Pavel Ošmera - Vortex-Fractal Physics: Page 123-129

EUROPEAN POLYTECHNICAL INSTITUTE KUNOVICE
P R O C E E D I N G S

FOURTH INTERNATIONAL CONFERENCE ON SOFT
COMPUTING APPLIED IN COMPUTER AND
ECONOMIC ENVIRONMENT

ICSC 2006

January 27, Kunovice, Czech Republic

Edited by:
Prof. Ing. Imrich Rukovansk, CSc, and Doc. Ing. Pavel Omera, CSc

Prepared for print by:
Bc. Andrea imonov, DiS., Bc. Pavel Kubala, DiS. and Ing. Petr Matuk

Printed by:
European Polytechnical Institute Kunovice, 2006

ISBN : 807314084-5
FOURTH INTERNATIONAL CONFERENCE ON SOFT COMPUTING
APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENT

ICSC 2006

Organized by

THE EUROPEAN POLYTECHNICAL INSTITUTE, KUNOVICE
THE CZECH REPUBLIC

Conference Chairman

Ing. Oldich Kratochvl, Dr.h.c., rector

Conference Co-Chairmen

Prof. Ing. Imrich Rukovansk, CSc.
Assoc. Prof. Ing.Pavel Omera, CSc.

INTERNATIONAL PROGRAMME COMMITEE

O. Kratochvl Chairman (CZ)
M. Baraski (Poland)
J. Batinec (Czech Republic)
J. Diblk (Czech Republic)
P. Dostl (Czech Republic)
U. K. Chakraborthy (USA)
B. Kulcsr (Hungary)
V. Mikula (Czech Republic)

P. Omera (Czech Republic)
J. Petrucha (Czech Republic)
I. Rukovansk (Czech Republic)
G. Vrtesy (Hungary)
W. Zamojski (Poland)
J. Zapletal (Czech Republic)
T. Walkowiak (Poland)

ORGANIZING COMMITEE

I. Rukovansk (Chairman)
P. Omera
A. imonov
P. Kubala
J. Kavka
P. Matuk
M. Balus
I. Polkov
J. chov
T. Chmela
J. Mek
. Mikul
R. Jura
M. Zlek

Session 1: ICSC
Chairman: Doc. RNDr. Josef Zapletal, CSc.

ICSC FOURTH INTERNATIONAL CONFERENCE ON SOFT COMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIRONMENT
EPI Kunovice, Czech Republic, January 27, 2006
5

OBSAH
A MESSAGE FROM THE GENERAL CHAIRMAN OF THE CONFERENCE...................................................7
POZNMKA K ROZHODOVN ZA RIZIKA A NEJISTOTY
Zapletal Josef...............................................................................................................................................................9
THE USE OF FUZZY LOGIC FOR SUPPORT OF DIRECT MAILING
Dostl Petr .................................................................................................................................................................21
THE COLLATION OF VARIOUS METHODS FOR THE SOLUTION OF TRANSPORTATION PROBLEMS
Abdurrzzag Tamtam..................................................................................................................................................27
SOLUTION OF STRUCTURAL INTERBRANCH SYSTEM OF A DYNAMIC MODEL
Batinec Jaromr, Diblk Josef ....................................................................................................................................35
PROBABILITY THEORY AND STATISTICS IN THE COMBINED FORM OF STUDY OF THE
BACHELOR STUDENT PROGRAMMES AT FEEC BUT
Novk Michal ............................................................................................................................................................43
APLIKACE FUZZY SYSTM PRO PODPORU ROZHODOVN A ZEN
Mikula Vladimr, Petrucha Jindich............................................................................................................................45
EXAMPLES OF USING CONCEPTS OF PROBABILITY THEORY IN MANAGEMENT DECISION
MAKING
Novk Michal, Fajmon Betislav................................................................................................................................51
OPTIMIZATION OF MATERIAL CHARACTERIZATION BY ADAPTIVE TESTING
Vrtesy Gbor, Tom Ivan, Mszros Istvn .............................................................................................................57
VYUIT KOMPLETNHO GENETICKHO ALGORITMU PRO EEN OPTIMALIZACE VROBNHO
PROCESU Z HLEDISKA MAXIMALIZACE ZISKU
Kostiha Ji ................................................................................................................................................................65
REVITALIZING COMPANY INFORMATION SYSTEMS AND COMPETITIVE ADVANTAGES
Lacko Branislav.........................................................................................................................................................73
MODEL LEARNING AND INFERENCE THROUGH ANFIS
Amalka Al Khatib......................................................................................................................................................81
GRAMMATICAL EVOLUTION WITH BACKWARD PROCESSING
Omera Pavel, Popelka Ondej, Rukovansk Imrich ...................................................................................................89
OBJECT RECOGNITION BY MEANS OF NEW AL
astn Ji, Minak Martin.......................................................................................................................................99
APLIKCIA TERIE GRAFOV V INTELIGENTNOM DOPRAVNOM SYSTME
Klietik Tom......................................................................................................................................................... 105
THE VORTEX-FRACTAL THEORY OF THE UNIVERSE STRUCTURES
Omera Pavel........................................................................................................................................................... 109
VORTEX-FRACTAL PHYSICS
Omera Pavel........................................................................................................................................................... 123
VZNAM MONITOROVN POTAOVCH ST
Rukovansk Imrich.................................................................................................................................................. 131
ANALZA DAT S VYUITM NEURONOVCH ST A KONTINGENNCH TABULEK
Petrucha Jindich...................................................................................................................................................... 137
DETECTION OF INITIAL DATA GENERATING BOUNDED SOLUTIONS OF LINEAR DISCRETE
EQUATIONS
Batinec Jaromr, Diblk Josef .................................................................................................................................. 143
ON SOME PROPERTIES OF FRACTIONAL CALCULUS
Krupkov Vlasta, marda Zdenk ............................................................................................................................ 157

6

ON THE STABILITY OF LINEAR INTEGRODIFFERENTIAL EQUATIONS
marda Zdenk........................................................................................................................................................ 163
EXISTENCE OF POSITIVE SOLUTIONS FOR RETARDED FUNCTIONAL DIFFERENTIAL EQUATIONS
WITH UNBOUNDED DELAY AND FINITE MEMORY
Diblk Josef, Svoboda Zdenk.................................................................................................................................. 169
APPLICATION OF NON SIMPLEX METHOD FOR LINEAR PROGRAMMING
Tomov Marie........................................................................................................................................................ 173
AUTHOR INDEX.................................................................................................................................................. 179

7

A M ESSAGE F ROM TH E GEN ERAL CH AI RM AN OF TH E CON F EREN CE

D ear guest s and part i ci pant s at t hi s conf erence

You are get t i ng t hi s ant hol ogy f rom t he 4t h sci ent i f i c conf erence I CSC 2006
I nt ernat i onal Conf erence on Sof t Comput i ng Appl i ed i n Comput er and Economi c
Envi ronment .

I t i s my pl easure t o gi ve t hanks f or t he preparat i on of t he conf erence t o Prof . I ng.
I mri ch Rukovansky, CSc.

Fuzzy l ogi c and neuron net works have become an i mport ant part of work at our
U ni versi t y duri ng t he l ast f our years. The conf erence part i ci pant s gave t hei r
papers on t hei r sci ent i f i c work and resul t s gai ned duri ng t he l ast year. Some
academi cs (D oc. Pavel Omera CSc, I ng. D ost l ), t hei r papers are a part of t hi s
ant hol ogy, wi l l send t hei r resul t s t o wel l known magazi nes abroad hi ghl i ght i ng
t hat t hei r research resul t s were publ i shed i n t hi s ant hol ogy f or t he f i rst t i me.

I am pl eased t hat t he academi cs f rom Czech, Sl ovak, Hungary, Pol and and Russi a
uni versi t i es t ook part i n t he conf erence.

Thi s conf erence was not onl y an i mport ant sci ent i f i c but al so a soci al event .

Al l ow me t o wi sh l ot s of success t o al l t he part i ci pant s of t he conf erence. I ki ndl y ask t hem t o be wi t h us at t he 5
t h

conf erence I CSC 2007.

K unovi ce, January 27, 2006 D i pl . I ng. Ol di ch K rat ochvl , D r.h.c.
rect or
Ing. Oldich Kratochvl, Dr.h.c.

8

9

POZNMKA K ROZHODOVN ZA RIZIKA A NEJISTOTY
Josef Zapletal

Evropsk polytechnick institut, s.r.o.

Abstrakt: V lnku se popisuj metody rozhodovac analzy, kter maj ponkud odlin rysy od
rozhodovacch metod operan analzy, kter pedstavuj pedevm optimalizan nstroje vhodn pro
een jednoduch, dobe strukturovanch rozhodovacch problm. Naopak, charakteristickm rysem
rozhodovac analzy je to, e sna skloubit exaktn postupy a modelov nstroje se znalostmi a
zkuenostmi eitel tchto problm. Heuristick metody vznamn ovlivuj postupy a vsledn een
problm. Uvedeme zkladn pojmy , metody a nstroje rozhodovac analzy, resp. rozhodovn za rizika a
nejistoty. Mezi n bude patit pojem subjektivn pravdpodobnost, funkce utility za rizika a nkter grafick
nstroje podpory een rozhodovacch problm za rizika a nejistoty.

Klov slova: Rozhodovac analza, operan analza, deterministick metody, stochastick metody,
subjektivn pravdpodobnost, pomr szek, funkce utility za rizika, sklon rozhodovatele k riziku, averze
k riziku, sklon k riziku, konkvn, linern a konvexn funkce utility, jistotn ekvivalent.

1 vod
Tento pspvek m bt jakmsi metodickm nvodem pro studenty EPI, kte se zabvaj ve svch projektech
problematikou rozhodovn a to zejmna rozhodovn zen nedeterministickch proces. Vchozm materilem se
mn stala skripta Jiho Fotra a Jiho Ddiny Manarsk rozhodovn a dle prce [2], [4], [5], [7], [13]. Pklad
v podkapitole Metoda relativnch velikost je pebrn z [6].

2. Subjektivn pravdpodobnosti.
2.1 Objektivn a subjektivn pravdpodobnost
Dleitou soust ppravy rozhodovn je vyjasnit si mon budouc situace. Zejmna oekvan ekonomicko
politick vlivy, kter maj vliv na dsledky uvaovanch variant rozhodovn. Nkter z tchto skutenost mohou bt
nepzniv(nadroda ve velkch geografickch oblastech, nadvroba uritho zbo ve velkch ekonomicky silnch
sttech atd.). Naopak nkter mohou bt pzniv (trn konjunktura, stup urit konkurence z trhu v dsledku
zmodernizovn vroby, zskn novch odbyti pro zabhnutou vrobu aj.).
Je proto nutn, njakm zpsobem pro dal vyhodnocovn stanovit mru nebezpe u nepznivch vliv a mru
nadjnosti u pznivch vliv. Takovou mrou bv zpravidla pravdpodobnost. Rizikov situace a jejich umstn
v konkrtnm ase vak nejsou obvykle zcela a pravideln opakovateln. Nejsme vtinou v situaci, kdy m manar
rozhodnout o nkupu jistho potu nhradnch dl do rezervy pro dan drah stroj a pitom v, kolik a s jakou
pravdpodobnost se tyto dly pokazily a kdy existuje jaksi objektivn systm pravdpodobnost, z nho lze vychzet.
Pi rozhodovn obecnch monch situac nem manar k dispozici minul statistick daje a pro pravdpodobnost
ohodnocen rizikovch situac uplatnit pouze tzv. subjektivn pravdpodobnosti. Tyto jsou zaloeny na pedpokladu, e
kad subjekt, kterm ovem nen pouh jedinec, m urit pedpoklad vvoje a vru v tento vvoj. Subjektivn
pravdpodobnost je pak vyjdenm mry osobnho pesvden subjektu (celho tmu prognostik a manar) na
nastoupen uritho jevu ppadn udlosti.

2.2. seln vyjden subjektivn pravdpodobnosti
Subjektivn pravdpodobnost meme vyjdit bu seln, nebo slovn. seln vyjden me mt dv formy:
Prvn forma pomoc sel od 0 do 1 ppadn od nula procent do 100%. Hodnota pravdpodobnosti nula vyjaduje, e
dan situace nebo jev urit nenastane, hodnota pravdpodobnosti 1, resp. 100% indikuje, e dan situace nebo jev
nastanou s jistotou.

Druh forma selnho vyjden subjektivn pravdpodobnosti je vyjden bu ve form pomru udvajcho poet

10

realizac danho jevu z celkovho potu monch ppad (nap. porucha uritho zazen nastane v prmru jedenkrt
za 100 dn), nebo pomoc tzv. pomru szek. V tomto ppad vyjaduje manar svoji vru ve vskyt danho jevu
(kterm me bt nap. spch nov vyvinutho vrobku na trhu) nap. vrokem typu: Vsadil bych 3:1, e vrobek
bude na trhu spn. Pravdpodobnost trnho spchu vrobku je pak
3
3 +1
= 0, 75.

seln stanoven subjektivnch pravdpodobnost obvykle probh ve spoluprci se specialisty z dan vcn oblasti. Pi
stanovovn spchu vvoje uritho vrobku to me bt vedouc vvojovho tmu, podobn pi urovn ve
prodeje zase marketingov odbornk. Pro ilustraci postup stanoven subjektivnch pravdpodobnost uvedeme dv
metody a to metodu relativnch velikost a metodu kvantil.

Metoda relativnch velikost
Tato metoda je vhodn pro urovn subjektivnch pravdpodobnost jev, kterch je pouze omezen poet. V tto
metod se uruje nejprve ten jev (situace), kterou povauje odbornk za nejpravdpodobnj. Tato pravdpodobnost
pak slou jako zklad pro stanoven pravdpodobnost dalch jev (situac). Pro nzornost uvedeme nsledujc
ilustrativn pklad.

Pedpokldejme, e podnik kupuje nov vrobn zazen, piem s nkupem je teba objednat urit poet kus
vznamnch a vtinou velmi drah nhradn sousti, kter se nhodn pokozuje. Jako podklad pro tuto objednvku
je teba urit pravdpodobnost jednotlivch hodnot potu poruch dan sousti (to jsou v naem ppad jevy, resp.
Situace, kter mohou nastat) bhem provozu doby ivotnosti kupovanho vrobnho zazen. Pokud je toto zazen ji
nkolikansobn v provozu a vedouli se statistiky poruchovosti spad tato loha do operan analzy, konkrtn do
teorie zsob. Vtinou ale nejsou k dispozici dostaten rozshl statistick vyhodnocen a pro stanoven subjektivnch
pravdpodobnost jednotlivch pot poruch se vyuije informac zskanch diskus analytika s odbornkem. Z takov
diskuse vyplynulo, e
maximln pedpokldan poet poruch je pt (poet poruch se me tedy pohybovat od nuly k pti);
nejpravdpodobnj poet poruch jsou dv;
pravdpodobnost jedn resp. t poruch je stejn velk a je piblin dvakrt men ne pravdpodobnost dvou
poruch;
pravdpodobnost dn, resp. pti poruch je stejn velk a je zhruba desetkrt men ne pravdpodobnost dvou
poruch;
pravdpodobnost ty poruch je piblin ptkrt men ne pravdpodobnost dvou poruch.

Jestlie nyn pravdpodobnost vzniku dvou poruch (tj. potu poruch s nejvt pravdpodobnost) ozname P a
pravdpodobnosti nastoupen i poruch jako p
i
, pak z ve uvedenho plyne

o
p P
P
p p
P
p p
P
p
=
= =
= =
=
2
1 3
5
4
2
10
5

Druh rovnice shora vyjaduje tvrzen, e pravdpodobnost vzniku jedn poruchy je stejn velk jako pravdpodobnost
nastoupen t poruch a ob jsou piblin dvakrt men ne pravdpodobnost nastoupen dvou poruch, kterou jsme
oznaili P.

Protoe cel pravdpodobnostn prostor je tvoen pti hodnotami, mus platit

o
p p p p p p + + + + + =
1 2 3 4 5
1 .

11

0
0,05
0,1
0,15
0,2
0,25
0,3
0,35
0,4
0,45
0 1 2 3 4 5
Jestlie nyn do tto rovnice dosadme za
o
P P
p , p , p P , = = =
1 2
10 2
atd., dostaneme

P P P P P
P + + + + + = 1
10 2 2 5 10

eenm tto rovnice je hodnota P , = 0 42 , pomoc n ji urme pravdpodobnosti jednotlivch pot poruch.
Dostaneme

o
P ,
p , = =
0 42
0 04
10 10
B

P ,
p ,
p P ,
P ,
p ,
P ,
p ,
P ,
p ,
= = =
= =
= = =
= =
= =
1
2
3
4
5
0 42
0 21
2 2
0 42
0 42
0 21
2 2
0 42
0 08
5 5
0 42
0 04
10 10
B
B

Stanoven subjektivn pravdpodobnost tvo rozdlen pravdpodobnosti potu poruch. Toto rozdlen meme zapsat
bu ve tvaru tabulky (viz prvn dek tabulky 1), kde kad hodnot potu poruch odpovd urit pravdpodobnost,
nebo graficky, pomoc histogramu, kde na ose x zobrazme jednotliv poty poruch a na ose y jim odpovdajc
pravdpodobnosti (viz obr.1).

Tab.1 Rozdlen pravdpodobnost

Obr.1 Rozdlen pravdpodobnosti potu poruch

Poet poruch

Pravdpodobnosti
0 1 2 3 4 5
Pravdpodobnost 0,04 0,21 0,42 0,21 0,08 0,04
Kumulativn
Pravdpodobnost

0,04

0,25

0,67

0,88

0,96

1

12

1
0,75
0,5
0
0,25
10 000
8 500
8 000
7 000
5 000
Pravdpodobnost Poptvka ( ks/rok )
Poznmka
Rozdlen pravdpodobnost potu poruch meme t vyjdit pomoc tzv. kumulativnch pravdpodobnost (viz tet
dek tab. 1). Tyto kumulativn pravdpodobnosti vyjaduj pravdpodobnost, e poet poruch bude men nebo roven
uritmu potu poruch. Napklad pravdpodobnost jevu, e poet poruch bude men nebo roven dvma (tj. nastanou
bu dv, jedna nebo dn porucha) je 0,67. Tuto pravdpodobnost zskme soutem (kumulac) pravdpodobnost
nastoupen dn, jedn nebo dvou poruch. Plat 0,67 = 0,04 + 0,21 + 0,42. Kumulativn pravdpodobnost nula a pti
poruch je jedna. Je mon tvrdit s jistotou, e bhem danho obdob provozu zazen poet poruch nepev pt.
Kumulativn pravdpodobnosti uveden v tab. 1 ve svm souhrnu definuj distribun funkci nhodn veliiny udvajc
poty poruch.

Metoda kvantil
Tato metoda je vhodn pro stanoven subjektivnho rozdlen pravdpodobnosti v ppad, e poet monch jev
(situac), kter mohou nastat je vysok, ppadn nekonen. Tento charakter m vtina faktor rizika, nap. nkupn a
prodejn ceny uritch produkt a surovin, ve poptvky, devizovch kurs aj.

Podstata metody kvantil vyplyne nzorn z pkladu stanoven rozdlen pravdpodobnosti budouc poptvky po
uritm vrobku nov uvdnm na trh.

Pedpokldejme, e z diskuse analytika s marketingovm odbornkem vyplynulo, e ron ve poptvky se me
pohybovat od pti tisc kus do deseti tisc (pesimistick odhad pt tisc, optimistick odhad deset tisc uruj hranici
intervalu, ve kterm se me poptvka pohybovat),

Dle se me postupovat dvma zpsoby. V prvnm marketingov odbornk uruje velikosti poptvky, kter odpovdaj
podle jeho nzoru uritm pevnm hodnotm pravdpodobnosti, nap. 0,25, 0,5, a 0,75. Pi druhm zpsobu uruje
marketingov odbornk hodnoty pravdpodobnost, kter podle jeho soudu odpovdaj uritm zvolenm hodnotm
poptvky, nap. esti tiscm kus, sedmi tiscm kus, osmi tiscm kus a devti tiscm kus.

Jestli nap. v prvnm ppad vedla diskuse analytika s marketingovm odbornkem k zvru, e pravdpodobnosti 0,25
odpovd velikost poptvky sedm tisc kus, pravdpodobnostem 0,5 a 0,75 velikost poptvky osm tisc kus a osm
tisc pt set kus, pak tyto dvojice sel spolu s dvojicemi 0; 5 000 a 1; 10 000 pedstavuj subjektivn rozdlen
pravdpodobnosti poptvky (viz obr.2).Jednotliv dvojice je teba chpat tak, e nap. pravdpodobnost, e ron
poptvka po danm produktu nepekro vi sedmi tisc kus, je 0,25 (je to tedy pravdpodobnost, e poptvka bude
men nebo nejve rovna sedmi tiscm kus). Pravdpodobnost, e poptvka nepekro osm tisc kus je 0,5 ,
pravdpodobnost nepekroen poptvky velikosti osm tisc pt set je 0,75 a konen pravdpodobnost velikosti 1
odpovdajc deseti tiscm znamen, e marketingov odbornk povauje za zcela jist , e ron poptvka po danm
produktu nepekro hodnotu deseti tisc kus.

Obr.2 Uren poptvky pro dan hodnoty pravdpodobnost

Subjektivn rozdlen pravdpodobnosti poptvky po danm produktu je nyn mon opt zobrazit graficky, kdy na x-
ov ose zobrazme hodnoty poptvky a na ose y-ov odpovdajc pravdpodobnosti. Tm dostaneme graf distribun
funkce.

13

5000 7000 8000 8500 10000
1
0,75
0,5
0,25
0
5000 7000 8000 8500 10000 6000
Na obr.3 uveden distribun funkce poptvky odpovd kumulativnm pravdpodobnostem potu poptvek
v pedchozm pkladu I u poptvky (a jinch faktor rizika s velkm potem jejich monch hodnot) je vak mon
urit nco, co odpovd pravdpodobnostem (nekumulativnm. Jde o tzv. hustotu pravdpodobnosti, jej grafick
zobrazen pro ppad na poptvky uvdme na obr. 4. Na x-ovou osu opt nanme hodnoty poptvky, ale na y-ov
ose nejsou odpovdajc pravdpodobnosti, ale ji zmnn hustota pravdpodobnosti. Graf hustoty pravdpodobnosti
lze interpretovat takto: Pravdpodobnosti uritch hodnot poptvky jsou dny velikost odpovdajcch ploch pod
kivkou hustoty pravdpodobnosti na obr. 4. Cel plocha pod touto kivkou je normovna a je rovna jedn (vyjaduje to
jistotu, e ron poptvka po danm produktu nebude ni ne pt tisc kus a souasn nepeshne hodnotu deseti tisc
kus). Z obr. 2 resp. 3 vidme, e pravdpodobnost toho, e poptvka nepekro sedm tisc kus, je 0,25. Stejn
pravdpodobnost (tj. e se poptvka bude pohybovat mezi pti tisci kusy a sedmi tisci kusy) je vyjdena plochou
obrazce pod kivkou na obr. 4 od potku s hodnotou pt tisc do kolmice v bod sedm tisc.

Obr.3 Distribun funkce poptvky

Stejn tak nap. pravdpodobnost poptvky v intervalu od sedmi tisc kus do osmi tisc pti set kus je dna plochou
vyrafovanho obrazce pod kivkou na obr.4 ohranienho kolmicemi v hodnot poptvky sedm tisc kus a osm tisc
pt set kus. Tuto pravdpodobnost meme t stanovit z odpovdajcch hodnot na obr. 3, resp. 4 jakoto rozdl
pravdpodobnosti, e poptvka bude men nebo rovna osmi tiscm pti stm kus a pravdpodobnosti, e poptvka
nepekro sedm tisc kus, tj. 0,75 0,25 = 0,5. Distribun funkce (viz obr. 3) a hustota pravdpodobnosti (viz obr. 4)
jsou tedy ve vzjemnm jednoznanm vztahu a pi znalosti jedn kivky lze urit druhou kivku a naopak.

14

2.3. Slovn vyjden subjektivnch pravdpodobnost
Pednost selnho vyjden subjektivnch pravdpodobnost je jeho jednoznanost. Pokud chceme tchto
pravdpodobnost vyut pi tvorb a een model podporujcch manarsk rozhodovn, nelze ut jinho ne
selnho vyjden subjektivnch pravdpodobnost.

Uritou nevhodou je, e se manaei mnohdy kvantitativnmu vyjden vyhbaj a radji pracuj se slovnmi popisy
subjektivnch pravdpodobnost, kter jsou veobecn srozumiteln a pijateln. Mezi selnmi hodnotami a slovnmi
popisy subjektivnch pravdpodobnost existuje urit vztah, kter meme vyjdit nap. pomoc tab. 2 (Tepper Kpl
1991).

Slovn vyjden subjektivnch pravdpodobnost m vak t znan nevhody. Nelze jej vyut pro tvorbu
matematickch model podporujcch ppravu manarskho rozhodnut. Krom toho praktick zkuenosti ukazuj, e
jednoznan vztah mezi selnm a slovnm vyjdenm subjektivnch pravdpodobnost, uvedench v tabulce 2, nen
uritou zvaznou normou a e rzn lid chpou uplatnn slovn popisy odlin a pikldaj jim nestejn obsahov
vznam) ble viz Moore, 1983). Nejednoznanost slovnch vyjden subjektivnch pravdpodobnost je proto jejich
znanm nedostatkem, kter me ztit komunikaci pi tmovm een. Vzhledem k tmto skutenostem me slovn
vyjden subjektivnch pravdpodobnost slouit jako urit prvn stupe, po nm nsleduje uplatnn nkter metody
selnho stanoven tchto pravdpodobnost.

Tab. 2 seln a slovn vyjden subjektivnch pravdpodobnost

Vyjden subjektivn pravdpodobnosti

seln Slovn
0 Zcela vyloueno
0,1 Krajn nepravdpodobn
0,2 - 0,3 Dosti nepravdpodobn
0,4 Nepravdpodobn
0.6 Pravdpodobn
0,7 0,8 Dosti pravdpodobn
0,9 Nanejv pravdpodobn
1 Zcela jist

3 Funkce utility
3.1 Postoj rozhodovatele k riziku
Pi rozhodovn za rizika a nejistoty, a to zvlt ve fzi hodnocen variant a vbru varianty uren k realizaci , hraje
vznamnou roli postoj rozhodovatele k riziku. Rozhodovatel (manar, podnikatel) me mt bu averzi
k riziku,ppadn sklon k riziku nebo neutrln postoj k riziku.

Rozhodovatel s averz k riziku se sna vyhnout volb znan rizikovch variant a vyhledv mlo rizikov varianty,
kter se znanou jistotou zaruuj dosaen vsledk, kter jsou pro nj pijateln. Rozhodovatel se sklonem k riziku
naopak vyhledv znan rizikov varianty (kter maj nadji na dosaen zvlt dobrch vsledk, ale jsou spojeny i
vym nebezpem patnch vsledk, resp. ztrt) a preferuj je ped variantami mn rizikovmi. U rozhodovatele
s neutrlnm postojem k riziku jsou averze a sklon k riziku ve vzjemn rovnovze.

Postoj rozhodovatele k riziku pat k jednomu ze zkladnch pojm teorie rozhodovn za rizika a nejistoty. Jeho
definice je zaloena na chovn rozhodovatele v situaci, kdy m monost volby mezi dvma variantami, z nich jedna je
rizikov a druh nerizikov. Pedpokldejme nap. e rizikov varianta vede s pravdpodobnost p
1
k vsledku x
1
a
s pravdpodobnost 1 - p
1
k vsledku x
2 .
Nerizikov varianta nech s jistotou zaruuje dosaen vsledku, kter je roven
oekvn (stedn) hodnot vsledku prvn varianty, tj. zaruuje dosaen vsledku x
1
. p
1 +
x
2
.( 1 - p
1 ) .

Podle definice m rozhodovatel averzi k riziku prv tehdy, dv-li v kad situaci ve uvedenho typu pednost druh
(tj. nerizikov) variant ped prvn (rizikovou) variantou. Jestlie rozhodovatel preferuje vdy prvn, rizikem zatenou
variantu ped druhou nerizikovou variantou, m sklon k riziku. Pro rozhodovatele s neutrlnm postojem k riziku jsou
ob ve uveden varianty indiferentn (tj. hodnot je stejn vysoko).

15

0 Kritrium
1
Utilita
1
2
3
Pedpokldejme, e rozhodovatel nap. vybr ze dvou variant. Prvn vede s pravdpodobnost 0,5 k zisku ve vi 10
milion K a se stejnou pravdpodobnost k nulovmu zisku. Druh varianta zaruuje s jistotou dosaen zisku ve vi 5
milion K (co se prv rovn oekvanmu zisku pi aplikaci prvn varianty nebo 0,5 . 10 mil. K + 0,5 . 0 = 5 mil,
K). V tomto ppad rozhodovatel s averz k riziku vol druhou variantu, kdy s jistotou doshne zisku 5 mil. K (tzn.,
e se sna vyhnout situaci, kter by v ppad volby prvn varianty mohla nastat a vst k nulovmu zisku).
Rozhodovatel se sklonem k riziku vol prvn variantu, u kter oceuje znanou nadji (50% pravdpodobnost)
doshnout zisku ve vi 10 mil. K (to je o pt mil. K vce, ne zaruuje druh varianta). Pro rozhodovatele
s neutrlnm vztahem k riziku jsou ob varianty stejn vhodn.

Postoj rozhodovatele k riziku ovlivuje vce faktor. K nejvznamnjm pat jeho osobn zaloen, minul zkuenosti
(tj. spnost nebo nespnost pedchozch rozhodnut), dle okol, ve kterm volba rizikovch variant probh.

3.2 Konstrukce funkce utility
Funkce utility za rizika (existuj terminologick nejednotnosti. Krom funkce utility za rizika se pouvaj i termny
funkce uitku za rizika, resp. uitkov funkce za rizika), slou jako nstroj pomoc kterho lze kvantitativn vyjdit
postoj rozhodovatele k riziku. Lze dokzat (Keeney Raiffa,1976), e pro rozhodovatele s averz k riziku je funkce
utility konkvn, rozhodovatel se sklonem k riziku m funkci utility konvexn. Funkce utility rozhodovatele
s neutrlnm postojem k riziku je linernho tvaru. (Grafick znzornn funkce utility v zvislosti na postoji
rozhodovatele k riziku pro kriteria vnosovho a nkladovho typu je uvedeno na obr. 5 a 6.

Obr. 5 Rostouc funkce utility kritria vnosovho typu
Vysvtlivky:
1 . . . Konkvn funkce utility rozhodovatele s averz k riziku.
2 . . . Linern funkce utility rozhodovatele s neutrlnm postojem k riziku.
3 . . . Konvexn funkce utility rozhodovatele se sklonem k riziku.

Pro nleit pochopen funkce utility je teba zdraznit, e tato funkce nevyjaduje celkov postoj rozhodovatele
k riziku, tj. vzhledem k celmu souboru kritri hodnocen variant, nbr postoj k riziku z hlediska danho kritria
hodnocen. Vzhledem k tomu se tato funkce oznauje t jako dl, resp. jednorozmrn funkce utility. Tvar tto
funkce pro danho rozhodovatele me bt (a z pravidla je) pro jednotliv kritria zsti odlin (nap. pro nkter
kritria jsou odpovdajc funkce utility konkvn, tzn. e vyjaduj averzi rozhodovatele k riziku, vzhledem k jinm
kritrim m rozhodovatel neutrln postoj k riziku a odpovdajc funkce utility jsou linern)

Dve ne si ukeme postup konstrukce funkce utility pro zvolen kritrium hodnocen, musme se jet seznmit
s jednm zkladnm pojmem, na kterm je tato konstrukce zaloena. Tmto pojmem je tzv. jistotn ekvivalent.
Jistotnm ekvivalentem (pro dan kritrium hodnocen) varianty, kter vede k dsledkm (vzhledem k tomuto kritriu)
velikosti x
1
, x
2
, . . . , x
n
s pravdpodobnostmi p
1
, p
2
, . . . , p
n
rozumme takovou hodnotu dsledku , jeho utilita je rovna
prv stedn (oekvan) utilit varianty vzhledem k tomuto kritriu. (Rozhodovatel se tedy cen dsledek rovn
jistmu ekvivalentu, resp. variantu, kter vede s jistotou k dsledku rovnmu jistotnmu ekvivalentu, resp. variantu,

16

0 Kritrium
1
1
2
3
Utilita
kter vede s jistotou k dsledku rovnmu jistotnmu ekvivalentu stejn vysoko, jako ve uvedenou variantu zatenou
rizikem).

Plat tedy
(1)

kde . . . jistotn ekvivalent ,
. . . utilita jistotnho ekvivalentu ,
. . . utilita dsledku
i
x .

Obr 6 Klesajc funkce utility kritria nkladovho typu

Vysvtlivky:
1 . . . Konkvn funkce utility rozhodovatele s averz k riziku.
2 . . . Linern funkce utility rozhodovatele s neutrlnm postojem k riziku.
3 . . . Konvexn funkce utility rozhodovatele se sklonem k riziku.

Jestlie se vrtme k pedchozmu pkladu dvou variant (prvn byla rizikov a vedla s pravdpodobnost 0,5 k zisku 10
mil. K a se stejnou pravdpodobnost k nulovmu zisku a druh zaruovala s jistotou zisk ve vi 5 mil. K.) pomoc
kterho jsme demonstrovali postoj rozhodovatele k riziku pak pro prvn rizikovou variantu plat x
1
= 10 mil., x
2
= 0 a p
1

= p
2
= 0,5. Pokud nyn rozhodovatel cen stejn vysoko tuto rizikovou variantu jako variantu, kter s jistotou zaruuje
zisk nap. ve vi 3 mil K (tj.utilita jistho zisku ve vi 3 mil. K je rovna oekvan utilit rizikov varianty, neboli
podle vztahu (1) plat u (3) = 0,5.u (10) + o,5 . u (0), je jistotn ekvivalent tto rizikov varianty roven prv 3 mil. K.

Jistotn ekvivalent meme tak interpretovat ponkud jinak. Budeme-li uvaovanou rizikovou variantu povaovat za
loterii s vhrami 10 a 0 mil.K (dosahovanmi se stejnou pravdpodobnost 0,5), pak je jistotn ekvivalent roven
minimln stce, za kterou je subjekt ochoten tuto loterii prodat. V naem ppad by tedy tato prodejn cena inila 3
mil. K.

Na zklad vztahu jistotnho ekvivalentu dan rizikov varianty a jejho oekvanho dsledku je rovn mon
vymezit postoj rozhodovatele k riziku. Jestlie v pedchozm pklad byl pro danho rozhodovatele jistotn ekvivalent
rizikov varianty (jej oekvan zisk je 0,5 . 10 + 0,5 . 0 = 5 mil. K) men ne tento oekvan zisk (plat 3 <5), m
rozhodovatel vzhledem ke kritriu tvoenmu ziskem averzi k riziku. Je-li jistotn ekvivalent dan rizikov varianty
vy ne jej oekvan zisk , tj. vy ne 5 mil. K, m rozhodovatel vzhledem k zisku klon k riziku. Jistotn
ekvivalent dan rizikov varianty rozhodovatele s neutrlnm postojem k riziku je v naem ppad roven prv 5 mil.
K. Rozdl mezi oekvanm dsledkem rizikov varianty a jejm jistotnm ekvivalentem se nkdy oznauje jako
rizikov prmie (podrobnji viz Fotr Pek, 1986)

17

3.3 Pklad uren funkce utility
Konstrukci funkce utility si ukeme na pklad jejho stanoven pro kritrium hodnocen, tvoenho ziskem.
Pedpokldejme, e pi een uritho rozhodovacho problmu investin povahy bylo formulovno nkolik variant,
kter jsou v odlin me rizikov a jejich mon zisk se pohybuje od 0,25 mil. K do 9,8 mil. K. Nam kolem je
pomoci stanovit rozhodovateli, kter je odpovdn za volbu varianty uren k realizaci, funkci utility zisku.

