0& what so ever - although they may be useful in certain applicative conte2ts*
!.3 ;valuatin* the 7roosed Fuzzification Al*orith'
%n the following section( we3re going to evaluate the proposed /A by applying it directly to
$hen3s model C9(97D for different orders* 0he first e2periment is apply the algorithm to $hen3s first
order model C9D and compare performance with the one in C'D( where the authors also apply their
algorithm directly to $hen3s model* )e2t( performance will be evaluated for different model orders
by comparing results to the ones reported by $hen in C97D*
0o evaluate performance across models( the mean s5uared error "4&# and mean absolute
percentage error "4AP# are used as performance measures* 0he respective measures are defined
by the e5uations
AA,B=
7
n
_
t =7
n

forecast
t
actual
t

actual
t
7,, ":9#
and
A#B=
7
n
_
i=7
n
( forecast
i
actual
i
)
+
* ":'#
/rom fuzzified data in table 7'( we get the following first order relationships6
A
7
K A
+
A
>
K A
77
A
;
K A
>
A
+
K A
9
A
77
K A
77
A
77
K A
7'
A
9
K A
:
A
77
K A
7,
A
7'
K A
7;
A
:
K A
<
A
7,
K A
<
A
7;
K A
7<
A
<
K A
<
A
<
K A
;
A
7<
K A
7<
A
<
K A
>
A
;
K A
;
A
7<
K A
7;
Tab%e ,3- First or+er re%ationsi(s-
4oreover( from the data in table 7:( we get the following first order /?RB3s6
Broup 76 A
7
K A
+
Broup <6 A
77
K A
7,
(

A
77
( A
7'
Broup +6 A
+
K A
9
Broup >6 A
7,
K A
<
Broup 96 A
9
K A
:
Broup -6 A
;
K A
;
( A
>
Broup '6 A
:
K A
<
Broup 7,6 A
7'
K A
7;
Broup :6 A
<
K A
;
(

A
<
(

A
>
Broup 776 A
7;
K A
7<
Broup ;6 A
>
K A
77
Broup 7+6 A
7<
K A
7;
(

A
7<
Tab%e ,5- First or+er F<">Bs-
';
0o defuzzify forecasted output( we can use the centroid( given by e5uation 79* /or a
symmetrical fuzzy number A
i
L "a( b( c( d#( this computation can be reduced to finding the midpoint
of the crisp interval of u
i
( given by

Cb( cD* Fased on the /?RB3s shown in table 7;( forecast results
are derived as shown in table 7<*
Aear Enro%%ment Cen
C9D
Ceng et a%
C'D
$ro(ose+ FA
7-<7 79,:: - - -
7-<+ 79:;9 7',,, 79:97 7'+9,
7-<9 79>;< 7',,, 79-7+ 7'+9,
7-<' 7';-; 7',,, 7';<9 7'+9,
7-<: 7:';, 7::,, 7:'9: 7::'7
7-<; 7:977 7;,,, 7:'9: 7::'7
7-<< 7:;,9 7;,,, 7:'9: 7::'7
7-<> 7:>;7 7;,,, 7:'9: 7;7-;
7-<- 7;>,< 7;,,, 7;-:> 7;7-;
7->, 7;-7- 7;>99 7<+77 7;7-;
7->7 7;9>> 7;>99 7<+77 7<:,<
7->+ 7:'99 7;>99 7:'9: 7;7-;
7->9 7:'-< 7;,,, 7:'9: 7::'7
7->' 7:7': 7;,,, 7:'9: 7::'7
7->: 7:7;9 7;,,, 7:'9: 7::'7
7->; 7:->' 7;,,, 7:'9: 7::'7
7->< 7;>:- 7;,,, 7;-:> 7;7-;
7->> 7>7:, 7;>99 7<+77 7<:,<
7->- 7>-<, 7-,,, 7>>;7 7>><+
7--, 7-9+> 7-,,, 7-+'+ 7>><+
7--7 7-99< 7-,,, 7-,:+ 7>><+
7--+ 7>><; 7-,,, 7-,:+ 7>><+
- - -
1SE ',<:,< +;77;+ 77-,-;
1A$E 9*77m +*;;m 7*'+m
Tab%e ,6- Com(aring *orecast resu%ts )en or+er C ,-
/orecasted results are calculated according to the same principles shown earlier in section 9*+(
in step ;* Referring to table 7<( we see that forecast error is reduced when applying the proposed /A
to $hen3s first order model( since the /A case has the lowest 4& and 4AP vis-E-vis the other
models referred to in the table*
%n the second e2periment( the /A has been applied directly to $hen3s high order model C97D for
'<
different orders* 2perimental results( in form of 4& and 4AP( are presented in table 7>*
:r+er 1SE
Cen C97D
1SE
(ro(ose+
FA
1A$E
Cen C97D
1A$E
(ro(ose+
FA
+ >-,-9 7,<>< 7*;+m ,*:;m
9 >;;-' 7,:'9 7*:;m ,*:'m
' >-9<; 77,-- 7*:<m ,*:;m
: -':9; 77<7: 7*;:m ,*:>m
; ->+7: 77'>; 7*;>m ,*:<m
< 7,',:; 7,9<7 7*<'m ,*:9m
> 7,+7<- 7,-;, 7*<,m ,*::m
- 7,+<>- 7,,'- 7*;>m ,*:+m
Tab%e ,2- Com(aring resu%ts *or iger or+ers-
Again( we see better results are obtained when using the proposed /A vis-E-vis the interval
partition used by $hen in C97D* Ae will briefly illustrate how some calculations are derived for the
second order case* 0he second and third order /?RB3s( obtained from table 7'( are shown in table
7- and +,( respectively*
Broup 76 A
7
( A
+
K A
9
Broup ;6 A
<
( A
>
K A
77
Broup 7+6 A
;
( A
;
K A
>
Broup +6 A
+
( A
9
K A
:
Broup <6 A
>
( A
77
K A
77
Broup 796 A
;
( A
>
K A
77
Broup 96 A
9
( A
:
K A
<
A
>
( A
77
K A
7'
Broup 7'6 A
77
( A
7'
K A
7;
Broup '6 A
:
( A
<
K A
<
Broup >6 A
77
( A
77
K A
7,
Broup 7:6 A
7'
( A
7;
K A
7<
Broup :6 A
<
( A
<
K A
<
Broup -6 A
77
( A
7,
K A
<
Broup 7;6 A
7;
( A
7<
K A
7<
A
<
( A
<
K A
>
Broup 7,6 A
7,
( A
<
K A
<
Broup 7<6 A
7<
( A
7<
K A
7;
A
<
( A
<
K A
;
Broup 776 A
<
( A
;
K A
;
Tab%e ,9- Secon+ or+er F<">Bs-
Broup 76 C, A
7
( A
+
K A
9
Broup ;6 A
<
( A
<
( A
<
K A
>
Broup 776 A
77
( A
7,
( A
<
K A
<
Broup 7;6 A
;
( A
>
( A
77
K A
7'
Broup +6 A
7
( A
+
(

