Vous êtes sur la page 1sur 3

# MATLAB SOFTWARE TOOL FOR NEURO-FUZZY IDENTIFICATION AND DATA

## ANALYSIS. Ver. 0.1

BY
Gianl!a B"n#e\$%i- PhD Student - IRIDIA ULB Bruxelles, Belgium e-mail:gbonte@ulb.ac.be
Mar" Bi#a##ari - Graduate Students , Dipartimento di Eletronia , Politenio di !ilano, Ital",
GENERAL DESCRI&TION
#his so\$t%are is the preliminar" &ersion o\$ a larger pro'et, %hose omplete &ersion %ill (e
a&aila(le in the next months) It has (een implemented in the !atla( language &er) *)+)
#his &ersion per\$orms the identi\$iation o\$ a input-output mapping, a&aila(le in the \$orm o\$ a set o\$
, input-output examples, %ith a \$u--" arhiteture, inspired to #a.agi-Sugeno approah /01 /+1
/*1) #he \$u--" arhiteture is harateri-ed (" a set o\$ rules, %hih are properl" initiali-ed and
tuned (" a learning algorithm) #he rules are in the \$orm:
02 i\$ input0 is A00 and input + is A0+ then output 3\$04input0,input+2
+2 i\$ input0 is A+0 and input + is A++ then output 3\$+4input0,input+2
)))))))))))))
%here Ai' are parametri mem(ership \$untion)
#he program pro&ides a set o\$ strutural alternati&es in the de\$inition o\$ the \$u--" model: the user
ma" hoose the shape o\$ the mem(ership \$untions o\$ the anteedents 4gaussian or triangular2, the
parametri \$orm o\$ the onse5uent 4onstant or linear \$2, the om(ination method o\$ the rules
4normali-ed or not2) 6urther, it is possi(le to hoose (et%een t%o di\$\$erent lustering poliies,
.7mean or 8"perplane 6u--" 9lustering /:1, to pro&ide the identi\$iation algorithm %ith good
initial &alues o\$ the enters and (ases o\$ the mem(ership \$untions)
#he program searhs also \$or the ;right; omplexit" 4num(er o\$ rules2 o\$ the arhiteture ("
adopting a proedure o\$ ross-&alidation on the a&aila(le data set) It starts %ith a minimal num(er
o\$ rules and at eah step inrease the num(er o\$ rules (" restarting the glo(al proedure, until a
maximum num(er o\$ rules is reahed 4the user is \$ree to set properl" %hat is the desired range o\$
omplexit" to range o&er2) #hen eah struture is harateri-ed (" its error in generali-ation,
estimated (" a proedure o\$ ross-&alidation and the optimal num(er o\$ rules is searhed ("
omparing the ross &alidation error o(tained at di\$\$erent le&els o\$ omplexit") At the end o\$ the
glo(al training phase, the ross &alidation error is plotted against the num(er o\$ rules used and the
user is as.ed to hoose the le&el o\$ omplexit" at %hih the \$u--" s"stem loo.s to per\$orm (etter
4\$ig)02: a\$ter this, the \$u--" s"stem o\$ the hosen omplexit" is identi\$ied \$rom srath (" using the
%hole data set) #he model is no% read" to (e used \$or predition on ne% samples)
0 5 10 15 20 25 30
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
\$ig) 0) 9ross-&alidation error &s) omplexit" diagram
REFERENCES'
/01 Bersini 8), Bontempi G), Deaeste.er 9) 40<<=2) C"\$%arin( RBF an) F**+ In,eren!e
S+-#e\$- "n T.e"re#i!al an) &ra!#i!al Ba-i- in F.Fogelman-Soulie', P. Gallinari (Eds.) ICA
'!",International Con#erence on Arti#icial eural et\$or%s, Paris , &ol) 0, pp) 0><-0?*
/+1 Bersini 8), Bontempi G) 40<<>2 N"/ !"\$e- #.e #i\$e #" )e,**i,+ ner"-,**+ \$")el-
IRIDIA - ULB #ehnial Report <>-? 4su(mitted to Fu&&' Sets and S'stems2
/:1 Be-de. @)9), Anderson I)!.0 An A%%li!a#i"n ", #.e !-Varie#ie- Cl-#erin( Al("ri#.\$- #"
&"l+("nal Cr1e Fi##in(, IEEE (rans. S'st., )an C'bern),&ol)S!9-0=,no =) pp >:?->*0
SeptABt 0<C=)
/*1 Sugeno !) , Dasu.a%a #), A F**+-L"(i!-Ba-e) A%%r"a!. #" 2ali#a#i1e M")elin(,
IEEE (rans. Fu&&' S'st., &ol)0 no 0,6e(r 0<<:)
NOTES AND TI&S 4names o\$ !atla( &aria(les are in 3"l)2:
#he desired \$eature must (e seleted (e\$ore running the program ;unommenting; the
orresponding line in the module \$u--"7()m) An analogous graphi &ersion is pro&ided in the \$ile
\$u--"7g)m)
#hree options 4mutuall" exlusi&e2 are a&aila(le :
%r3;omp7err;E
%r3;one7mod;E
%r3;e&aluate;E
Fith %r3Gomp7errH the program searhes \$or the (est struture 4in terms o\$ num(er o\$ rules2 ("
ranging the models o&er the set [!"\$%4\$in !"\$%4\$a5]) !"\$%4\$in and !"\$%4\$a5 represent
the extremes o\$ the omplexit" inter&al 4minimum and maximum num(er o\$ rules, respeti&el"2)
Fith %r3Gone7modH a partiular struture %hose num(er o\$ rules is pro&ided in the &aria(le
!"\$%, is identi\$ied starting \$rom the a&aila(le training data set)
6or these t%o modes, the user an hoie the desired model on\$iguration in terms o\$ mem(ership
shape 4gaussian or triangular2, loal model struture 4onstant or linear2, rule o\$ model
omposition 4%eigthed or not2, initiali-ation proedure 4.-means or \$u--" lustering2)
Fith %r3Ge&aluateH the program e&aluates the per\$ormane o\$ a model 4ontained in matla( \$ile
spei\$ied %ith the &aria(le #\$")4i)2 on a test set 4ontained in the matla( \$ile spei\$ied %ith the
&aria(le in4i)6. #he predited output is sa&ed in the \$ile %re)4i).
#he samples matrix must (e named )a#a and must (e sa&ed as a matla( .mat \$ile %hose name and
path must (e pro&ided to the program through the reser&ed line in the module #u&&'*b.m)
#he data matrix is a matrix /n"4)a#a, n"41arI01 %here n"4)a#a is the num(er o\$ ro%s and
n"41arI0 is the num(er o\$ olumns) n"4)a#a represents the num(er o\$ samples a&aila(le, n"41ar
is the num(er o\$ inputs %hile the num(er o\$ outputs is al%a"s restrited to one)
All the inputs and the output should (e saled (et%een -0 and 0) Automati saling %ill (e a
\$eature o\$ the next &ersion)
#he results o\$ the omputation are the : matries: !en#er- 3a-e- and %ar, %hih, ontaining
respeti&l" the oordinates o\$ the enters and the (ases o\$ the mem(ership \$untions and the
parameters o\$ the output model o\$ the rules, gi&e a desription o\$ the \$u--" s"stem 'ust identi\$ied)
#he : matries !en#er-, 3a-e- and %ar are not automatiall" sa&ed to dis. at the end o\$ the
omputation (ut the" are a&ila(le in the memor" spae o\$ !atla()
N.B. 6or an" 5uestions,(ugs, pro(lesm do not hesitate to ontat the authors