Prvnm krokem pi stanoven funkce utility pro dan kritrium je vymezit jej definin obor. Jako krajn body
defininho oboru meme zvolit bu nejmen a nejvt hodnotu kritria v danm souboru variant (v naem ppad
0,25 mil.K a 9,8 mil. K) nebo veliiny, kter vznikly jejich vhodnm zaokrouhlenm. My se rozhodneme pro
zaokrouhlen a funkci utility zisku budeme stanovovat pro interval, jeho doln mez je 0 mil. K a horn mez 10 mil. K.

Pro stanoven defininho oboru funkce utility meme stanovit hodnoty utility v krajnch bodech tohoto oboru.
Vyuijeme zde toho, e funkce utility za rizika (stejn jako funkce utility za jistoty) nevyjaduje absolutn ocenn
(vhodnost pro rozhodovatele) monch hodnot kritria, ale ocenn relativn, a proto je volba hodnot utility v krajnch
bodech defininho oboru arbitrrn. (arbitrrnost volby hodnot funkce utility v jejch krajnch bodech vyplv z toho,
e funkce utility je jednoznan a na pozitivn linern transformaci, tj. pvodn stanoven funkce utility u (x) me
bt nahrazena libovolnou jinou funkc v (x), pro kterou plat vztah v (x) = a.u (x) + b, piem a > 0). Zvykem vak je
volit pro kritria vnosovho typu utilitu doln meze defininho oboru rovnou nule a utilitu horn meze tohoto oboru
rovnou jedn (nkdy stu), piem pro kritria nkladovho typu je to prv naopak. V naem ppad zvolme tedy
utilitu nulovho zisku jako nula a utilitu zisku ve vi 10 mil. K jako jedna (plat tedy u (0) = 0, u (10) = 1) , take
znme ji dva body hledan utility zisku.

Dle stanovme nkolik dalch bod hledan funkce utility a to pomoc jistotnch ekvivalent. Urovn jistotnch
ekvivalent probh v dialogu analytika (konzultanta)s rozhodovatele, kdy analytik klade postupn rozhodovateli
dotazy, piem prvn dotazy me rozhodovatel zodpovdt snadno, avak nronost dalch dotaz (tak jak se dialog
bl k uren jistotnho ekvivalentu ) se zvyuje.

Dialog analytika s rozhodovatele by mohl mt v naem ppad asi nsledujc prbh. Analytik polo rozhodovateli
nejdve dotaz. Preferujete vce variantu, kter Vm s jistotou zaru zisk ve vi 1 mil. K, nebo rizikovou variantu,
kter s pravdpodobnost 0,5 vede k zisku 10 mil. K a se stejnou pravdpodobnost k nulovmu zisku? Rozhodovatel
(pokud jeho averze k riziku nen zvl vrazn) asi odpov, e preferuje vce vce danou rizikovou variantu. V tomto
ppad dal dotaz analytika na rozhodovatele zn nap. Preferujete vce variantu s jistm ziskem ve vi 8 mil. K
nebo danou rizikovou variantu? Rozhodovatel, (pokud nem nem vysoce vrazn sklon k riziku ) patrn odpov, e si
vce cen nerizikov varianty s jistm ziskem velikosti 8 mil. K.

Charakter dalch dotaz analytika na rozhodovatele je obdobn. Analytik v druhm kroku zjiuje, zda rozhodovatel
preferuje vce variantu s jistm ziskem 1,5 mil. K (resp. dle 6 mil. K.) ne danou rizikovou variantu. Pokud n
rozhodovatel pedstavuje pevldajc typ rozhodovatele s uritou averz k riziku, bude patrn jet preferovat danou
rizikovou variantu ped variantou s jistm ziskem 1,5 mil. K. A zcela jist si bude vce cenit varianty s jistm ziskem 6
mil. K ne dan rizikov varianty.

Dalmi obdobnmi dotazy analytika na rozhodovatele dospjeme v dalch krocch nap. k situaci, kdy si rozhodovatel
o nco mlo vce cen dan rizikov varianty ne varianty s jistm ziskem 2,5 mil. K a stejn tak povauje za o nco
mlo lep variantu s jistm ziskem 3,5 mil. K ne danou rizikovou variantu.Jestlie analytik dle zjist, e
rozhodovatel cen stejn vysoko variantu s jistm ziskem 3 mil. K jako danou rizikovou variantu, urili jsme jistotn
ekvivalent rizikov varianty, kter s pravdpodobnost 0,5 poskytuje zisk 10 mil. K a se stejnou pravdpodobnost
nulov zisk. Tento jistotn ekvivalent in 3 mil. K.

Tento nepm zpsob stanoven jistotnho ekvivalentu dan rizikov varianty v dialoga analytika s rozhodovatelem je
pro rozhodovatele mn nron ne pm zpsob, tj. poloen dotazu: Jakou vi jistho zisku si cente stejn
vysoko jako jako rizikovou variantu vedouc se stejnou pravdpodobnost 0,5 k zisku 10 mil. K a k nulovmu zisku?
Krom vy nronosti vede tento pm zpsob i k mn spolehlivm vsledkm. (Podrobnji k problematice
k urovn jistotnch ekvivalent viz Fotr-Pek, 1986.)

Stanoven jistotn ekvivalent nm umouje urit tet bod (krom dvou ve charakterizovanch krajnch bod) funkce
utility zisku. Vzhledem ke vztahu (1) mus toti platit

18

u (3) = 0,5 . u (0) + 0,5 . u (10) (2)

a po dosazen za u (0) = 0 a za u (10) =1 do vztahu (2) dostaneme

u (3) = 0,5 .0 + 0,5 . 10 = 0,5 (3)

Tm jsme urili utilitu jistotnho ekvivalentu dan rizikov varianty, kter je 0,5 a mme tedy dal (ji tet) bod
hledan funkce utility zisku, kter m souadnice 3 a 0,5.

Dal dva body funkce utility urme stejnm zpsobem, a to pomoc jistotnch ekvivalent dalch rizikovch variant,
vytvoench s vyuitm znalosti jistotnho ekvivalentu velikosti 3 mil. K. Stanovme (opt v dialogu analytika
s rozhodovatelem) jistotn ekvivalenty dvou rizikovch variant, z nich prvn vede s pravdpodobnost 0,5 k nulovmu
zisku a se stejnou pravdpodobnost k zisku 3 mil. K a druh s pravdpodobnost 0,5 k zisku ve vi 3 mil. K a 10
mil. K.

in-li jistotn ekvivalent prvn rizikov varianty 1,3 mil. K a jistotn ekvivalent druh rizikov varianty 5,5 mil. K,
mus opt vzhledem k (1) platit:

u (1,3) = 0,5 . u (0) + 0,5 . u (3 (4)

u (5,5) = 0,5 . u (3) + 0,5 . u (10) (5)

Protoe vak ji vme, e u (3) = 0,5 , dostaneme po dosazen za u (0) = 0, u (3) = 0,5 a u (10) = 1 do vztah (4) a
(5)

u (1,3) = 0,5 .0 + 0,5 . 0,5 = 0,25 (6)

u (5,5) = 0,5 . 0,5 + 0,5 . 1 =0,75 (7)

Nyn ji znme souadnice dalch dvou bod funkce utility zisku, kter jsou (1,3; 0,25) a (5,5; 0,75), a celkem tedy je
znmch ji pt bod tto funkce. Zskan body funkce utility zobrazme v grafu (na ose x vyneseme velikosti zisku a
na ose y jim odpovdajc utility viz. obr. 7). Prolome-li tyto body vhodnou kivkou, zskme aproximaci grafickho
zobrazen funkce utility zisku pro danho rozhodovatele (a dan rozhodovac problm). Z obr. 7 vidme, e funkce
utility zisku danho rozhodovatele je konkvn. Tento rozhodovatel m tedy pro dan kritrium averzi k riziku.

Pro praktick uplatnn funkce utility pi hodnocen a vbru rizikovch variant me bt vhodn stanovit funkn tvar
funkce utility a urit jeho parametry.
V ppad neutrlnho postoje rozhodovatele k riziku, kdy funkce utility je linern, m jej vzorec tvar

( )
*
x x
u x ,
x x
-
=
-
0
0

kde u
0
je nejhor a
*
x nejlep hodnota danho kritria (tj. krajn body defininho oboru). Jej kvantifikace
nevyaduje odhad parametr.

19

1,3 5,5 10
1
0,75
0,5
0,25
0
Utilita
Zisk 3
1
0
Utilita
Kritrium
Averze k riziku
Inflexn bod
Sklon k riziku
Oblast zisku Oblast ztrty

Obr. 7 Funkce utility zisku

V ppad averze, resp. sklonu rozhodovatele k riziku, lze odpovdajc konkvn, resp. konvexn funkci utility asto
zobrazit exponenciln funkc tvaru

( )
a . x b
u x e
+
= (9)
K uren parametr a a b tto funkce uijeme vech funknch hodnot funkce utility ve stanovench bodech , tvoench
krajnmi body jejho defininho oboru a jistotnmi ekvivalenty. Uren koeficient se provd aproximan metodou
nejmench tverc pes logaritmovn dan rovnice (9).

Empirick vzkumy chovn rozhodovatel za rizika a jejich funkce utility ukazuj, e znan pevld averze k riziku.
Souasn se vak ukazuje, e postoj subjektu k riziku je asto odlin v zvislosti na tom, zda jde o zisky nebo ztrty.
Zatmco v oblasti zisku je pevldajc averze k riziku, pak v oblasti ztrt pevld sklon k riziku (to vak plat sp o
mench ztrtch, nebo v ppad znanch a katastrofickch ztrt, vedoucch k ruinovn subjektu opt vrazn
pevld averze k riziku).Postoj rozhodovatele s averz k riziku v oblasti zisku a sklonem k riziku v oblasti ztrt lze pak
zobrazit funkc utility s inflexnm bodem, kter oddluje jej konkvn st pro kladn hodnoty ziskovho kritria od
sti konvexn pro zporn hodnoty tohoto kritria(viz obr. 8).

Obr. 8 Funkce utility s inflexnm bodem

K funkci utility a k jej konstrukci je teba jet poznamenat, e tato funkce vyjaduje vdy subjektivn postoj
rozhodovatele k riziku vzhledem k danmu kritriu. dn objektivn funkce utility (pro dan kritrium) neexistuje.
Funkce utility rznch rozhodovatel se mohou liit (a tak se obvykle li, jak ukazuj vsledky empirickch studi).

20

Anii funkce utility tho rozhodovatele pro dan kritrium, zjiovan v rznch obdobch, nemus bt stejn.
Empirick vzkumy t ukazuj, e konstrukce funkce utility je obecn obtn zleitost (z hlediska informac, kter je
teba od subjektu pro jej kvantifikaci zskat). Rozhodovatel s n proto v mnoha ppadech nejsou ochotni pracovat.

Seznam literatury
[1] BATA, A. Plnov rozhodovac procesy a jejich systm. Praha : Academia, 1977.
[2] ERN, J.; GLCKAUFOV, D. Vcekriteriln vyhodnocovn v praxi. Praha : SNTL, 1982.
[3] EDEN, C.; JONES, S.; SIMS, D. Messing About in Problme. Oxford : Pergamon Press, 1983.
[4] FOTR, J. Pprava a hodnocen podnikatelskch projekt. Praha : VE, 1993.
[5] FOTR, J. Manarsk rozhodovac analza. Praha : VE, 1992.
[6] FOTR, J.; DDINA, J. Manarsk rozhodovn. Praha : VE.
[7] FOTR, J.; PEK, M. Exaktn metody ekonomickho rozhodovn. Praha : Academia, 1986.
[8] IVANCEVICH, J. M.; DONESLY, J. H.; GOBBON, J. L. Management. Principles and Functions. Homewood
: R. D. Irvin, 1989.
[9] MOORE, P. G. The Business of Risk. Cambridge : University Press, 1983.
[10] NOVK, M. Examples of using concepts of probability theory in managementdecision makinng. Mezinrodn
konference. Kunovice : EPI, s.r.o., 2006.
[11] NOVK, M. Probability theorz in combined form of study at FEEC BUT. Mezinrodn konference. Kunovice :
EPI, s.r.o., 2006.
[12] PEK, M.; VOBOIL, J. Vybran metody dlouhodobho prognzovn a jejich vyuit. Praha : Ekonomick
stav SAV, 1981.
[13] STCHLE, W. H. Management. Mnchen : Verlag Franz Valen, 1989.
[14] VLEK, R. Hodnotov management. Praha : Management Press, 1992.
[15] VLEK, R. Pruka hodnotov analzy. Praha : SNTL, 1983.
[16] WATSON, S. R.; BUDGE, J. R. Decision Synthesi. Cambridge : Cambridge University Press, 1987.
[17] ZAPLETAL, J. Operan analza. Kunovice : Skriptorium VO, 1995.
[18] ZRUBA, P. aj. Zklady podnikovho managementu. Praha : Aleko, 1991.

Adresa:
Doc. RNDr. Josef Zapletal, CSc.
Osvobozen 699
686 04 Kunovice
esk republika
Tel./fax.: +420 572 549 018/ +420 572 548 788
e-mail: zapletal@vos.cz

21

THE USE OF FUZZY LOGIC FOR SUPPORT OF DIRECT MAILING
Petr Dostl

Brno University of Technology

Abstrakt: The article deals with the use of fuzzy logic for the support of direct mailing. The brief
description of fuzzy logic, the process of calculation, the scheme of models, rule blocks, attributes and their
membership functions are mentioned. The use of fuzzy logic is an advantage especially for support of
direct mailing where evaluation is very complicated.

Klov slova: fuzzy logic, direct mailing, model, rule block, membership function, attributes

1. INTRODUCTION
The use of fuzzy logic is an advantage especially for the support of direct mailing where evaluation is very complicated.
The advantage is that linguistic variables are used. Fuzzy logic measures the certainty or uncertainty of membership of
an element of the set. Analogously man makes decisions during mental and physical activities. The solution of a certain
case is found in the principle of rules that were defined by fuzzy logics for similar cases. Fuzzy logics belong among
methods that are used in the area of direct mailing.

2. The fuzzy processing
The calculation of fuzzy logics consists of three steps: fuzzification, fuzzy inference and defuzzification.

The fuzzification means that the real variables are transferred to linguistic variables. The definition of linguistic
variable goes out from basic linguistic variables, for example, at the variable risk there are set up the following
attributes: none, very low, low, medium, high, very high. Usually there are used from three to seven attributes of
variable. The attributes are defined by the so called membership function, such as , , Z, S and some others. The
membership function is set up for input and output variables.

The fuzzy inference defines the behavior of the system by means of rules of type <When>, <Then> on the
linguistic level. The conditional clauses evaluate the state of input variables by the rules. The conditional clauses
are in the form
<When> Input
a
<And> Input
b
.. Input
x
<Or> Input
y
.. <Then> Output
1,

it means, when (the state occurs) Input
a
and Input
b
, .., Input
x
or Input
y
, , then (the situation is) Output
1
.

The fuzzy logic represents the expert systems. Each combination of attributes of variables, incoming into the system
and occurring in condition <When>, <Then>, presents one rule. Every condition behind <When> has a corresponding
result behind <Then>. It is necessary to determine every rule and its degree of supports (the weight of rule in the
system). The rules are created by the expert himself.

The defuzzification transfers the results of fuzzy inference on to the output variables, that describes the results
verbally (for example, whether the risk exists or not).

The system with fuzzy logics can work as an automatic system with entering of input data. The input data can be
represented by many variables.

3. DIRECT MAILING
This case presents the use of fuzzy logic for direct mailing, whether the client is visited personally, sent a letter or not to
speak to him. See the model on fig. 1.

22

Fig. 1 Project chart

The input variables and their attributes are Loan (fig. 2) (none, small, medium, high), Salary (fig. 3) (low, medium,
high), Age (fig. 4) (young, medium, old, very old), Children (no, a few, many), State (single, married, divorced) and
Place (big city, city, village).

Fig. 2. The attributes and membership functions of variable Loan

Fig. 3. The attributes and membership functions of variable Salary

23

Fig. 4. The attributes and membership functions of variable Age

The rule blocks with attributes are Finance (excellent, good, bad), Personality (unsuitable, suitable, good, excellent).
The fig. 5 shows the attributes and membership functions of the Finance.

Fig. 5. The attributes and membership functions of variable Finance

The output variable Mailing with the attributes evaluates whether the client will be visited or a letter will be sent to him
or he will not be spoken to him. See fig. 6.

Fig. 6 The attributes and membership functions of variable Mailing

24

Fig. 7 shows one from two rule blocks Finance with their rules and degree of support that set up the relation between
input and output variables.

Fig. 7 Rule block

When the model is made, it is necessary to tune it (to set up the inputs on known values, evaluate the results and to
change the rules or weights, if necessary). If the system is tuned, it is possible to use it in practice.

Fig.8 The attributes and membership functions of output variable Mailing

The set up of the rule block distinguish single cases. For example, the result of decision making is the no contact with
client in case when the person has a low salary and a lot of loans, he lives in a village, he is of old age, he is single and
without children. Fig. 8 shows this result, where the mailing is evaluated not to contact the client.

The effort is to bring the profit in the future from the investment into the marketing. The evaluation, whether the
marketing project is profitable or loss-making, is possible to evaluate after a certain time.

4. CONCLUSION
The mentioned case is only the fraction of possible variants of the use of fuzzy logic in various areas of decision
making. The theory of fuzzy logic contributes to the quality of decision making. The decision making process is an
important activity of firms. It is possible to say, that the successful decision making make the firm successful.

It is necessary to emphasize, that these methods support the decision making and that the responsibility of optimal
variant or variants are on those, who make the decision.

Fuzzy logic as well artificial neural networks and genetic algorithms belongs to relative strong methods as a tool of
artificial intelligence for the support of decision making.

25

LITERATURE
[1] ALIEV, A.; ALIEV, R. Soft Computing and Its Applications, World Scientific Pub. Ltd, UK2002, 444 p.,
ISBN 981-02-4700-1.
[2] ALTROCK, C. Fuzzy Logic & Neurofuzzy Applications in Business & Finance. USA : Prentice Hall, 1996,
375p., ISBN 0-13-591512-0.
[3] DOSTL, P. Modern metody ekonomickch analz Finann kybernetika. Zln : UTB, 2002, 110p., ISBN
80-7318-075-8.
[4] DOSTL, P. Soft Computing and Stock Market. Brno : VUT, 2003, p.258-262, ISBN 80-214-2411-7.
[5] DOSTL P.; K, L. Fuzzy Logic and Financial Time Series. Kunovice : EPI, s.r.o., 2004, International
Conference on Soft Computing, s.93-97., ISBN 80-7314-025-X.
[6] DOSTL, P.; RAIS, K. Operan a systmov analza II. Brno : VUT FP, 2005, Skripta, 160s., ISBN 80-
214-2803-1.
[7] DOSTL, P.; MACH, E. The Use of Fuzzy Logic in Pedagogy of Gifted Students. Brno : 2005, Business and
Economic Development in Central and Eastern Europe, Konference, s.18, 5s., ISBN-214-3012-5.
[8] DOSTL, P. Vybran metody rozhodovn v podnikov sfe. Brno : VUT-FP, 2005, Habilitan prce, 22 s.
188 s. ISBN 80-214-3083-4, ISSN 1213-418X.
[9] DOSTL, P. Vyuit fuzzy logiky v risk managementu. In Progressive Methods and Tools of Management
and Economics of Companies. Brno : 2005, 5s., ISBN 80-214-3099-0.
[10] DOSTL, P.; RAIS, K.; SOJKA, Z. Pokroil metody manaerskho rozhodovn. Grada, 2005,168s, ISBN
80-247-1338-1.
[11] FANTA, J. Psychologie, algoritmy a uml inteligence na kapitlovch trzch. Praha : Grada, 2001, 168p.,
ISBN 80-247-0024-7.
[12] KAZABOV, K.; KOZMA, R. Neuro-Fuzzy Techniques for Intelligent Information Systems Physica-Verlag.
Germany, 1998, 427p., ISBN 3-7908-1187-4.
[13] KLIR, G.J.; YUAN, B. Fuzzy Sets and Fuzzy Logic, Theory and Applications. New Jersy : Prentice Hall, USA,
1995, 279p., ISBN 0-13-101171-5.
[14] KOLEKTIV FuzzyTech Users Manual, Inform. GmbH, Germany, 2002, 258p.
[15] NOVK, V. Fuzzy mnoiny a jejich aplikace. Praha : SNTL, 1990, 297p., ISBN 80-03-00325-3.
[16] RAIS, K.; SMEJKAL V. zen rizik. Praha : Grada, 2003, 270p., ISBN 80-247-0198-7.
[17] RIBEIRO, R.; YAGER, R. Soft Computing in Financial Engineering. A Springer Verlag Copany, 1999, 590p.,
ISBN 3-7908-1173-4.

Address:
Ing. Petr Dostl, CSc.
Kolejn 4
612 00 Brno, Czech Republic
Tel. +420 541 143 714, Fax. +420 541 142 692
e-mail: dostal@fbm.vutbr.cz, dostal@iqnet.cz
http://www.iqnet.cz/dostal

26

27

THE COLLATION OF VARIOUS METHODS FOR THE SOLUTION OF TRANSPORTATION
PROBLEMS
Abdurrzzag Tamtam


Abstrakt: The paper solves the problem of effectiveness of various methods with respect to their sizes. We
study and compare the North West Corn method,the index method, the Vogel method. It is obvious that the
first allocation required values of goods by customers will be most expansive when using North West Corn
method which is based on geographical principle without any economical conditions. We show that the
index method gives a better result than North West Corn method but the Vogel method brings a better
solution with respect to sizes. We modified Vogel method and we show by example that for greater
matrixes among suppliers and customers, the modified method the gives optimum solution.

Klov slova: Supplier, customer, cell, free cell water, occupied cell stone, North West Corn method,
index method, Vogel method, modified Vogel method.

1) Introduction
A special class of linear programming problems are the so called distributive problems. Such a distributive problem can
be formulated in full generality as follows:

Minimise the objective function
z
m
i 1
n
j 1
c
i j
x
i j

under the conditions

n
j 1
x
i j
a
i
m
i 1
k
i j
x
i j
b
j

and non-negative of x
i j

where
i = 1, 2, . . . , m , j = 1, 2, . . . , n

These problems can be solved also with the Simplex method. But such a manner of solving is usually very lengthy and
laborious. Special properties of distributive problems make it possible to use special methods such as transportation
problem.

There we suppose m suppliers S
i
( i = 1, 2, . . . , m) with inventories s
i
of units of identical commodity and n
customers K
j
( j = 1, 2, . . . , n) with requirements of k
j
units of the same commodity.
We suppose that:
m
i 1
s
i

n
j 1
k
j

and we say that a balanced transport problem is given. Further the expanses for transporting a unit from the supplier S
i
to the customer K
j
are known. We denote them c
ij
.. The number of transported units of commodity from i-th supplier to
the j-th customer will be denoted by x
ij
. The problem is to find m.n dimensional vector

28

[ x
11
, x
12
, . . . , x
1n
, x
21
, x
22
, . . . , x
2n
, . . . , x
m1
, x
m2
, . . . , x
mn
] Satisfying the following restrictions:
x
11
+ x
12
+ x
13
+ x
1n

= s
1

x
21
+ x
22
+ x
23
+ x
2n

= s
2
x
m1
+ x
m2
+ x
m3
+ x
mn

= s
m
x
11
+ x
21
+ x
31
+ . . . . . . + x
m1

= k
1

x
12
+ x
22
+ x
32
+ .. . . . . . . . + x
m2

= k
2
x
1n
+ x
2n
+ x
3n
. . . . . . . . . ... x
mn

= k
n

This minimises the objective function
z = c
11
x
11
+ c
12
x
12
+ ... + c
1n
x
1n
+ c
21
x
21
+ c
22
x
22
+ ... + c
2n
x
2n
+ ... + c
m1
x
m1
+ c
m2
x
m2
+ ... + c
mn
x
mn

Coefficients standing at variables are equal to one or zero. The number of zero coefficients is n . m . (m + n - 2), the
number of unity coefficients is only (2. m . n). The system of equations is dependent. We see that after summing the
first m equations we obtain the same result as after summing the last n equations. Therefore the solution has at most (m
+ n 1) non-zero components.

If the number of non-zero components is just (m + n - 1), then the solution is called non-degenerated, in case of smaller
number of solutions we say that the solution is degenerated. That is the reason for which we do not solve the transport
problem by Simplex method and we use simpler methods.

2 Solving of Transportation Problem
Hereafter we study the optimising process on a concrete example. The comparison of the methods for the solving of
transportation problem will be done at one example which contains three suppliers and four customers. The following
example with a small matrix of 3x4 cells shows the advantages of index method, Vogel method and modified Vogel
method against the North West Corn method, and also shows that the Vogel method exceeds the index method but this
small example does not show any difference between Vogel method and our modified method.

Example: Let us suppose that we have three suppliers S
1
, S
2
and S
3
with inventories 310, 200 and 190 units of
commodity and four customers K
1
, K
2
, K
3
, K
4
with demands for 250, 100, 150 and 200 of units. The transport costs c
ij

from the i-th supplier to the j-th customer are given by table 1.
Table 1
Customers
Suppliers
K
1
K
2
K
3
K
4
Inventories
S
1
20 14 11 12

310
S
2

6 15 18 15

200
S
3

17 12 19 23

190

Demands

250 100 150 200 700

North West Corner Method
The name arose from the geographical point of view. NW method is a geographical method of occupation of fields
(cells) in the table, which has nothing common with economical point of view. We are beginning from the cell P
11

determined by the row D
1
and the column S
1
(From North and West on a map) the value c
11

has no sense for the
construction. This cell will be occupied by maximum possible part of required commodity from S
1
, it is 250. The

29

supplier S
1
still has 310 - 250 = 60 units. We continue with occupations of cells in the east course till the inventories of
the first supplier are spent. So the cell P
12
determined by double S
1
, K
2

obtains 60 units. Inventories of the first supplier
are equal to zero, but the second customer is not satisfied. We pass on to the second supplier S
2
who still has 40
remaining units, hence the cell P
21
will be occupied by 40 units. At the supplier S
2
there are still 160 units of
commodity. It is possible to satisfy the whole requirement of the customer K
3
, and we put it into the cell P
23
. The last 10
units from S
2
will be given to the forth customer in the cell P
24
. The customer K
4
claims 200 units of commodity and he
has only 10 units. The rest 190 units will be transported from the third supplier. Six cells are occupied from the total
number of 12 cells. In this example (m+n-1) is just equal to 6, and hence the problem is not degenerated.

The value of the objective function is z = 250.20+60.14+14+40.15+150.18+10.15+190.13=13660 financial units
We obtained the first solution. It is not the best solution of transportation so far.

This fact can be proved by calculus using table 2 containing the first solution obtained by NW method. The occupied
cells are called stones and the empty cells are called waters.

Table 2
Customers
Suppliers
K
1
K
2
K
3
K
4
Inventories

S
1

20
250
14
60

11 12

310

S
2

6 15
40
18
150
15
10

200

S
3

17 12 19 23
190

190
Demands 250 100 150 200 700

Index Method
It is obvious that the economical point of view has its important position. This method works with costs c
ij
and begins
with occupation of the cell with the smallest costs and continues over greater and greater costs of cells to the maximum
cost. Simultaneously the sum of values in stones in every row is identical with the initial inventory of the relevant
supplier. Similarly the sums of values in stones are the same as the requirements of relevant customers. We explain this
method in our example.

We begin with the cell P
21
, which has the smallest cost c
21
= 6. We occupy it with the maximum requirement which can
be delivered from the supplier S
2
, it is 200 units of commodity. This does not satisfy the demand of K
2
. For the next
smallest cost is c
13
, we occupy the cell P
13
with the maximum amount again, this one is given by the demand of S
3
. The
next smallest cost equal to 12 occurs in two cells namely in P
14
and in P
32
. We see that they are in different rows and
also in different columns. Therefore it is not necessary to choose the order of occupation of these cells. We put into the
first cell 160 units, there is not more at S
1
. The amount needed by for the customer K
2
is transported from K
3
into the
cell P
32
. The cells with even higher costs c
12
=14, c
22
=c
24
=15, are not occupied because the inventories of the second
supplier S
2
are exhausted. We come to the cost c
31
= 17. The customer K
1
asks for the total amount 250 units and he has
200 units distributed in the stone P
21
thus he receives in the cell P
31
the amount of 50 units of commodity. For analogous
reasons (the inventories of the supplier are exhausted or the demands of the customer are satisfied) we omit the costs 18,
19 and 20. We finish the allocation in the cell P
34
with 40 units. The general visualisation can be found in the table 3.
The value of the objective function for the solution obtained by the index method rewritten in the table 3 is equal to:

z = 200.6 + 150.11 + (100 + 160).12 + 50.17 + 40.23 = 7740

30

Table 3
Customers
Suppliers
K
1
K
2
K
3
K
4
Inventories

S
1
20 14 11
150

12
160

310

S
2
6
200

15 18 15

200

S
3
17
50

12
100
19 23
40

190
Demands 250 100 150 200 700

We see that the index method brings a better result than the North West corner method. The index method
demonstrative applies the economical specifications and it requires speculation during all of the allocation process. On
the other hand the North West corner method can be called mechanical we use a given algorithm without extra
cogitation. However nor even the solution obtained by index method need be optimal. Especially for larger problems we
come to results which have very far from the optimal solution. Better results can be received using a method of
approach called the Vogel method (VAM method). We illustrate this one on the same example and such we will be able
to compare all these methods. Using the Vogel method we even receive the optimal solution for our example.

Vogel Approach Method (VAM Method)
In the course of the index method we occupy a cell with the smallest cost the earliest possible which is lying in concrete
row and concrete column. It may be that after the exhaustion of inventories in this row or by refilling the demands of a
customer who is in that concrete column. We must occupy the cell with a very high cost. For this reason costs for all the
solution are growing abnormal. VAM method uses a process which excludes these situations, unfortunately not
absolutely.

Table 4

Customers
Suppliers
K
1
K
2
K
3
K
4
Inventories Difference

S
1

20 14 11 12
310

1

S
2
6 15 18 15
200

9

S
3
17 12 19 23
190

5

Demands 250 100 150 200 700
Differences 11 2 7 3

The basic pattern of VAM method is to prevent from such a system of occupation of cells. It is to prevent from a
situation that to the end of dispatching of inventories to customers the differences among costs are growing
inappropriately. This unwelcome situation approves oneself such that in corresponding rows or columns are great
differences between the minimal cost and the nearest higher one.

31

Table 5
Customers
Suppliers
K
1
K
2
K
3
K
4
Inventories Difference

S
1

20 14 11 12
310

1,1

S
2
6
200
15 18 15
200

9,x

S
3
17 12 19 23
190

5,5

Demands 250 100 150 200 700
Differences 11,3 2,2 7,8 3,11

We keep away from such situations namely as that at the beginning of allocation of inventories to customers we prefer
those rows and columns where there is the maximum difference between the smallest value of cost and the nearest
greater one. After finding such an array (row or column) we implement the maximum possible inventory into the cell
with minimum cost. We apply this method on our example again. We calculate differences in all the rows and columns.
The results are in the last row and last column of table 4. The greatest difference is in the first column. We find the cell
with minimum cost in that column. It is the cell P21. We occupy it by the maximal possible amount i.e. by 200. The
inventories of the supplier S2 are exhausted. This fact will be denoted by a lying cross at differences in the second row
where characteristics of the second supplier are described. We do an analogous conclusion when the demands of the
customer are satisfied. We write the lying cross into a cell for differences which is lying in the column of the
corresponding customer. After every operation by which either the supplier is empty or the customer is satisfied we
omit this row or column. Hence it is necessary to re-count all the remaining differences. The situation as it is after the
first assignment into the cell P21 is given in the table 5. (Simultaneously the difference is re-counted. The highest
difference is in the fourth column now. The bottom cost is in the cell P14 . This cell will be occupied by the demand of
customer K4 i.e. by 200 units. The amount of the fourth customer is satisfied. We put it as that we write the lying cross
at differences in the fourth column. After the next recounting of differences the maximal difference (8) is in the third
the capacities of the second supplier S2 are run out, therefore the costs c
2
j , j = 1,2,3,4 are not used for the
computational procedure of new differences. The lowermost cost is in the cell P13 .We occupy this cell by the
remaining maximum amount from the first supplier S1, i.e. 110 units. Now both the suppliers S1 and S2 are exhausted
and simultaneously the costs from the first two rows are not considered for the next calculation of the differences. As
the column differences, we overwrite the according lying costs. We complete desired amounts of commodity of
customers in a row K2 which receives 100 units of commodity, K1 which receives 50 units of commodity, and K3
which receives 100 units of the same commodity.

The whole process of dispatching of commodity and the sequences of differences for adequate rows and columns are
given in table 6.
The value of the objective function obtained by the Vogel method described in the table 6 is equal to
z = 200.6 + 110.11 + (200 + 100).12 + 50.17 + 40.19 = 7620 .

Simultaneously it is the minimum value. Every other solution produces a larger value of the objective function by most
and hence larger costs for transportation. In the end we can say that the Vogel method brings usually the best results
from all those last given methods.

32

Table 6
Customers
Suppliers
K
1
K
2
K
3
K
4
Inventories Differences

S
1

20 14 11
110
12
200

310

1,1,3,x

S
2

6
200

15 18 15
200

9,x

S
3

17
50
12
100
19
40

23
190

5,5,5,5
Demands 250 100 150 200 700
Differences
11,3,3,
17,17,x
2,2,2,1
2,x
7,8,8,19,
19,19,x
3,11,x

Very often the Vogel method gives directly the optimal solution. It is usually for not extensive problems, for example in
our case. For more complicated problems the solution which is received by Vogel method is near to optimum.

Modified Vogel Method
This method can be used when the matrix of cells for suppliers and customers is of the type m . n where m >= 3 and n
>=3. In the account when the numbers m, n are near to three there we receive the same results as for not modified Vogel
method. Now to the modified Vogel method.