A
9
K A
:
Broup <6 A
<
( A
<
( A
>
K A
77
Broup 7+6 A
7,
( A
<
( A
<
K A
;
Broup 7<6 A
>
( A
77
( A
7'
K A
7;
Broup 96 A
+
(

A
9
( A
:
K A
<
Broup >6 A
<
( A
>
( A
77
K A
77
Broup 796 A
<
( A
<
( A
;
K A
;
Broup 7>6 A
77
( A
7'
( A
7;
K A
7<
Broup '6 A
9
( A
:
( A
<
K A
<
Broup -6 A
>
( A
77
( A
77
K A
7,
Broup 7'6 A
<
( A
;
( A
;
K A
>
Broup 7-6 A
7'
( A
7;
( A
7<
K A
7<
Broup :6 A
:
( A
<
( A
<
K A
<
Broup 7,6 A
77
( A
77
( A
7,
K A
<
Broup 7:6 A
;
( A
;
( A
>
K A
77
Broup +,6 A
7;
( A
7<
( A
7<
K C
Tab%e 20- Tir+ or+er F<">Bs-
Benerally the forecasting part is trivial( e2cept for a few special cases* /or e2ample( when
forecasting year 7-<<( an ambiguity occurs* 0o see this( consider the following second order
relationship from table 7'( for t L 7-<<6

" (t +)( " (t 7)-"(t )=" (7-<:)( " (7-<;)-" (7-<<)=A
<
( A
<
-A
<
*
'>
0he above relationship matches group : in table 7-* Fut two other identical left hand sides e2ist for
group :* 0herefore( we need to find the corresponding third order relation* Again( for t L 7-<<( we
get the following third order /?RB from table 7'6
" (t 9)( "(t +)( " (t 7)-"(t )=" (7-<')( " (7-<:)( " (7-<;)-" (7-<<)=A
:
( A
<
( A
<
-A
<
*
At this point( no ambiguities are found in table +, for the corresponding /?RB* 1ence the
forecasted result( + (7-<<)( is computed as the midpoint of the crisp interval( u
<
( in the fuzzy set A
<
L "7:7'-(7:99-(7::9,(7:<+,#* 0he computation yields
+ (7-<<)=
7:99-+7::9,
+
-7:'9:*
%n the current e2ample( we have demonstrated how forecast accuracy is influenced by different
interval partitions* /rom my own standpoint( this is an additional drawbac= of current /0& models*
0he reason for this is rooted in the assumption that intervals should reflect the same characteristics
as the data they represent( as this is more useful from a data analytical perspective* &o( although
comparative results presented in the current section favour the proposed /A in terms of forecast
accuracy( this has not been the primary goal* Rather the goal has been to develop a method which
ob!ectively determines interval partitions without re5uiring any user intervention*
!.! Defuzzifyin* $utut
%n the previous sections( the fuzzification part has been discussed* %n the following section we
will present a novel approach to defuzzify forecasted output* 0he defuzzification method presented
here comprises the following steps6
Ste( ,: stablish fuzzy set groups "/&B3s#*
Ste( 2: $onvert the /&B3s into corresponding if statements*
Ste( .: 0rain the if-then rules*
Ste( 4: Derive forecasts*
Fefore we go into details with the individual steps( it is important to understand how
defuzzified output is computed* /irst recall from the definition provided in section +*79( that an n-
order fuzzy relationship is denoted as " (t n) (.( " (t +)( "(t 7)-" (t )( where " represents a
fuzzified forecast value at time t* %n traditional /0&( it is assumed that the left hand side of the fuzzy
relation is fuzzified in the same manner as the right hand side* /or e2ample( if "( on the left hand
side( represents a trapezoidal set( then "( on the right hand( side represents a trapezoidal set as well*
'-
%n the modified version introduced here( this notion has been revised such that " (t ) is given by the
following defuzzification operator( + (t )( defined by
+ (t )=
_
i=7
n
a
t i
w
i
where
w
i
| ,(7
"::#
and a
t=i
denotes the actual value at time t = i* .therwise stated( the defuzzified output is the weighted
sum of the actual values from time (t n) to t 7( where n depends on the time series span* /or
e2ample( if n L +( we have
+ (t )=(a
t7
w
7
)+(a
t +
w
+
)* ":;#
.ne 5uestion needing to be addressed is how the defuzzification operator deployed here should
be interpreted from a fuzzy logical point of view* 0he thought here is simply to consider the weights
as a fuzzy relationship between past values "inputs# and the future value "output#* ach w
i
represents
the strength of the causal relationship between a given input and some un=nown output* 0he closer
w
i
is to 7( the stronger the relationship and vice versa*
%t has to be stressed that the defuzzification operator introduced here is not an aggregation
operator from a traditional point of view( since it does not satisfy all of the basic conditions of
aggregation "see Appendi2 %%#* 0he proposed operator has been specifically adapted to solve the
problem at hand because none of the other operators discussed earlier have been found useful in this
conte2t* Averaging operators( for e2ample( never produce outputs less than the minimum value of
arguments or larger than the ma2imum value of arguments* %n the current situation( this re5uirement
is undesirable due to the fact that future demand patterns often fluctuate beyond the boundaries of
previous min and ma2 values* 0o illustrate this( we need to ta=e a closer loo= at the enrollment data
in table 7<* /or t L 7-<9 and n L +( we get( a
7-<+
L 79:;9 and a
7-<7
L 79,::* Assuming +"t# is an
averaging operator( output is restricted to the interval C79,::(79:;9D* 1owever actual output for t L
7-<9 is 79>;< which is out of reach by any averaging operator* $onsider another case for t L 7->7
and n L +* Ae then get a
7->,
L 7;-7- and a
7-<-
L 7;>,<* %f +"t# is the min operator( we get min"a
7->,
(
a
7-<-
# L 7;>,<( and( if +"t# is the ma2 operator( we get ma2"a
7->,
(

a
7-<-
# L 7;-7-* Fut actual output for
t L 7->7 is 7;9>> which also is unreachable by any averaging operator* As a conse5uence( a basic
re5uirement for the defuzzification operator proposed here( is that it covers a broader interval than
min and ma2* A reasonable assumption with regards to the bounds of arguments( a
t - i
( is that they
are within the limits of the defined universe set*
:,
!.!.1 ;stablish'ent Fuzzy Set 4rous 5FS49s6
%n conventional /0&( fuzzy relationships are identified immediately after data have been
fuzzified* 1owever( in the model presented here( the right hand side of the fuzzy relation is not
=nown until the weights have been determined* &o( instead of identifying relationships and
establishing /?RB3s( we establish fuzzy set groups "/&B3s#* 0he purpose of the /&B establishment
is to partition historical data into uni5ue sets of sub patterns which subse5uently are converted into
corresponding if statements* During the first pass of the algorithm( consecutive sets are grouped
pairwise* 0able +7 show the fuzzified data in table 7' grouped in this manner* very /&B appears in
chronological order*
<abe% FS> <abe% FS>
7 MA7( A+N 7+ MA<( A<N
+ MA+( A9N 79 MA<( A;N
9 MA9( A:N 7' MA;( A;N
' MA:( A<N 7: MA;( A>N
: MA<( A<N 7; MA>( A77N
; MA<( A<N 7< MA77( A7'N
< MA<( ADN 7> MA7'( A7;N
> MA>( A77N 7- MA7;( A7<N
- MA77( A77N +, MA7<( A7<N
7, MA77( A7,N +7 MA7<( A7;N
77 MA7,( A<N
Tab%e 2,- Estab%isment o* FS>Bs-
0o e2emplify the principles of grouping( consider year 7-<7( 7-<+ and 7-<9 which respectively
are fuzzified as A
7
( A
+
and A
9
"see table 7'#* 0he pairwise grouping of sets is carried out in the
following order6
M " (t +) ( " (t 7)N=M A
i , t+
, A
i , t7
N*
0he above group( with two elements( is referred to as a second order /&B* Fy following this
principle( the following two second order /&B3s are derived6
M " (7-<7)( " (7-<+) N=M A
7
, A
+
N
and
M " (7-<+)( " (7-<9) N=M A
+
, A
9
N*
%n table +7( these groups are labelled as 7 and +( respectively*
Ultimately( the goal of grouping sets in this manner is to obtain a series of /&B3s free of
ambiguities* An ambiguity occurs( in this conte2t( if two or more /&B3s contain the same
combination of elementsGsets - i*e* they are not uni5ue* /rom table +7( it can be seen that not all
/&B3s are uni5ue* )ote that the /&B3s labelled as :( ; and 7+ are identical( as is the case with > and
7;* %n order to obtain a series of disambiguated /&B3s( we e2tend the ambiguous /&B3s to third
:7
order /&B3s( by including the previous set in the corresponding time series* /or a second order /&B(
the combination (M " (t +) ( " (t 7)N( is e2tended to include " (t 9)( so the respective /&B now
e5uals the following third order /&B6 M " (t 9)( " (t +)( " (t 7)N* 0able ++ shows the e2tensions
of the ambiguous /&B3s identified in table +7*
<abe% =F?t & 2@8 F?t & ,@D F?t & .@ E4ten+e+ FS>
=F?t & .@8F?t & 2@8 F?t & ,@D
: MA<( A<N A: MA:( A<( A<N
; MA<( A<N A< MA<( A<( A<N
> MA>( A77N A< MA<( A>( A77N
7+ MA<( A<N A7, MA7,( A<( A<N
7; MA>( A77N A; MA;( A>( A77N
Tab%e 22- E4ten+ing ambiguous FS>Bs-
0he e2tension process is continued until a uni5ue combination of elements is obtained for each
/&B* /rom table ++( we see that only a single e2tension is re5uired to obtain a uni5ue combination
of elements in this particular case* An updated overview of the /&B3s in table +7 is shown in table
+9*
<abe% FS> <abe% FS>
7 MA7( A+N 7+ MA7,( A<( A<N
+ MA+( A9N 79 MA<( A;N
9 MA9( A:N 7' MA;( A;N
' MA:( A<N 7: MA;( A>N
: MA:( A<( A<N 7; MA;( A>( A77N
; MA<( A<( A<N 7< MA77( A7'N
< MA<( ADN 7> MA7'( A7;N
> MA<( A>( A77N 7- MA7;( A7<N
- MA77( A77N +, MA7<( A7<N
7, MA77( A7,N +7 MA7<( A7;N
77 MA7,( A<N
Tab%e 2.- 'isambiguate+ FS>Bs-
!.!.2 Convertin* FS49s into if state'ents
Defuzzified output( + (t )( is obtained by matching historical patterns with a corresponding if-
then rule* 0he if statements are generated on basis of the content of the /&B3s* 0his tas= is fairly
simple as the se5uence of elements of each /&B is the same as they appear in time* 0hat is( for any
/&B of size n( the elements appear in the same se5uence as in the corresponding time series6
" (t n) ( " (t n+7)(.( " (t 7)*
ach /&B can be therefore easily be transformed into if-then rules of the form6
:+
if ( "(t 7)=A
i , t7
" (t +)=A
i , t+
." (tn+7)=A
i , tn+7
"(t n)=A
i ,t n
) 8
then w
7( t7
=Hw
+( t +
=H.w
n7( t n+7
=Hw
n, tn
=H
/or practical reasons( the se5uence of conditions in the if-statement appear in reversed order
compared to their e5uivalent /&B3s* /or e2ample( an /&B of the form6
M A
i ,t +
( A
i ,t 7
N(
is converted into an e5uivalent if-rule of the form6
if ( "(t 7)=A
i , t7
" (t +)=A
i , t+
)*
Ahen a rule is matched( the resultant weights are returned and the forecasted value( + (t )( is
computed according e5uation ::* 0o illustrate this( suppose we need to find a matching if-then rule
when forecasting the enrollment for year 7-<9* /rom table 7'( we get
" (7-<7)=A
7
and
" (7-<+)=A
+
for t L 7-<9* )ow( assume the following if-then rule already e2ists in the current rule
base6
if ( "(t 7)=A
7
" (t +)=A
7
)8
then w
7( t7
=,*;'>>w
+(t+
=,*9>>+*
0he above rule is then matched as6
if ( "(7-<97)=A
7
" (7-<9+)=A
7
) 8
then w
7(7-<+
=,*;'>>w
+(7-<7
=,*9>>+*
Using e5uation ::( the forecasted enrollment for year 7-<9 is computed as
+ (7-<9)=(79:;9,*;'>>)+(79,::,*9>>+)=79>;<*;+-79>;>*
Fy processing all of the data in table +9( a series of incomplete if statements are generated as
shown in table +'* %n order to determine the weights( we utilize P&. "see section +*7+# to train the
rules individually to match the data they represent*
:9
"u%e 1atcing (art
7 if ( " (t 7)=A
+
"(t +)=A
7
)
+ if (( "7)=A
9
" (t +)=A
+
)
9 if ( " (t7)=A
:
"(t +)=A
9
)
' if ( " (t7)=A
<
"(t +)=A
:
)
: if ( " (t 7)=A
<
" (t +)=A
<
" (t 9)=A
:
)
; if ( " (t7)=A
<
"(t +)=A
<
" (t9)=A
<
)
< if ( " (t7)=A
>
"(t +)=A
<
)
> if ( " (t 7)=A
77
" (t +)=A
>
" (t9)=A
<
)
- if ( " (t7)=A
77
"(t +)=A
77
)
7, if ( " (t 7)=A
7,
" (t +)=A
77
)
77 if ( " (t 7)=A
<
" (t +)=A
7,
)
7+ if ( " (t 7)=A
<
" (t +)=A
<
" (t 9)=A
7,
)
79 if ( " (t 7)=A
;
" (t +)=A
<
)
7' if ( " (t 7)=A
;
" (t +)=A
;
)
7: if ( " (t 7)=A
>
" (t +)=A
;
)
7; if ( " (t 7)=A
77
" (t +)=A
>
" (t9)=A
;
)
7< if ( " (t 7)=A
7'
" (t +)=A
77
)
7> if ( " (t 7)=A
7;
" (t +)=A
7'
)
7- if ( " (t 7)=A
7<
" (t +)=A
7;
)
+, if ( " (t 7)=A
7<
" (t +)=A
7<
)
+7 if ( " (t 7)=A
7;
" (t +)=A
7<
)
Tab%e 24- >enerate+ i* ru%es in crono%ogica% or+er-
!.!.3 Trainin* the if1then rules 8ith 7S$
%n the following we are going to provide an e2ample of how P&. is utilized to tune the weights
in the defuzzification operator in e5uation ::* 0he user defined parameters are set as follows
7
6
e 0he inertial coefficient( V( e5uals 7*'*
e 0he self confidence and social confidence coefficient( c
7
and c
+
(