Table 7
Customers
Suppliers
K
1
K
2
K
3
K
4
Inventories Differences

S
1

20 14 11
110
12
200

310

3,9,9,x

S
2

6
200

15 18 15
200

9,x

S
3

17
50
12
100
19
40

23
190

7,7,7,7
Demands 250 100 150 200 700
Differences
14,3,3,
17,17,x
3,2,2,12,
x
8,8,8,19,
19,19,x
11,11,x

We take the three smallest c
ij
in every row and column. We order these three numbers with respect to greatness. Let us
suppose that we are working in the i
o
row and the smallest numbers are
i j i j i j
o o o
c c c
1 2 3
.
Now we do differences
o o
i j i j
-
3 2
2
d c c = and
1
o o
i j i j
-
2 1
d c c = hence we define modified difference (max difference) as
the sum of differences
1
d and
2
d , We obtain modified difference as
M
= +
1 2
d d d . We apply this modified method to
our example. The final table is the table 7. We see that the modified method gives the same dislocation as the VAM.
We can say that the modified method did not bring anything new and better. But it is also necessary to state, that it is
not worse. The modified differences can be more favourable for examples with bigger matrices. We show it on the
following example:

33

Example. Let us suppose that we have seven suppliers S
1
, S
2
, . . . , S
7
with inventories 100, 200, 300, 400, 500, 600
and 700 units of commodity and four customers K
1
, K
2
, K
3
, K
4
with demands for 250, 100, 150 and 200 of units. The
transport costs c
ij
from the i-th supplier to the j-th customer are given in table 8 :

Table 8
Customers
Suppliers
K
1
K
2
K
3
K
4
Inventories
Max
differences

S
1

9
10
11 6

7
90

100

3,3,3,3,3,3,x

S
2

8
150

9 3
50
7
200

5,5,5,5,5,x

S
3

6
300

8

4

5
300

2,2,2,2,2,2,2,
x

S
4

9 2
400
3 7
400

4,4,4,x

S
5

3
500
3 5 9
500

2,2,x

S
6

2 9 1
600
9
600

8,x

S
7

9 6
340
7 3
360

700

4,4,4,4,6,x
Demands 960 740 650 450
Max
differences
4,6,3,3,3
,3
4,4,4,5,x 2,2,3,3,3
,3,x
4,4,4,4,4
,2,x

Vogel method accomplishes the starting allocation that the cost is equal to 11580 financial units. Our modified method
brings the starting allocation in the cost 9890 financial units (see table 8). It is obvious that the modified Vogel method
presented here brings more better results than the classic Vogel method. But it is necessary to admit that at an other
choice of column or row with the same w-differences we can come to other and worst result.

REFERENCES
[1] ACKOFF, R.; SASSIENI, M. Fundamentals of operations research. N.Y. : Wiley, 1968.
[2] BELLMAN, R.; DREYFUS, S. Dynamic programming. Princeton : PUP, 1962.
[3] DUDIRKIN, J. Opereach research. Praha : SNTL, 1994.
[4] CHURCHMAN, C. W.; ACKOFF, R.; ARNOFF, E. Introduction to operation research. N.Y. : Wiley, 1975.
[5] SAATY, T. Mathematical Methods of Operations Research. N.Y. N.Y. : Mac grave, 1959.
[6] ZAPLETAL, J.; ZSTRA, B. Vybran kapitoly z operanho vzkumu. Zln : VUT-FT, 1983.
[7] ZAPLETAL, J. Operan analza. Skriptorium VO, Kunovice : EPI, s.r.o., 1995.
[8] ZAPLETAL, J.; VACULK, J. Podprn metody rozhodovacch proces. Brno : MU, 1998.

34

Address:
Abdurrzzag Tamtam
Faculty of Electrical engineering and Communication
Purkynova 118,
e-mail: xtamta00@stud.feec.vutbr.cz,

35

SOLUTION OF STRUCTURAL INTERBRANCH SYSTEM OF A DYNAMIC MODEL
Jaromr Batinec, Josef Diblk


Abstrakt: The whole range of problems by mutual deliveries among various manufacturing branches of
industry and a lot of sizes at on market place is given by crude productions of individual branches. In this
paper, it will be shown that this dynamic problem can be solved with special mathematical methods.

Klov slova: Branches of industry, crude production, technical coefficient,dynamic model.

Let us suppose that the economical system is divided into n manufacturing branches. We denote
i
x the whole amount
produced by i th branch of industry. Further we denote
ij
X the amount of production of i th branch supplied to the
j th branch. At the end we denote
i
y the amount of products of the i th branch for final usage (ii. market,
export). The whole relations among producers and their customers can be given as follows:

1 11 12 1 1
2 21 22 2 2
1 2
. . .
. . .
. . . . . . .
. . .
n
n
n n n nn n
x X X X y
x X X X y
x X X X y
+ + + +
+ + + +
+ + + +
(1)

The common cognitions allow us to do the following assumption: The supply
ij
X of the i th branch to the j th one
is direct proportional to the crude production of the j th branch
j
x . Then we have:

i j i j j
X a x (2)
where the coefficient of the direct proportion is called the technical coefficient.
If we know the supplies among all branches of industry and the amount for the final usage from the previous seasons
then we can calculate the technical coefficient such as we calculate all 1, 2, . . . ,
i
x i n from the system (1) and
hence the technical coefficient can be obtained from the equation

i j
i j
j
X
a
x
(3)
Calculating
ij
X from (3) for 1, 2, . . . , , and 1, 2, . . . , i n j n and putting them into (1) the following system of
linear equations is received:

1 11 1 12 2 1 1
2 21 1 22 2 2 2
1 1 2 2
. . .
. . .
. . . . . . . . . . . .
. . .
n n
n n
n n n nn n n
x a x a x a x y
x a x a x a x y
x a x a x a x y
+ + + +
+ + + +
+ + + +
(4)

The system (4) can be rewritten into the matrix form:

. X A X Y + . (5)

36

Vector X is called vector of crude production of branches, vector Y is called vector of final usage and the matrix A is
called the matrix of technical coefficients. The system of equations can be rewritten into the form
( ). E A X Y (6)
where E is a unit matrix of the type
n
n
it is n rows and n columns.
The matrix E A can be supposed as an operator of transformation which for the input vector X (vector of crude
production of branches) Y assigns the vector of output (vector of final usage). This result is not eminent for economy.
More interesting and more important for economy is to search out the operator which for given Y assigns the vector X.
This operator can be received by multiplication of the equation (6) by the inverse matrix ( )
1
E A

and we receive
( )
1
. X E A Y
(7)
The matrix ( )
1
E A

is denoted by B and it is called the matrix of the full material burden. Matrix equations are called
fundamental form of open static model of inter branches relations. The elements of the matrix

i j
B b 1
]
(8)
are called coefficients of the full material costs. The coefficient
i j
b indicates inverse consumption of the production of
the i th branch which is necessary for the delivery of the production of the j th branch for the final usage. Among
i j
a and
i j
b holds the following relation:
i j i j
a b for , 1, 2, . . . , i j n .
For the expression of economical interpretation of the coefficients
i j
b it is useful to restore the model (7) as the system
of the equations again:

1 1 2 2
. . . 1, 2, . . .,
i i i in n
x b y b y b y i n + + + (9)
For the new vector of usage
' ' '
1 2
' , , . . .,
n
Y y y y the system ( 9 ) is of the form:

' ' ' '
1 1 2 2
... 1, 2, ...,
i i i in n
x b y b y b y i n + + + (10)
Let us denote ,
i i i
x x x & ,
i i i
y y y & . , . 2 , 1 n i L
and we do a subtraction of the left and right sides of suitable equations of the systems ( 9 ) and (10). We receive the
system
. , , 2 , 1 ,
2 2 1 1
n i y b y b y b x
n in i i i
L L + + + (11)
The magnitude
i
x tells the difference of the change of the whole crude production of the i th branch if the change of
the components of the final usage is . , , ,
2 n i
y y y L It is possible to prove that the coefficients
i j
b are non negative
numbers and hence . 0
i
x We put 1
i
y and 0
i
y for all k j . Then from (11)
. , , 2 , 1 , 1 n i b x
ij i
L (12)
The coefficient
i j
b of the full material costs sets the value for which must be the production increased in the
i th branch that the j th branch increases supply for final usage upon the unit.
The coefficient
i j
b of the full material costs includes in itself at first the value of the direct supply from the i th branch
into the j th branch which is necessary to the production of a unit in the j th branch, secondly the values of
supplies of the i th branch which contracts the j th branch mediate by instrumentality of the others branches.
We supposed at the model (6) that we know the vector of the crude production X and we gain the vector of final usage
Y . Contrary at the model (7) we know the vector of final usage Y and with the aid of it we look for the vector of the
crude production X . As the third type of problems we solve the following situation. We have given for some branches
the crude production and for the other the amount of the vector of final usage. At this treatment of the model the
required vector contains k elements formed by crude production and n k elements formed by final trade outlets from
the system ( n k > ). The branches of the model can be transformed so that we obtain two families:
The first family contains branches for which the capacities of crude production are given.
The second family of branches for which we know the amount of the full final usage.
We compose our model such that we calculate from the first family the full final usage and from the second one the
crude production for complementary branches. The first family will be denoted by the index 1, the second by the index

37

2. We look for such matrix R for which the following equality relation holds.

1 1
2 2
.
X Y
R
Y X
1 1
1 1
] ]
(13)
It is obvious that we must construct four sub-matrixes with the following property:

11 12
1 1
2 2
21 22
.
R R
X Y
Y X
R R
1
1
1 1
1
1 1
1
] ]
1
]
(14)

where
11
R is the matrix of the type
n k
n k
,
12
R is of the type
n k
k
,
21
R is of the type
k
n k
and
2 2
R is of the type
k
k
.
From (14) there follows:

11 1 12 2 1
21 1 2 2 2 2
R X R Y Y
R X R Y X
+
+
(15)
We infer the matrixes
11 12 21 22
, , , R R R R as follows: We divide the matrix A of technical coefficients with the agreement of
the allocation of branches of industry into two families:

11 12
1 1 1
2 2 2
21 2 2
.
A A
X Y X
X Y X
A A
+
1
1
1 1
1
1 1
1
] ]
1
]
(16)

We extend the equation (16) into two equations with respect to the rules for multiplication of matrixes and we have:

11 1 12 2 1 1
21 1 2 2 2 2 2
A X A X Y X
A X A X Y X
+ +
+ +
(17)
After the rearrangement of the equations (17) we obtain

( )
( )
1 11 1 12 2
2 21 1 2 2 2
Y E A X A X
Y A X E A X

+
(18)
When we put
2
X from the second equation of (18) we get
( ) ( )
1
2 2 2 2 2 2 21 1
X E A Y E A A X
+
and hence

( ) ( ) ( )
1 1
1 11 1 12 2 2 2 2 2 21 1
Y E A X A E A Y E A A X

+
1
]
(19)
After comparing with the first equation from (15) we get

( ) ( )
( )
1
11 11 12 2 2 21
1
12 12 2 2
R E A A E A A
R A E A

(20)
Similarly as (20) we get from the second equation of (15)

( )
( )
1
21 2 2 21
1
2 2 22
R E A A
R E A

(21)
While the interpretation of the elements of matrixes ( ) E A and ( )
1
E A

in models (6) and (7) is simple, the
interpretation of the elements in the matrixes , 1, 2 ;
i j
R i j is more complicated task. For the understanding of the
content of the elements of sub-matrixes , 1, 2 ;
i j
R i j we study fractional products.

38

Dynamic systems
Production in a certain interval depends on the accumulation in the previous interval. The statistic models do not
express this dependence.

Accumulation is a component of final product. In the balance of the relations between the branches, we denoted the
final products of the individual branches as follows: y
1
, y
2
, ... , y
n
, - This final product basically consists of two parts:
y
i
(1)
... the consumed part, and
y
i
(2)
... the accumulated part, i.e. y
i
= y
i
(1)
+ y
i
(2)
.

For distinguishing the individual periods of time, we shall denote gross production of the i-th branch in year t by the
symbol x
i
(t) and the final product in year t by the symbol y
i
(t).

Accumulated product of i-th branch becomes a part of the means of other branches. We denote the part of accumulated
product of t i-th branch in year t which is invested into the j-th branch as y
ij
(t). Then
. ) (
1
) 2 (

n
j
ij i
t y y
If the accumulated product itself consists only of floating means that are consumed in the following year (t+1), the
following relation obviously holds between the increment of production in the j-th branch [x
j
(t+1)-x
j
(t)] and the
investment into the products of the i-the branch y
ij
(t):
) ( ) 1 ( )] ( ) 1 ( [ ) ( t X t X t x t x a t y
ij ij j j ij ij
+ + .
However, part of the accumulated product is of the form of basic funds that are not consumed within a single year.
Suppose the consumption of investments y
ij
(t) is divided into T
ij
years. This means that only T
ij
-part of the investment
y
ij
(t) is consumed within the following year. The reality is hus better characterised by the relation
)) ( ) 1 ( (
) (
t x t x a
T
t y
j j ij
ij
ij
+
,
which after multiplication gives
)). ( ) 1 ( ( ) ( t x t x T a t y
j j ij ij ij
+
The relation between the increment in the year (t+1) and the accumulation in the preceding year is therefore given by a
system of technical coefficients a
ij
and by the system of average periods of usability of T
ij
, that are also of technical
nature. Therefore we substitute them with the so-called "investment coefficient", denoted by c
ij
:
ij ij ij
T a c
The system of investment coefficients, and similarly the system of technical coefficients, forms square matrix C.
Using the investment coefficients, the relation between the increment of production in the j-th branch and and the extent
of investments into the production of the j-th branch may be written in the form
)) ( ) 1 ( ( ) ( t x t x c t y
j j ij ij
+
The whole of the accumulated production is thus equal to

+
n
j
n
j
j j ij ij i
t x t x c t y t y
1 1
) 2 (
. )) ( ) 1 ( ( ) ( ) (
This equation connects the accumulation of the i-th branch with the increment of the production in the individual
branches. Similar equations may be obtained for all the branches
). ( )) ( ) 1 ( ( )) ( ) 1 ( ( )) ( ) 1 ( (
), ( )) ( ) 1 ( ( )) ( ) 1 ( ( )) ( ) 1 ( (
), ( )) ( ) 1 ( ( )) ( ) 1 ( ( )) ( ) 1 ( (
) 2 (
2 2 2 1 1 1
) 2 (
2 2 2 2 22 1 1 21
) 2 (
1 1 2 2 12 1 1 11
t y t x t x c t x t x c t x t x c
n n n nn n n
n n n
n n n
+ + + + + +
+ + + + + +
+ + + + + +
L
O
L
L

39

From this system of equations, we may directly determine how much we need to accumulate from the production of the
individual branches in the given year to reach the planned increment of the production in the individual branches. When
we write the system of equations in the matrix form
,
) 2 (
Y X C
and after adaptation
) 2 ( 1
Y C X

we may determine the increment of the individual branches in the following year, when the level and structure of
accumulation is given.

Now we take into account the period t = 1, 2, ..., T and adopt the following notation:
+ +
n
j
i ij ij i
n i t y t z t X t x
1
, , , 2 , 1 ), ( )) ( ) ( ( ) ( L (22)
where X
ij
(t) is the supply of the i-th branch to the j-th branch for consumption in the t-th period of time, z
ij
(t) is the
supply of the i-th branch to the j-th branch for investment during the t-th period of time, y
i
(t) is the final product of the
i-th branch in the t-th period of time. Suppose, similarly as with the static model, that the
), ( ) ( t x a t X
j ij ij

where a
ij
is the technical coefficient that does not change in time.

Further, suppose that the supply of the i-th branch to the j-th branch for investment during the period of time t is
proportional to the increment of the production of the j-th branch in one period of time:
), ( )) ( ) 1 ( ( ) ( t x c t x t x c t z
j ij j j ij ij
+
where x
j
(t) is the increment of the total production of the j-th branch within one period, c
ij
is the investment coefficient
, again independent of time.

By substituting into the balance equation (22), we obtain, in the matrix form,
), ( ) ( ) ( ) 1 ( t Y t X E C A t CX + +
We obtained a system of difference equations
), ( ) ( ) ( ) 1 ( t Y t X C A E t CX + + +
). ( ) ( ) ( ) 1 (
1 1
t Y C t X C A E C t X

+ + +
From this system, we may determine how the inter-branch relationships should look in order to obtain the required
growth.
Denote N C M C A E C +

: , : ) (
1 1
. We receive
). ( ) ( ) 1 ( t NY t MX t X + + (23)
We obtained a system of difference equation defined for all t. If |M| is not equal to 0, then a unique solution of the
system exists.

By subsequent substitution, we transform the system (23) into a single difference equation of the n-th order, which may
be solved e.g. with the aid of eigenvalues of the characteristic equation.

If M has a small order, then the solution was described in [17].
Moreover, we can use the next Theorem (proof see [3], p. 124).

Theorem:
The unique solution of the initial value problem

( ) ( ) ( ) ( ), 1 n G n X n M n X + +
( )
0 0
X n X
is given by

40

( ) ( ) ( ) ( ). , ,
1 1
1
0
1
0 0
0 0
r G i M X i M X n n X
n
n r
n
r i
n
n i

,
_
,
_

If M is a constant matrix, then the solution is given by
( ) ( ). , ,
1
1
0 0 0
0
0
r G M X M X n n X
n
n r
r n n n

+
Example:
We consider the system

( ) ( ) . 0 0 , 1 0
. 1 ) ( 2 ) 1 (
, ) ( ) ( 2 ) 1 (

+ +
+ + +
y x
n y n y
n n y n x n x

Solution:
In this case we have
,
_
1 0
1 2
M ,
,
_
1
) (
n
n G ,
,
_
0
1
) 0 ( X .

Then
,
_
n
n n
n
n
M
2 0
2 2
1
.
Hence
( )
,
_
,
_
,
_
,
_

1 2 0
2 1 2
0
1
2 0
2 2
) (
1
0
1
2 1 1
r r n n
n X
n
r
r n
r n r n
n
n n

( )
( ) *
2
2 1 2
0
2
1
2 1
1
0
,
_
+
+
,
_
r n
r n r n
n
r
n
r n r

We use the formula
( ) ( )
( )
1
1
2
1
1
1 1
n
r
n
n r
a
a na a a
ra .

(*)
( )
,
_
,
_
,
_
+
1
1
]
1
,
_
,
_
,
_
,
_
,
_
,
_
,
_

n
n n
n
n
n
r
r
n
r
n
r
r r
n
n
n
n
2
1
1
2
1
2
1
1
2
1
2
0
2
2
1
2
1
2
1
4
1
2
1
4
1
2
0
2
2
1
0
1
1
1
0

,
_
,
_
,
_
,
_
,
_
+
1
1
]
1
,
_
,
_
n
n n
n
n
n
n
n n
n n n
2
1
1
4
3
2 2
2
1
1
2
1
2 2 2
1
4
2
0
2
1
.

41

So we have
( )
( ) .
2
1
1
,
4
3
2 2
1
n
n n
n y
n n n x
,
_

+

Acknowledgement
This research has been supported by the Czech Ministry of Education in the frame of MSM002160503
Research Intention MIKROSYN New Trends in Microelectronic Systems and Nanotechnologies.

REFERENCES
[1] ACKOFF, R. L. Progress in Operation Research. New York : John Wiley & Sons, Inc. 1961.
[2] CHURCHMAN, Ch. W.; ACKOFF; R. L.; ARNOFF, L. Introduction to Operations Research. New York :
John Wiley & Sons, Inc. 1957.
[3] ELAYDI, S. N. An introduction to dufference equations, Second Edition, Springer, 1999.
[4] HABR, J.; VEPEK, J. Systmov analza a syntza. Praha : SNTL, 1972 .
[5] BECK, J.; LAGOV, M.; ZELINKA, J. Linern modely v ekonomii. Praha : SNTL, 1982.
[6] KLAPKA, J.; DVOK, J.; POPELKA, P. Metody operanho vzkumu. Brno : VUTIUM, 2001.
[7] PRGEROV, A. Diferenn rovnice. Praha : SNTL, 1971.
[8] RAIS, K. Vybran kapitoly z operan analzy. Brno : PGS, 1985.
[9] ROCCAFERRERA, G. M. F. Operation Research Models for Business and Indusry. Chicago, New York :
S.W. publishing company, 1964.
[10] TER-MANUELIANC, A. Modelovn problm zen. Praha : Institut zen, 1977.
[11] VACULK, J.; ZAPLETAL, J. Podprn metody rozhodovacch proces. Brno : Masarykova univerzita 1998.
[12] WALTER, J. a kol. Operan vzkum. Praha : SNTL, 1973.
[13] WALTER, J. Stochastick modely v ekonomii. Praha : SNTL, 1970.
[14] ZAPLETAL, J.; ZSTRA, B. Vybran kapitoly z operanho vzkumu. Zln : VUTFT, 1983.
[15] ZAPLETAL, J. Operan analza. Kunovice : Skriptorium VO, 1995.
[16] ZAPLETAL, J. Structural Interbranch System of Static Model. International conference of EPI Kunovice,
2005, 303 307. ISBN 80-7314-052-7.
[17] BATINEC, J. Structural Interbranch System of Dynamic Model. International conference of EPI Kunovice,
2005, 317 322. ISBN 80-7314-052-7.

Address:
Doc. RNDr. Jaromr Batinec, CSc.
Department of Mathematics
Faculty of Electrical Engineering and Communication
Technick 8,
616 00 Brno,
bastinec@feec.vutbr.cz

Address:
Prof. RNDr. Josef Diblk, DrSc.
Technick 8,
616 00 Brno,
diblik@feec.vutbr.cz

42

43

PROBABILITY THEORY AND STATISTICS IN THE COMBINED FORM OF STUDY OF THE
BACHELOR STUDENT PROGRAMMES AT FEEC BUT

Michal Novk


Abstrakt: At FEEC BUT the teaching of bachelor student programmes in the combined form of study
began in the academic year 2004/2005. This contribution discusses how the basics of probability theory
and statistics are included in the programmes. Some general information on the course as well as first
experience from it are also given.

Klov slova: distance learning, combined form of study, probability, statistics, teaching mathematics

Introductory information context and prerequisites
Teaching in the combined form of study began at FEEC BUT in the academic year 2004/2005. The courses as well as
their outlines and requirements are the same as in the attended form of study (with the exception of one subject offered
in the attended form but not in the combined one). Mathematics in the bachelor student programmes is therefore taught
in four subjects: Mathematical seminar, Mathematics 1, Mathematics 2 and Mathematics 3.

Mathematical seminar is meant to revise secondary school knowledge of mathematics necessary for further studies. In
the combined form of study it is a long weekend course at the beginning of the term; the students can either attend it or
submit exercises only.

Mathematics 1 (in the first term) includes basics of linear algebra, basics of differential and integral calculus of
functions of one variable and basics of differential calculus of more variables. Mathematics 2 (in the second term)
includes solving differential equations and basics of theory of complex functions and integral transformations. Both
courses consist of five tutorials.

The course on probability and statistics
Basic concepts of probability and statistics are taught during the third term in Mathematics 3. Its outline, however,
includes basic numerical methods as well. There are four tutorials (3 lessons each) out of total number of six in the third
term which include Mathematics 3. The outline of the first Mathematics 3 course was as follows:
Tutorial 1
introduction, revision, classical and geometrical probabilities, discrete and continuous random variable, issues
of expected value and dispersion
Tutorial 2
some basic distributions of probability (binomial, Poisson, exponential, uniform, normal), idea of statistical
testing, basic statistical tests (sign test, z-test & mean expected value test), queuing theory (time permitting)
Tutorial 3
numerical methods (not to be discussed here)
Tutorial 4
final tutorial, revision, sample tests

The tutorials introduce students to the basic concepts of probability theory and statistics. Students learn that a relatively
small number of formulas and theorems have profound effects in a number of situations. The choice of tasks emphasises
the way of decoding the respective word problems and finding the way of applying the mathematical means to solve the
problems. The part on statistical testing stresses the general concept of testing and its applicability in various contexts.

44

Students are required to submit exercises for each tutorial. These are assessed by a maximum of 30 points in total. The
written only exam at the end of term is assessed by a maximum of 70 points. In order to pass the subject students are
required to acquire the minimum of 50 points. Students have access to sample tests these can be downloaded from
teachers homepage. The final tutorial deals with the exam as well.

Students can study the subject matter from [2], which respects the needs of distance form of study and was targeted at
this subject. The text, which is also used by students in the attended form of study, is available as a PDF file from a
number of links including the faculty website and teachers homepage. Special office hours in convenient time are set
for the combined students only consulting subject matter by telephone is widely used.

Another contribution in the proceedings of this conference, [3], gives examples of tasks solved throughout the course.
Applications of these problems as well as the necessary mathematical knowledge are included there as well.

Probability and statistics in the master study programmes
Since teaching in the combined form of study started only as late as 2004/2005, there are no students in the master study
programmes yet. The master study programmes in the attended form of study, however, offer a course Probability,
statistics, operations research, which deepens the knowledge of probability theory and statistics. The topics dealt with
in the course include: basic statistical tests t-test, F-test; confidence intervals; linear regression; post-hoc tests;
goodness of fit test; nonparametric tests; mathematical methods in economics - linear programming, the transport
problem; dynamic programming, recursive algorithm, inventory models.

Conclusion
Knowledge of probability theory and statistics is necessary for dealing with a great number of situations which can
occur in various contexts. The combined form of study as a means designed to provide access to this knowledge to
students who are already employed can help to improve position and status not only of such students but also of their
employers.

References
[1] BATINEC, J. Vuka matematiky na FEKT VUT Brno (v bakalskm i magisterskm studiu). In 37.
konferencia slovenskch matematikov. ilina : Slovensk matematick spolenost, 2005.
[2] FAJMON, B.; RIKOV, I. Matematika 3. Brno : UMAT FEKT VUT, 2003, available from
https://www.feec.vutbr.cz/et/skripta/umat/Matematika_3_S.pdf.
[3] NOVK, M. Examples of using concepts of probability theory in management decision making. This
proceedings.

Address:
Mgr. Michal Novk, Ph.D.
stav matematiky,
FEKT VUT v Brn
Technick 8
616 00 Brno
tel.: +420-541143135
e-mail: novakm@feec.vubtr.cz

45

APLIKACE FUZZY SYSTM PRO PODPORU ROZHODOVN A ZEN
Vladimr Mikula, Jindich Petrucha


Abstrakt: Tento pspvek pojednv o pouit fuzzy logiky pro podporu rozhodovacho procesu. Zkladn
poznatky o fuzzylogice byly publikovny v poetnch literrnch pramenech- viz pehled literatury na konci
tohoto lnku. Strun pehled teorie fuzzy systm me ten nalzt v tutorilnm lnku publikovanm
autorem v Proceedings of the International Conference of EPI, s.r.o. Kunovice, leden 2005 [1] Popsno je
uspodn fuzzy systmu s vysvtlenm elu a funkce individulnch blok systmu a strun popis vyuit
pro rozhodovac a dic procesy.

Klov slova: fuzzy sets, fuzzy systems, universe of fuzzy sets,degree if membership, fuzzification of input
and out-put variables, fuzzy rules, fuzzy associative memory FAM (bank of rules), fuzzy inferences,
MAXMIN and MAXPROD methods (MAMDANIs and LARSENs method), centroid, defuzzification of
centroid, crisp value of output variable.

vod
V relnm svt jsou jevy a aktivity popisovny pokud mono exaktn na zklad idealizovanch matematickch
model (bn v prodnch a technickch vdch). Avak v nkterch oblastech, nap. v ekonomice, v organizovn
chodu systm spoleenskho charakteru) nejsou vdy k dispozici exaktn matematick modely, nebo jsou obtn
formulovateln, anebo, i kdy by se dal takov model zformulovat, byl by velmi komplikovan a tkopdn a mono i
prakticky nepouiteln. V tchto situacch se m dl vce zanaj uplatovat metody, popisujc systmy a jevy na
zklad expertnch znalost sledovan problematiky. Pro een se vyuvaj postupy a metody z oblasti uml
inteligence, kam pat uml neuronov st (napodobujc mylen) a fuzzy systmy, zaloen na fuzzy logice (a na tzv.
approximate reasoning,) a vyuvajc tzv.lingvistick promnn, tedy vyjadovn pomoc vgnch jazykovch
prostedk, hodnotcch velikost parametr njak veliiny, odstupovan v jistm rozsahu pomoc vstinch slov
(nlepek, label), nap: velmi mal, mal, stedn, velk, velmi velk, apod..) I kdy tyto pojmy nejsou oste vymezeny,
jsou tzv neostr, nebo-li fuzzy, vyjaduj obvykle v pijateln a srozumiteln form pslunou oblast platnosti njakho
tvrzen a daj se modelovat pomoc fuzzy mnoin. Pouit ostr (binrn, booleovsk) logiky, uznvajc jenom
pravdivost (vyjdenou symbolem logick jedniky), nebo nepravdivost (vyjdenou logickou nulou) jistho tvrzen je
nkdy pli hrub a nepijateln a lpe vyhovuje odstupovn mry pravdivosti hodnotami kontinualn rozloenmi
v definovanm intervalu (nap. od nuly do jedniky). Proto je ve fuzzy logice zaveden pojem stupn pslunosti prvku
do dan fuzzy mnoiny, pro kter plat < 0, 1 >. Vyuvn vgnch pojm pro popis skutenost pomoc
lingvistickch prostedk je odedvna bn a mnohdy velmi vhodn a je soust kadodennho ivota. Proces zen
anebo rozhodovn je pak expertn popsn pravidly typu
JESTLI-E < pedbn podmnka > PAK < nsledek, innost >, nebo, jak je veobecn pouvno IF <antecedent >
THEN < consequent >. Soubor tchto pravidel pak tvo tzv. znalostn banku ( fuzzy associative memory,
FAM) a postihuje pomoc patin kvantifikovanch lingvistickch prostedk uvaovan dic, nebo rozhodovac
proces je to obdoba programu sekvennho digitlnho potae, operujcho v pevn vtin na bzi ostr,
booleovsk logiky. Poznatky o fuzzy logice, kter je obecnj ne ostr binrn logika (je je specilnm ppadem
fuzzy logiky) jsou popsny ve velmi poetn literatue - viz reference na konci tohoto lnku.

46

Uspodn fuzzy systmu pro zen a pro podporu rozhodovn
je uvedeno na Obr. 1. Algoritmus a funkci jednotlivch blok fuzzy systmu lze strun popsat takto:

Obr. 1. Uspodn fuzzy systmu

1. Zskn hodnot vstupnch promnnch x
1
a x
n
(nap. pomoc senzor na zenm objektu (nebo
z vhodndatabanky)
2. prava tchto hodnot (normovn, prava na bezrozmrn seln hodnoty)
3. Fuzzifikace vstupnch a vstupnch veliin, tj. jejich rozdlen na dl fuzzy podmnoiny v pslunch
univerzech a pidlen nzv (nlepek, label) tmto podmnoinm ve smyslu ve uvedench vah o jazykov
promnn. Otzkou je stanoven optimlnho potu tchto podmnoin. m vce jich zvolme, tm jemnj zen,
nebo rozhodovn doshneme,ale prodlou se tm vpoetn as. Ukazuje se, e rozumn poet podmnoin je 3 a
9, obvykle se vol lich poet (symetrick rozloen kolem stedn hodnoty. Sousedn podmnoiny se mus
sten pekrvat, aby bylo dosaeno plynul pechzen univerzem. Stupe pekryt (overlapping) si vyaduje
hlub rozbor, ale piblin lze zvolit koeficient pekryt jako pomr k
p
= A / B, kde A je ka intervalu pekryt
dvou sousednch podmnoin na ose x a B je celkov interval na ose x , zabran tmito podmnoinami . asto se
vol pekryt tak, aby prsek obou pekrvajcch se podmnoin byl na hodnot = (0,3 a 0,5).
4. Sestaven banky pravidel (FAM- fuzzy associative memory) na zklad expertnch znalost ee-nho
problmu. Pravidla (rules) jsou ji zmnnho typu IF < antecedent > THEN < consequent >. Rozmr banky
pravidel je roven potu vstupnch promnnch n a poet pravidel P je roven souinu potu podmnoin p
i
na kter
jsou patin univerza vstupnch veliin rozdlena, tedy P = p
1
.

p
2
. ... p
n
. Uvaujme nap. systm, kter m dv
vstupn veliiny, tedy n = 2 (ozname je x
1
a x
2
) a jednu vstupn veliinu y. Banka pravidel bude tedy
dvourozmrn (obdlnkov). Nech univerzum veliiny x
1
je rozdleno do pti podmnoin, oznaench
lingvisticky na stupn: velmi nzk (VN), nzk (N), stedn (S), vysok (V) a velmi vysok (VV), univerzum
veliiny x
2
nech m ti stupn: nzk (N), stedn (S ) a vysok (V ), a nech vstupn veliina y bude mt tak ti
stupn: mal (M) stedn (S) a velk (V).. Poet pravidel P = p
1
. p
2
= 5 . 3 = 15. Nech expertn stanoven
pravidla, reprezentovan doporuenm obsazenm jednotlivch polek znalostn banky maj rozdlen podle Obr.
2a. Je-li poet vstupnch promnnch roven tem, banka pravidel bude trozmrn hranol, obecn pro n vstupnch
promnnch to bude n rozmrn tleso (Obr.2b). Dle plat zsada, e pro kadou vstupn veliinu musme
vytvoit samostatnou banku pravidel
5. Fuzzy inference. Fuzzifikovan vstupn a vstupn veliiny pivdme do bloku inferenc, kde na zklad pravidel
uloench ve znalostn bance jsou vykonvny operace, zvan fuzzy inference. Vsledkem tchto operac je
zskn tzv. centroid vstupnch veliin, co jsou obvykle subnormln fuzzy mnoiny (nedosahujc rovn =
1). Inference jsou typu MAXMIN (Mamdaniho metoda), nebo MAXPROD (Larsenova metoda). Pro objasnn
algoritmu inferenc uvaujme konkrtn fuzzifikovan veliiny podle Obr.3.

Blok defuzzifikace
6
Blok prav
vstupnch veliin
7
Blok fuzzy
inferenc
5
Banka pravidel IF-
THEN (FAM)
4
Blok fuzzifikace
3
Blok prava
vstupnch veliin
2
y
1
... y
m

x
1
...x
n

y
1
...y
m

Vstupn promnn
zen nebo
rozhodujc objekt
Upraven vst. prom. y
1
, ...y
m

ostr
hodn.
1

47

x
1
y
VN N S V VV

x
2
N V V V V V x
3
x
1
S M S S S S x
2

V M M M M M

a b

Obr.2. Pklad banky pravidel pi dvou (a) a pi tech (b) vstupnch promnnch.

U inferenc typu MAXMIN postupujeme takto:
pro definovan hodnoty vstupnch veliin, nap. x
1A
a souasn x
2A
zjistme do kterch fuzzy podmnoin tyto veliiny
pat a s jakm stupnm pslunosti (x
1A
), (x
2A
). Tak nap. x
1A
le v mnoin V s hodnotou = 0,7 a tak
v sousedn mnoin S, kde dosahuje hodnotu = 0,35. Veliina x
2A
le pitom v mnoin N, kde dosahuje hodnoty
= 0,45 a tak v mnoin S s hodnotou = 0,15. Na zklad banky pravidel (pouijeme FAM banku uvedenou ve)pro
tuto situaci zapeme nsledujc pravidla:
R1 : IF x
1
= x
1A
V ( 0,7) AND x
2
= x
2A
N (0,45) THEN y V (0,45)
OR
R
2 :
IF x
1
= x
1A
S (0,35) AND x
2
= x
2A
S (0,15) THEN y S (0,15).
OR next rule.
Protoe v antecedentu je pouita spojka AND, znac prnik, pouvme ve smyslu pravidla o prniku fuzzy mnoin
[1]:
V
= min (
V
,
N
). Ob pravidla R1 a R2 plat souasn, take je spojme spojkou OR, nebo-li sjednocen.
Souasn plat i ostatn pravidla, ale ta se v uveden situaci neprojev..