respectively( both e5uals +*
e 0he minimum and ma2imum velocity is limited to C-,*,7(,*,7D*
e 0he minimum and ma2imum position is limited to C,(7D*
e 0he number of particles e5uals five*
0he fitness function employed here is the s5uared error "&#( defined by
7 0he parameters are selected based e2perimental results*
:'
#B=( forecast
t
actual
t
)
+
":<#
Fasically the idea is to evaluate the aggregated result( + (t )( against the actual outcome at time
t( and ad!ust the weights in the defuzzification operator such that the s5uared error is minimized* Fy
minimizing & for each t( 4& is minimized as well* %n the following e2ample( the stopping
criteria is defined by setting the minimum & to 9 and the ma2imum number of iterations to :,,
+
*
During the first step of the algorithm( the weights "positions# are initialized* )ote we assume
the e2istence of a stronger relationship between actual output and the more recent observations in
the time series data* &o( if " (t 7) is fuzzified as A
i
and " (t +) as A
&
( a stronger relationship is
assumed to e2ist between A
i
and + (t ) than between A
&
and + (t )* 0herefore( relatively higher
weights are assigned to the most recent observations when positions are initialized* Applying this
approach( w
t=i
will usually remain larger than w
t=iS7
at the point of termination* 0able +: and +;
respectively show the initial positions and velocities of all particles for a matching rule !*
$artic%e $osition , ?w,@ $osition 2 ?w2@ SE
7 ,*<: ,*: >(,+'('<9
+ ,*<: ,*: >(,+'('<9
9 ,*<: ,*: >(,+'('<9
' ,*<: ,*: >(,+'('<9
: ,*<: ,*: >(,+'('<9
Tab%e 23- Initia% (ositions o* a%% (artic%es-
%n the e2ample above( rule 7 in table +' is trained* As can be seen from table +:( the personal best
positions are the same for all particles during initialization* 1ence the personal best positions e5uals
the global best position for all particles*
$artic%e v, v2
7 ,*,,'- ,*,,77
+ ,*,,9+ ,*,,;:
9 ,*,,9' ,*,,>7
' ,*,,+9 ,*,,,-
: ,*,,,< ,*,,'>
Tab%e 25- "an+omize+ initia% ve%ocities o* a%% (artic%es-
Ahen all particles and velocities have been initialized( the velocities are updated before positions
are incremented* @elocities are updated according to e5uation ',* 0he computations yield6
+ &topping criteria is determined on basis of e2perimental results*
::
v
7(7
=(7*',*,,'-)++r
7
(,*<:,*<:)++r
+
(,*<:,*<:) =,*,,;-
v
7(+
=(7*',*,,77)++r
7
(,*:,*:)++r
+
( ,*:,*:) =,*,,7:
v
+(7
=(7*',*,,9+)++r
7
( ,*<:,*<:)++r
+
(,*<:,*<:) =,*,,':
v
+(+
=(7*',*,,;:)++r
7
(,*:,*:)++r
+
(,*:,*:) =,*,,-7
v
9(7
=(7*',*,,9')++r
7
( ,*<:,*<:)++r
+
(,*<:,*<:) =,*,,'>
v
9(+
=(7*',*,,>7)++r
7
(,*:,*:)++r
+
(,*:,*:) =,*,779
v
'(7
=(7*',*,,+9)++r
7
( ,*<:,*<:)++r
+
(,*<:,*<:) =,*,,9+
v
'(+
=(7*',*,,,-)++r
7
(,*:,*:)++r
+
(,*:,*:) =,*,,79
v
:(7
=(7*',*,,,<)++r
7
(,*<:,*<:)++r
+
(,*<:,*<:) =,*,,,7
v
:(+
=(7*',*,,'>)++r
7
(,*:,*:)++r
+
( ,*:,*:) =,*,,;<*
Positions are incremented according to e5uation '7* %ncremented positions after the first
iteration are shown in table +<*
$artic%e w, w2 SE
7 ,*<:'- ,*:,77 >('>>(>>:
+ ,*<:9+ ,*:,;: >(<;<(:<'
9 ,*<:9' ,*:,>7 >(-,<(>-:
' ,*<:+9 ,*:,,- >(+;-(;7>
: ,*<:,< ,*:,'> >('9>('-7
Tab%e 26- Te (ositions o* a%% (artic%es a*ter te *irst iteration-
After the first iteration( none of the computed & values in table +< are less than >(,+'('<9* 0hus no
personal best positions nor global best positions are reached at this point* At some point( the
stopping criteria is met and the algorithm terminates* 0he personal best positions of all particles
after termination are listed in table +>*
$artic%e w, w2 SE
7 ,*;<9> ,*9;-- 7,7:-
+ ,*;>:' ,*9:,+ 7
9 ,*;;>; ,*9;;+ 9+:
' ,*;<+' ,*9'>+ ',:-<
: ,*;><- ,*99>9 7':++
Tab%e 22- Te (ersona% best (ositions o* a%% (artic%es a*ter termination-
According to table +>( particle + has the global best position* 1ence the weights associated to rule !
e5uals ,*;>:' and ,*9:,+. 0he pseudo code for the training algorithm is shown on page :<*
:;
$S: a%goritm *or training o* te i*&ten ru%es
,recondition: a set of untrained if=then rules
,ostcondition: a set of trained if=then rules
*or all rules !
id
M
7* i* a matching pattern id L MA
id,t - 7
non A
id,t - n =7
n A
id,t - n
N is found for rule !
id
M
7*7* retrieve actual value( a
t=i
( from dataset( from inde2 i L , to n*
7*+* initialize global best fitness value as #B
'lobalEbest
L SQ
*or each particle( p
i
( from inde2 i L 7 to . M
7*9* initialize position( w
i&
( from inde2 & L 7 to n*
7*'* initialize velocity( v
i&
( from inde2 & L 7 to n*
7*:* compute defuzzified output( + (t )
i
( by
+ (t )
i
=
_
& =7
n
a
t &
w
i&
*
7*;* compute s5uared error( #B
i
( by
#B
i
=(+ (t )
i
actual
t
)
+
*
7*<* initialize local best fitness value( #B
localEbest
( as
#B
localEbest
L #B
i
7*>* initialize local best position( localEbest
i&
( as
localEbest
i&
@ w
i&
from & L 7 to n
7*-* i* #B
i
p #B
'lobalEbest
M
7*-*7* update global best fitness value( #B
'lobalEbest
( value as
#B
'lobalEbest
@ #B
i

7*-*+* update global best position( 'lobalEbest
&
( as
'lobalEbest
&
@ w
i&
from & L 7 to n
NGGif
NGGfor
)i%e stopping criteria is unsatisfiedM
*or each particle( p
i
( from inde2 i L 7 to . M
7*7,* update velocity( v
i&
( from & L 7 to n by
v
i&
L

V*qv
i&
F c
7
qr
7
"localEbest
i&
- w
i&
# S c
+
qr
+
"'lobalEbest
&
- w
i&
#
7*77* i*

:
min
r v
i&
7*77*7* set v
i&
@ :
min
7*7+* i* :
ma2
p v
i&
7*7+*7* set v
i&
@ :
ma2
7*79* update position from inde2 i L 7 to n by
w
i&
L w
i&
S v
i&
7*7'* goto step 7*:*
7*7:* goto step 7*;*
7*7;* i* #B
i
p #B
localsbest
M
7*7;*7* update local best fitness value by
#B
localsbest
L #B
i
7*7;*+* update local best position by
localEbest
i&
@ w
i&
from i L 7 to n
NGGif
7*7<* i* #B
i
p #B
'lobalEbest
M
7*7<*7* update global best fitness value by
#B
'lobalEbest
@ #B
i
7*7<*+* update global best position by
'lobalEbest
i&
@ w
i&
from i L 7 to n
NGGif
NGGfor
NGGwhile
+* update then-part of rule !
id
as
(w
id , t 7
='lobal s best
7
.w
id , t n7
='lobal s best
n7
w
id , n
='lobal sbest
n
)
NGGfor
:<
After the weights have been optimized via P&.( the 3blan=s3 in the then part can be filled* 0he
partially completed if-rules from table +' are shown in fully completed form in table +-* 0he fi2ed
parameters supplied to the training algorithm are e5uivalent to the those listed on page :'*
<abe% 1atcing (art /eigts
7 if ( " (t 7)=A
+
"(t +)=A
7
) then w
7
L ,*;'>> and w
+
L ,*9>>+
+ if (( "7)=A
9
" (t +)=A
+
) then w
7
L ,*;:>; and w
+
L ,*'7,+
9 if ( " (t7)=A
:
"(t +)=A
9
) then w
7
L ,*;;< and w
+
L ,*',>
' if ( " (t7)=A
<
"(t +)=A
:
) then w
7
L ,*;9-: and w
+
L ,*9;-
: if ( " (t 7)=A
<
" (t +)=A
<
" (t 9)=A
:
) then w
7
L ,*''77( w
+
L ,*97:> and w
9
L ,*+;--
; if ( " (t7)=A
<
"(t +)=A
<
" (t9)=A
<
) then w
7
L ,*';9>( w
+
L ,*';': and w
9
L ,*,-<>
< if ( " (t7)=A
>
"(t +)=A
<
) then w
7
L ,*;;-: and w
+
L ,*9-;<
> if ( " (t 7)=A
77
" (t +)=A
>
" (t9)=A
<
) then w
7
L ,*'9<-( w
+
L ,*9>-+ and w
9
L ,*+7<7
- if ( " (t7)=A
77
"(t +)=A
77
) then w
7
L ,*7;,' and w
+
L ,*>79<
7, if ( " (t 7)=A
7,
" (t +)=A
77
) then w
7
L ,*:'-< and w
+
L,*9<->
77 if ( " (t 7)=A
<
" (t +)=A
7,
) then w
7
L ,*:--< and w
+
L,*9>,-
7+ if ( " (t 7)=A
<
" (t +)=A
<
" (t 9)=A
7,
) then w
7
L ,*'7:7( w
+
L ,*9-;; and w
9
L ,*7:>+
79 if ( " (t 7)=A
;
" (t +)=A
<
) then w
7
L ,*;7-' and w
+
L ,*9<97
7' if ( " (t 7)=A
;
" (t +)=A
;
) then w
7
L ,*<:+' and w
+
L ,*9,+
7: if ( " (t 7)=A
>
" (t +)=A
;
) then w
7
L ,*9>;- and w
+
L ,*<,'
7; if ( " (t 7)=A
77
" (t +)=A
>
" (t9)=A
;
) then w
7
L ,*';;>( w
+
L ,*9>'< and w
9
L ,*+<+:
7< if ( " (t 7)=A
7'
" (t +)=A
77
) then w
7
L ,*;:' and w
+
L ,*'+7+
7> if ( " (t 7)=A
7;
" (t +)=A
7'
) then w
7
L ,*;9: and w
+
L ,*',7+
7- if ( " (t 7)=A
7<
" (t +)=A
7;
) then w
7
L ,*;+,+ and w
+
L ,*9><'
+, if ( " (t 7)=A
7<
" (t +)=A
7<
) then w
7
L ,*:-9+ and w
+
L ,*9>97
+7 if ( " (t 7)=A
7;
" (t +)=A
7<
) then w
7
L H and w
+
L H
Tab%e 29- >enerate+ i*&ten ru%es in crono%ogica% or+er-
!.# Conclusion
0his section has presented a modified high order /0& model* %ntroductory( a novel /A was
proposed based on the trapezoid fuzzification approach C'D* 0he proposed algorithm can be applied
to any /0& model incorporating interval partitions* 0he algorithm is regarded as an improvement of
similar wor= C'D in the sense that fuzzification is carried out automatically* 2perimental results
indicate that forecast accuracy can improved using the proposed /A although this was not the goal
per se* Actually the main intention has been to develop an approach where interval partitions are
determined ob!ectively without the need of user intervention* Fased on current test results( it is
believed this goal has been achieved*
:>
0he emphasis of the wor= presented here has been to improve consistency between forecast
rules and the data they derive from* %n order to achieve this( a defuzzification operator is proposed
ad hoc* %n traditional models( output is defuzzified via interval "or fuzzy set# operations( whereas in
the proposed model( defuzzified output is a weighted sum of actual values* Fy utilizing P&. and
aggregation( forecast rules can be individually tuned to match the data they represent( regardless of
the chosen interval partitions* .verall e2perimental results are presented in the ne2t section*
:-
# ;0eri'ental "esults
%n section '( the concepts of a new /0& model were discussed but we have not yet
demonstrated overall model performance in terms of forecast accuracy* 0he purpose of this section
is to evaluate the performance of the proposed model vis-E-vis other related prediction models*
Performance is compared by the same principle as previously shown in the thesis( namely by
evaluating the performance of forecast rules( using the same dataset they derive from*
#.1 Co'arin* different FTS 'odels
Aear Actua%
Enro%%ment
Cen
?or+er C .@
C97D
<i E Ceng
C>D
Sing
?or+er C .@
C9+D
CenE9su
C7<D
CenECung
?or+er C 9@
C7:D
FU: et a%
?or+er C 9@
C9,D
$ro(ose+
mo+e%
7-<7 79,:: - - - - - - -
7-<+ 79:;9 - 79:,, - 79<:, - - -
7-<9 79>;< - 79:,, - 79><: - - 79>;>
7-<' 7';-; 7':,, 7':,, 7'<:, 7'<:, - - 7';-;
7-<: 7:';, 7::,, 7::,, 7:<:, 7:9<: - - 7:';,
7-<; 7:977 7::,, 7::,, 7::,, 7:979 - - 7:9,-
7-<< 7:;,9 7::,, 7::,, 7::,, 7:;+: - - 7:;,+
7-<> 7:>;7 7::,, 7::,, 7::,, 7:>79 - - 7:>;7
7-<- 7;>,< 7;:,, 7;:,, 7;:,, 7;>9' 7;>'; - 7;>,;
7->, 7;-7- 7;:,, 7;:,, 7;:,, 7;>9' 7;>'; 7;>-, 7;-7-
7->7 7;9>> 7;:,, 7;:,, 7;:,, 7;'7; 7;'+, 7;9-: 7;9-,
7->+ 7:'99 7::,, 7::,, 7::,, 7:9<: 7:';+ 7:'9' 7:'9'
7->9 7:'-< 7::,, 7::,, 7::,, 7:9<: 7:';+ 7::,: 7:'-<
7->' 7:7': 7::,, 7::,, 7:+:, 7:7+: 7:7:9 7:7:9 7:7'9