Vslednou fuzzy mnoinu vstupn veliiny (vsledn centroid) dostaneme tak, e mnoinu V vstupn veliiny y
oeeme ve vce = 0,45 a mnoinu S oeeme ve vce = 0,15. Takto zskan dl centroidy, ozname je jako C
1
a
C
2,
spojme podle pravidla o sjednocen fuzzy mnoin:
C
= max (
C1
,
C2
), jak je ukzno na obr 3a.

VN N S V VV M S V
1
0,7 0,45
MAXMIN
0,35 0,15

0 x
1A
x
1
0 y
C
2
C1

N S V M S V
1 1
MAXPROD
0,45 0,45
0,15
0,15
0 x
2A
0 y
C
2
C
1

Obr.3. Rozdlen fuzzifikovanch veliin uvaovanho fuzzy systmu a ukzka inferenc typu MAXMIN a
MAXPROD.

48

Podle metody MAXPROD (Larsenova metoda) postupujeme tak, e pi zskvn dlch centroid neoezvme
pslun podmnoiny vstupn veliiny, ale je snme vynsobenm (product, odtud PROD) hodnotou minima
zskanho aplikac jednotlivch pravidel. Takto zskan dl centroidy pak spojme operac sjednocen (MAX) a
dostaneme vsledn centroid (Obr. 3.).
1. Defuzzifikace centroidu vstupn veliiny je operace pi n se sname vyhodnocenm centroidu zskat ostrou
(crisp) hodnotu vstupn veliiny y. Je nkolik metod defuzzifikace, ale nejpouvanj je metoda nalezen
souadnice tit centroidu (centre of gravity, COG), tedy y
COG
podle vztahu
n
(y)y dy (y
i
) y
i
- i = 1
y
COG
=
n
(y) dy (y
i
)
- i =1

Druh st tohoto vztahu pouv msto integrace souty vraz (y
i
) y
i
a (y
i
) zskanch vzorkovnm centroidu
(Obr. 4.).

1

( y
i
) tit ( COG)

.

0
y
1
y
2
y
i
y
n
y

y
COG

Obr. 4. Centroid vstupn veliiny vyjaden pomoc vzorkovn.

Urme-li vechny hodnoty y
COG
pro vechny hodnoty vstupnch veliin a znzornme je v souadnm systmu y
COG
= f
(x
1,
x
2
. x
n
), dostaneme tzv. rozhodovac (u dicho procesu dic) plochu uvaovanho fuzzy systmu v (n +1)
rozmrnm prostoru.
2. Blok prav vstupnch hodnot. Vstupn ostr hodnoty y
COG
zskan na vstupu defuzzifiktoru mohou bt
pouh seln hodnoty bez rozmru. Pro reln systm me bt nutn dodat k nim rozmr a pizpsobit je
k doucm rozsahm veliin pro kter je systm sestaven (nap. v relnch dicch systmech veliiny y mohou bt
elktrick napt, nastavovan v definovanch rozsazch, nebo me jt o dlku asovho intervalu po kter m
funkce systmu probhat, atd.). K tomu slou blok prav vstupnch veliin.

Aplikace poznatk
Ve uveden postupy lze aplikovat v rznch oborech. Uveme zjednoduen pklad z oblasti rozho-dovn
managementu pi financovn urit akce, kterou m realizovat subjekt S, jeho jakostn parametry P
1
, P
2
... P
n
jsou
uloeny v databzi DB a na zklad tchto dat lze expertnm zpsobem vyhodnotit daje potebn pro rozhodnut, zda
uvaovanou akci v poadovan vi nklad financovat nebo ne.

Ale rozhodovn jenom mezi dvma krajnmi hodnotami tedy ANO nebo NE me bt pli hrub. Jemnj a snad
pijatelnj by bylo pouit jet dvou stup mezi nimi, tedy nap. SPE ANO a SPE NE. Pouit dalho stupn
uprosted by mohlo vst k neuritmu stavu, proto se piklonme k uvaovanm tyem stupm doporuen zda
uvaovanou akci realizovat, tedy ANO, SPE ANO, SPE NE a NE .

Vi nklad na akci lze tak kvantifikovat vhodnmi lingvistickmi stupni nap. VYSOK, STEDN a NZK.
Kvalitativn parametr realizanho subjektu (uvaujme pro nzornost jen jeden (vznamn) parametr, a to P
1
, i kdy

49

tuto metodiku lze zavst i pro vce parametr) tak rozdlme expertn na douc poet lingvisticky odstupovanch
hodnot, nap. VELMI MAL, MAL, STEDN, VELK, VELMI VELK. Algoritmus tohoto procesu znzoruje
Obr. 5.

Finln
Databze umoujc urit Expertn stanoven Fuzzy systm pro podporu rozhodnut
kvalitativn parametry fuzzy podmnoin rozhodovacho procesu
realizanho subjektu vst. a vst.veliin podle schmatu na Obr. 1
a banky pravidel

x
1
y VM M S V VV N SN SA A
x
2
VM M S V VV 1 1
N N SN SA A A
S N SN SA A A 0 x
1
0 y
V N N SN SA A
N S V
Banka pravidel IF - THEN 1 x
1
kvalitativn parametr P
1
realiz. subjektu
x
2
vka nklad
y rozhodnut managementu zda akci
0 x
2
financovat pi uvaovan vi nklad

Obr. 5. Algoritmus rozhodovacho procesu, pklad banky pravidel IF- THEN a rozvren fuzzifikovanch vstupnch a
vstupnch veliin.

Zvr
Tento lnek podv strun pehled o metod vyuit fuzzy logiky pro podporu rozhodovacho procesu.
Navazuje na pedchoz tutoriln lnek [1], uveden ve sbornku mezinrodn konference EPI Kunovice, konan
v lednu 2006. Vyuit fuzzy mnoin umouje zskat vhodn vsledky i v ppadech, kdy nen k dispozici exaktn
matematick model systmu, nebo je velmi komplikovan a nevhodn pro een v relnm ase, ale kdy existuj
expertn znalosti o een problematice.Potebn teoretick zklady jsou v uvedench literrnch pramenech. V lnku
jsou strun popsny jednotliv bloky fuzzy systmu, jejich funkce, ppadn nejdleitj zsady nvrhu. Na zvr je
ukzn pklad koncepce fuzzy systmu pro podporu rozhodovacho procesu managementu pi financovn urit akce.
Metodika se d vyut i pro rzn dal lohy z oblasti bankovnictv, ekonomiky, prmyslu, atd. Vstup tohoto systmu
lze brt spe jako kvalifikovan doporuen pi rozhodovn. Finln rozhodnut, samozejm, zle na nzoru
rozhodujcho subjektu, jm je uveden management.

Literatura:
[1] MIKULA, V. Exploitation of fuzzy logic in control and decision processes. Proceedings of the International
Conference of Kunovice : EPI, 2005.
[2] ZADEH, L. A. Fuzzy Sets. Inf. and Control, 8, 1965, pp. 338- 353.
[3] KOSKO,B. Neural Networks and Fuzzy Systems. Prentice Hall Inc., 1992.
[4] NOVK, V. Fuzzy mnoiny a jejich aplikace. Praha : Matematick semin, SNTL, 1992.
[5] NOVK, V. Zklady fuzzy modelovn. Praha : BEN Technick literatura, 2000.
[6] POKORN, M. Uml inteligence v modelovn a zen. Praha : BEN- Technick literatura, 1996.
[7] ZADEH, L. A.; LANGARI, R.; YEN, R.; COX, E. Fuzzy Logic Educational Program. Motorola Co. 1992.
[8] KAUFMANN, A. Initiation lementaire aux Sous- ensambles Flous l Usage des Dbutants. cole
Polytechnique Fderal de Lausanne, 1992.
[9] KONEN, V.; PEZLAR, R.; REJNU, O. Fuzzy expertn systmy systmy a rozhodovn. Brno : Acta
Universitatis Agriculturae et Silviculturae Mendelianae Brunensis, 2001.

50

Adresa:
Prof. Ing. Vladimr Mikula, CSc.
Osvobozen 699,
686 04 Kunovice
te./fax.: +420 572 549 018, +420 572 548 788
e-mail: mikula@vos.cz

Adresa:
Ing. Jindich Petrucha, Ph.D.
Osvobozen 699,
686 04 Kunovice
te./fax.: +420 572 549 018, +420 572 548 788
e-mail: petrucha@vos.cz

51

EXAMPLES OF USING CONCEPTS OF PROBABILITY THEORY IN MANAGEMENT
DECISION MAKING
Michal Novk, Betislav Fajmon

FEKT VUT v Brn

Abstrakt: Probability theory and statistics play an important part in everyday company and management
life and decision making. In the contribution we show that even very basic concepts can be used in solving
practical problems. The choice of tasks and ways of solving them conform to the curriculum of a course on
probability theory and statistics offered in a combined form of study of bachelor student programmes at
FEEC BUT, which is referred to elsewhere in this proceedings.

Klov slova: probability theory, statistics, combined form of study, teaching mathematics

Introductory information
Elsewhere in the proceedings of this conference, [6], there is mentioned the curriculum of a course on probability theory
and statistics offered in the combined form of study at FEEC BUT. This course is a brief and introductory one only it
includes classical and geometrical probabilities, discrete and continuous random variable, basic terms of statistics, some
very basic distributions of probability and introduction to the issue of statistical testing. Yet even these concepts only
can be used in solving some important practical tasks.

We are going to discuss various problems which can occur in various contexts in a number of variations and we are
going to show how the knowledge of basic concepts only can help in solving them. The variability of the choice is
intentional we want to show that the basics of probability theory and statistics can be applied in a number of situations
at almost any position in the company without any special deep education, long training or use of specialised software.

The issue of guarantee period
Let us consider the following situation:

The operating life of a product can be described by a certain distribution of probability. We are ready to tolerate a
certain number of legitimate complaints in the guarantee period. What length of the guarantee period shall we set?

Once we know the distribution of probability describing the operating life of our product, the task can be solved in a
simple way. Let us denote the continuous random variable describing the operating life of our product by X , the
expected value of the operating life by EX, the relative number of legitimate complaints we are ready to tolerate by
and the length of the guarantee period by G. Then in fact we need to find such G that < ) ( G X P . For the sake of
simplicity, let us consider the exponential distribution
1
. Then we have

EX
G
e 1 , which results in
) 1 ln( EX G .

Designing a special offer
Let us imagine the following special offer:

We are going to offer something for free. Every eligible person is entitled to do something (cast a die, turn a lottery-

1
The choice follows from the fact that the exponential distribution is one of those taught in the course on probability
and statistics referred to in [6]. It is to be noted, however, that the choice of exponential distribution in this context may
be rather special.

52

wheel, etc) with a relatively small number of possible outcomes. If certain states are reached, the person is given the
advertised thing and entitled to another chance. This keeps repeating as long as the desired states are reached.

We may consider the following specific example:
Beers for free! If you can squeeze in, turn our lottery-wheel! Six numbers only! 5 and 6 mean a beer for free and
another chance! As long as 5 and 6 keep falling, you keep drinking! The only way to stop drinking for free is to pray for
1 to 4!

The nature of this offer can be revealed using the concept of expected value. Let X be a discrete random variable
denoting the number of beers drunk for free by one person and p(x) the probability mass function of X. Let us except the
cases of stopping drinking for other reasons than turning out the wrong numbers. We get that
1
3
2
) (
+
x
x p for
,...} 2 , 1 , 0 { x and 0 ) ( x p otherwise. The expected number of beers drunk for free by one person in such a special
offer can be computed as
0
1
3
2
x
x
x EX and it turns out that EX=0.5, which is definitely less than the offer suggests.

Designing board games
Many board games contain fields known as function fields, i.e. fields which require some action or direct the game.
The flow of the game can be controlled or directed by a suitable choice of positions of these fields, or rather the number
of fields between them. It becomes apparent if the
game is played with more than one die and the
length of each players move is the sum of numbers
on the dice. The graph shows values of probability
mass function of a discrete random variable denoting
the possible sums on two six-sided dice:

Seven is the most likely sum on two dice it is three
times more probable than three or eleven and six
times more probable than two or twelve.

The only mathematical concept used in this example
is the notorious formula of classical probability
| |
| |
) (
A
A P , where |A| is the number of positive
outcomes of an experiment and || is the number of
all possible outcomes.

The issue of guessing in multiple choice tests
There are many objections to multiple choice tests. It has been often suggested that the results may be influenced by
simple guessing. Let us consider the following conditions:
There is a given number of questions (let us denote it by n), each of which offers the same number of options (let us
denote it by k), out of which r options are always correct. Let us suppose that the respondent does not know anything
about the subject matter of the test yet knows how many options are correct and guesses accordingly. What is the
expected number of correctly answered questions if the question is regarded as answered correctly if all correct options
are marked only?

Distribution of probability of sums on two dice
0,000
0,050
0,100
0,150
0,200
2 3 4 5 6 7 8 9 10 11 12
possible sums
p
r
o
b
a
b
i
l
i
t
y

53

The probability that a question is answered correctly under
these circumstances is the same for every question and
equals
,
_
r
k
p
1
. The table shows the respective
probabilities for given k and r. The expected number of
correctly answered questions is p n EX . . The example of
100 questions of six options out which three are always
correct (which can be considered a reasonable compromise)
immediately negates the objection on guessing.

Reading specifications
Reading and understanding specifications of products is an important part of everyday company life. Let us e.g.
consider a test with the following specifications:
Point span: 0 100; minimum points to pass the test: 50; random variable X, which describes the results of the test, has
normal distribution, with expected value 62 and dispersion 25
2
.
It is obvious that such a test is useless, since if X~No( ,
2
), we have 9973 . 0 ) 3 3 (
+ < < X P , which

means that failing the proposed test is almost impossible because in our case 9973 . 0 ) 77 47 (
< < X P .

Let us consider another simple example:
Random variable X describing the time before a problem with a machine occurs has exponential distribution of
probability. The problem occurs once in H hours. We denote T the operating time before the problem occurs for the first
time and p probability that the machine works without a problem for longer that T. For H=2000 and p=0.99 we get that
T is as short as 20 hours.

This could make us reject such a machine since 20 hours of non-problematic operating time is indeed not acceptable.
Yet with some background knowledge of probability theory we could easily object to such reasoning since with p
decreasing T rapidly increases. The general formula for counting T is in our case of exponential distribution of
probability p H T ln , which follows from p T X P ) ( , where ) 1 ( ~ Exp X describes the time before the problem
occurs in case that the problem occurs once in H hours
2
. The following table gives values of T for some values of p and
H.
H 2000
p 0,99 0,98 0,97 0,96 0,95 0,94 0,93 0,92 0,91 0,90 0,85 0,80
T 20,1 40,4 60,9 81,6 102,6 123,8 145,1 166,8 188,6 210,7 325,0 446,3

H 3000
p 0,99 0,98 0,97 0,96 0,95 0,94 0,93 0,92 0,91 0,90 0,85 0,80
T 30,2 60,6 91,4 122,5 153,9 185,6 217,7 250,1 282,9 316,1 487,6 669,4

H 4000
p 0,99 0,98 0,97 0,96 0,95 0,94 0,93 0,92 0,91 0,90 0,85 0,80
T 40,2 80,8 121,8 163,3 205,2 247,5 290,3 333,5 377,2 421,4 650,1 892,6

Comparing results
Comparing data is a necessity in a number of situations. However, wrong conclusions are often drawn from comparing
the incomparable. Let us consider the following data sets with the same span of possible results, where
} 20 , 19 ,..., 11 , 10 {
i
x :
Set A:
i
x : 12, 12, 12, 13, 14, 14, 15, 15, 15, 16, 18, 19, 20

2
Naturally, another piece of knowledge would be necessary to support or question the fact that exponential distribution
is used in this respect.
number of options (k)
3 4 5 6 7 8
1 0,33 0,25 0,20 0,17 0,14 0,13
2 0,33 0,17 0,10 0,07 0,05 0,04
3 0,25 0,10 0,05 0,03 0,02
4 0,20 0,07 0,03 0,01
5 0,17 0,05 0,02
6 0,14 0,04
n
u
m
b
e
r

o
f

c
o
r
r
e
c
t

o
p
t
i
o
n
s

(
r
)

7 0,13

54

Set B:
i
x : 10, 10, 11, 11, 11, 13, 16, 17, 18, 19, 19, 20, 20

Value 18
i
x occurs in both data sets. The average value for both data sets is 15 x , therefore for both data sets we
have that 18
i
x is better than the average. Yet after we employ the standardized z-values, which take into account
both average and dispersion, we get that in set A the respective z-value is 22 , 1
16
z while for set B we have
77 , 0
16
z , which means that value 18
i
x is relatively much better in set A even though its distance from the (same)
average is the same in both sets.

The issue of statistical testing
Since z-tests only are dealt with throughout the course in the combined form of study at FEEC BUT, which is referred
to in [6], let us consider only this type of statistical tests. For the sake of simplicity let us deal with the test
constant. The mathematical background of this test is very simple: once we set the hypotheses, we are looking for
such
k
x that
2
) (

k
x X P , where X~No( ,
2
) and is the given significance level. The
k
x can be easily
obtained with the help of the respective z-value, since the parameters and
2
(thus also ) are known.
This simple background can be applied in a number of various situations. Typically:
We know that the operating time of a machine (random variable X) can be described by the normal distribution of
probability as X~No( ,
2
). We are going to test a new technique designed to increase the operating time. Some
machines have been enhanced with the technique and their operating time measured. Given the significance level
what conclusion on the quality of the technique can be drawn from the values?

Statistical testing can be easily abused in order to manipulate the recipient into accepting wrong conclusions. This is
especially true for the choice of the significance level and the misinterpretation of constant and > constant or
< constant tests. Given suitable significance level and disregarding the nature of the test almost any hypothesis may
seem acceptable.

Conclusion
Knowledge of probability theory and statistics is an integral part of responsible decision making in many aspects of
company routine. It is often believed that these theories are too abstract or too complex to be used by non-trained staff
or without specialised software. The above contribution does not challenge this assumption, which is naturally valid in a
great many contexts. It rather complements the idea by suggesting that there exist real-life situations which can be
solved by a surprisingly modest amount of knowledge of probability theory.

References:
[1] BATINEC, J. Matematika pro bakale na FEKT VUT. In Matematika na vysokch kolch. Praha : 2003.
[2] BATINEC, J.; DIBLK, J. Vuka matematiky v magisterskm studiu na FEKT VUT. In XXIII International
Colloquium on The Acquisition Process Management, Sbornk abstrakt a elektronickch verz pspvk na
CD-ROM. Brno : Univerzita obrany, 2005.
[3] CASELLA, G.; BERGER, R. L. Statistical Inference, 2
nd
ed. Duxbury Thompson Learning, 2002.
[4] FAJMON, B.; RIKOV, I. Matematika 3. Brno : UMAT FEKT VUT, 2003, available from
https://www.feec.vutbr.cz/et/skripta/umat/Matematika_3_S.pdf
[5] FOTR, J.; DDINA, J. Manaersk rozhodovn. Praha : Vysok kola ekonomick, Fakulta
podnikohospodsk, 1994.
[6] NOVK, M. Probability theory and statistics in the combined form of study of the bachelor student
programmes at FEEC BUT. This proceedings.
[7] ZAPLETAL, J. Poznmka k rozhodovn za rizika a nejistoty. This proceedings.
[8] ZAPLETAL, J. Zklady potu pravdpodobnosti a matematick statistiky. Brno : PC-DIR, 1995.

55

Address:
Mgr. Michal Novk, Ph.D.
stav matematiky,
FEKT VUT v Brn
Technick 8
616 00 Brno
tel.: +420-541143135

Address:
RNDr. Mgr. Betislav Fajmon, PhD.
stav matematiky,
FEKT VUT v Brn
Technick 8
616 00 Brno
tel.: +420-541143135

56

57

OPTIMIZATION OF MATERIAL CHARACTERIZATION BY ADAPTIVE TESTING
Gbor Vrtesy
1
, Ivan Tom
2
, Istvn Mszros
3

1
Research Institute for Technical Physics and Materials Science Hungarian Academy of Sciences,
2
Institute of Physics, Academy of Sciences of the Czech Republic
3
Department of Materials Science and Engineering, Budapest University of Technology and Economics

Abstract: A new procedure, called Adaptive Testing was applied for non-destructive characterization of
cold-rolled austenitic stainless steel samples. The flat samples were magnetized by an attached yoke, and
sensitive, reliable descriptors of their plastic deformation strain were obtained from the proper evaluation,
based on the measurements of series of magnetic minor hysteresis loops, without magnetic saturation of
the samples. The results were compared with the results of conventional, reference measurements.
Significant increase of sensitivity was found if Adaptive Testing was applied.

Keywords: Adaptive testing, nondestructive material evaluation, material parameter optimization

1. Introduction
Magnetic measurements are frequently used for characterization of changes in structure of ferromagnetic materials,
because their magnetization processes are closely related to the microstructure of the materials. This fact also makes
magnetic measurements an evident candidate for non-destructive testing, for detection and characterization of any
modification and/or defects in materials and in products manufactured from such materials [1,2]. Majority of traditional
magnetic investigations of variation of structural material properties simply make use of several parameters of the
saturation-to-saturation major hysteresis loop (coercive force, remanent induction, saturation magnetization,
permeability), see e.g [3,4]. These traditional parameters were very suitably established for general account of magnetic
properties of ferromagnetic samples, but they were never optimized for magnetic reflection of various structural
properties of the measured specimens and for their current alterations.

An alternative, sensitive and experimentally friendly approach to this topic, the Adaptive Testing (AT) method,
suggests a procedure of accumulation of data on the selected physical process, whose parameters are systematically
modified in as broad ranges of values as to get the most complex picture of the behavior. Next analysis of the recorded
data leads to a large family of degradation curves, i.e. of potential calibration curves, the most satisfactory of which is
then picked up by a software algorithm as the optimally adapted calibration curve for next tests of unknown samples of
the inspected material altered in the expected way. Based on the magnetic minor loops measurement the method of
Magnetic Adaptive Testing (MAT) was considered recently in [5] and [6]. MAT introduced general magnetic
descriptors to diverse variations of non-magnetic properties of ferromagnetic materials, optimally adapted to the just
investigated property and material. According to this method the sets of minor hysteresis loops are scrutinized, and
sensitive descriptors of the property variation of the material are identified.

In this work an example of application of AT is given. We describe an experimental search for the calibration
curve/curves, best adapted to magnetic examination of a particular steel material, subjected to cold rolling. Next
testingof any unknown sample of the same kind of steel would be expected to indicate the level of the currently applied
plastic strain through magnitude of the chosen magnetic feature of the material.

2. ADAPTIVE TESTING OF COLD ROLLED AUSTENITIC STEEL
Titanium stabilized austenitic stainless steel, 18/8 type, was studied. Stripe-shaped specimens were annealed at 1100 C
for 1 hour. Then they were quenched in water in order to prevent any carbide precipitation, and to achieve
homogeneous austenitic structure as the starting material structure. The as-prepared stainless steel specimens were cold-
rolled at room temperature to different strains (from 33 to 63% strain in the case of the investigated specimens). For the
reference measurement, a major (saturation-to-saturation) magnetic hysteresis loop was taken from each sample.

A specially designed Permeameter [6] with a magnetizing yoke was applied for measurement of families of minor loops

58

differential permeability of the magnetic circuit. The magnetizing coil wound on the ferrite yoke gets a triangular
waveform current with step-wise increasing amplitudes and with a fixed slope magnitude in all the triangles. This
produces time-variation of the effective field, h
a
(t), in the magnetizing circuit and a signal is induced in the pick-up coil.
As long as h
a
(t) sweeps linearly with time, the voltage signal U(h
a
,h
b
), in the pick-up coil is proportional to the
differential permeability, (h
a
,h
b
), of the magnetic circuit
t h h h h B const h h U const h h
a a b a b a b a
/ * / ) , ( * ) , ( * ) , (
Permeameter works under full control of a PC computer, which sends the steering information to the function generator,
and collects the measured data. An input-output data acquisition card accomplishes the measurement. The computer
registers data-files for each measured family of the minor permeability loops, corresponding to each measured
sample.

The experimental raw data are processed by a data-evaluation program, which divides the originally continuous data of
each measured sample into a family of individual permeability half loops. Then the family, either of the top half-loops
or the bottom half-loops or their average is chosen for next processing. The program filters experimental noise and
interpolates the experimental data into a regular square grid of elements,
ij
(h
ai
,h
bj
), of a -matrix with a pre-
selected field-step. The consecutive series of -matrices, each taken for one sample with strain, , of the consecutive
series of the more-and-more deformed material, describes the magnetic reflection of the material plastic deformation.

The matrices are processed by a matrix-evaluation program, which normalizes them by a chosen reference matrix, and
arranges all the mutually corresponding elements
ij
of all the evaluated -matrices into a
ij
() table. Each
ij
()-
column of the table numerically represents one
ij
()-degradation function of the material. The matrix-evaluation
program calculates sensitivity of each degradation function and draws their sensitivity map in the plane of the field
coordinates (h
ai
,h
bj
)(i,j). This map shows the relative sensitivity of each
ij
()-degradation function with respect to the
plastic deformation strain, , of the investigated material. Sensitivity of each degradation function is computed as the
slope of its linear regression and it is expressed by a shade in the sensitivity map figure.
-750 -500 -250 0 250 500 750 1000 1250
-250
0
250
500
750
1000
1250
1500
1750
h
a
[A/m]
h
b

[
A
/
m
]

Fig. 1
Map of relative sensitivity of the -degradation functions,
ij
()(h
ai
,h
bj
)(). (The crossing point of the lines indicate
the most sensitive
ij
()-degradation function.)

Permeability matrices of all the samples were calculated from the measured data, and the matrices-evaluation process
was applied to compare sensitivity of all the individual degradation functions,
ij
(), each corresponding to a pair of the
field-coordinates (h
ai
,h
bj
). The sample having the lowest strain was used for the normalization. Fig. 1 shows the map of
the relative sensitivity of the
ij
()-degradation functions. The elements, depicted in the sensitivity map as the whitest,
correspond to the most sensitive
ij
()-degradation functions. The most sensitive element, characterized by h
a
=700 A/m
and h
b
=1200 A/m corresponds to the top of the white area. Its location is shown by the two, crossing perpendicular
lines in Fig. 1. It is also seen from the sensitivity map, that within the whitest region the
ij
()-degradation functions
vary only very slightly, so the neighbouring elements of the chosen (h
a
700, h
b
1200) provide practically the same value.
This makes the choice of the proper, sensitive descriptor to be very reliable. The most sensitive
ij
()-degradation
function is shown in Fig. 2. The B
max
values, which were determined from reference measurement on the major loops,
are also indicated in the figure.

59

30 35 40 45 50 55 60 65
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
5.5

d
e
g
r
a
d
a
t
i
o
n

f
u
n
c
t
i
o
n
s
Plastic strain [%]
h
a
700,h
b
1200
Reference

Fig. 2
The most sensitive
ij
()-degradation function and the result of the reference measurement

Integrating the permeability along the field h
a
, hysteresis loops and hysteresis loop B-matrices can be obtained. The B-
matrices contain the same information as the matrices, however, presentation of the -dependences of the
corresponding B
ij
()-degradation functions is different and sometimes advantageous. After the same procedure of the
matrices-evaluation and the corresponding normalization, all the B
ij
()-degradation functions for 0h
bj
2000 A/m, -
h
bj
h
aj
+h
bj
, are shown in Fig. 3. It is worth of mentioning, that all the B
ij
()-degradation functions show the same
shape-type of dependence on plastic deformation.
30 35 40 45 50 55 60 65
-200
-150
-100
-50
0
50
100
150
200
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F
a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6
A B C D
E F
a b c d e f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6
A B C D E F
a b c d e f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6
A B C D E F
a b c d e f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6
A B C D E F
a b c d e f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6
A B C D E F
a b c d e
f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6
A B C D E F
a b c d e f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6
A B C D E F
a b c d e f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6
A B C D E F
a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6
A B C D E F
a b c d e f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f
1 2 3 4 5 6
A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6
A B C D E F
a b c d e f
A B C D E F
a b c d e f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5
6
A B C D
E
F
a
b c
d
e
f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4
5
6
A
B C D
E
F
a
b c
d
e
f
A B C D
E
F
a
b c d
e
f
A B C D E F
a b c d e f
A B C D E F
a b c d e f
A B C D E F
a b c d e f
1
2
3
4
5
6
A
B C D
E
F
a b c d e
f
A B C D E F
a b c d e f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f
1 2 3 4
5
6
A
B C D
E
F
a
b c
d
e
f
A B C D E F
a b c d e f
A B C D E F
a b c d e f
A B C D E F
a b c d e f
A B C D E F
a b c d e f
A B C D E F
a b c d e f
A B C D E F
a b c d e f
1 2 3 4 5 6 A B C D E F a b c d e f 1 2 3 4 5 6 A B C D E F a b c d e f
B
-
d
e
g
r
a
d
a
t
i
o
n

f
u
n
c
t
i
o
n
s
Plastic strain [%]

Fig. 3
B
ij
()-degradation functions, as functions of the plastic strain.

The sensitivity map for the B
ij
()-degradation functions is presented in Fig. 4. In contrast to Fig. 1, here the white (and
also black) areas indicate those regions where the matrix elements varies magnitudes vary rapidly, jumping from one
element to the neighbouring one. For instance, elements corresponding to the steepest slopes in Fig. 3 are located in the
black and white areas of Fig. 4. These descriptors are very sensitive, but their reliability is questionable, because
moving from one element to the neighbouring one, high jumps of values can happen (even from a large positive value
to a large negative one). Because of this reason, the choice of descriptors from the homogeneously gray areas seems to
be the most reliable one. The optimal choice of the B
ij
()-degradation functions are shown in Fig. 5.

60

-600 -400 -200 0 200 400 600
600
800
1000
1200
1400
1600
1800
h
a
[A/m]
h
b

[
A
/
m
]

Fig. 4
Map of relative sensitivity of the B-degradation functions, B
ij
() B(h
ai
,h
bj
)().
30 35 40 45 50 55 60 65
1
2
3
4
5
6
7
8
B
-
d
e
g
r
a
d
a
t
i
o
n

f
u
n
c
t
i
o
n
s
Plastic strain [%]
h
a
100h
B
1000

Fig. 5
The optimal choice of the B
ij
()-degradation functions.

A third type of matrices (-matrix) can also be obtained. This is the matrix of the derivative of permeability with
respect to the field, h
a
, (the first derivative of permeability, /h
a
, or the second derivative of magnetic induction,
2
B/h
a
2
). The sensitivity map of the
ij
()-degradation functions is shown in Fig. 6. The optimal
ij
()-degradation
function (a compromise between sensitivity and reliability, as explained above) is shown in Fig. 7.

61

-200 0 200 400 600 800 1000 1200
0
200
400
600
800
1000
1200
1400
1600
h
a
[A/m]
h
b

[
A
/
m
]

Fig. 6
Map of relative sensitivity of the
ij
()-degradation functions.

30 35 40 45 50 55 60 65
0
2
4
6
8
10
12
14
'

d
e
g
r
a
d
a
t
i
o
n

f
u
n
c
t
i
o
n
s
Plastic strain [%]
h
a
450,h
b
950

Fig. 7
Optimal choice of
ij
()-degradation functions, as a function of plastic strain.

3. DISCUSSION
As it was already mentioned the originally paramagnetic austenite specimens became more and more ferromagnetic, as
a consequence of the applied cold-rolling. All austenitic stainless steels are paramagnetic in the annealed, fully
austenitic condition, and the only magnetic phase, which can be induced (e.g. by cold-rolling) in the low carbon
austenitic stainless steels, is the bcc -martensite, which is highly ferromagnetic. This process can be followed easily
by magnetic measurements.

By applying the above described adaptive testing method, the relatively small difference between the magnetic
characteristics of the investigated sample series can be determined much more sensitively, than by the conventional
methods. By using the
ij
()-degradation functions derived from the permeability matrices, it is possible to increase the
sensitivity of the determination of the appearance of ferromagnetic -martensite by about a factor 3, as compared with
the classical major loop approach, if proper descriptors are used. (For comparison see Fig. 2.) The reliability of this
determination is very good, which is illustrated by the sensitivity matrix. Here a wide plateau is seen, where sensitivity
of the
ij
()-degradation functions is varied only very slightly, if their field coordinates are miss-positioned. In other
words, the sensitivity (and also the reliability) of the measurement is not influenced by the exact choice of the matrix
element.

62

The sensitivity with respect to the applied strain is increased, if the hysteresis loop B-matrix parameters are considered
instead of permeability matrix. The sensitivity map shows the area, where the most sensitive and/or most reliable
descriptors are found. If the descriptors are very carefully chosen, more than 1:7 ratio can be obtained between the less
and the most ferromagnetic piece of the investigated samples. The most sensitive area in the B
ij
-sensitivity map is not as
large (there is not such a wide plateau), as in the case of sensitivity of the
ij
()-degradation functions, but there is a
well defined area, where the most sensitive descriptors are positioned.

Scatter of the parameters is very low, as can be seen in Fig. 3. It is possible to increase the sensitivity: more than 1:100
ratio can be reached, but in this case the exact choice of matrix elements becomes crucial. The shapes of the
dependences of all the B
ij
-matrix elements vs. plastic deformation are very similar to each other. So, the descriptors,
which were evaluated from the hysteresis loop B-matrices, seem to be especially suitable for very sensitive
characterization of the changes, which are introduced by the cold rolling in the austenitic material.

It is possible to increase the sensitivity even more substantially, and at the same time to avoid the possible mistake of
any miss-positioning by choosing the first derivative of permeability (-matrix) as the source of descriptors. In the case
of -matrices, the largest sensitivity can be obtained if the most reliable area is taken. For the presented series of
samples it is 1:14 (see Fig. 7). On the other side, the scatter of the -matrix elements is the largest, if we take into
account all the elements, but even in this case an area of the elements can be found, from where reliable enough
elements can be taken.

4. CONCLUSIONS
This paper was dedicated to the indirect experimental measurement of variable material properties, in particular to the
question how from all available features of the chosen physical process (here from the process of magnetization of the
samples by an external field), which was employed for description of the investigated material variation to determine
that one, which is the best adapted to the particular material under inspection, to the particular variation/degradation of
the material, and to the particular demands declared by the examiner.

The introduced way of the Adaptive Testing of materials suggests a procedure of collection of data on the selected
physical process, whose parameters are systematically modified in as broad ranges of values as to get the most complex
picture of the behavior. Next analysis of the recorded data leads to a large family of degradation functions, the most
satisfactory of which is then defined as the optimally adapted calibration curve for next tests of unknown samples of
the investigated material altered in the expected way.

As shown in the presented experimental example, an optimum selection among the degradation functions has to take
into consideration not only the desired high sensitivity of the calibration curve, but needs to demand also low
experimental error of the curve-constituting measured values and low curvature of the sensitivity surface (referred to as
stability of the calibration curve) around the selected AT-coordinates point.