7->: 7:7;9 7::,, 7::,, 7::,, 7:7+: 7:7:9 7:7:9 7:7;9
7->; 7:->' 7::,, 7::,, 7::,, 7:-9> 7:-<< 7:-<7 7:->+
7->< 7;>:- 7;:,, 7;:,, 7;:,, 7;>9' 7;>'; 7;>-, 7;>:-
7->> 7>7:, 7>:,, 7>:,, 7>:,, 7>+:, 7>799 7>7+' 7>7:,
7->- 7>-<, 7>:,, 7>:,, 7>:,, 7>><: 7>-7, 7>-<7 7>-<7
7--, 7-9+> 7-:,, 7-:,, 7-:,, 7-+:, 7-99' 7-99< 7-9+>
7--7 7-99< 7-:,, 7-:,, 7-:,, 7-+:, 7-99' 7-99< 7-99;
7--+ 7>><; 7>:,, 7>:,, 7><:, 7>><: 7>-7, 7>>>+ 7>><:
- - - - - - -
1SE >;;-' >:,', <;:,- :;77 77,7 +9' 7
1A$E 7*:9 7*:9 7*'7 ,*9; ,*7: ,*,7' ,*,,;
Tab%e .0- Com(aring +i**erent *uzzy time series mo+e%s-
Different /0& models are compared in terms of 4& and 4AP in table 9,* All of the /0&
models referenced in the table( are among those with the highest forecasting accuracy found in
literature* 0he 4& and 4AP of the proposed model is 7 and ,*,,;( respectively* Foth measures
;,
are lower than for any of the referenced models in the table* Fased on these results( it can be
concluded that the proposed model outperforms any e2isting /0& model in the training phase*
%ntroductory( in section 7( it was argued that the number forecast rules decreases as the order
increases in high order models* 0his is evident from table 9,( when considering the forecasted
enrollments by the two models by $henG$hung C7:D and Uuo et al C9,D( as it can be noted that the
first < - > years of enrollment are not forecasted* 4oreover( by increasing the order( additional
combinations of patterns "fuzzy sets# have to be matched which reduces the probability of finding
e5uivalent pattern combinations in future data*
#.2 Conclusion
$omparative e2periments conducted so far show that the proposed model outperforms its
counterparts* 1owever there is not sufficient evidence to conclude whether this is a good
forecasting method in general( since this re5uires more e2tensive research* Fecause this method of
forecasting is inherently rule based( its practical usefulness highly depends on its abilities to derive
matching forecast rules( and consistency of those( under un=nown conditions* 0his( however( has
not been the focus area of the current pro!ect nor any other related research found in literature* &o(
as for now( this aspect of /0& remains virtually une2plored*
;7
% Final Conclusion
0his pro!ect contributes to current research in two ways* /irst( a novel approach has been
developed which combines aggregation and P&.* Fy combining these techni5ues( forecast rules can
be individually tuned to match the data they represent( regardless of selected interval partitions* %t
has been found that the individual tuning of rules reduces the need to increase the model3s order to
improve forecast accuracy( as opposed to the models recently published by the authors in C7:D and
C9,D* As a conse5uence( better data utilization is achieved in form of6 "7# an increased number of
forecast rules8 "+# fewer pattern combinations to be matched with future time series data* 0he
second contribution to current research( is a fuzzification algorithm( developed as a byproduct* 0he
algorithm( which is a further improvement of the wor= published in C'D( uses an ob!ective measure
to automatically generate interval partitions* 2perimental results indicate that forecast accuracy
may be improved( using the proposed fuzzification approach* All in all( comparative e2periments
confirm the proposed model3s superiority over its counterparts under =nown conditions* 1owever
true performance under un=nown conditions has yet to be confirmed*
;+
( "eferences
C7D P* &ong and F*&* $hissom( /orecasting enrollments with fuzzy time series - part %( /uzzy &ets
and &ystems :' "7--9#( pp* 7--*
C+D P* &ong and F*&* $hissom( /orecasting enrollments with fuzzy time series - part %%( /uzzy &ets
and &ystems ;+ "7--'#( pp* 7->*
C9D &*4* $hen( /orecasting enrollments based on fuzzy time series( /uzzy &ets and &ystems >7
"7--;#( pp* 977-97-*
C'D $*1* $heng( J*R* $hang and $*A* jeh( ntropy-based and trapezoid fuzzification fuzzy time
series approaches for forecasting %0 pro!ect cost( 0echnological /orecasting t &ocial $hange
<9 "+,,;#( pp* :+'-:'+*
C:D &*R* &ing( A robust method of forecasting based on fuzzy time series( Applied 4athematics
and $omputation 7>> "+,,<#( pp* '<+-'>'*
C;D U*1* 1uarng( 0*1*U ju and j*A* 1su( A multivariate heuristic model for fuzzy-time series
forecasting( &ystems( 4anagement and $ybernetics 9< "+,,<#( pp* >9;->';*
C<D 0*A* Jilani and &*4*A* Furney( A refined fuzzy time series model for stoc= mar=et forecasting(
&tatistical 4echanics and its Applications 9>< "+,,>#( pp* +>:<-+>;+*
C>D &*0* ?i and j*$* $heng( Deterministic fuzzy time series model for forecasting enrollments(
$omputer and 4athematics with Applications :9 "+,,<#( pp* 7-,'-7-+,*
C-D $*1* $heng( J*A* Aang and $*1* ?i( /orecasting the number of outpatient visits using a new
fuzzy time series based on weighted transitional matri2( 2pert &ystems with Applications 9'
"+,,>#( pp* +:;>-+:<:*
C7,D U* 1uarng and 0*1*U* ju( Ratio-based lengths of intervals to improve fuzzy time series
forecasting( &ystems( 4anagement and $ybernetics 9; "+,,;#( pp* 9+>-9',*
C77D &*4* $hen and J*R* 1wang( 0emperature prediction using fuzzy time series( &ystems(
4anagement and $ybernetics 9, "+,,,#( pp* +;9-+<:*
C7+D 0*A* Jiliani( &*4*A* Furney and $* Ardil( 4ultivariate high order fuzzy time series forecasting
for car road accidents( %nt* Journal of $omputational %ntelligence ' "+,,>#( pp* *
C79D &*0* ?i( j*$* $heng and &*j* ?in( A fcm-based deterministic forecasting model for fuzzy time
series( $omputers and 4athematics with Applications :; "+,,>#( pp* 9,:+-9,;9*
C7'D R*$* 0saur( J*$* . jang and 1*/* Aang( /uzzy relation analysis in fuzzy time series model(
$omputers and 4athematics with Applications '- "+,,:#( pp* :9--:'>*
C7:D &*4* $hen and )*j* $hung( /orecasting enrollments using high-order fuzzy time series and
genetic algorithms( %nt* Journal of %ntelligent &ystems +7 "+,,;#( pp* '>:-:,7*
;9
C7;D &*4* $hen( J*R* 1wang and $*1* ?ee( 1andling forecasting problems using fuzzy time series(
/uzzy &ets and &ystems 7,, "7-->#( pp* +7<-++>*
C7<D &*4* $hen and $*$* 1su( A new method to forecast enrollments using fuzzy time series( %nt*
Journal of Applied &cience and ngineering + "+,,'#( pp* +9'-+''*
C7>D U* 1uarng and 1*U* ju( A dynamic approach to ad!usting lengths of intervals in fuzzy time
series forecasting( %ntelligent Data Analysis > "+,,'#( pp* 9-+<*
C7-D U* 1uarng( ffective lengths of intervals to improve forecasting in fuzzy time series( /uzzy
&ets and &ystems 7+9 "+,,7#( pp* 9><-9-'*
C+,D J*&* &ullivan and A*1* Aoodall( A comparison of fuzzy forecasting( /uzzy &ets and &ystems
;' "7--'#( pp* +<--+-9*
C+7D 1*&* ?ee and 4*0* $hou( /uzzy forecasting based on fuzzy time series( %nt* Journal of
$omputer 4atemathics >7 "+,,'#( pp* <>7-<>-*
C++D U* 1uarng and 1*U* ju( A type + fuzzy time series model for stoc= inde2 forecasting( Physica
A 9:9 "+,,:#( pp* '':-';+*
C+9D 1*U* ju( Aeighted fuzzy time series models for 0A%u forecasting( Physica A 9'- "+,,:#( pp*
;,--;+'*
C+'D &*0* ?i and j*P* $hen( )atural partioning-based forecasting model fore fuzzy time series(
/uzzy &ystems 9 "+,,'#( pp* 79::-79:-*
C+:D $*$* 0sai and &*J* Au( /orecasting enrollments with high-order fuzzy time series( /uzzy
%nformation Processing &ociety "+,,,#( pp* 7-;-+,,*
C+;D U* 1uarng and 0*1*U* ju( 0he application of neural networ=s to forecast fuzzy time series(
Physica A 9;9 "+,,;#( pp* '>7-'-7*
C+<D 0*?* $hen( $*1* $heng and 1*J* 0eoh( 1igh-order fuzzy time series based on multi period
adaption model for forcasting stoc= mar=ets( Physica A 9>< "+,,>#( pp* ><;->>>*
C+>D &*0* ?i and j*P* $hen( )atural partitioning based forecasting model for fuzzy time-series( +,,'
% %nt* $onference on /uzzy &ystems 9 "+,,'#( pp* 79::-79:-*
C+-D $*1* $heng( B*A* $heng and J*A* Aang( 4ulti-attribute fuzzy time series method based on
fuzzy clustering( 2pert &ystems with Applications 9' "+,,>#( pp* 7+9:-7+'+*
C9,D %*1* Uuo( &*J* 1orng( 0*A* Uao( 0*?* ?in( $*?* ?ee and j* Pan( An improved method for
forecasting enrollments based on fuzzy time series and particle swarm optimization( 2pert
&ystems with Applications 9; "+,,-#( pp* ;7,>-;77<*
C97D &*4* $hen( /orecasting enrollments based on high-order fuzzy time series( $ybernetics and
&ystems6 An %nt* Journal 99 "+,,+#( pp* 7-7;*
C9+D &*R* &ing( A simple time variant method for fuzzy time series forecasting( $ybernetics and
;'
&ystems6 An %nt* Journal 9> "+,,<#( pp* 9,:-9+7*
C99D 1*U* ju( A refined fuzzy time-series model for forecasting( Physica A 9'; "+,,'#( pp* ;:<-
;>7*
C9'D B*J* Ulir and F* juan( /uzzy sets and fuzzy logic6 theory and applications( %&F)6 ,-79-
7,77<7-:( Prentice 1all( 7--:*
C9:D J* Janzen( 0utorial on fuzzy logic(
http6GGwww*iau*dtu*d=Gv!!GpubsGlogic*pdf "validated .ctober +> +,,-#*
C9;D ?*A* wadeh( /uzzy sets( %nformation and $ontrol > "7-;:#( pp* 99>-9:9*
C9<D D* Dubois and 1* Prade( .perations on fuzzy numbers( %nt* Journal of &ystems &cience "7-<>#(
pp* ;79-;+;*
C9>D 0*&* ?iou and 4* J* Aang( Ran=ing fuzzy numbers with integral value( /uzzy &ets and
&ystems :, "7--+#( pp* +'<-:::*
C9-D R* R* jager( Ran=ing fuzzy subsets over the unit interval( Proc* 7-<> $D$ "7-<>#( pp* 7'9:-
7'9<*
C',D &*)* &ivanandam( &* &umathi and &*)* Deepa( %ntroduction to fuzzy logic using 4A01?AF(
%&F)6 7,9-:',-9:<>,-<( &pringer-@erlag( +,,<*
C'7D 4* Detyniec=i( /undamentals on aggregation operators(
http6GGwww-poleia*lip;*frGvmarcinGpapersGDetysAB.Ps,7*pdf "validated .ctober +> +,,-#*
C'+D B* 4ayor and * 0rillas( .n the representation of some aggregation functions( Proc* of %&4@?
"7->;#( pp* 777-77'*
C'9D 1*?* ?arsen( %mportance weighted .AA aggregation of multicriteria 5ueries( Proc* of the
)orth American /uzzy %nformation Processing &ociety $onference "7---#( pp* <',-<''*
C''D 1*?* ?arsen( /uzzy =nowledge operators6 averaging operators "AAU lecture note#(
http6GGaaue*d=GvlegindG/?s+,,;G/?-,'G/?-,'m+,noter*pdf "validated .ctober +> +,,-#*
C':D 1*?* ?arsen( /uzzy =nowledge operators6 triangular norm operators "AAU lecture note#(
http6GGaaue*d=GvlegindG/?s+,,;G/?-,9G/?-,9m+,)oter*pdf "validated .ctober +> +,,-#*
C';D J* Uennedy and R* berhart( Particle swarm optimization( Proc* of % %nt* $onference on
)eural )etwor= "7--:#( pp* 7-'+-7-'>*
C'<D J* Uennedy( R* berhart and j* &hi( &warm intelligence( %&F)6 7-::>;,-:-:--( 4organ
Uaufman( +,,7*
C'>D P* &ong and F*&* $hissom( /uzzy time series and its models( /uzzy &ets and &ystems :'
"7--9#( pp* +;--+<<*
;:
+ Aendi0 I
%n this section we will show how the multi-argument form of the algebraic sum( in e5uation 9>( is
derived as a proof of induction*
#tep $.
/irst we must show that statement #
i=7
n
( x
i
)
as
=7
]
i=7
n
(7x
i
) is true for the base case( n=+ * /or
n=+ we get6
#
i=7
+
( x
i
)
as
= x
7
+x
+
x
7
x
+
=7(7x
7
)(7x
+
) L 7
]
i =7
n
(7x
i
) *
0hat is( the statement holds for n L +*
#tep %.
Assume that the same statement holds for n L /* Ae must then show that the statement is true for n
L / S 7* )ow let (=#
i =7
/
( x
i
)
as
=7
]
i=7
n
(7x
i
)* 0hen we get #
i=7
/ +7
( x
i
) = (+x
/+7
(x
/+7
( or6
#
i=7
/ +7
( x
i
) =
(
7
]
i=7
/
(7x
i
)
)
+x
/ +7