The presented example obviously demonstrated that AT, focusing on the explored concrete material and the explored
concrete degradation, introduces kind of a trade involving sensitivity, stability, smoothness, shape, and experimental
friendliness of the optimum calibration curves. This focus usually leads to excellent results of AT, which are
substantially more advantageous for description of the explored material variations than the plain use of the traditional
descriptors, which are focused on the employed physical process itself.

ACKNOWLEDGEMENTS
The financial support by the Hungarian Scientific Research Fund (T-035264 and T-062466) and by the Academy of
Science of the Czech Republic (projects No. 1QS100100 and AVOZ 10100520) is appreciated.

REFERENCES
[1] JOHNSON, M. J.; LO, C. C. H.; ZHU, B.; CAO, H.; JILES, D.C.; Nondestruct. Eval. 20 (2000) 11.
[2] JILES, D.C. Magnetic methods in nondestructive testing. K. H. J. Buschow et al., Ed., Encyclopedia of
Materials Science and Technology, Oxford : Elsevier Press, 2001. p. 6021.
[3] JILES, D.C. NDT Int. 21 (1988) 311.
[4] DEVINE, M. K. Min. Met. Mater. (JOM) (1992) 24.
[5] TOM, I.; Magn. Magn. Mat. 268 (2004) 178.
[6] TOM, I.; PEREVERTOV, O. JSAEM Studies in Applied Electromagnetics and Mechanics. T. Takagi and
M. Ueasaka (Eds.), 9 (2001) 533.

63

Address:
Dr. Gbor Vrtesy, Dr.Sc.
Research Institute for Technical Physics and Materials Science
Hungarian Academy of Sciences,
H-1525 Budapest,
P.O.B. 49, Hungary
phone:+3613922677
fax: +3613922226
e-mail: vertesyg@mfa.kfki.hu

Address:
RNDr. Ivan Tom, CSc.
Institute of Physics, Academy of Sciences of the Czech Republic
Na Slovance 2,
18221 Praha, Czech Republic
phone: +420266052177
fax: +420286890527
e-mail: tomas@fzu.cz

Address:
Dr. Istvn Mszros
Department of Materials Science and Engineering,
Budapest University of Technology and Economics,
H-1111 Budapest,
Goldmann ter 3, Hungary
phone: +3614632883
fax: +3614633250
e-mail: meszaros@eik.bme.hu

64

65

VYUIT KOMPLETNHO GENETICKHO ALGORITMU PRO EEN OPTIMALIZACE
VROBNHO PROCESU Z HLEDISKA MAXIMALIZACE ZISKU
Ji Kostiha

VUT Brno

Abstrakt: lnek obsahuje popis kompaktnho genetickho algoritmu, jeho specifika a odlinosti od
klasickho genetickho algoritmu. Soust je ukzka efektivnosti na standardn testovac funkci a aplikace
algoritmu na konkrtn ekonomickou optimalizan lohu s kolem maximalizace zisku vrobnho procesu.

Klov slova: kompaktn, genetick, algoritmus, evolun, optimalizace, vcerozmrn, problm, een,
maximalizace, zisk, vrobn, proces

vodem
Evolun algoritmy (EA) jsou svm principem zaloeny na technick interpretaci biologickch dj. Zkladn
mylenkou popsanou Darwinem je pevn nejsilnjch jedinc. Bhem ivota se jedinci mezi sebou k, mutuj a
tm vznikaj nov jedinci, kte nsledovn vstupuj do procesu vbru a dalho ken. Nejlpe vybaven jedinci
k ivotu maj nejvt anci ke ken, zatmco slab jedinci maj malou anci. Tm je zarueno velmi pravdpodobn
pokraovn genetickch vtv silnch jedinc, zatmco vtve slabch jedinc se ukonuj. Na zklad principu EA se
daj nalzt jejich spolen rysy:
Inspirovny prodnmi evolunmi procesy.
Pracuj s populac jedinc.
Iteran charakter pstupu k een. Jeden iteran cyklus je dn jednou populac.
Stochastick a heuristick charakter.
Poskytuj een, kter nemus bt optimln, ale je vhodn. een je poskytnuto v krtkm ase.

Genetick algoritmy (GA) jsou dosud nejspnj z evolunch algoritm vbec. Pro svou efektivitu a velmi dobr
vsledky nachzej uplatnn v mnoha odvtvch lidsk innosti. Pouvaj se napklad pi vytven rozvrh prce pro
stroje v tovrnch, v teorii her, v ekonomii managementu, pro een optimalizanch problm multimodlnch funkc,
pi zen robot, v rozpoznvacch systmech a v lohch umlho ivota. Dle genetick algoritmy nachzej
uplatnn pi een tzv. NP-plnch problm, kde tm vechny ostatn algoritmy selhvaj, tj. kde vpoetn as je
exponenciln nebo faktoriln zvisl na potu promnnch.

Za svou dobu existence vzniklo mnoho modifikac EA a stle se vyvj nov. Kompaktn genetick algoritmus (CGA) je
speciln ppad algoritmu. ad se do skupiny GA, avak je postaven na zcela odlinm zklad. Hlavn rozdly jsou v
technice generovn populace a rekombinanch opertorech.

Obrzek 1: Hierarchie evolunch algoritm

Srovnn CGA a klasickho GA
Principem klasickho GA i CGA, stejn jako vech EA je een loh, kter se daj popsat tzv. fitness funkc a hledn
jejich globlnho extrmu. Podle typu lohy me jt o minimalizaci nebo maximalizaci. Fitness funkce se skld z
vlastn funkce a omezujcch podmnek. Funkce i podmnky jsou charakteristick pro dan problm a je nutn je
vytvoit podle zadn konkrtnho problmu.
Evolun algoritmy
Genetick algoritmy
Klasick genetick algoritmus
Kompaktn genetick algoritmus

66

Zkladem GA jsou chromozmy, kter v matematick interpretaci odpovdaj jedincm, nkdy tak nazvanmi
agenty. Jedinci tvo populaci, kter se mn s pibvajcmi generacemi. Populace reprezentuje informaci o stech
prohledvanho prostoru, kter byly v dosavadnm vpotu vzorkovny. Standardn pouvan opertory ken,
mutace a selekce potom uruj, jak m bt s touto informac naloeno pi generovn dalch potenciln slibnch
een. Tyto opertory tvo zklad klasickho GA.

CGA algoritmus pouv pravdpodobnostn popis aktulnho stavu vpotu (model popisujc souasnou populaci).
Opertory ken, mutace a selekce jsou zde nahrazeny vzorkovnm prohledvanho prostoru na zklad danho
pravdpodobnostnho modelu. Jeho zkladem je pravdpodobnostn chromozm (PCh), kter je zapsn jedinm n-
rozmrnm relnm vektorem. Nepracuje se s relnou populac, ale s vektorem, na zklad kterho se v dob vpotu
generuje reln chromozm (jedinec). Kad pozice vektoru vyjaduje pravdpodobnost vskytu hodnoty 1 na dan
pozici chromozomu. Poten stav je dn hodnotou 0,5 na vech pozicch PCh.

Na obrzku 2 je znzornn PCh pro jednoduchou tybitovou lohu o dvou neznmch parametrech reprezentovanch
geny 1 a 2. V prav sti je nejpravdpodobnj binrn hodnota generovanch jedinc. ern sla pedstavuj
ustlen hodnoty - ji se nemohou mnit. ed neustlen - mohou se jet v prbhu vpotu mnit.

Obrzek 2: Pravdpodobnostn chromozm

Vvoj vpotu spov v modifikaci PCh podle vygenerovan populace njakm vhodnm zpsobem, aby dochzelo k
pibliovn hodnot generovanch jedinc ke globlnmu extrmu. Postup vpotu klasickho GA a CGA je na obrzku
3.

Obrzek 3: Srovnn klasickho genetickho algoritmu a) a kompaktnho genetickho algoritmu b)

Modifikace PCh se provd podle uritch pravidel, kterch me bt cel ada. Pouit pravidlo v rmci zde
uvedench vsledk je nsledujc:
vyber nejlepho a nejhorho jedince z populace
jestlie se bity jedinc li a u lepho jedince je 1, pak k hodnot bitu PCh na dan pozici piti 0,01 (1%)
jestlie se bity jedinc li a u lepho jedince je 0, pak k hodnot bitu PCh na dan pozici odeti 0,01 (1%)
jestlie maj bity stejnou hodnotu, nedlej nic

Elitismus
Elitismus je technika, kter se pouv k zapamatovn nejlepho, nebo nkolika nejlepch jedinc pedchzejc
populace a jejich automatick zaazen do nsledujc populace. Tm je stoprocentn zarueno peit nejlepch jedinc.
Toto vede v mnoha ppadech ke zlepen efektivnosti algoritmu. Pi aplikaci v CGA m smysl, jak vyplv z principu
algoritmu, zaazovat pouze jednoho nejlepho jedince. V nsledujc generaci se srovnvaj vichni nov vytvoen
jedinci tj. cel nov populace a elitn jedinec z pedchoz generace. Vybere se nov nejlep jedinec a je oznaen za
elitnho. Modifikace bit PCh se provd podle nejlepho a nejhorho jedince, piem nejlep jedinec v tomto
ppad odpovd elitnmu jedinci.

0 0,8 0,6 1 0,3 0,5 0,2 1
Gen 1 Gen 2
1 1 1 0 1 0 1
Pravdpodobnostn chromozm Reprezentace PCh chromozmu v binrnm tvaru
Gen 1 Gen 2
0
populace
ken
mutace
selekce
pravdpodobnostn chromozm (PCh)
populace
modifikace PCh
a) b)

67

Stabilizan hranice
Stabilizan podmnka, viz. podmnky ukonen vpotu, me vst k pedasnmu ustlen v loklnm extrmu nebo
na njak hodnot dan pedasnm ustlenm nkterho nebo nkterch bit. V takovch ppadech me dochzet k
zskvn nesprvnch hodnot nejen pi uvznut algoritmu v loklnm extrmu, ale tak k uvznut "nkde" na funkci,
zapinn ustlenm bitem, kter ji neme zmnit svou hodnotu. Proto jsem algoritmus doplnil o stabilizan
hranice, zobrazen na obrzku 4. Interval b znzoruje modifikan oblast, ve kter se mou pohybovat hodnoty PCh.
Naopak intervaly a na okrajch jsou zakzny. K stabilizaci potom dochz na tchto hranicch. Je tm stle zaruena
pravdpodobnost generovn jedinc s opanou hodnotou bitu a monost peklopen a ustlen bitu k opan hodnot.
Takto je zamezeno mon pedasn nedouc stabilizaci bit PCh.

Obrzek 4: Stabilizan hranice

Podmnky ukonen vpotu
Podmnek ukonen vpotu me bt mnoho. U klasickho GA jsou dv zkladn. Ukonen pi doshnut zadanho
potu generac a ukonen pi doshnut zadan pesnosti vpotu. Druh podmnka se tk pedevm testovacch
funkc, kde znme hodnotu hledanho extrmu. U CGA pibv jet jedna dleit podmnka a to ukonen pi
stabilizaci hodnot PCh na hodnotch 0 a 1. Tato podmnka je zajmav tm, e algoritmus jakoby sm pozn kdy doel k
extrmu a ji neme vygenerovat jinou hodnotu. Vypotan hodnota nemus bt pmo hodnota globlnho extrmu,
toto zle na sloitosti problmu, ale je to njak vhodn hodnota, kterou lze povaovat za velmi dobr een
zkoumanho problmu. Pi stabilizaci je douc ukonen algoritmu, protoe ji neme dojt k nalezen lep hodnoty
a vichni generovan jedinci jsou toton a odpovdaj stabilizovanmu PCh. Pi pouit stabilizanch hranic toto
neplat, ale pravdpodobnost, e bude nalezeno lep een je velmi mal a algoritmus je dle neefektivn.

Testovac loha
Pro testovn CGA jsem pouil standardn testovac funkci Rastriginovu, zobrazenou na obrzku 6. Vhodou tto
funkce je stejn zpis pi pouit libovolnho potu neznmch parametr tzn. testovn je mon na libovoln
rozmrnosti bez nutnosti zsahu do matematickho zpisu tto funkce. Dal vhodnou vlastnost funkce je poloha
globlnho extrmu v bod nula na vech osch pi vech rozmrnostech.

Parametry:
interval -5,12 a +5,12 na kad ose
16-ti bitov kdovn gen
maximln mon pesnost 0,00015625 pi pouitm kdovn a velikosti intervalu
5
n
1 loklnch extrm, n = poet dimenz
vsledn hodnoty potebnho potu generac k nalezen globlnho minima prmrovny aritmeticky z 10 een
pouit vpoetn prostedek: pota s procesorem AMD Athlon 1600+

Na obrzku 5 je vidt zvislost potu generac na velikosti populace. Tenk ra nahoe znzoruje een s nejvtm
potem potebnch generac k nalezen minima a naopak tenk ra dole s nejmenm potem potebnch generac.
Tlust ra vyznauje aritmetick prmr z 10 een. Krok zvyovn populace na ose x je 10. Jedno een je
dosahovno piblin bhem jedn sekundy.

0 1
a
a b

68

Obrzek 5: Zvislost potu generac potebnch pro nalezen globlnho minima Rastriginovi funkce na potu jedinc v
populaci pro dva neznm parametry

Vliv elitismu
U testovac lohy se dvma a temi neznmmi parametry je patrn pozitivn vliv pi pouit mench populac, dov
do 40 jedinc. S rostouc rozmrnost problmu, tj.zvyovn potu neznmch parametr, se elitismus ukazuje jako
nevhodn pro zefektivnn algoritmu. Prmrn mnostv generac potebn pro nalezen globlnho een je vt.
Toto je zpsobeno astou konvergenc k loklnmu extrmu a uvznut v nm.

Vliv stabilizan hranice
S pouitm stabilizanch hranic u tohoto problmu nedochz pli ke zlepen prmrnho potu potebnch generac
k nalezen globlnho minima. Dochz vak k vt spolehlivosti doshnut lepho een a omezuje se ast uvznut
algoritmu. Jako vhodn hodnota stabilizan hranice se ukazuje mal slo v du procent.

69

Obrzek 6: Standardn Rastriginova testovac funkce pro dva neznm parametry, a) ez rovinou xz v y = 0, b) nhled
na rovinu xy ze shora, c) , d) 3D nhled

70

Pklad een lohy maximalizace zisku vrobnho procesu pomoc CGA
Zadn
V podniku se vyrb pt druh vrobk (A, B, C, D, E) ze t druh materilu (S1, S2, S3). Materil je k dispozici pro
plnovan obdob v omezenm mnostv: 1500kg S1, 300kg S2, 450kg S3. Vrobn procesy jednotlivch druh
vrobk probhaj nezvisle. Jin omezen nepichz v vahu. Spoteba materilu na jeden kus vyrbnch druh
vrobku (kg na 1kus) a velkoobchodn ceny jednotlivch vrobk jsou nsledujc:
Vrobek
A B C D E
S1 - 0,4 0,3 0,6 0,6
S2 0,05 0,2 0,1 0,1 -
Spoteba (kg na 1kus) materilu:
S3 0,1 0,2 0,2 0,1 0,2
Velkoobchodn cena (K na 1kus) 20 120 100 140 40
Tabulka 1: Zadn lohy

kol: Vypotejte kolik je teba vyrobit jednotlivch druh vrobk pi danch omezench, aby v plnovanm obdob
bylo dosaeno maximlnch treb za plnovan obdob.

Sestaven ekonomicko matematickho modelu
Funkce maximalizace zisku:
max 1 2 3 4 5 1 2 3 4 5
( , , , , ) 20 120 100 140 40 z x x x x x x x x x x + + + +

Omezujc podmnky:
1 2 3 4 5
1 2 3 4 5
1 2 3 4 5
S1: 0 0, 4 0, 3 0, 6 0, 6 1500
S2: 0, 05 0, 2 0,1 0,1 0 300
S3: 0,1 0, 2 0, 2 0,1 0, 2 450
x x x x x
x x x x x
x x x x x
+ + + +
+ + + +
+ + + +

een
Pouil jsem 16 bitov kdovn gen z testovac lohy, kter je svm rozmezm hodnot (0 a 65535) pro jednotliv
geny vce ne dostaujc. Funkce z
max
pedstavuje fitness funkci, do kter se pidaj omezujc podmnky s postihy za
nedouc een.

Z deseti provedench een oznail algoritmus hodnoty v tabulce 2 jako optimln a na nich tak setrval. Devadest
procent hodnot bylo oznaeno bhem prvnch 1000 generac a do 10 sekund na potai s procesorem AMD Athlon
1600+. Prmrn hodnota spotanho een je 379674 K. Optimln een odpovd 380000 K.

slo
een
hodnota oznaenho
maximlnho zisku [k]
1 379900
2 379400
3 379680
4 379560
5 379960
6 380000
7 379900
8 379880
9 378820
10 379640
Tabulka 2: Hodnoty spotan CGA

71

Zvrem
CGA, stejn jako ostatn EA poskytuj v pomrn velmi krtkm ase een, kter je nebo se bl k optimlnmu.
Pouit algoritmu je vhodn tam, kde ostatn algoritmy selhvaj nebo nejsou schopn podat een v dostaten krtkm
ase. Piem nen zcela nutn znt nejlep een a postauje znalost nkterho z velmi dobrch een. Tomuto
charakteru loh odpovdaj ekonomick problmy, proto je pouit algoritmu v tchto lohch vhodn.

Pouit zkratky a pojmy
CGA Compact Genetic Algorithm (Kompaktn genetick algoritmus), speciln ppad GA, kter je
zaloen na PCh
EA Evolutionary Algorithm (Evolun algoritmus), matematick algoritmy inspirovan prodou
Elitn jedinec Nejlep jedinec za celou dobu vpotu, pop. sti vpotu
Fitness Sla jedince, vyjaduje mru vhodnosti een danho jedince
Fitness funkce Funkce popisujc zadanou lohu, charakteristick pro danou lohu, hled se na n globln
extrm odpovdajc optimlnmu een lohy
GA Genetic Algorithm (Genetick algoritmus), skupina algoritm spadajc pod EA
Gen Prvek vektoru parametr, zkladn stavebn jednotka chromozomu, odpovd zakdovan
promnn
Generace Krok iteranho cyklu pi hledn optimlnho een pomoc GA
Chromozm Genetick informace ve form etzce, skld se z gen, v matematick interpretaci odpovd
jedinci
Jedinec (Agent) Nositel genetick informace, jedinci tvo populaci, v matematick interpretaci odpovd
chromozmu
Ken Konstrukce novch jedinc (potomk) dle pvodnch jedinc vybranch z generace (rodi)
Mutace Nhodn zmna hodnot bit v chromozmu
PCh Probability Chromozome (pravdpodobnostn chromozm), zvltnost CGA, na zklad PCh se
konstruuj jedinci populace, podle danch pravidel je zptn modifikovn
Populace Mnoina jedinc v danm kroku iteranho cyklu zvanm generace
Potomek Jedinec, jen je vsledkem rekombinace rodie()
Rekombinace Proces generovn novho jedince, obvykle odpovd ken a mutaci
Rodi Jedinec vstupujc do rekombinace, vsledkem je(jsou) nov jedinec(i)
Selekce Vbr jedinc do nov populace podle danch pravidel

Literatura
[1] MAK, V.; TPNKOV, O.; LAANSK, J. a kol. Uml Inteligence (4)
[2] PETERKA, I. Genetick algoritmy. Matematicko-fyzikln fakulta Univerzity Karlovy, duben 1999.
[3] KALTOV, E.; DOBI, J. Evolun algoritmy. http://lucifer.fav.zcu.cz/uir/
[4] HARIK, a kol., 1997.

Adresa:
Ing. Ji Kostiha
VUT Brno
Kolejn 4
612 00 Brno
tel.: +420 775 303 543
e-mail: kostiha@email.cz

72

73

REVITALIZING COMPANY INFORMATION SYSTEMS AND COMPETITIVE ADVANTAGES
Branislav Lacko

VUT v Brn

Abstract: The contribution deals with innovation of computer based information system

Key words: Information system, revitalization, innovation of information system, grow of company

1 Introduction
At present, there is hardly any company having no information system working. Nevertheless, we can still read in our
magazines and brochures articles on the information systems design many of those books reasoning extensively the
necessity of introducing information systems. Current problems, however, do not lie in the absence of information
systems in Czech companies. The problem is in the quality of our existing information systems.

From this perspective, it is suitable to talk about the necessity of the information systems innovation. This paper's title
used the word revitalizing instead of the expression "innovating" for the following reasons:
Revitalization of enterprises, which many of our companies have been trying to do must necessarily comprise
revitalizing those enterprises information systems as well
The word "revitalization" covers more precisely the goal of the change which the information systems of our
enterprises should go through

After having been designed and implemented, each information system was handed over to its users in a certain
condition, size, having characteristic features. Since the moment of being handed over for use, it can develop, stagnate
or decline. By the development of an information system we understand improving its qualitative and quantitative
parameters in the course of its use. If those parameters remain unchanged during the system's use, stagnation occurs. In
case of impairing the information system parameters, we call it decline - a degradation of the information system.

The qualitative parameters describe certain characteristic values of the information system, especially of technological,
programme, system and operation type, e.g.: number of computers, external memories capacity, processor operation
speed, etc. The qualitative parameters can be considered similarly, such as: the way of using database concept,
automation degree of tasks, and the like.

For each i-th parameter out of n parameters we can stipulate a so-called increase index r
i,j
for moment t
j
as follows.

We introduce for each parameter its unit value h
i,0
in time t=0 i.e. at the moment of its handover for operation. We will
make it equal one. Then we can determine the value of the increase index for another time period compared to the initial
value of the parameter. Consequently we can easily calculate the average increase index Rj, for a certain moment of
time j. Then we can even make a graph of dependence of the information system development in time by the means of
the average index of the information system development.

What we will call the information system revitalization is a situation when the average index of the information system
development changes by a decisive leap, resulting mainly of qualitative parameters, especially those oriented to the
benefits of the information system and supporting the decisive processes in the enterprise directly.

2 Implication statement on the information system development
The causes of the information system development must be seen in the requirements of the company staff to the
information system. Those are derived from the following facts:
the company size increases
the experience in the information system use grows making thus the demands for the information system to

74

increase as well
the requirements for information increase in consequence of the increasing information barrier in the society
technological and programming means of computing technology develop
new information technologies develop
the theory and practice of the information systems develop
the demands for quality increase generally,
etc.

The development of the enterprise, however, is a factor to be identified as the most important one.
Under the term of development of the enterprise we will understand a situation, in which the size of the enterprise
extends, the production volume grows and profits increase.

We can declare the following statement:
If an enterprise develops, the information system of the enterprise develops as well.

The enterprise development represents an antecedent of implication here, while the information system development
represents a consequent of this implication.

Let us try to comment individual combinations possible of the implication presented. The zero therein represents "false"
and one represents "true" of the respective affirmation.
0 0
An enterprise, which does not develop or even stagnates, has no resources for the information system development. It is
probably managed incorrectly; therefore its management is not probably interested in developing the information
system.
1 1

A developing enterprise must support its successful development by developing its information system to meet the
information demands of its employees and ensure relevant and top-quality information to support decision-making
processes.
0 1

This true combination of the implication says that even a bad performing enterprise can develop its information system.
It may happen in some of the following situations:
An enterprise revealed bad functioning of its information system as an obstruction in its development and decided
to improve it to a necessary level. Unfortunately, this situation comprises even the alternative of investing into the
information system only as "a wonder" process which is automatically to prevent any further stagnation of the
company or even its decline.
A stagnating enterprise makes use of the progress made in the information technologies, improving its system by
various, especially intensifying methods as a consequence of general progress in this sphere.
The milieu of the enterprise (state administration bodies, state legislation, competitors, client demands, public
opinion, etc.) drive the enterprise to develop its information system.
1 0

A false combination signals a situation which may have a negative impact on the enterprise prosperity and further
development by the decline of the stagnating information system which will stop providing support to the management
and employees of the company. It may turn to be one of the factors which at first will cause problems and later might
even become a cause of the company decline.

3 A sentence on the information system development pace
The fact itself that the information system develops is not enough if we want it to support the decision making activities
in the company really well.
We must find out the validity of the following sentence:
An information system must develop in a pace corresponding to the pace of the enterprise development

Similarly, as for the information system development, we can determine an average index of the company development
R` and its parameters for certain moments of time j (the equal ones will make the best results).

75

The pace of development the speed can be expressed in both cases as the first derivation of the function of the
average index increase of the information system development or respectively of the average increase index of the
enterprise depending on time.

It should be applied, that if an information system of suitable size was handed over for use, supporting management of a
company within the scope necessary, the pace of the information system development should be equal or faster
compared to the economic development of the company:

If the pace is slower, the information system development pace lags behind the company development pace, resulting
in a disproportion which may mean problems again, and possibly even stagnation and decline of the company in case of
non-solving such situation.

4 A problem of information systems revolutionary development
In the previous contemplations, gradual growth of both information system and the enterprise parameters i.e. so-called
evolutionary development was assumed. In practice, revolutionary development of the information system or the
enterprise often occurs, or revolutionary growth in both cases, such as an exchange of an existing outmoded computer
for a new and progressive one, two firms merger, increase in production through opening a new-built hall, etc.
Expressed in graphics, this will be manifested by a leap within the respective growth curve (see Fig.1).

R

t
Fig.1

In a book by C. Gray [1], the author explains why the enterprise's growth is a necessary precondition for its successful
existence. It does not mean that the number of employees would have to increase permanently, but after all, the
company must develop still more and more perfect products, decrease its costs, increase profits, etc.

Company stagnation in today's dynamic world of market economy heads towards its breakdown very fast.

At present era of global computerization, it is more important than ever to develop its information system as well.

Current computer producers meet this demand, designing their individual computer models of standard computer lines
in such a way that their efficiency covers the large scope of the performance and external disc memories capacities.
That means the user may gradually extend his computer fluently according to his increasing requirements and financial
capacities, which is very advantageous for him. For the start of his information system implementation, he does not
have to pay unnecessarily for his computer capacity or memory capacity of the external memories capacity, which he
would not use. Moreover, he can schedule the increase in his computer's capacity as his gradually earned financial
resources allow. And he constantly works with the same operational system, within the same database and
communication setting, with the same standard and application programme equipment, valorizing and protecting the
resources invested into the technological equipment and personnel training before. If a computer does not have a
concept like that, the individual modules increase their capacity with certain delay. Figure No.2 shows a solution, where
the user bought a new model having higher capacity immediately after the capacity of the old model exhausted. He pays
unnecessarily for this capacity improvement for the entire period B, before he can really use this potential. Fig. 3 shows
a situation where the user is waiting and does not cover his increased information requirements until they reach the

76

possibilities of a new model.

R

A B C t
Fig.2

Demand
Satisfaction

R

t
A B C
Fig.3

There are losses in both cases, however. In the former case due to unexploited and paid capacity. In the latter, in
consequence of insufficient computer support in steering and decision taking. Fig. 4 demonstrates the course of
information system benefits in individual periods A,B,C. Therefore the companies offering a limited number of
computer models design their parameters to cover mutually. That enables using a more efficient model before it
reaches the limits of its capacity. Such situation is ideal, in which the user works with a supplier ready and willing to
accept the old model back as a counter value during the innovation. If the user does not have an opportunity like that,
the benefits decrease and losses occur caused by the value of the computer discarded.

77

t
A B C

Fig.4

Another situation, much worse, is that of a user who finds out that at the start of the information system implementation
he had decided for a computer having no further capacity extension possibilities, and the new computer offered requires
substantial financial resources to be invested into programmes modifications. Similar situation may happen in case of
communication network exchange, which may comprise cables redesign to achieve higher transmission speeds,
exchange of the still working network cards for more efficient ones together with the adjustment of the operational
system adaptation which must communicate now with the new network programme equipment. The whole event
postpones the possibility of solving the entire problem and due to the overhaul and modifications; the parameters may
even impair for a certain amount of time. The increase curve may run as shown in Fig.5. Sections B and C are critical as
the users requirements are not covered there. In addition, in section C, technical complications often occur, requiring
extra costs. Therefore the costs curve's line is often that shown in Fig. 6. An irregular, immediate increase in costs is
very unfavourable from the company finances point of view, as any financial expert can explain and confirm.
Everybody thus try to avoid such situation; the users should design their concepts of the information system
implementation in such a way that similar situations do not happen.

IS Capacity

t
A B C D

Fig. 5

78

A B C D

Fig.6

"Leap" changes in the information system development may occur for other than technological reasons, of course
(financial situation of the company, changes in the company's information policy, etc.).

Revitalization always represents a leap change in efficiency. Therefore the aforementioned curves must be
contemplated very well in relation to its implementation.

5 Discussion on the information system revitalization goals
It is a common phenomenon in the information technologies field, that modern information technologies and the
application thereof are a motor of innovations within the information systems sphere. In spite of that, it is necessary to
state in relation to the information system s revitalization, that main driving force should be in striving for improving
the information system quality of the with regard to company management support and company processes efficiency
improving so that the competitive abilities of the company can grow.

Implementation of information technologies on itself cannot be considered a competitive advantage anymore. The
reality is that the inability to use modern information technologies effectively to maximum extent turns out to be a
retardation factor in company development leading to the company's competitive abilities impairment.

The goals of information system revitalizing must be conceived in relation to overall improvement in company
processes effectiveness and efficiency (Business Process Reengineering). Information, especially of economic type, is
essential for proper company management. [8, 9] Parallel economic data processing intercepted by modern controlling,
and using the information obtained from those data cohere to the possibility of good company development [7].

6. Conclusion
The knowledge, quoted at the end of the preceding paragraph should be understood as an urgent stress on the
importance of the right information system revitalization strategy choice. Creating no information system strategy
causes a lot of troubles to Czech companies. After all, it is one of the most frequent reasons why most contributions of
the existing information systems are very moderate.

This paper would like to point out some important facts:
1. When revitalizing a company, its information system revitalization must be solved as well
2. The information system revitalization goals must support directly the company revitalization
3. Properly chosen strategy of the information system revitalization implementation, coherent to the company's
development planned enables correct setting of the so useful information system development dynamics

Literature:
[1] GRAY, C. Rst podniku. Publikace edice Business Guide pro mal a stedn podnikatele, Praha : Readers
International Prague, 1993.
[2] STRASSMAN, P. A. Stages of Growth. Datamation, October 1976, str. 46 50.

79

[3] LACKO, B. Analza zkuenost z inovace potaovho systmu v k. p. TOS KUIM. Kuim : Intern
publikace TOS KUIM, 1982.
[4] LACKO, B. Restrukturalizace bze dat. Sbornk refert semine DATASEM 95, CS COMPEX Brno : 1995.
[5] LACKO, B. Vvojov trendy v informanch a dicch systmech. Abstrakt habilitan pednky, Brno : VUT
FS, 1994, 16 stran.
[6] LACKO, B. Analza dynamika rozvoje informanho systmu. Sbornk mezinrodn konference Systmov
integrace 95, Praha : VE KIT, 1995, str. 133 144.
[7] FEDOROV, A. Some connections between the accountant management and economic development in Czech
Republic. In: International Congress, Business and Economic Development in Central and Eastern Europe,
Brno : TU of Brno, 2001.
[8] TVRDKOV, M. Zavdn a inovace informanch systm. Praha : Grada, 2000.
[9] MERUNKA, V. ; POLK, J. ; CARDA, A. Umn systmovho nvrhu. Praha : Grada, 2003.

Address:
Doc. Ing. Branislav Lacko, CSc.
VUT v Brn
Technick 2
CZ-616 69 Brno
e-mail: lacko @ fme.vutbr.cz

80

81

MODEL LEARNING AND INFERENCE THROUGH ANFIS
Amal Al Khatib


Abstract: This paper discusses the map and architecture of a learning procedure called ANFIS (adaptive-
network-based fuzzy inference system), ANFIS is a fuzzy inference system implemented in the framework of
adaptive networks. By using a hybrid learning procedure, ANFIS can construct an input-output mapping
based on both human knowledge (in the form of fuzzy if-then rules) and using input-output data pairs.

1. INTRODUCTION
Intelligent systems have appeared in many technical areas, such as consumer electronics, robotics and industrial control
systems. Many of these intelligent systems are based on fuzzy control strategies which describe complex systems
mathematical model in terms of linguistic rules.

Fuzzy set theory derives from the fact that most natural classes and concepts are fuzzy rather than crisp nature. On the
other hand, people can approximate well enough to perform many desired tasks. The fact is that they summaries from
massive information inputs and still function effectively. For complex systems, fuzzy logic is quite suitable because of
its tolerance to some imprecision.

2. ANFIS
2.1. Model Learning and Inference Through ANFIS
Assuming that we already have a collection of input/output data and would like to build a fuzzy inference system that
approximate the data, this system would consist of a number of membership functions and rules with adjustable
parameters similarly to that of neural networks.

Rather than choosing the parameters associated with a given membership function arbitrarily, these parameters could be
chosen so as to tailor the membership functions to the input/output data in order to account for these types of variations
in the data values.

2.2. What Is ANFIS?
ANFIS (adaptive-network-based-fuzzy inference system) is considered to be an adaptive network which is very similar
to neural networks, using a given input/output data set, ANFIS constructs a fuzzy inference system (FIS) whose
membership function parameters are adjusted using either a backpropagation algorithm alone, or in combination with a
least squares method. This allows the fuzzy systems to learn from the data they are modeling.

2.3. ANFIS Objective
The purpose of ANFIS is to integrate the best features of Fuzzy Systems and Neural networks. From Fuzzy Systems it
is a representation of prior knowledge into a set of constraints to reduce the optimization search space. From Neural
networks it is an adaptation of back propagation to structured network to automate the parametric tuning.

2.4. ANFIS Architecture
For simplicity, Ill assume the fuzzy inference system under consideration has two inputs x and y and one output z.
Suppose that the rule base contains two fuzzy if-then rules of Takagi and Sugenos type.

Rule 1: If x is A1 and y is B1, then f1 = pl x + q1y + rl
Rule 2: If x is A2 and y is B2, then f2 = p2x + q2y + r2

82

Then the fuzzy reasoning is illustrated in Fig. 2.4.1.(a), and the corresponding equivalent ANFIS architecture is shown
in Fig. 2.4.1.(b).

Fig. 2.4.1.(a)

Fig. 2.4.1.(b)

The node functions in the same layer are of the same function family as described below:
Layer 1: Every node i in this layer is a square node with a node function

( ) x A O
i i

1

Where x is the input to node i, and A
i
is the linguistic label associated with this node function. In other words,
1
i
O is
the membership function of A
i
and it specifies the degree to which the given x satisfies the quantifier A
i.
Usually
( ) x A
i
is chosen to be bell-shaped such as:
( )
i
i
i
i
b
a
c x
x A
1
1
]
1
,
_
2
1
1

Where {a
i
,b
i
,c
i
} is the parameter set. As the values of these parameters change, the bell-shaped functions very
accordingly, thus exhibiting various forms of membership functions to linguistic label A
i
.
Parameters in this layer are referred to as premise parameters.

Layer 2: Every node in this layer is a circle node labeled II which uses the logic operation that the user chooses
(AND,OR), example:

w
i
= ( ) x A
i
AND ( ) x B
i

Each node output represents the firing strength of a rule.