(
7
]
i =7
/
(7x
i
)
)
x
/+7
=7
]
i=7
/
(7x
i
)+x
/ +7

(
x
/ +7
x
/ +7

]
i=7
/
(7x
i
)
)
=7
]
i=7
/
(7x
i
)+x
/+7
x
/+7
+x
/+7

]
i =7
/
(7x
i
)
=7
]
i=7
/
(7x
i
)+x
/+7

]
i=7
/
(7x
i
)
=7
(]
i =7
/
(7x
i
)
)
(7x
/+7
)
=7
]
i=7
/+7
(7x
i
)*
&ince #
i=7
n
( x
i
)
as
=7
]
i=7
n
(7x
i
) applies for n L / S 7( it remains true for every positive integer n
by the induction principle which concludes the proof*
;;
. Aendi0 II
%n this section we will e2amine whether the defuzzification operator in e5uation :: is an
aggregation operator in classical fuzzy logic sense* Recall from section +*77 that an aggregation
operator is a real function h( mapped over the unit interval6
h6 | ,(7
n
-| ,(7
which as minimum satisfies the following conditions6
7* h(,(., ,)=, and h(7(.,7)=7 "boun+ary con+itions#8
+*
h( x
7
,., x
n
) 4 h( (
7
,., (
n
)(
if
x
i
4 (
i
for all i N8 "monotonic increasing#
9* h is continuous with respect to each of its arguments8
0o see whether condition 7 holds( we assume the operator in :: only accepts arguments
between , and 7* /rom e5uation :: we =now that the weights( w
i
( are sub!ected to the condition
,<w
i
<7*
0his implies that the largest possible output is obtained when w
i
L 7 for all i N*
0herefore we select w
i
L 7 for all arguments a
i
* )ow we can easily see that the lower boundary
condition is satisfied since
a
7
7+a
+
7+.+a
n
7=,(
if a
i
L , for all i N* Regarding the upper
boundary condition( it can easily be seen that it holds for the unary case "i*e* n L 7#( since
a
7
7=7(
if
a
7
L 7* 1owever for n R +( it does not hold since the statement
a
7
7+a
+
7+.+a
n
7=7
is not true(
if a
i
L 7 for all i N* 1ence the operator in e5uation :: is not an aggregation operator with regards
to condition 7* 4onotonicity and continuity are trivially satisfied*
;<