Layer 3: Every node in this layer is a circle node labeled N. the ith node calculates the ratio of the ith rules firing
strength to the sum of all rules firing strengths:
2 1
w w
w
w
i
i
+

83

Layer 4: Every node i in this layer is a square node with a node function

) (
4
i i i i i i I
r y q x p w f w O + +

Where
i
w is the output of layer 3, and {p
i
,q
i
,r
i
} is the parameter set. Parameters in this layer will be referred to as
consequent parameters.

Layer 5: The single node in this layer is a circle node labeled that comuptes the overall output as the summation of all
incoming signals, i.e.,

i
i
i
i i
i
i i
w
f w
f w output overall O
5
1

Fig 2.4.2. shows a 2-input, ANFIS with 2 rules. Two membership functions are associated with each input, so the input
space is partitioned into four fuzzy subspaces, each of which is governed by a fuzzy if-then rules. The premise part of a
rule delineates a fuzzy subspace while the consequent part specifies the output within this fuzzy subspace.

Fig. 2.4.2. anfis map

2.5. Hybrid Learning Algorithm
From the anfis architecture (Fig. 2.4.2) it is observed that given the values of premise parameters, the overall output can
be expressed as linear combinations of the consequent parameters. The output f in layer 5 can be expressed as:
2
2 1
2
1
2 1
1
f
w w
w
f
w w
w
f
+
+
+

=
2 2 1 1
f w f w +
=
2 2 2 2 2 2 1 1 1 1 1 1
) ( ) ( ) ( ) ( ) ( ) ( r w q y w p x w r w q y w p x w + + + + +
which is linear in the consequent parameters (p
i
,q
i
,r
i,
p
2
,q
2
,r
2
) as a result we have:

S = set of total parameters
S
1
= set of premise parameters
S
2
= set of consequent parameters

After finding the initial parameters by the generation of the fuzzy inference system (FIS), we can directly apply the
hybrid learning rule which will be discussed in the coming section.

More specifically, in the forward pass of the hybrid learning algorithm, functional signals go forward till layer 4 and the
consequent parameters are identified by the least squares estimate.
In the backward pass, the error rates propagate backward and the premise parameters are updated by the gradient
descent. The following table (Table 2.5.1.) summarizes the activities in each pass.

premise
parameters
consequent
parameters
A1
A2
B1
B2
x
y
w1
w2
w1*f1
w2*f2
wi*fi
x y
w1
w2
Layer 1 Layer 2 Layer 3 Layer 4 Layer 5
ANFIS
premise
parameters
consequent
parameters
A1
A2
B1
B2
x
y
w1
w2
w1*f1
w2*f2
wi*fi
x y
w1
w2
premise
parameters
consequent
parameters
A1
A2
B1
B2
x
y
w1
w2
w1*f1
w2*f2
wi*fi
x y
w1
w2
ANFIS

84

forward pass backward pass
MF param. (premise) fixed back propagation
Rule param. (consequence) least-squares fixed

Table 2.5.1. Tow pass in Hybrid Learning procedure

The consequent parameters thus identified are optimal (in the consequent parameters space) under the condition that the
premise parameters are fixed. Accordingly the hybrid approach is much faster than the strict gradient descent and it is
worthwhile to look for the possibility of decomposing the parameter set.

However, it should be noted that the computation complexity of the least squares estimate is higher than that of the
gradient descent. In fact, there are four methods to update the parameters, as listed below according to their computation
complexities:
1. Gradient descent only: all parameters are updated by the gradient descent.
2. Gradient descent and one pass of least squares estimate (LSE): the LSE is applied only once at the very beginning
to get the initial values of the consequent parameters and then the gradient descent takes over to update all
parameters.
3. Gradient descent and LSE: this is the proposed hybrid learning rule.
4. Sequential (Approximate) LSE only: The Anfis is linearized with respect to the premise parameters.

The choice of the above methods should be based on the trade-off between computation complexity and resulting
performance.

3.9. Hybrid Learning Rule (Forward pass)
Though we can apply the gradient method to identify the parameters in an adaptive network, the method is generally
slow and likely to become trapped in local minima. Thats why a Hybrid Learning Rule is proposed which combines the
gradient descent method and the least squares estimate (LSE) to identify parameters.

For simplicity, it is assumed that the adaptive network under consideration has only one output
Output = F ( ) , S I
Where I is the set of input variables and S is the set of parameters. If there exists a function H such that the composite
function HF is linear in some of the elements of S, then these elements can be identified by the least squares method.
More formally, if the parameter set S can be decomposed into two sets

S = S
1
S
2

Where

represents direct sum such that HF is linear in the elements of S
2
, then upon applying H to the output
equation we have
H(output) = H F ( ) , S I

Which is linear in the elements of S
2
. Now given values of elements of S
1
, we can plug P training data into the previous
equation and obtain a matrix equation:
AX = B
Where X is an unknown vector whose elements are parameters in S
2
. Let
2
S = M, then the dimensions of A, X and B
are PxM, Mx1 and Px1, respectively. Since P (number of training data pairs) is usually greater than M (number of linear
parameters), this is an over determined problem and generally there is no exact solution to the previous equation (AX =
B). Instead, a least squares estimate (LSE) of X, X
*
, is sought to minimize the squared error
2
B AX .

This is a standard problem that forms the grounds for linear regression, adaptive filtering and signal processing. The
most well-known formula for X
*
uses the pseudo-inverse of X:

X
*
= (A
T
A)
-1
A
T
B

85

Where A
T
is the transpose of A, and (A
T
A)
-1
A
T
is the pseudo-inverse of A if A
T
A is non-singular. The sequential
formulas are used to compute the LSE of X. this sequential method of LSE is more efficient (especially when M is
small) and can be easily modified to an on-line version for systems with changing characteristics. Specifically, let the
ith row vector of Matrix A be
T
i
a and the ith element of B be
T
i
b , then X can be calculated iteratively using the
sequential formulas:
) (
1 1 1 1 1 i
T
i
T
i i i i i
X a b a S X X
+ + + + +
+
,
1
1 1
1 1
1
+ +
+ +
+
+

i i
T
i
i
T
i i i
i i
a S a
S a a S
S S i = 0,1,.,P-1
Where
i
S is often called the covariance matrix and the least squares estimate X
*
is equal to X
p
. The initial conditions
are X
0
= 0 and S
0
= I , where is a positive large number and I is the identity matrix of dimension MxM.
Now we can combine the gradient method and the least squares estimate to update the parameters in an anfis structure.
Each epoch of this hybrid learning procedure is composed of a forward pass and a backward pass. In the forward pass,
we supply input data and functional signals go forward to calculate each node output until the matrices A and B are
obtained, and the parameters in S
2
are identified by the sequential least squares formulas.

After identifying parameters in S
2
the functional signals keep going forward till the error measure is calculated. In the
backward pass, the error rates (the derivatives of the error measure with respect to each node output) propagate from the
output end and toward the input end, and the parameters in S
1
are updated by the gradient method (as will be explained
in the coming section).

For given fixed values of parameters in S
1
, the parameters in S
2
thus found are guaranteed to be the global optimum
point in the S
2
parameter space due to the choice of the squared error measure. Not only can this hybrid learning rule
decrease the dimension of the search space in the gradient method, but in general it will also cut down substantially the
convergence time.

3.10. Gradient descent (Backward pass)
The objective of applying this step is to update the premise parameters (membership functions parameters). And in
this paper Im using the Back propagation algorithm.
ANFIS Back propagations basic idea is based on the Error measure E (Overall error measure)
2 ) (
5
) (
1
) (
1
) (
2
1
p p
P
p
p
P
p
O d E E

where:
p = number of nodes in a layer
d = p
th
component of desired output vector
O
5
= actual output in layer 5 of the ANFIS structure (see Fig.4)

Assuming that the symbol represents each membership functions parameters update then for each parameter
the update formula is:
i
p P
p i
i
E E
) (
1

where

i
i
E
k
is the learning rate

is the step size

i
E

is the derivative update

86

The chain rule is used in order to calculate the partial derivatives. Based on the Sugeno inference mechanism, the error
rate for consequence parameters can be calculated as follows:

i
p
p
p
p
p
j
p
j
p
j
p
j
p
p p
j i
p
O
O
O
O
O
O
O
O
O
O d
E

) (
1 , 1
) (
1 , 1
) (
1 , 2
) (
1 , 2
) (
, 3
) (
, 3
) (
, 4
) (
, 4
) (
5 ) (
5
) (
2
1
) (
) (
The derivation of
j
O
O
, 4
5
is as follows
1
, 4
, 4
, 4
5
,
_
j
k
k
j
O
O
O
O

for
j
i
O
O
, 3
, 4
is
( )
j i
j i r q x p
O
r q x p O
O
O
i i i
j
i i i i
j
i
'
+ +
+ +
0
) (
, 3
, 3
, 3
, 4

for
j
i
O
O
, 2
, 3
is
'
,
_
j i if
O O
O
j i if
O O
O
O
O O
O
O
O
j
i
j
i
,
,
2 , 2 1 , 2
1 , 2
2 , 2 1 , 2
1 , 2
, 2
2 , 2 1 , 2
, 2
, 2
, 3

and for
i
i
a
O
, 1
,
i
i
b
O
, 1
, and
i
i
c
O
, 1
are
2 2
2
, 1
] ) ( 1 [
) ( 2
i
b
i
i
i
b i
i
i
i
a
c x
a
a
c x
b
a
O
i
,
2 2
2 2
, 1
] ) ( 1 [
) ( ) (
i
i
b
i
i
i
i
b
i
i
i
i
a
c x
a
c x
In
a
c x
b
O
,
2 2
2
, 1
] ) ( 1 )[ (
) ( 2
i
i
b
i
i
i
b
i
i
i
i
i
a
c x
c x
a
c x
b
c
O

3. CONCLUSION
The hybrid system ANFIS with inference mechanism based on the adaptive network is being considered in this paper.
An important property of this system is that its possible to tune the parameters of the fuzzy system that has been
described with the help of ANFIS. The same methods that are used to tune the parameters are also used to tune the
weights in the neural networks. As the ANFIS is a multilayer network, it can be concluded, that the tuning of its
parameters is a non-linear task (the parameters of the inner layers cant be expressed linearly). Hence, new algorithms
inherit problems that are related to the training algorithms of the multilayer neural. Moreover, new algorithms require
more computational power; ANFIS is a learning method that is computationally more complex yet more effective than
some of the other methods proposed in the neural network field.

References
[1] JANG, J. S.; ANFIS, R. Adaptive-Network-Based Fuzzy Inference System, IEEE Transactions on Systems,
Man and Cybernetics, Vol. 23, No. 3, May/June 1993, pp. 665-683.
[2] The MathWorks, Inc.: Neural Network Toolbox (Matlab Toolbox), in reference to the handbook of The
MathWorks, Inc., Boston 1998
[3] VALISHEVSKY, A. Comparative Analysis of Different Approaches towards Multilayer Perceptron Training,
Scientific Proceedings of Riga Technical University, 2001.

87

[4] Neuro-Fuzzy Modeling and Control, J.S.R. Jang and C.-T. Sun, Proceedings of the IEEE, 83(3):378-406
[5] The Fuzzy Logic Toolbox for use with MATLAB, J.S.R. Jang and N. Gulley, Natick, MA: The MathWorks
Inc., 1995

Address:
Ing. Amal Al Khatib
VUT v Brn
Technick 2
CZ-616 69 Brno
e-mail: akhatib@seznam.cz

88

89

GRAMMATICAL EVOLUTION WITH BACKWARD PROCESSING
Pavel Omera, O. Popelka,
1
Imrich Rukovansk
2

1
Brno University of Technology,
2
European Polytechnical Institute Kunovice

Abstract: This paper describes Parallel Grammatical Evolution (PGE) that can evolve complete programs
using a variable length linear genome to govern the mapping of a Backus Naur Form grammar definition.
To increase the efficiency of Grammatical Evolution (GE) the influence of backward processing was
tested. The significance of backward coding (BC)and the comparison with standard coding of GEs is
presented. BC can speed up Grammatical Evolution with high quality features. The adaptive significance
of Parallel Grammatical Evolution with male and female populations has been studied.

1 INTRODUCTION
Grammatical Evolution (GE) [1] can be considered a form of grammar-based genetic programming (GP). In particular,
Kozas genetic programming has enjoyed considerable popularity and widespread use. Unlike a Koza-style approach,
there is no distinction made at this stage between what he describes as function (operator in this case) and terminals
(variables). Koza originally employed Lisp as his target language. This distinction is more of an implementation detail
than a design issue. Grammatical evolution can be used to generate programs in any language, using Backus Naur Form
(BNF). BNF grammars consist of terminals, which are items that can appear in the language, i.e. +, -, sin, log etc. and
non-terminal, which can be expanded into one or more terminals and non-terminals. A non-terminal symbol is any
symbol that can be rewritten to another string, and conversely a terminal symbol is one that cannot be rewritten. The
major strength of GE with respect to GP is its ability to generate multi-line functions in any language. Rather than
representing the programs as parse tree, as in GP, a linear genome representing is used. A genotype-phenotype mapping
is employed such that each individuals variable length byte strings, contains the information to select production rules
from a BNF grammar. The grammar allows the generation of programs, in an arbitrary language that are guaranteed to
be syntactically correct. The user can tailor the grammar to produce solutions that are purely syntactically constrained,
or they may incorporate domain knowledge by biasing the grammar to produce very specific form of sentences.

GE system in [1-3] codes a set of pseudo random numbers, which are used to decide which choice to take when a non-
terminal has one or more outcomes. Because GE mapping technique employs a BNF definition, the system is language
independent, and theoretically can generate arbitrarily complex functions. There is quite an unusual approach in GEs, as
it is possible for certain genes to be used two or more times if the wrapping operator is used. BNF is a notation that
represents a language in the form of production rules. It is possible to generate programs using the Grammatical Swarm
Optimization (GSO) technique [2] with a performance similar to the GE. Given the relative simplicity of GSO, the
small population sizes involved, and the complete absence of a crossover operator synonymous with program evolution
in GP or GE. Grammatical evolution was one of the first approaches to distinguish between the genotype and
phenotype. GE evolves a sequence of rule numbers that are translated, using a predetermined grammar set into a
phenotypic tree.

Our approach uses a parallel structure of GE (PGE). A population is divided into several subpopulations that are
arranged in the hierarchical structure [4]. Every subpopulation has two separate parts: a male group and a female group.
Every group uses quite a different type of selection. In the first group a classical type of GA selection is used. In the
second group only different individuals can be included. It is a biologically inspired computing similar to a harem
arrangement. This strategy increases an inner adaptation of PGE. The following text explains why we used this
approach. Analogy would lead us one step further, namely, to the belief that the combination of GE with a sexual
reproduction [5-6]. On the principle of the sexual reproduction we can create a parallel GE with a hierarchical structure.

2 PARALLEL GRAMMATICAL EVOLUTION
The PGE is based on the grammatical evolution GE [1], where BNF grammars consist of terminals and non-terminals.
Terminals are items, which can appear in the language. Non-terminals can be expanded into one or more terminals and
non-terminals. Grammar is represented by the tuple {N,T,P,S}, where N is the set of non-terminals, T the set of

90

terminals, P a set of production rules which map the elements of N to T, and S is a start symbol which is a member of N.
For example, below is the BNF used for our problem:
N = {expr, fnc}
T = {sin, cos, +, -, /, *, X, 1, 2, 3, 4, 5, 6, 7, 8, 9}
S = <expr>
and P can be represented as 4 production rules:
1. <expr> := <fnc><expr>
<fnc><expr><expr>
<fnc><num><expr>
<var>
2. <fnc> := sin
cos
+
*
-
U-
3. <var> := X
4. <num> := 0,1,2,3,4,5,6,7,8,9

The production rules and the number of choices associated with each are in Table 1. The symbol U- denotes an unary
minus operation.

Table 1: The number of available choices for each production rule.
rule no choices
1 4
2 6
3 1
4 10

There are notable differences when compared with [1]. We dont use two elements <pre_op> and <op>, but only one
element <fnc> for all functions with n arguments. There are not rules for parentheses; they are substituted by a tree
representation of the function. The element <num> and the rule <fnc><num><expr> were added to cover generating
numbers. The rule <fnc><num><expr> is derived from the rule <fnc><expr><expr>. Using this approach we can
generate the expressions more easily. For example when one argument is a number, then +(4,x) can be produced, which
is equivalent to (4 + x) in an infix notation. The same result can be received if one of <expr> in the rule
<fnc><expr><expr> is substituted with <var> and then with a number, but it would need more genes.

There are not any rules with parentheses because all information is included in the tree representation of an individual.
Parentheses are automatically added during the creation of the text output.

If in the GE is not restricted anyhow, the search space can have infinite number of solutions. For example the function
cos(2x), can be expressed as cos(x+x); cos(x+x+1-1); cos(x+x+x-x); cos(x+x+0+0+0...) etc. It is desired to limit the
number of elements in the expression and the number of repetitions of the same terminals and non-terminals.

3 BACKWARD PROCESSING OF THE GE
The chromosome is represented by a set of integers filled with random values in the initial population. Gene values are
used during chromosome translation to decide which terminal or nonterminal to pick from the set. When selecting a
production rule there are four possibilities, we use gene_value mod 4 to select a rule. However the list of variables has
only one member (variable X) and gene_value mod 1 always returns 0. A gene is always read; no matter if a decision is
to be made, this approach makes some genes in the chromosome somehow redundant. Values of such genes can be
randomly created, but genes must be present.

The figure Fig. 1 shows the genotype-phenotype translation scheme. Body of the individual is shown as a linear
structure, but in fact it is stored as a one-way tree (child objects have no links to parent objects). In the diagram we use
abbreviated notations for nonterminal symbols: f - <fnc>, e - <expr>, n - <num>, v - <var>.
The column description in Fig. 1:
A. Objects of the individuals body (resulting trigonometric function),

91

B. Genes used to translate the chromosome into the phenotype,
C. Modulo operation, divisor is the number of possible choices determined by the gene context,
D. Result of the modulo operation,
E. State of the individuals body after processing a gene on the corresponding line,
F. Blocks in the chromosome and corresponding production rules,
G. Block marks added to the chromosome.

Fig.1: Relations between genotype (column B) and phenotype (column A)

Since operation modulo takes two operands, the resulting number is influenced by gene value and by gene context (Fig.
1C = see Fig. 1 column C). Gene context is the number of choices, determined by the currently used list (rules,
functions, variables). Therefore genes with same values might give different results of modulo operation depending on
what object they code. On the other hand one terminal symbol can be coded by many different gene values as long as
the result of modulo operation is the same (31 mod 3) = (34 mod 3) = 1. In the example (Fig. 1A) given the variables
set has only one member X. Therefore, modulo divider is always 1 and the result is always 0, a gene which codes a
variable is redundant in that context (Fig. 1D). If the system runs out of genes during phenotype-genotype translation
then the chromosome is wrapped and genes at the beginning are reused.

4 PROCESSING THE GRAMMAR
The processing of the production rules is done backwards from the end of the rule to the beginning (Fig. 2). E.g.
production rule <fnc><expr1><expr2> is processed as <expr2><expr1><fnc>. We use <expr1> and <expr2> at this
point to denote which expression will be the first argument of <fnc>.

Fig. 2: Proposed backward notation of a function tree structure

92

The main difference between <fnc> and <expr> nonterminals is in the number of real objects they produce in the
individuals body. Nonterminal <fnc> always generates one and only one terminal; on the contrary <expr> generates
an unknown number of nonterminal and terminal symbols. If the phenotype is represented as a tree structure then a
product of the <fnc> nonterminal is the parent object for handling all objects generated by <expr> nonterminals
contained in the same rule. Therefore the rule <fnc><expr1><expr2> can be represented as a tree (Fig. 3).

To select a production rule (selection of a tree structure) only one gene is needed. To process the selected rule a number
of n genes are needed and finally to select a specific nonterminal symbol again one gene is needed. If the processing is
done backwards the first processed terminals are leafs of the tree and the last processed terminal in a rule is the root of a
subtree. The very last terminal is the root of the whole tree. Note that in a forward processing (<fnc><expr1><expr2>)
the first processed gene codes the rule, the second gene codes the root of the subtree and the last are leafs.

When using the forward processing and coding of the rules described in [1] its not possible to easily recover the tree
structure from genotype. This is caused with <expr> nonterminals using an unknown number of successive genes. The
last processed terminal being just a leaf of the tree. The proposed backward processing is shown in Fig. 1E.

4.1 PHENOTYPE TO GENOTYPE PROJECTION
Using the proposed backward processing system the translation to a phenotype subtree has a certain scheme. It begins
with a production rule (selecting the type of the subtree) and ends with the root of the subtree (in our case with a
function) (Fig. 1F). In the genotype this means that one gene used to select a production rule is followed by n genes
with different contexts which are followed by one gene used to translate <fnc>. Therefore a gene coding a production
rule forms a pair with a gene coding terminal symbol for <fnc> (root of the rule). Those genes can be marked when
processing the individual. This is an example of a simple marking system:
BB Begin block (a gene coding a production rule)
IB Inside block
EB End block (a gene coding a root of a subtree)

The EB and BB marks are pair marks and in the chromosome they define a block (Fig. 1G). Such blocks can be nested
but they dont overlap (the same way as parentheses). The IB mark is not a pair mark, but it is always contained in a
block (IB marks are presently generated by <num> nonterminals). Given a BB gene a corresponding EB gene can be
found using a simple LIFO method.

A block of chromosome enclosed in a BB-EB gene pair then codes a subtree of the phenotype. Such block is fully
autonomous and can be exchanged with any other block or it can serve as completely new individual.
Only BB genes code the tree of individuals body, while EB and IB genes code the terminal symbols in the resulting
phenotype. The BB genes code the structure of the individual, changing their values can cause change of the applied
production rule. Therefore change (e.g. by mutation) in the value of a structural gene may trigger change of context of
many, or all following genes.

This simple marking system introduces a phenotype feedback to phenotype; however it doesnt affect the universality of
the algorithm. Its not dependent on the used terminal or nonterminal symbols; it only requires the result to be a tree
structure. Using this system its possible to introduce a progressive crossover and mutation.

4.2 CROSSOVER
When using grammatical evolution the resulting phenotype coded by one gene depends on the value of the gene and on
its context. If a chromosome is crossed at random point, it is very probable that the context of the genes in second part
will change. This way crossover causes destruction of the phenotype, because the newly added parts code different
phenotype than in the original individual.

This behavior can be eliminated using a block marking system. Crossover is then performed as an exchange of blocks.
Fig. 3: Production rule shown as a tree

93

The crossover is made always in an even number of genes, where the odd gene must be BB gene and even must be EB
gene. Starting BB gene is presently chosen randomly; the first gene is excluded because it encapsulates (together with
the last used gene) the whole individual.

The operation takes two parent chromosomes and the result is always two child chromosomes. It is also possible to
combine the same individuals, while the resulting child chromosomes can be entirely different.

Given the parents:
1) cos( x + 2 ) + sin( x * 3 )
2) cos( x + 2 ) + sin( x * 3 )

The operation can produce children:
3) cos( sin( x * 3 ) + 2 ) + sin( x * 3 )
4) cos( x + 2 ) + x

This crossover method works similar to direct combining of phenotype trees, however this method works purely on the
chromosome. Therefore phenotype and genotype are still separated. The result is a chromosome, which will generate an
individual with a structure combined from its parents. This way we receive the encoding of an individual without
backward analysis of his phenotype. To perform a crossover the phenotype has to be evaluated (to mark the genes), but
it is neither used nor know in the crossover operation (also it doesnt have to exist).

4.3 MUTATION
Mutation can be divided into mutation of structural (BB) genes and mutation of other genes. Mutation of one structural
gene can affect other genes by changing their context therefore structural mutation amount should be very low. On the
other hand the amount of mutation of other genes can be set very high and it can speed up searching an approximate
solution.

Given an individual:
sin( 2 + x ) + cos( 3 * x )

and using only mutation of non-structural genes, it is possible to get:
cos( 5 x ) * sin( 1 * x )

Therefore the structure doesnt change, but we can get a lot of new combinations of terminal symbols. The divided
mutation allows using the benefits of high mutation while eliminating the risk of damaging the structure of an
individual.

4.4 POPULATION MODEL
The system uses three populations forming a simple tree structure (Fig. 4). There is a Master population and two slave
populations, which simulate different genders. The links among the populations lead only one way - from bottom to top.
Fig. 4: The population model

4.5 FEMALE POPULATION
When a new individual is to be inserted in a population a check is preformed whether it should be inserted. If a same or
similar individual already exists in the population then the new individual is not inserted. In a female population every
genotype and phenotype occurs only once. The population maintains a very high diversity; therefore the mutation
operation is not applied to this population. Removing the individuals is based on two criterions. The first criterion is the
age of an individual - length of stay in the population. The second criterion is the fitness of an individual. Using the
second criterion a maximum population size is maintained. Parents are chosen using the tournament system selection.

94

4.6 MALE POPULATION
New individuals are not checked so duplicate phenotypes and genotypes can occur, also the mutation is enabled for this
population. Mutation rate can be safely set very high (30%) provided that the structural mutation is set very low (less
then 2%). For a couple of best individuals the mutations are nondestructive. If a protected individual is to be mutated a
clone is created and added to the population. If the system stagnates in a local solution the mutation rate is raised using
a linear function depending on the number of cycles for which the solution wasnt improved. Parents are chosen using a
logarithmic function depending on the position of an individual in a population sorted by fitness. For every selected
male parent a new selection of female parent is made.

4.7 MASTER POPULATION
The master population is superior to the male and female populations. Periodically the subpopulations send over their
best solutions. Moreover the master population performs another evolution on its own. Parents are selected using the
tournament system. The master population uses the same system of mutations as the male population, but for removing
individuals from the population only the fitness criterion is used. Therefore master population also serves as an archive
of best solutions of the whole system.

4.8 FITNESS FUNCTION
Around the searched function there is defined an equidistant area of a given size. Fitness of an individuals phenotype is
computed as the number of points inside this area divided by the number of all checked points (a value in <0,1>). This
fitness function forms a strong selection pressure; therefore the system finds an approximate solution very quickly.

4.9 RESULTS
Given sample of 100 points in the interval [0,2] and using the block marking system described in 5.1, PGE has
successfully found the searched function sin(2*x)*cos(2+x) on the majority of runs. The graph (Fig 5.) shows
maximum fitness in the system for ten runs and an average (bold). On the other hand the same system with phenotype
to genotype projection disabled (Fig. 6). The majority of runs didnt find the searched function within 120 generations.

We have simplified the generation of numbers by adding a new production rule, thus allowing the generation of
functions containing integer constants. The described parallel system together with phenotype to genotype projection
improved the speed of the system. The progressive crossover and mutation eliminates destroying partial results and
allowed us to generate more complicated functions (e.g. sin(2 * x)*cos(2 + x)).

0
0 , 2
0 , 4
0 , 6
0 , 8
1
1 2 5 5 0 7 5 1 0 0
G e n e r a t i o n
F i t n e s s

Fig. 5: Convergence of the PGE using backward processing (average in bold)

We have described a parallel system, Parallel Grammatical Evolution (PGE) that can map an integer genotype onto a
phenotype with the backward coding. PGE has proved successful for creating trigonometric identities.
Parallel GEs with the sexual reproduction can increase the efficiency and robustness of systems, and thus they can track
better optimal parameters in a changing environment. From the experimental session it can be concluded that modified
standard GEs with two sub-populations can design PGE much better than classical versions of GEs.

95

0
0 , 2
0 , 4
0 , 6
0 , 8
1
1 2 5 5 0 7 5 1 0 0
G e n e r a t i o n
F i t n e s s

Fig. 6: Convergence of the PGE using forward processing (average in bold)

Fig.7: Convergence of the PGE with 5 PC using backward processing (average in bold)

Fig.8: The parallel structure of PGE with 6 computers
0
0,2
0,4
0,6
0,8
1
0 50 100
Generation
Fitness

96

The PGE algorithm was tested with the group of 6 computers in the computer network (see Fig. 8). Five computers
calculated in the structure of five subsystems MR1, MR2, MR3, MR4, and MR5 and one master MR. The male
subpopulation M of MR in the higher level follows the convergence of the subsystem. In Fig. 7 is presented 10 runs of
the PGE- program. The shortest time of computation is only 10 generation. All calculation were finished before 40
generation. This is better to compare with backward processing on one computer (see Fig. 5). The forward processing
on one computer was the slowest (see Fig. 6).

5 CONCLUSIONS
The increased awareness from other scientific communities, such as biology and mathematics, promises new insights
and new opportunities. There is much to accomplish and there are many open questions. Interest from diverse
disciplines continues to increase and simulated evolution is becoming more generally accepted as a paradigm for
optimization in practical engineering problems.

The parallel grammatical evolution can be used for the automatic generation of programs. This can help us to find
information as a part of complexity. We are far from supposing that all difficulties are removed but first results with
PGEs are very promising.

REFERENCES
[1] ONEILL, M.; RYAN, C. Grammatical Evolution: Evolutionary Automatic Programming in an Arbitrary
Language Kluwer. Academic Publishers 2003.
[2] ONEILL, M.; BRABAZON, A.; ADLEY, C. The Automatic Generation of Programs for Classification
Problems with Grammatical Swarm. Proceedings of CEC 2004, Portland, Oregon (2004) 104 110.
[3] PIASCZNY, W.; SUZUKI, H.; SAWAI, H. Chemical Genetic Programming Evolution of Amino Acid
Rewriting Rules Used for Genotype-Phenotype Translation, Proceedings of CEC 2004, Portland, Oregon
(2004) 1639 - 1646.
[4] OMERA, P.; IMONK, I.; ROUPEC, J. Multilevel distributed genetic algorithms. In Proceedings of the
International Conference IEE/IEEE on Genetic Algorithms, Sheffield (1995) 505510.
[5] OMERA, P.; ROUPEC, J. Limited Lifetime Genetic Algorithms in Comparison with Sexual Reproduction
Based Gas. Proceedings of MENDEL2000, Brno, Czech Republic (2000) 118 126
[6] OMERA, P. Evolution of System with Unpredictable Behavior, Proceedings of MENDEL2004, Brno, Czech
Republic (2004) 1 - 6.
[7] OMERA, P. Genetic Algorithms and their Aplications, the habilit work, in Czech language 2002.

Address:
Doc. Ing. Pavel Omera, CSc.
Institute of Automation and Computer Science
Technicka 2,
Tel.: +420 541 142 294
Fax: +420 541 142 490
e-mail: osmera @fme.vutbr.cz

Address:
Bc. Ondej Popelka
Technicka 2,
Tel.: +420 541 142 294
Fax: +420 541 142 490
e-mail: odinuv@gmail.com,

97

Address:
European Polytechnical Institute, s.r.o.
Osvobozen 699,
686 04 Kunovice, Czech Republic
e-mail: rukovansky@vos.cz

98

99

OBJECT RECOGNITION BY MEANS OF NEW AL
Ji tastn, Martin Minak


Abstract: This document provides an overview of algorithms for object recognition. Three basic
algorithms are described - recognition with the aid of the momentum, recognition with the aid of grammar
and back propagation algorithm, thus recognition with the aid of neural network. Finally, speed and
applicability of these algorithms are compared.

Key-Words: Back Propagation Algorithm, Momentum, Grammar

1 Introduction
Pattern recognition consists in sorting objects into classes. Class is a subset of objects whose elements have common
features from the classification standpoint. Object has a physical character, which in computer vision is most frequently
taken to mean a part of segmented image.

Methods for the classification of objects constitute last and upper-most step in computer vision theory.

The following methods were mutually compared:
Recognition with the aid of moments
Recognition with the aid of grammar describing the edges of object
Recognition with the aid of neural network (back propagation)

A real technological scene for object classification was simulated by digitizing
five selected objects (see Fig. 1). For this purpose, two-dimensional images of
three-dimensional objects were prepared. The aim was to test such objects that
resemble two-dimensional images of real objects. The choice of objects of
similar shape was also intentional.
Fig. 1

2 Recognition with the aid of the momentum method
The resultant moment characteristics for object detection, which are used in the program, will be in the form:

02 20 1
+ (1)
2
11
2
02 20 2
4 ) ( + + (2)
2
03 21
2
12 30 3
) 3 ( ) ( + (3)
2
03 21
2
12 30 4
) ( ) ( + + + (4)
[ ]
[ ]
2
03 21
2
12 30 03 21 03 21
2
03 21
2
12 30 12 30 12 30 5
) ( ) ( 3 ) )( 3 (
) ( 3 ) ( ) )( 3 (

+ + + +
+ + + +
(5)
[ ] ) )( ( 4 ) ( ) ( ) (
03 21 12 30 11
2
03 21
2
12 30 02 20 6
+ + + + + (6)
[ ]
[ ]
2
03 21
2
12 30 03 21 12 30
2
03 21
2
12 30 12 30 03 21 7
) ( ) ( 3 ) )( 3 (
) ( 3 ) ( ) )( 3 (

+ + +
+ + +
(7)

100

The moment description represents binary and grey shade areas. This method (see [7]) is based on computing seven
moment object flags. These moments are invariant with respect to repositioning, rotation and size of the object.

Recognition with the aid of the moment method has yielded very good results. This method faultlessly classified most
objects already at the stage of learning on some model etalon. The moments method can be used for a different edge
detector than for which the moments were calculated since they are not too sensitive to changes in object edges, e.g. the
Canny detector (at higher sigma it rounds the edges) and the Sobel operator. This method is unfit for applications
requiring the recognition of minimum dissimilar shapes, because this method is not sensitive to minor shape changes.
This method is fit for applications requiring fast recognition of dissimilar objects in different rotation.

3 Recognition with the aid of grammar
While in flag methods of pattern recognition use is made of quantitative description of objects by numerical parameters,
the flag vector, in syntactical methods the input description is of quantitative nature reflecting the structure of the
object. The elementary properties of syntactically described objects are referred to as primitives. Primitives are edge
parts of a certain shape or a graph or relation description of areas when the primitives are sub-areas of a certain shape.
The task of syntactical pattern recognition of an image is to determine whether the image under analysis corresponds to
the images of a given grammar, i.e. whether this grammar can generate this image. The image is represented by a
language string given by the grammar.

The simplest way of pattern recognition is comparison with model. A string representing the image is compared with
elements of sentences set representing single model images. In the comparison, either complete or partial agreement
with the model is necessary but on the basis of a certain adapting criterion. This method is simple and rapid. If a
complete image description is necessary for pattern recognition, syntactical analysis is required. In object analysis tasks,
the aim is to obtain a description that would include not only the listing of recognized objects and their mutual
arrangement (structural information) but also their dimensions and distances found between them (semantic
information). In the design of syntactical analyzer we must expect random effects such as image distortion.

Primitives are the basic building element of image. When choosing them, their easy recognition must be taken into
consideration. For images that are characterized by an edge or skeleton it is suitable to have parts of line as primitives.
For example, a line segment can be characterized by its beginning and end, its length or angle. The same holds for
curves. The choice of primitives depends on the application being solved.
It generally holds:
Primitives must be easy to recognize also by existing non-syntactical methods
Primitives must provide a compact and sufficient description of images by means of specified relations

If primitives of greater complexity are used we obtain a simpler structural description of objects and these results in
applying a simpler grammar for the description of objects. But it also leads to a greater complexity when seeking such
primitives in an image. On the contrary, simpler primitives lead to a more complex grammar but they are easy to
identify in the image. After the choice of primitives the next important step is to set up transcription rules for the
grammar, based on experience and knowledge.

Example of setting up grammar for object and choosing primitives is on Fig 2). If from a point marked on Fig. 2) we
proceed anticlockwise, the string will be:
d f b c a j b c a g
Due to the poor quality of input image it may happen that short straight segments exhibited on Fig. 2) will not be
detected and thus the objects string will also be changed
d f b c j b c a g
d f b c a j c a g
d f b c j c a g
For this case it is suitable to modify the grammar such that it also generates the above strings. This type of approach is
suitable for the expected deformations. In the case of random deformations it is advantage to use the classification by
means of distance (the Levenshtein distance).

101

Fig. 2

In the program generated, a syntactical analyzer is applied, which operates using the following algorithm:
1. If not all the strings have been analysed, read a new string and proceed with step 2, otherwise proceed with step 7
2. Perform the bottom-to-top analysis for the class selected
3. If the string belongs to the language of the grammar of selected class, proceed with step 6
4. If the number of string rotations is less than the string length, rotate the string and proceed with step 2, otherwise
proceed with step 5
5. If the number of string rotations is less than (360 / angle step), rotate the object by the angle step given and
proceed with step 2
6. enter the result and proceed with step 1
7. write the message about pattern recognition

The syntactical analyzer has been designed for the left linear grammar

String rotation: This mean shifting the last terminal symbol to the beginning

1.rotace
abcde eabcd L

Object rotation: This means rotating the object by a given angle and thus obtaining a different string.

If we need to classify N objects, we must create N classes, N grammars for them, and the respective languages L(G
1
),
L(G
2
), ..., L(G
N
). For example, if grammar G
x
generates words containing only one terminal symbol b, then all the
objects containing just this one symbol b will belong to class X pertaining to this grammar. Objects containing more
than one symbol b will be further analysed using the remaining grammars. In the case that no grammar is found that
corresponds to the given string, the object will be suppressed.

In the case of primitives marking single edge segments, the grammar is very sensitive to small mistakes in edge
detection. It is necessary to tailor the grammar for a definite type of edge detector. For example, it is a mistake to set up
the grammar for objects to which the current-zero operator was applied, and then to recognize objects by means of the
Canny edge detector with sizeable Sigma (it modifies edges).

Grammars are suitable to use in applications which require recognizing differently rotated objects and when the
emphasis is on recognizing small changes in the segment edge. Setting up a grammar requires time and the knowledge
of grammar description of edges. Preparing the rules for grammar must be done manually, it is not done automatically
as in other methods.

102

4 Back Propagation Algorithm
Back-propagation algorithm is an iterative method where the network gets from an initial non-learned state to the full
learned one (see [10]). It is possible to describe the algorithm in the following way:

random initialization of weights;
repeat
repeat
choose_pattern_from_training_set;
put_chosen_pattern_in_input_of_network;
compute_outputs_of_network;
compare_outputs_with_required_values;
modify_weights;
until all_patterns_from_traning_set_are_chosen;
until total_error < criterion;

The learning algorithm of back-propagation is essentially an optimization method that is able to find the weight
coefficients and thresholds for the given neural network and training set. The network is assumed to be made up of
neurons the behaviour of which is described by the formula:
,
_
N
i
i i
x w S y
1
(17)
where the output nonlinear function S is defined by the formula:
S
e
( )

+

1
1
(18)

where determines the curve steepness in the origin of coordinates. Input and output values are assumed to be in the
range < 0, 1 >.

In the following formulas the parameter o denotes the output layer, h the hidden layer, and i,j the indexes. Index i
indexes output neurons and index j their inputs. Then y
i
h
means i-th neuron output of the hidden layer and w
ij
o
means
the weight connecting i-th neuron of the output layer and j-th neuron of the previous hidden layer.

The appurtenant back-propagation algorithm can be written in the following steps:
1. Initialization. You set at random all the weights in the network at values in the recommended range < -0.3, 0.3 >.
2. Pattern submitting. You choose a pattern from the training set and put it in network inputs. Then you compute
outputs of particular neurons by relations (17) and (18).
3. Comparison. First you compute the neural network energy (SSE) under relation (19).
E y d
i i
i
n

1
2
2
1
( ) (19)
Then you compute an error for output layer by the relation:

i
o
i i
o
i
o
i
o
d y y y ( ) ( ) 1 (20)
4. Back-propagation of an error and weight modification. You compute for all neurons in the layer:
w t t y t w t
ij
l
i
l
j
l
ij
l
( ) ( ) ( ) ( ) +

1
1 (21)

i
l
i
l
i
l
t t t ( ) ( ) ( ) + 1 (22)
By the relation:

i
h
i
h
i
h
ki
h
k
h
k
n
y y w

1 1 1
1
1 ( ) (23)
you back-propagate an error in the layer nearer the inputs. Then you modify the weights:

103

w t w t w t
ij
l
ij
l
ij
l
( ) ( ) ( ) + + 1 (24)

i
l
i
l
i
l
t t t ( ) ( ) ( ) + + 1 (25)
You apply step number 4 to all the layers of network. You start with output layer followed by hidden layers.
5. Termination of pattern selection from the training set. If you have submitted all patterns from the training set to
network then continue with step number 6 else you go back to step number 2.
6. Termination of learning process. If neural network energy in the last computation has been less than the criterion
selected then terminate the learning process else you continue with step number 2.

Flag vectors have been submitted to network and arm lengths have been transformed into values from interval < 0, 1 >.
The number of vector components has been put in position 140 in the implemented computer program.

This method is the fastest of all the methods under comparison. For the description of objects using this method, 70
symptomatic vectors were used that went from the centre of gravity to object edges. This method can recognize objects
with considerably modified shapes but it may identify incorrectly objects of similar shape. This method recognizes
differently rotates objects. The error of the method increases with decreasing size of objects.

5 Conclusion
Recommended application of pattern recognition methods:
Grammar recognition is suitable where the recognition of rotated objects is required and where single edge segments
need to be detected with high accuracy without the risk of the occurrence of significant errors, and where high-speed
classification is required. Regarded as significant is an error that cannot be implied in the rules.

Recognition with the aid of moments is suitable where the edge course is not very important, where a rough division
into single classes is sufficient. For example, it does not matter whether there is a sharp transition or a short curve
between two edges.

Recognition with the aid of neural network is suitable where high-speed classification with randomly rotated objects
is required and where we need to tolerate some differences between learned etalons and classified objects.

The fastest methods for pattern recognition are recognition with the aid of grammar and recognition with the aid of
neural network. The moment method is the slowest.

Acknowledgement
This research was supported by the grants:
No 102/03/0434 Limits for broad-band signal transmission on the twisted pairs and other system co-existence. The
Grant Agency of the Czech Republic (GACR)
No 102/03/0260 Development of network communication application programming interface for new generation of
mobile and wireless terminals. The Grant Agency of the Czech Republic (GACR)
No 102/03/0560 New methods for location and verification of compliance of quality of service in new generation
networks. The Grant Agency of the Czech Republic (GACR)
No CEZ: J22/98: 261100009 Nontraditional methods for investigating complex and vague systems
No CZ 400011(CEZ 262200011) Research of communication systems and technologies (Research design)
Grant 1570 F1 New approach to the subject High-speed Communication Systems (grant of the Czech Ministry of
Education, Youth and Sports)
Grant 1563 F1 Restructure of telecommunication objects for third age university (grant of the Czech Ministry of
Education, Youth and Sports)

References:
[1] FISHER, R.: World and Scene Representations. [Online], 2002.
<www.dai.ed.ac.uk/CVonline/repres.htm>
[2] HEALTH M. and SARKAR S.: Edge detection comparison. [Online], 1996.
<marathon.csee.usf.edu/edge/edge_detection. html>

104

[3] GONZALES, R.,C. and WOODS, R., E.: Digital Image Processing, Addison-Wesley Publishing Co., New
York 1993.
[4] RA, J. and BENE, B.: Model Computer Graphic, Computer Press, Prague 1996.
[5] SMITH, M.,W. and Davis, W.,A.: A new Algorithm for Edge Detection. John Wiley, New York 1974.
[6] SVITK, R.: Edge Detection on Images. [Project-online]. ZU FAV. Plze 2001.
<http://herakles.zcu.cz/~rsvitak/school/zvi_rsvitak. pdf>
[7] ONKA, M. and HLAV, V. and BOYLE, R..: Image Processing, Analysis and Machine Vision. PWS,
Boston 1998.
[8] JEEK, B.: Computer Graphics II. [Lectures -online]. UHK FIM
<http://lide.uhk.cz/home/fim/ucitel/fujezeb1/www/>. Hradec Krlov 2002.
[9] BULB, M. : Programming [Online]. Prague 2000 <www.freesoft.cz/projekty/vyhen/clanky/prog/bres. html>
[10] NOREK, M. and JIINA, M.: Neuronov st a neuropotae, Prague, 1998.
[11] ONKA, M: Course Digital Image Processing. [Online]. <css.engineering.uiowa.edu/~DIP>, Prague 2002.
[12] ZAHN, T., CH.: Fourier Descriptors for Plane Closed Curves. In: IEEE Trans. on Computers, vol. C-21, No.3,
1972.
[13] LOPEZ-CAVIEDES, M. and SANCHEZ-DIAZ, G.: A New Clustering Criterion in Pattern Recognition.
International Journal WSEAS Transactions on Computers, Issue 3, Volume 3, July 2004, ISSN 1109-2750.
[14] RODRIGUEZ, J. N. and CO.: An Artificial Vision System for Identify and Classify Objects.
International Journal WSEAS Transactions on Computers, Issue 2, Volume 3, April 2004, ISSN 1109-2750.

Address:
Ing. RNDr. Ji astn, CSc.
Technicka 2,
Tel.: +420 541 142 294
Fax: +420 541 142 490
e-mail: stastny@fme.vutbr.cz

Address:
Ing. Martin Minak
Technicka 2,
Tel.: +420 541 142 294
Fax: +420 541 142 490
e-mail: yminar02@stud.fme.vutbr.cz

105

APLIKCIA TERIE GRAFOV V INTELIGENTNOM DOPRAVNOM SYSTME
3

Tom Klietik

ilinsk univerzita

Abstrakt: lnok pojednva o aplikcii terie grafov v inteligentnom dopravnom systme. Konkrtne
o aplikcii Hamiltonovskch cykloch v grafe t. j. riei lohu obchodnho cestujceho. Pomocou danej
lohy me dopravn podnik optimalizova (minimalizova) prepravn nklady. A to tak, e minimalizuje
prepravn trasu. Dan problm je ilustrovan na modelovom prklade a rieen pomocou optimalizanho
software LINGO.

Kov slov: graf, cyklus v grafe, minimalizcie, elov funkcia,, ohranienia, inteligentn dopravn
systm,

Relne systmy opisujeme a skmame pomocou idelnych matematickch objektov. Jednm z takchto idelnych
objektov, ktor vytvorila matematika a ktor sli vyjadreniu mnohch, obsahovo asto celkom odlinch situci, je
graf.

Teria grafov patr medzi najmladie matematick disciplny, ako systematick veda sa sformovala iba v tridsiatych
rokoch minulho storoia. Za jednho z prvch priekopnkov terie grafov sa povauje Leonard Euler, ktor sa
preslvil okrem inho rieenm problmu siedmych mostov mesta Krovca. loha spovala v navrhnut okrunej
cesty, ktor prechdza cez vetky mosty, ale cez kad iba raz. Euler dokzal, e takto okrun cesta neexistuje. Teriu
grafov alej rozpracovali G. R. Kirchhoff, K. Appel, W. Haken, W. R. Hamilton a in. V prspevku budem podrobnejie
rozobera poznatky terie grafov, ktor ako prv definoval William Rowan Hamilton. Jeho poznatky budem aplikova
v tzv. lohe obchodnho cestujceho resp. okrunom problme.

Tto lohu je mon riei dvoma spsobmi a to: heuristickmi metdami alebo pomocou linerneho programovania.
Ja budem okrun problm riei metdami linerneho programovania. Z vpotovho hadiska ide o asovo
mimoriadne nron lohu, a preto nartneme na modelovom prklade monos rieenia pomocou optimalizanho
software LINGO.

V lohe o obchodnom cestujcom treba uri, v akom porad obchodn cestujci, ak vyjde z uritho miesta, navtvi
prve raz ostatn mest a vrti sa nasp do vchodiskovho mesta. Predpoklad sa, e je znmy poet miest v sieti n
a e znme s aj vzdialenosti medzi jednotlivmi mestami c
ij
(i = 1,2,....n, j = 1,2,....n). Uveden problm mono
zapsa ako lohu linerneho programovania kde premenn s bivalentn t. j. ak sa cesta na prslunej trase zrealizuje,
hodnota premennej je 1, ak sa cesta neuskuton, hodnota premennej sa rovn 0.

Uvaujeme, e sa cesta medzi ubovonmi dvoma miestami na t-tom kroku (t = 1,2,....n) uskuton. Do lohy
zavedieme bivalentn premenn x
ijt
. Ak sa cesta na t-tom kroku z miesta i do miesta j uskuton, x
ijt
= 1, ak nie, x
ijt
= 0.
Pre koeficienty c
ij
plat:

,
0,
,
ij
ij
c ak existujecesta z miestai domiesta j
c ak i j
M ak neexistujecesta z miestai domiesta j

'
(1)

3
prspevok je vstupom vedeckho projektu CISKO, . a kol. :Ekonomick aspekty inteligentnho dopravnho systmu
(dopravnej telematiky) v odbore cestnej dopravy, projekt VEGA 1/12349/04, U v iline, FPEDaS

106

lohu mono matematicky potom formulova nasledovne:

1 1 1
min , , , 1, 2,....,
n n n
ij ijt
i j t
c x i j t n

(2)
za podmienok

1 1
1, 1, 2,....,
n n
ijt
i j
x t n

(3)

1 1
1, 1, 2,....,
n n
ijt
j t
x i n

(4)

1 1
1, 1, 2,....,
n n
ijt
i t
x j n

(5)
( 1)
1 1
0, 1, 2,...., , 1, 2,...., 1
n n
ijt jk t
i k
x x j n t n +

(6)
1
1 1
0, 1, 2,....,
n n
ijn jk
i k
x x j n

(7)
{ } 0,1 ijt X (8)

Podmienky (3) a (5) zabezpeuj, e na kad miesto me obchodn cestujci prs len raz aj z kadho miesta odde
ten raz, priom vykon n ciest. Podmienky (6) a (7) zabezpeuj, e ak obchodn cestujci vykon na k-tom kroku
cestu do niektorho miesta s, tak me na k+1-om kroku vyjs ten z toho istho miesta s. Podmienky (3) a (7)
zabezpeuj nepreruenos okrunej cesty a zamedzuj vzniku cyklov. Podmienka (8) je podmienkou bivalentnosti.

V lohe linerneho programovania potom vystupuje n.n.n premennch a n+n+n+n+(n-1)+n ohranien
4
. Rozmery
lohy nie s sce problmom z vpotovho hadiska, pomocou niektorho z optimalizanch programovch balkov
ich mono rchlo vyriei, prinaj vak znan komplikcie pri zostavovan lohy.

Dan lohu meme previes na tzv. Tuckerovu formulciu lohy o obchodnom cestujcom. Do lohy sa zavdzaj
bivalentn premenn x
ij
, ktor oznauj cestu medzi miestom i a miestom j. Aby sa zabrnilo vytvoreniu cyklov
v dopravnej sieti, zavdzaj sa do lohy linerneho programovania alie podmienky v tvare:
1 , 2, 3,...., , i j ij u u nx n i j n i j +
v ktorch premenn ui a uj mu nadobda ubovon hodnoty (s to relne sla priraden miestu i, resp. miestu j).
Potom mono problm obchodnho cestujceho formulova ako lohu linerneho programovania nasledujcim
spsobom:

1 1
min
n n
ij ij
i j
c x

(9)
za podmienok

1
1, 1, 2,....,
n
ij
i
x j n
(10)

1
1, 1, 2,....,
n
ij
j
x i n
(11)

4
V prspevku budem problematiku ilustrova na modelovom prklade s piatimi miestami t.j. loha by mala 125
premennch a 29 ohranien.

107

1 , 2, 3,...., , i j ij u u nx n i j n i j + (12)
{ } 0,1 , 1, 2,.... ij X i j n (13)

V takto naformulovanej lohe bude poet premennch a poet ohranien u vrazne redukovan.
5
. Ale aj tak bude
loha u aj pri 5 miestach pomerne rozsiahla a preto je vhodn poui niektor z optimalizanch softwarovch
balkov. lohu som sa rozhodol riei v programe LINGO, ktor obsahuje pecilny jazyk, pomocou ktorho meme
zpis lohy ete o nieo zjednodui. Nevhodou je to, e vone k dispozcii je iba vubov demo verzia, ktor sce
funguje ako ostr verzia, m vak urit obmedzenia. Jednm z obmedzen je aj poet bivalentnch premennch a to
iba 30 t. j. meme optimalizova iba p miest.

lohu zapeme nsledovne:

MODEL:
!loha obchodnho cestujceho;

SETS:
MIESTO/A,B,C,D,E/:U;
MATICA(MIESTO,MIESTO):X,KM;
ENDSETS

!minimalizcia potu ubehnutch kilometrov;
MIN= @SUM( MATICA:KM*X);
!riadkov a stpcov sty s rovn 1;
@FOR( MIESTO(I):@SUM(MIESTO(J):X(I,J))=1);
@FOR( MIESTO(J):@SUM(MIESTO(I):X(I,J))=1);
!Premenn U mu by ubovoln;
@FOR( MIESTO:@FREE(U));
!Kuhn-Tuckerove podmienky;
@FOR( MIESTO(I)|I#GT#1:@FOR( MIESTO(J)|J#GT#1:
U(I)-U(J)+@SIZE(MIESTO)*X(I,J)<=@SIZE(MIESTO)-1));
!Podmienky bivalentnosti;
@FOR(MATICA:@BIN(X));
DATA:
!Matica vzdialenost;
KM=0 25 60 87 42
25 0 68 12 58
60 68 0 33 40
87 12 33 0 71
42 58 40 71 0;

ENDDATA
END

5
V modelovom prklade to bude iba 25 premennch a 17 ohranien

108

Po spusten prkazom Solve dostaneme nsledujce rieenie:

Global optimal solution found at iteration: 65
Objective value: 152.0000

Variable Value Reduced Cost
U( A) 0.000000 0.000000
U( B) 4.000000 0.000000
U( C) 2.000000 0.000000
U( D) 3.000000 0.000000
U( E) 0.000000 0.000000
X( A, A) 0.000000 0.000000
X( A, B) 0.000000 25.00000
X( A, C) 0.000000 60.00000
X( A, D) 0.000000 87.00000
X( A, E) 1.000000 42.00000
X( B, A) 1.000000 25.00000
X( B, B) 0.000000 0.000000
X( B, C) 0.000000 68.00000
X( B, D) 0.000000 12.00000
X( B, E) 0.000000 58.00000
X( C, A) 0.000000 60.00000
X( C, B) 0.000000 68.00000
X( C, C) 0.000000 0.000000
X( C, D) 1.000000 33.00000
X( C, E) 0.000000 40.00000
X( D, A) 0.000000 87.00000
X( D, B) 1.000000 12.00000
X( D, C) 0.000000 33.00000
X( D, D) 0.000000 0.000000
X( D, E) 0.000000 71.00000
X( E, A) 0.000000 42.00000
X( E, B) 0.000000 58.00000
X( E, C) 1.000000 40.00000
X( E, D) 0.000000 71.00000
X( E, E) 0.000000 0.000000

Z tohto vyplva, e optimlna trasa je nsledovn A-E-C-D-B-A, prepravn nklady na dan trasu bud 152 jednotiek
a rieenie dosiahneme do 65 itercich t. j. neexistuje pri danch vstupnch podmienkach in trasa, ktor by viedla ku
nim nkladom, ako trasa ktor sme zistili pomocou lohy o obchodnom cestujcom.

Literatra:
[1] UNOVSK, L. Modely sieovej analzy. Bratislava : Alfa, 1991.
[2] BREZINA, I.; IVANIOV, Z. Kvantitatvne metdy v logistike. Bratislava : Ekonm, 1999.
[3] DADO, M. et al. TASID technolgie a sluby inteligentnej dopravy, vedecko technick projekt .
AV/819/2002, Cisko, . et al: iastkov projekt Ekonomick a mimoekonomick efekty a hodnotenie
investci, U v iline, 2002-2005
[4] GREGOV, E. Regionlne aspekty globalizcie: nov loha konkurennej schopnosti reginu. In:
Globalizcia a jej socilno-ekonomick dsledky 05: Zbornk z medzinrodnej vedeckej konferencie. Rajeck
Teplice 2005. ISBN: 80-8070-463-5

Adresa:
Ing. Tom Klietik, PhD.
ilinsk univerzita,
010 26 ilina
t..:041/5133221
e-mail: tomas.kliestik@fpedas.utc.sk

109

THE VORTEX-FRACTAL THEORY OF THE UNIVERSE STRUCTURES
Pavel Omera


Abstract: The strength of physical science lies in its ability to explain phenomena as well as make
predictions based on observable, and repeatable phenomena according to known laws. Science is
particularly weak in examining unique, nonrepeatable events. We try to piece together the knowledge of
evolution with the help of biology, informatics and physics to describe a complex vortex structure of the
universe. Evolution is a procedure where matter, energy, and information come together. We would like to
find the plausible unifying mechanisms for an explanation of the vortex systems. Investigators with
specialized training in overlapping disciplines can bring new insights to the area of study, enabling them
to make original contributions. This paper is an attempt to explain a vortex-fractal principle of universe
structures, vortex light rays and what is gravitation.

Keywords: evolution, universe, a basic particle structure, light, gravitation

1. Introduction
Matter has an innate tendency to self-organizing and generating complexity [1-9]. This tendency has been at work since
the birth of the universe, when a pinpoint of featureless matter budded from nothing at all [11]. Irreversibility and
nonlinearity characterize phenomena in every field of complexity. Nonlinearity causes small changes on one level of
organization to produce large effects (anomalies) at the same or higher levels. The smallest of events can lead to the
most massive consequences. We can see an emergent property, which manifests as the result of positive and negative
feedback. But global features of the system cannot be understood only by analyzing the parts separately. Deterministic
chaos arises from the infinitely complex fractal structure (see Fig.1). A fractals form is the same no matter what length
scale we use. By using the techniques of parallelism and massive parallelism in computer simulations we come a little
closer to explaining of basic principles of complex systems. Our attention is directed to the most efficient algorithms of
turbulence simulation, which can help us understand a behavior of very complex fractal objects as a whirl. Chaotic
systems are exquisitely sensitive to initial conditions, and their future behavior can only be reliably predicted over a
short time period. Moreover, the more chaotic system, the less compressible its algorithmic representation.

Turbulence is regarded as one of the grand challenge problems in contemporary high-performance computing.
Despite this astonishing progress during the fifty years since the visionary work of von Neumann, simulating turbulent
fluid flow in realistic way is still largely beyond the capability of today computers. In essence, the common underlying
theme linking complexity of nature with computation, which depends on the emergence of a complex organized
behavior from many simpler cooperative and conflicting interactions between the microscopic components, whether
they are spinning electrons, atoms etc.
Fig. 1 A spiral structure as a fractal

110

Earthquakes, avalanches, and financial crashes do have a common fingerprint: the distribution of events follows a
simple power law [3], [11]. This power law means that the physics of small avalanches is the same as that of large ones.
Self-organization is a natural consequence of time evolution of vast aggregates of simple agents (particles). By making
these agents interact in a more complex way we could create an even greater variety of behavior, such as spiral
structures (see Fig.2b) reminiscent of galaxies (see Fig. 3a,b), hurricanes (see Fig. 3c), tornado and particles of matter.
Nonliving things, for instance crystals, are capable of self-reproduction during growth. Evolution on the edge of chaos
can be extended for nonliving systems [6 8]. The negative forces are caused by negative fluctuation and positive
forces are caused by positive fluctuation and by selection as an influence of boundary conditions.

Fractals seem to be very powerful in describing natural objects on all scales. Fractal dimension and fractal measure, are
crucial parameters for such description [12 - 15]. Many natural objects have self-similarity or partial-self-similarity of
the whole object and its part. Different physical quantities describing properties of fractal objects in E-dimensional
Euclidean space with a fractal dimension D [12]. Fractal dimension D depends on the inter-relation between the number
of repetition and reduction of individual object. There is relationship between the dimensionality and fractal properties
of the matter, which contains the constant of golden mean = ( 5 1)/2 = 0.6180339887. Constant is a special case
of fractal dimension D defined by the condition D (D E + 2) = 1 for E = 3 [12]. Links between inverse coupling
constants of various interactions (gravitational, electromagnetic, weak and strong) in the three-dimensional Euclidean
space are discussed in [13]. Different properties of particles (and interactions between them) correspond to the specific
values of a fractal dimension. Following values (D = 0, E 2, E 1, E) play the most important role in such analysis
[13].

There exists a large body of knowledge about the process of natural evolution that can be used to guide simulations.
This process is well suited for solving problems with unusual constrains where heuristic solutions are not available or
generally lead to unsatisfactory results. Often revolution has an interdisciplinary character. Its central discoveries often
come from people straying outside the normal bounds of their specialties.

Naturalistic explanations of universes origin are speculative [1,9,11]. But does this mean such inquiries are impotent or
without value? The same criticism can be made of any attempt to reconstruct unique events in the past. We cannot
complete our knowledge without answering some of the fundamental question about nature. How does universe begin?
What is turbulence? Above all, in a universe ruled by entropy, drawing inexorably toward greater and greater disorder,
how does order arise? Although the various speculative origin scenarios may be tested against data collected in
laboratory experiments, these models cannot be tested against the actual events in question, i.e., the origin of complex
structures. Such scenarios, then, must ever remain speculation, not knowledge. There is no way to know whether the
results from these experiments tell anything about the way universe itself evolved. In a strict sense, these speculative
reconstruction are not falsifiable; the may only be judged plausible or implausible. In the familiar Popper sense of what
science is, a theory is deemed scientific if it can be checked or tested by experiment against observable, repeatable
phenomena. Behavior of complex nonlinear systems with unpredictable behavior can be demonstrated by a relatively
simple and transparent system as a magnetic pendulum [8]. The idea is to set the pendulum swinging and guess which
attractor will win. Even with just three magnets placed in a triangle, the pendulums motion cannot be predicted. The
unexpected behavior can be extended to physiological and psychiatric medicine, economic forecasting, and perhaps the
evolution of society. A physicist could not truly understand turbulence or complexity unless he understood pendulums.
The chaos began to unite the study of different systems. A simulation brings its own problem: the tiny imprecision build
into each calculation rapidly takes over, because this is a system with sensitive dependence on initial conditions. But
people have to know about disorder if they are going to deals with it. Classical scientists want to discover regularities. It
is not easy to find the grail of science, the Grand Unified Theory or theory of everything. On the other hand there is a
trend in science toward reductionism, the analysis of system only in terms of their constituent parts: quarks,
chromosomes, or neuron. Some scientists believe that they are looking for the whole.

Magnetic fields are most easily understood in terms of magnetic field lines. These field lines define the direction and
strength of the magnetic field at any location in 3D nonlinear space. These magnetic lines have both direction and
strength the closer we are to a magnetic source, then stronger the field lines. The magnetic field lines always begin on
the north poles of a magnet, and end on the south poles. The magnetic field of a magnetic dipole is approximately
proportional to the inverse cube of the distance from the dipole. Therefore, if we double the distance from the magnet,
then the magnetic field strength will be reduced by a factor of 8. Magnetic system of a magnetic pendulum is very
complex [8]. If we know the initial state we cannot predict the final state. Even with just three magnets on the base
plate, we cannot predict the motion. On the other hand, if we know the final state we cannot derive history to the initial
state. The same problem is with universes origin.

111

2. Self-organization of complex systems
Complex systems share certain crucial properties (non-linearity, complex mixture of positive and negative feedback,
nonlinear dynamics, emergence, collective behavior, spontaneous organization, etc.). In the natural world, such systems
include universe, brains, immune systems, ecology, cells, developing embryos, and ant colonies. In the human world,
they include cultural and social systems [5, 6, 8]. Each of these systems is a network of a number of agents acting in
parallel. In a brain, the agents are nerve cells; in ecology, the agents are species; in a cell, the agents are organelles such
as the nucleus and the mitochondria; in an embryo, the agents are cells, and so on. Each agent finds itself in the
environment produced by its interactions with the other agents in the system. It is constantly acting and reacting to what
the other agents are doing. There are emergent properties, the interaction of a lot of parts, the kinds of things that the
group of agents can do collectively, something that the individual cannot. There is no master agent - for example - a
master neuron in the brain. Complex systems have a lot of levels of organization (hierarchical structures), with agents at
any level serving as building blocks for agents at a higher level. An example of a self-organized structure is a whirlpool
(see Fig. 2a). Nonlinearity in feedback processes serves to regulate and control. Evolution is chaos with feedback [17].

3. Vortex structures
Perhaps vortex structures with vortex lines, such as are created approximately in a whirlpool or in a tornado are a
plausible speculation of elementary particle structures. The whirlpool-structure (a turbulent eddy) with a funnel shape
can have for example a water outlet of the bath or in the PET-bottle (see Fig. 2a). The streamlines are spirals (or circles)
about a vortex axis, similar to the lines of the magnetic field round a wire carrying a current. The velocity v of the flow
is inversely proportional to the distance from the vortex axis as can be observed at the drain hole of a bath-tub [16,18].
Speed v depends on friction. In the bath-tub the core is replaced by air. For a hurricane, the core is called the eye. If two
or more vortex lines are parallel side by side in the fluid, the core of each vortex line must move in the velocity field
arising from other vortex lines. So two parallel vortex filaments with opposite rotation (spin) follow straight lines
course side by side (see Fig. 6a), whereas with the same spin they dance round each other (Fig. 6b). If three or more
vortexes are working together (see Fig. 6c), a more complex structure can be created. If one bends a vortex line into a
closed ring, then the vortex ring moves with unchanging shape in a strait line: each part of the ring must move in the
velocity field of all the other parts. For example experienced smokers can blow smoke-rings (see Fig. 4).
a) b)
Fig.2 a) The vortex in the PET-bottle
b) The vortex model

112

Fig. 3 Examples of spiral structures: a), b) galaxies, c) the Earths hurricane

Fig. 4 The annular structure that can be created from vortexes (for example the electron)

The presented theory is based on the following assumptions and hypotheses:
1. There is a hidden substance in the universe that contains very small sub-particles with unmeasurable mass.
2. We will call these sub-particles osmerons (osmero was the the name of the deity in ancient Egypt for 4
pairs of gods).
3. Vortices and annular structures (rotational structures) can be created from these sub-particles.
4. There are two types of the vortices V
B
, V
T
with opposite flow of the energy E (see Fig. 5a,b)
5. The vortex pair with two V
B
, or two V
T
can create two types of the pair: the same rotation with parallel
rotational axes or the contra rotation with the parallel rotational axes (see Fig 6a,b and 13b,c).
6. The vortex V
B
and the vortex V
T
can create the pair with fore head orientation on the same rotational axis.
7. Vortices V
B
and vortices V
T
can create the chain (string) structure (see Fig. 14).
8. The vortex or annular structures can change from one structure to another very quickly.
9. There is sufficient amount of the accessible energy E.

If a semi-fractal description of nature is plausible for us, we can imagine that many objects of the universe are the
fractal-vortices [19]. If we see vortex structures in a macro-world (as spiral galaxies) and in real world (as the whirlpool
of bath-tub shown in Fig. 5a and as the tornado in Fig. 5b or hurricane) it can be probable that particles in micro-world
have similar fractal-vortex structures. The flow of energy E in the tornado-vortex V
T
and in the bath-tub-vortex V
B
has
opposite direction (see Fig. 5a and Fig. 5b). The pressure p is higher at the bottom of vortices. To create a vortex
structure we need a minimum value of energy a quantum of energy and the sufficient number of sub-particles
(osmerons). Perhaps there is a relation between Plancks constant and the minimum energy of the vortex structure V
B
or

113

V
T
.

a) b)
Fig. 5 Vortex structures: a) the vortex V
B
at the drain hole of bath-tub, b) the tornado-vortex V
T

The forces between two vortices and motion of two types of the vortex pairs are shown in Fig. 6. The behavior of the
vortex pair shown in Fig. 6a can help us to explain the expansion of the universe (the Hubbles law with anti-
gravitational forces F
ag
). The behavior of the vortex pair shown in Fig. 6b can help us to explain the disc and spiral
shape of the Milky Way with vortex-gravitational forces F
g
. But it cannot help us to explain the spherical shape of the
universe bodies as the Sun or the Earth. There are an another forces that can occur at the vortex structures (see Fig. 16a,
Fig. 17a). More then two vortices can form the complex structures (for example three vortices shown in Fig. 6c or more
vortices in Fig. 16).

Fig. 6 The motion and gravitational forces F
g
and anti-gravitational forces F
ag
of the vortex pair V
1
and V
2
:
a) with the contra rotation of
1
,
2

b) with the equal rotation of
1
,
2

c) the motion of 3 vortex particles p
1
, p
2
, and p
3
(all particles have the same direction of rotation).

114

We can build the chain structure from the vortices (see Fig. 10, 14). The forces and motion of two vortex pair is shown
in Fig. 3a. There are two possible lines between pairs, which depends on the direction of rotation P
1
and P
2
. The lines
between vortex pair P
1
and P
2
(see Fig. 7a) with the same rotation are shown in Fig. 7b. The arrangement of lines for the
contra rotation of pairs is shown on Fig.7c. More vortices can create a vortex ray (see Fig.8a).
a) b) c)
Fig. 7 a) The forces and motion of two vortex pairs with opposite flow of the energy E
b) Lines of hidden sub-particles between pairs with the same rotation of vortices
c) Lines of hidden sub-particles between pairs with the opposite rotation of vortex pairs

The value of the frequency of vortexs vibrations along the rotational axis increases the accumulated energy in the
complex vortex-structure, for example in photon rays (see Fig. 8). The photon-ray can be the vortex row (chain) with a
very small mass [1]. Every photon-ray can have opposite rotation with regard neighbor photon-rays (see Fig. 8a). The
number of vortex-rays in circle structure of the stream must be even (see Fig. 8a) to form divergent rays (see Fig 8b).
Figures 8a,b,c can help us to explain the behavior of vortex-rays: as the energy flow of particle structures (vortices) and
the wave transport of energy. Because the side rays in Fig.8b have no neighbor rays, they are deviated by forces F from
the neighbor that is near to the center of ray-flow. It can explain the wave behavior of the light flow of photons as
vortex rays behind the hole. The laser beam structure in Fig.8c can be explained with particle motion shown in Fig. 6c.
Photon-vortex-structure V (or V
p
) can be created from annular vortex-electron structure e by two ways:
a) by the change of the shape of vortex-electron-structure (see Fig. 9a),
b) by cutting the closed electron-structure (see Fig.9b).
a) b) c)
Fig. 8 Vortex rays of the light: a) the spin example of vortex-rays (if f = f
1
= 1 Hz then E = E
p
= h . f
1
),
b) the spreading of vortex-rays behind the hole (the wave refraction of light),
c) the spreading of vortex-rays with the same rotation (the laser beam).

115

Fig. 9 The release of vortex pair V
p
from vortex-electron structure e:
a) with change of the shape of vortex-electron-structure,
b) by cutting the closed electron-structure.

4. Elementary particles
Macroscopic matter consists of molecules, which are built out of atoms; these atoms define the elements. From the four
elements of the ancient Greeks we moved to 92 natural and about 20 artificial elements, each of which may appear in
several isotopes [16]. We learned that each atom consists of a small nucleus and a large electron cloud around it. The
nucleus is again composed of P charged protons and N neutral neutrons. A neutral atom thus has P electrons
determining its chemical behavior. For the same, P + N different atomic weights describe isotopes. In 1932 we had just
three basic particles, the proton, the neutron, and the electron, to build all known tangible matter from. The neutrino
(little neutron) was proposed in 1930 by Pauli to take away some of the energy, momentum and spin arising in beta-
decay of a neutron into a proton and electron. Neutrinos show very little interest in any reactions and have zero or very
small mass. Zero neutrino mass means smaller than measurable. Each particle seems to have an antiparticle of the
same mass but with a different sign of the electric charge. For example positron balances the electron, and the
antiproton was found in a particle accelerator built particularly to produce such antiprotons according to E = mc
2
. Basic
particles consist from smaller parts. The proton consists of two up and one down quark, or in short: proton = (uud), the
neuron = (udd). Quarks have fractional electric charges t1/3, t2/3, which explains the existence of the double particles
consisting of three quarks with 2/3 charge. Quarks appear in six types (six flavors): u, d, c, s, t, b (=up, down, charm,
strange, top, bottom). These six flavors are grouped into three generations, which correspond to the three leptons. Thus
the mesons with two quarks are Bosons, and the baryons with three quarks are Fermions. Each of the quarks and leptons
has its antiparticles; mesons are formed with one quark and one antiquark. Quarks, in contrast to leptons, appear in three
colors. We have at present 36 different quarks of various colors and favors, and 12 different leptons, or 48
fundamental particles all together. The masses of the three (up or down) quarks forming the proton or neutron are much
smaller than the mass of that nucleon; most of the mass is hidden in the interaction energy due to the enormous color
forces between quarks [16]. Perhaps it has something with a rotational vortex structure of nucleons. One very
speculative imagination of the electron structure is presented in Fig. 4. Experimental investigations of possible types of
reactions show that certain particle numbers are conserved in the sense that the number of incoming particles of this
type must equal the number of particles of the same type after the reaction is over. Each of the three-lepton generations
has the own conserved number of particles, and so do the quarks for all generation together. Perhaps during lepton
generation is a rotational structure changed (see Fig. 9). Leptons do not consists of quarks. Also the electric charge is
always conserved, whereas the mass can be transformed into energy and back. Also, electric charge is conserved with
antiparticles having opposite charge. However, antiparticles always count negative for the particle number and charge
(not for mass). Thus radiation energy can form an electron-positron pair since then the number of e-leptons is still zero.
So normal matter needs only the electron, u, and d as constituents, a nice simplification.
Investigators with specialized training in overlapping disciplines can bring new insights to an area of study, enabling
them to make original contributions. It can present the ways universes structure could have arisen. An open question
remains; what is gravity [16].

116

With these conservation laws we now understand a crucial difference between mesons (quark + antiquark) and baryons
(three quarks). The meson number is not conserved since the quark and antiquark can annihilate each other. The baryon
number is conserved since their quark number. For example the free neutron decays after 15 minutes into a proton, an
electron, and an anti-electron-neutrino. This is allowed since both the proton and the neutron has the same baryon
number of one. But a neutron star or pulsar does not decay into protons because of the strong forces between neutrons.
It seems the force between two quarks is quite strong and for large distances independent of distance. We cannot
observe quarks isolated somewhat like north and south poles of a magnetic dipole. So if we try to pull quarks apart, we
need so much energy that we merely create new particles. Only white combinations of quarks, where the color forces
have cancelled each other (like quark-antiquark, or three quarks with the three fundamental colors), are observed as
isolated particles [16].

We still feel gravitation since there are no negative masses, in contrast to positive and negative electric charges, which
cancel each other in their force over long distances. Electric forces do not propagate with infinite velocity but only with
the large but finite light velocity c. Perhaps c is maximum velocity of spreading in the hidden substance. Light waves
are called photons in quantum theory or Coulomb forces are transmitted via the quasi-particles called photons.
Similarly, gravitational forces propagate with velocity c, perhaps with the help of quantized gravity waves called
gravitons (not yet detected as quantized quasi-particles) [16]. Quite generally at present, forces are supposed to come
from the exchange of intermediate Bosons (virtual particles). Virtual particles are packets of energy E = mc
2
with a
short lifetime t, such that the energy-time uncertainty relation E t h/2 allows their creation. The color forces
between quarks are transmitted by gluons (i.e., by particles glueing the quarks together) of zero mass. They bind three
quarks together as a nucleon (proton or neutron). At some distance from this nucleon some remnant of color forces is
felt, since they have not canceled each other exactly. Coulomb forces and gravitation are felt over infinitive distances
without exponential cut-off and thus have zero mass. Color forces also must have infinite range since otherwise we
could isolate single quarks; thus also the gluons are massless. The weak interaction covers only very short distances
because of the large mass of the corresponding intermediate Bosons. The iteration energy remains the same if all spin
reverse their orientation.

What has been described here is the so called standard model, which includes color forces. The Grand Unified Theory
(GUT) combines it with electromagnetic and weak forces, and the Theory of Everything would include gravity.

Magnetism we understand on the basis of suitable models. How a spontaneous magnetization can be formed? A very
speculative imagination how an electromagnetic field can be created, during a jump of an electron between two atoms,
is presented in [19]. Magnetic and electric lines are presented as a vortex flow of hidden substance (subparticles with a
mass smaller than measurable).

Finally I would like to present my very speculative origin scenario. It is possible that all very complex systems exist in
anomalous states, as vortex structures. These anomalous states have a hierarchical structure. May be 3D-matter is a first
anomalous stage after a collision of supervortex spaces. At the second level is the origin of living systems. At the
third level is a brain with a consciousness. There is no greater anomaly in nature than matter that can live and can have
a consciousness.
a) b)
Fig. 10 A speculative structure of a photon flow (electromagnetic lines)

117

If a fractal description with the fractal dimension is plausible for us, we can imagine that almost all objects of the
universe are fractal vortices with a different fractal dimension. If we see vortex structures in a macro-world (as spiral
galaxies) and real world (as tornado, whirlpools, and hurricanes) it can be probable that particles in micro-world have
the same fractal-vortex structure [20]. To create a fractal vortex we need a minimum value of energy a quantum of
energy. There can be a relation between Plancks constant h and the fractal dimension of the vortex. The value of
frequency of vortexs vibrations (see Fig. 14) increases the accumulated energy of a vortex structure in coincidence
with physical law for photons energy. A photon flow can be vortex row with a very small mass (see Fig. 10, 14). It can
be an opened structure created by cutting the closed electron structure in Fig.10a). May be it is a better model than a
classical planetary model. Perhaps our universe is not a superstring space but a supervortex space. Vortex structures
can explain magnetism, perhaps gravity etc. Vortices can attract each other using their different polarities (see Fig. 6b).
Vortices with their rotation have inertia, which explains what the mass of matter can be to compare with a hidden
substance (subparticles without mass). We can see, for example, the fractal vortex structure on Jupiters weather or
Earths weather (see Fig. 3c). Perhaps vortex structures will be a plausible speculation but research is needed to test it.

The increased awareness from other scientific communities, such as biology and mathematics, promises new insights
and new opportunities. There is much to accomplish and there are many open questions. Interest from diverse
disciplines continues to increase and evolution of complex structures is becoming more generally accepted as a
paradigm for imagination of basic principles of nature. We can make our model more complex, and more faithful to
reality, or we can make it simpler and easier to handle (to generalize and abstract). Some patterns are fractal, exhibiting
structures self-similar in scale.

5. The annular structure of the basic particles
What are the shape and the structure of the basic particles as the electron, the proton, and the neutron? All these
particles have the spin. We use the definition of the spin s as the ratio the sum of the threads (coils) c
1
and c
2
to the
number of electron-threads

C
e
(see Fig. 11 and Fig. 12). The proton and the neutron are described as the group of three
quarks [20]. Our attempt to use the quarks: u and d to form the annular shape of the proton and the neutron is presented
in Fig. 11, 16 (not in right scale the electron is smaller and the proton and the neutron are thicker and larger).

Fig. 11 The annular and close energy structure of the basic particles with their spins s ( fractional electric charges)
(quarks u and d are only abstract and open substructures building blocks of the proton and the neutron)

The spin structure of basic particles can be explained with the description in Fig. 12 where circle structures are opened
to could be easily drawn (from Fig. 12a to Fig.12d). The closed structure from Fig. 12a is in Fig. 12e and from Fig. 12b
in Fig. 12f. Zero spin of the neutron can be form from nonzero threads (coils) c
1
and c
2
: c
1
= - c
2
.

118

Fig. 12 The spin structure of particles (Fig12e is closed structure from Fig. 12a, Fig. 12f is closed structure from Fig.
12b)

Forces between two vortex pairs with different axes and directions of the energy flow are presented in Fig. 13.
Fig. 13 Forces between two vortex pairs in the different arrangement

Forces between vortices in the photons (or in the gluons) flow of the energy are presented in Fig. 14.
Fig. 14 Forces and oscillations of the photon flow

119

Fig. 15 Forces between two electrons

The forces between two electrons and their trajectories t
r
are shown in Fig. 15. Electrons with the same direction of
trajectory t
r
form the electron rays. It occurs in two cases of electron-electron orientations (it is shown in Fig. 15 on the
right). Two neighbor electrons in the electron-ray slightly attract each other (with the force F
a
) due to the opposite
direction of magnetic lines. All reaction forces in the electron-ray have the same direction (the same as the trajectory t
r
).
The behavior of electron rays is similar to the behavior of photon-rays described in Fig. 8a. The strong nucleus forces
can be explained with the vortex bonds V
p1
and V
p2
between protons (see Fig. 16a) and neutrons.

c)
Fig. 16 a) The strong nucleus forces by vortex bonds V
p1
and V
p2
between protons
b) The spin structure of the proton or the neutron)
c) The model of the protons (or neutrons) structure (the same as in Fig. 16b)

120

a) b)

Fig. 17 a) The vortex bonds V
1a
, V
1b
and V
2a
, V
2b
between two electrons that can transport energy
b) Gravitation forces between two particles with mass m
1
and m
2
.

The forces F
e
(Coulombs low) between the electron and the proton depends on line density in the area S which is
inverse proportional to square distance d
2 [
[20].

The light beam (the photon flow) is a complex structure that can translate the energy by excitation of the vortex row.
There is the distance between two vortices where the couple force F has maximum value (see Fig. 14). Around this
position every vortex can oscillate as was presented in the center of Fig 14 (the wave theory of light). The vortices
(photons) in the light flow oscillate (vibrate) to translate energy (it is not similar to particle translation). But one vortex
pair (photon) can move and translate the energy separately. The strong nucleus forces can be explained with the vortex
bonds between vortex pairs V
p1
and V
p2
(see Fig. 16a).

The gravitational forces F
G
depend proportional on the density of magnetic lines in the area S (see Fig.17b) which
decreases indirectly with square distance d
2
and increases proportional with the mass m
1
and m
2
(Newtons low). The
higher number of protons vortex bonds is between protons in the nucleus (see Fig. 16a) the higher density of the
magnetic field will be in the area S and the gravitational forces will be higher. The higher energy has nucleus the higher
number of vortex bonds will be created and stronger forces are between protons and neutrons (the strong interaction).
The main component of gravitational force F
G
is the complex magnetic field with two-way magnetic lines (see Fig. 17b)
that alternate each another. There is an analogy between electric forces (Coulombs low) and magnetic forces for a
gravitational influence (Newtons low). All the universe is fill up with the magnetic lines. To explain the structure of the
magnetic lines we need lower sub-subparticles than are the basic particles as electron, proton, and neutron. This sub-
subparticles can form flexible magnetic lines with similar structure as photons but they have to be smaller (perhaps
they can be gluons).

121

Fig. 18 Forces between the proton p and electron e

Position of the electron from the proton during the levitation depends on two different types of forces. The magnetic
force F
m
repels the electron from the proton (see Fig. 18a) and the charge-reaction force F
r
of the electron attracts him.
The charge forces F
r
work on the principle of activity and reaction (the same principle as the rocket engine but with the
hidden substance as is shown in Fig. 15. The magnetic repulsion forces are stronger when the particles are closer. It
hangs where this upward repulsion balances the downward force of the charges, that is, at the point of equilibrium
where the total force is zero. If the electron were not spinning, the magnetic torque would turn it over. When the
electron is spinning, the torque acts gyroscopically and the axis does not overturn but rotates about the direction of the
magnetic protons field. This rotation is called precession. For the electron to remain suspended, equilibrium is not
enough. The equilibrium must also be stable, so a slight horizontal or vertical displacement produces a force pushing
the electron back toward the equilibrium point. The reaction force F
r
of the electron and the strength F
m
of
magnetization between the proton and the electron determine the equilibrium distance d where magnetism balances
rocket force F
r
. Slight changes of temperature alter the magnetization of particles.

6. Conclusions
The annular-vortex model might be better than a classical planetary one. Our universe might be considered as
supervortex space. Vortex structures can explain the electromagnetic field, perhaps gravitation too. Vortices can
attract each other using their different polarities (see Fig. 13a). Plancks constant h might be the energy E
p
of one vortex
pair V
p
(see Fig. 8a). Close vortex structures (Fermions) with their rotation have the inertia, which explains what the
mass of matter can be compared with a hidden substance (the sub-particles osmerons with very small and
unmeasurable size in the vortex structure). The radiation is an open vortex structure (Bosons, for example light with
photons) and matter is a close vortex structure with mass (for example: electrons, protons, and neutrons and follows
complex structures as the nucleus {see Fig. 16a} etc.). Electron structures rotate and proton (neutron) structures need
not rotate to have a rotating magnetic field. Both (the electron and the proton) have the rotating magnetic field. Vortex
structures might be a plausible speculation for a computer models and calculation. Fractals seem to be very powerful in
describing natural objects on all scales.

References
[1] THAXTON, Ch. B.; BRADLEY, W. L.; OLSEN, R. L. The Mystery of Lifes Origin: Reassessing Current
Theories, New York : Philosophical Libery, 1984.
[2] DAWKINS, R. The Selfish Gene. Oxford : Oxford Univrsity Press, 1976.
[3] KAUFFMAN, S. A. Investigations. New York : Oxford University Press, 2000.

122

[4] PRIGOGINE, I.; STENDERS, I. Order out of Chaos. Flamingo, 1985.
[5] OMERA, P. Complex Adaptive Systems. Proceedings of MENDEL2001, Brno : Czech Republic (2001) 137
143.
[6] OMERA, P. Complex Evolutionary Structures. Proceedings of MENDEL02, Brno : Czech Republic (2002)
109 116.
[7] OMERA, P. Evolvable Controllers using Paralel Evolutionary Algorithms. Proceedings of MENDEL2003,
Brno : Czech Republic (2003) 126 - 132.
[8] OMERA, P. Evolution of System with Unpredictable Behavior, Proceedings of MENDEL2004, Brno : Czech
Republic (2004) 1 - 6. Omera, P.: Genetic Algorithms and their Aplications, the habilit work, in Czech
language 2002.
[9] WADROP, M. M. Complexity The Emerging Science at Edge of Order and Chaos. Viking 1993.
[10] OMERA, P.; POPELKA, O.; PANACEK, T. Parallel Grammatical Evolution, Proceedings of
MENDEL2005, Brno : Czech Republic (2005).
[11] COVENEY, P.; HIGHFIELD, R. Frontiers of Complexity. Faber and Faber, 1996.
[12] ZMEKAL, O.; NEZADAL, M.; BUCHNICEK, M. Fractal-Cantorial geometry. Hausdorf dimension and
fundamental laws of physics, Chaos, Solitons and Fractals 17 (2003) 113-119.
[13] ZMEKAL, O.; NEZADAL, M.; BUCHNICEK, M. Coupling constants in fractal and cantorian physics.
Solitons and Fractals (2005) article in press
[14] EL NACHIE MS. On the exact mass spectrum of quark. Chaos, Soliton & Fractals 2002,14;369-76
[15] EL NACHIE MS. Quantum gravity. Clifford algebras and fundamental constant of nature. Chaos, Soliton &
Fractals 2002,14;437-50
[16] STAUFFER, D.; STANLEY, H. E. From Newton to Mandelbrot. A Primer in Theoretical Physics with Fractal
for the Personal Computer, Springer-Verlag Berlin Heidelberg, 1996.
[17] GLICK, J. Chaos - Making a New Science. Vintage, 1998.
[18] CAPRA, F. The Web of Life. HarperCollins Publishers, 1996.
[19] OMERA, P. Evolution of the univers structures. Proceedings of MENDEL 2005, Brno : Czech Republic
(2005) 1-6.
[20] OMERA, P. The Vortex-fractal Theory of the Gravitation, Proceedings of MENDEL 2005, Brno : Czech
Republic (2005) 7-14.

Address:
Technicka 2,
Tel.: +420 541 142 294
Fax: +420 541 142 490

123

VORTEX-FRACTAL PHYSICS
Pavel Omera


Abstract: We would like to find the plausible unifying mechanisms for an explanation of the vortex
systems. This paper is an attempt to attain a new and profound understanding of natures behavior as a
vortex-fractal principle for everything. There is a vortex explanation of polarization, the diffraction
grating, and we compare quantum electrodynamics (QED) with the vortex-fractal description. This new
approach can be called physics of vortex structures (FVS).

Keywords: vortex, polarization, diffraction grating, basic particle structure, light, gravitation.

1. Introduction
The electrical force, like a gravitational force, decreases inversely as the square of distance between charges. This
relationship is called Coulombs law. There are two kinds of matter, which we can all positive and negative. Like
kinds repel and unlike kinds attract unlike gravity where there is only attraction. But it is not precisely true when
charges are moving the electrical forces depends also on the motion of charges in a complicated way [2]. One part of
the force between moving charges we call the magnetic force. It is really one aspect of a vortex effect. That is why we
call the subject electromagnetism. We find, from experiment, that the force that acts on a particular charge no
matter how many other charges there are or how they are moving depends only on the position of that particular
charge, on the velocity of the charge, and on the amount of charge [2]. We can write the force on a charge q moving
with a velocity v as
F = q (E + v x B). (1.1)

We call E the electric field and B the magnetic field at the location of the charge. There is still something there when
the charge is removed. The field we consider as mathematical function of position and time. For an arbitrary closed
surface, the net outward flow or flux is the average outward normal component of the velocity, times the area of the
surface:
Flux = (average normal component).(surface area). (1.2)

In the case of an electric field, we can mathematically define something analogous to an outflow, and we again call it
flux, but of course it is not the flow of any substance, because the electric field is not the velocity of anything [2]. In the
vortex-fractal hypothesis of electron structure [6] it can be velocity of osmerons [8]. Osmerons are sub-particles that
create for example a vortex structure of an electron [7]. The name osmeron was derived from the name of the Egyptian
deity with 4 pairs of gods for primary creative forces (from a chaos beginning). Osmerons are too small that is why they
have unmeasurable size and mass (see Fig. 1).
Fig. 1 A vortex structure of light rays

124

Physics has a history of synthesizing many phenomena into a few theories. For example it was discovered that heat
phenomena are easily understandable from the law of motion. The theory of gravitation, on the other hand, was not
understandable from the other theories. Gravitation is, so far, not understandable in terms of other phenomena [4].
Quantum mechanics thus supplied the theory behind chemistry. So, fundamental theoretical chemistry is really physics.
The theory of interaction of light and matter is called quantum electrodynamics QED. There constants in quantum
electrodynamics, that have been measured and calculated with very high accuracy. QED theory is probably not too far
off these calculations. It is necessary to distinguish two questions: how Nature works and why Nature works that way.

There is another question: What holds the nucleus together? [2]. In nucleus there is several protons, all of which are
positive. Why dont they push themselves apart? It turns out that in nuclei there are, in edition to electrical forces,
nonelectrical forces, called nuclear forces, which are greater than the electrical repulsion. The nucleus forces, however,
have a short range their force falls off much more rapidly than 1/r
2
[2]. It seems to me that this is one complex energy
structure created from protons and neutrons connected by vortex bonds [7, 8]. In this complex nucleus structure energy
is running in one complex loop {7]. We may ask, finally, what holds a negatively charged electron (since it has no
nuclear forces). If an electron is all made of one kind of substance, each part should repel the other parts [2]. If we
accept the vortex electron-structure it can be vortex forces between photons from which the electron is created [7, 8].
What is the charge? It can be something that has relation to the flow of osmerons though annular electron structure
(ring). Electrical force, like gravitational force, decreases inversely as the square of distance between charges. There
must be the same principle. Perhaps there is very small escape of energy from nucleus complex loop in vortex bonds
(gravitons small number of osmerons) that creates gravitational field.

2. Diffracting grating
A particular color o f light can be split one more time in a different way, according to its so-called polarization. Thus
light is something like raindrops each little lump of light is called a photon - and if the light is all one color, all the
raindrops are the same size and vortex structure (see Fig. 1). The human eye is a very good instrument: it takes only
about five or six photons to active a nerve cell and send a message to the brain [4]. Light goes in straight lines; it bends
when it goes into water; when it is reflected from a surface like a mirror, the angle at which the light hits the surface is
equal to the angle at which it leaves the surface. Light can be separated into color; you can see beautiful colors on a
mud puddle when there is a little bit of oil on it (because the oil films thickness is not exactly uniform), lens focuses
light, and so on [4]. When a photon comes down on the surface of the glass, it interacts with electrons throughout the
glass, not just on the surface. The photon and electron do some kind of a dance, the net result is the same as if the
photon hit only the surface [4].

There is the relationship between the thickness of a sheet of glass and partial reflection [4]. It appears that partial
reflection can be turned off or amplified by the presence of an additional surface. It demonstrates a phenomenon
called interference. As the thickness of the glass increases, partial reflection goes a repeating of zero to 16%, with no
signs of dying out [4]. This strange phenomenon of partial reflection by two surfaces can be explained for intense light
by a theory of waves, but the wave theory cannot explain how the detector makes equally loud clicks as the light gets
dimmer. Quantum electrodynamics resolves this wave/particle duality by the probability that a photon will hit a
detector. Grand principle of QED: The probability of an event is equal to the square of the length of arrow called
probability amplitude. General rule of QED: Draw an arrow for each way and then combine the arrows (add them)
by hooking the head of one to the tail of the next [4] (see Fig. 2c). Every phenomenon about light that has been
observed in detail can be explained by the theory of quantum electrodynamics (QED) [4]. In Fig. 2a,b the same
diffraction on DVD surface is explained by the vortex structures with the same result like in Fig. 2c. Some osmerons
trajectory are changed or absorbed, due to symmetry of vortex structure is changed to asymmetric (compare with the
symmetrical vortex structure in Fig. 1). In the blue rays the diameter D
2
is greater then D
2
at the red rays. The diffraction
for red rays is greater then for blue rays because the asymmetry of vortex structure of red light is higher then at blue
light (see Fig. 2b). A diffracting grating with grooves at the right distance for red light also works for other colors. If
you shine white light down onto the grating, red light comes out at one place, orange light comes out slightly above it,
followed by yellow, green, and blue light all the colors of rainbow. Where there is a series of grooves close together,
you can often see colors when you hold a CD disc or better DVD disc under bright light at correct angels (see Fig.
2d and Fig. 3). What is interesting that one light ray doesnt really travel only in a strait line; it smells the neighboring
paths around it, and uses s small core of nearby space [4]. When the one slot b is smaller, the detector D starts clicking
not only in the position on strait line (photon 1 in Fig. 4). When we have two slots and the distance d between them is
decreasing (see Fig. 4) we can see interference between photon 1 and photon 2. This is an example of the uncertainty
principle; there is a kind of complementary between knowledge of where light goes trough two holes and where it

125

goes afterwards precise knowledge of both is impossible. So the idea that light goes in a straight line is a convenient
approximation to describe what happens in the world that is familiar to us [4].
Fig. 2 Diffraction on DVD surface

126

Fig. 3 Example: how we can measure the wave length of light (for example: red laser)

Fig. 4 Photons (or electrons] coming though sheet with two holes

Sometimes our observations (measurements) involve condition that are special and represent in fact a limited
experience with nature. It is a small section only of natural phenomena that one gets from direct experience. It is only
through refined measurements and careful experimentation that we have a wider vision [4]. And then we see
unexpected things; we see things that are far from what we would guess - far from what we could have imagined but
just to comprehend those things, which are there. For two slots problem in Fig. 4 there is one simplification at least.
Electrons behave in this respect in exactly the same way as photons; they are both screwy in exactly the same way [4].

3. Polarization
Polarization produces a large number of different possible couplings. All possible combinations of polarized electrons
and photons do not couple (see Fig. 6). So far, no fundamental spin 0 particles have been found. But we can see the
vortex rings in the water [9], in the air making a travelling vortex ring [2]. It is clear form Fig. 5 that in the points 1,2
there is not phase change. But osmerons in the points 3,4 have an opposite phase shift which change the symmetry of
vortex structure in the light ray.

127

Fig. 5 Polarization of light ray that has a vortex structure

Fig. 6 Forces that are between the electron and the proton (a, b) and their outer vortex structures (c, d)

Fig. 7 Vortex structure of electron and anti-electron with two face orientation

128

Fig. 8 Vortex structure of the proton and the antiproton

4. Conclusions
The vortex models might be better than a classical one. Our universe might be considered as supervortex space.
Vortex structures can explain the electromagnetic field, perhaps gravitation too. Vortices can attract each other using
their different polarities. Plancks constant h might be the energy E
p
of one vortex pair V
p
. Vacuum is a space full of
osmerons. There can be non-homogeneous density of osmerons from which vortex structures are created. Close vortex
structures with their rotation have the inertia, which explains what the mass of matter can be compared open structures
that create radiation. The radiation is an open vortex structure (for example light with photons) and matter is a close
vortex structure with mass (for example: electrons, protons, and neutrons and followed with complex structures as the
nucleus. All things are only complex vortex structures. Because the electron structure has the face and the back we can
distinguish two states 0 and 1. This can be used in coding of electron computers and electron memories
(analogy to quantum computers). Fractal dimensions seem to be very powerful in describing natural objects on all
scales. The behavior and creation of annular-vortex structures with zero spin was described in [2], [9]. Water forms a
spiraling, funnel-shaped vortex as it drains from 1.5 or 2-liter soda PET-bottle. A simple connector device from two
original lids with 1cm hole allows the water drain into a second bottle. Fill only one of soda bottles about two-thirds full
of water. Place the two bottles on a table with the filled bottle on top. Watch the water slowly drip down into the lower
bottle as air simultaneously bubbles up into the top bottle. The flow of water may come to a complete stop. To create
vortex structure it is necessary add chaotic movement (shacking) or better rapidly rotate the top bottle in a circle a few
times. Notice the shape of the top vortex and there is second vortex in the lower bottle (in principle it is tornado
structure destroyed by gravity). The whole assembly can then be inverted and the process repeated. This simple model
demonstrates the basic principle of vortex structures and how we can come from chaos to self-organized structure (the
basic principle of evolution for nonliving systems). This knowledge can be used when we trying quickly get the liquid
from a tank (canister).

References
[1] FEYNMAN, R. P.; LEIGHTON, R. B.; SANDS, M. The Feynman Lectures on Physics, volume I, Addison-
Wesley publishing company, 1977.
[2] FEYNMAN, R. P.; LEIGHTON, R. B.; SAMDS, M. The Feynman Lectures on Physics, volume II, Addison-
[3] FEYNMAN, R. P.; LEIGHTON, R. B.; SANDS, M. The Feynman Lectures on Physics, volume III, Addison-
[4] FEYNMAN, R. P. QED The Strange Theory of Light and Matter. Princeton University Press, 1988.
[5] FEYNMAN, R. P. The Character of Physical Law, Penguin Books, 1992.

129

[6] OMERA, P. Evolution of univers structures, Proceedings of MENDEL 2005, Brno : Czech Republic (2005)
1-6.
[7] OMERA, P. The Vortex-fractal Theory of the Gravitation, Proceedings of MENDEL2005, Brno : Czech
Republic (2005) 7-14.
[8] OMERA, P. The Vortex-fractal Theory of Universe Structures. Proceedings of the 4
th
International
Conference on Soft Computing ICSC2006, January 27, 2006, Kunovice, Czech Republic
[9] LIM, T. T.; NICLES, T. B. Instability and reconnection in thehead on collision of two vortex rings, letter to
Nature, vol. 357, May 1992.
[10] WALLRAFF, A.; LUKASHENKO, A.; LISENFELD, J.; KEMP, A.; FISTUL, M. V.; KOVAL, Y. &
USTINOV, A.V. Quantum dynamics of a single vortex, letters to nature, vol.425, September 2003.

Address:
Technicka 2,
Tel.: +420 541 142 294
Fax: +420 541 142 490

130

131

VZNAM MONITOROVN POTAOVCH ST
Imrich Rukovansk


Abstrakt: Zda potaov s pracuje efektivn, meme prokazateln zjistit pouze sledovnm jeho aktivit
a to menm.Tuto innost lze provdt vyuitm monitor (sovch analyztor), ktermi lze sledovat a
destky rznch vkonnostnch parametr (propustnost, vyuit CPU, vyuit komunikanch linek mezi
uzly, zk msta) a to na rznch rovnch st. Avak charakter men se li ppad od ppadu a zvis
na tom, jak vkonnostn parametry sledujeme, zda sledujeme pouze vybranou st st, nebo s jako
celek, zda vyuvme hardwareov, softwarov monitor, nebo kombinaci obou.Dleitost monitorovn
potaovch st, jako i rznorodost praktickch men dokladuje tento pspvek.

Klov slova:potaov s, monitor, men, bezdrtov spoje, WiFi, pepna, smrova, server,
propustnost, zt, vkonnost.

vod
Sloitost, rozlehlost a rznorodost potaovch st neustle roste.Trvale se zvyuj penosov rychlosti mezi uzly st,
lokln st se propojuj v geograficky rozshl celky , vznikaj modern sov operan systmy zajiujc maximln
prchodnost toku loh st. S clem maximlnho zuitkovn nkladnch prostedk technickho i programovho
vybaven st nastupuj rzn filosofie sdlen zdroj, prokldanch innost a soubnch aktivit rznch uivatel,
vstavba dmyslnch databzovch systm, a po schopnost sdruovn uzl st dle charakteru eench loh
(clustering).

Avak na otzku, zda vbec a do jak mry potaov s pracuje efektivn, do jak mry je vyuita kapacita pota
v uzlech st, jak jsou vyuity komunikan prostedky mezi jednotlivmi uzly, zda je efektivn zvolena topologie st,
jak je vyuit databz vetn rovn soubn pracujcch uivatel, aktivity na rovni vstupnch a vstupnch
jednotek, a na celou adu dalch skutenost zodpovdt nelze bez vyuit monitorovn (men) st, resp.nkterch
jejich st.

Monitorovn (men) potaov st provdme vyuitm hardwarovch, nebo softwarovch monitor, resp.
kombinac obou forem. Provst veker druhy men na potaov sti najednou je prakticky nemon. V souasn
dob se vyuv nkolik stovek druh men. Navc men se provd zpravidla jen ve vytipovan sti st (switche,
routery, servery,asti LAN, apod.), nebo innost ( toky paket, propustnost, peten, apod ).

Proto ped zamlenm monitorovnm je teba zodpovdt zsadn otzku, a to co chceme monitorovnm doshnout,
k emu poslou namen hodnoty eventuln ve kter sti st (sloky) oekvme zven efektivnosti nebo
vkonnosti.

Problematika optimalizace innosti a men potaovch st je eena v rmci vzkumnch kol katedry Aplikovan
informatiky na EPI Kunovice. Nkter konkrtn vsledky pro ilustraci uvdme. [1], [2], [6].

1 Uplatnn men pi zavdn nov technologie do stvajc st
Za elem zven propustnosti potaov st EPI Kunovice bylo rozhodnuto zalenit do n monost bezdrtovho
pstupu .Byl proveden nvrh takto koncipovan st vetn vytipovanch mst pro pstupov body, zvolen konkrtn
technologie WiFi a dal nleitosti souvisejc s technickm i programovm zabezpeenm st. V procesu realizace a
nslednho zprovoznn takto upraven st bylo zapoteb provst men a ovit tak reln pnos zvolen koncepce
een.

132

Men pstupovch bod
V prvn fzi implementace WiFi bylo teba naladit pstupov body tak aby si v dosahu vzjemn nekolidovaly a
souasn, aby jejich vyslac vkon odpovdal normm TU. Za elem zjitn skutenho stavu bylo provedeno
men u vech pstupovch bod v dan lokalit.

Menm bylo zjitno, e v souvislosti se zavedenm WiFi v budovch koly bude nutn pro zajitn sprvnho chodu
st velmi asto kontrolovat frekvence pstupovch bod a ppadn doladit podle toho, kter z kanl bude vykazovat
nejlep signl.

Men bylo provedeno pomoc monitoru na Linuxu distribuce Debian na softwaru Cacti, kter je voln distribuovn
v rmci open source licenc. Tento software se na EPI,s.r.o bn vyuv ke sledovn stavu st, zvlt pak
k monitorovn zte a prtoku dat v jednotlivch uzlech st. Bez vyuit vsledk monitorovn by zajitn
efektivnho chodu st vzhledem k jej rozmanitosti a intranetovch spoj mezi uzly rznmi technologiemi od
bezdrtov mikrovlnn a po optick metalick by bylo nereln. [3].

Men propustnosti st po zaveden technologie WiFi
Pro zjitn pnosu bezdrtov technologie WiFi do st koly, bylo teba vyjt z pvodn namench statistik
poskytnutch ve zmnnm monitorem st a porovnat je s nov namenmi hodnotami.

Testovnm nhodnho uivatele bezdrtovho pipojen se ukzalo, e se propustnost intranetu koly se dle oekvn
zvila. Porovnnm statistickch dat z dvjho provozu a souasnho reimu byl zjitn penos vtho objemu dat,
navc v kratch intervalech. Pzniv vliv zven propustnosti lze vypozorovat nejen v samotnm intranetu, ale tak u
internetovch penos st EPI. Uveden skutenosti ilustruje piloen obrzek (Obr.1).

Obr.1 Zatenost penosovch cest u vybranho uivatele.

Z obrzku je patrn rychlost stahovn v kilobajtech za sekundu. Od msce ervence je penos dat do intranetu
mnohonsobn vy ne v pedchozm msci. Rovn meme vypozorovat nrst zte jak na odchozm smru do
internetu (zelen kivka) tak i pchozm smru (modr kivka); zvenou penosovou kapacitu pak ilustruje na
obrzku fialov kivka (Total traffic)

Prezentovan vsledky zskan menm jednoznan potvrdily pnos zalenn nov technologie WiFi do koln st
[3].

2 Men sovch prvk a propustnosti LAN
Zkouman lokln potaov s spolenosti Branson je soust celosvtov WAN st spolenosti Emerson a je
urena pro sven plast, pro sven ultrazvukem, teplem a vibracemi. Firma vtinu vrobk produkuje pro
automobilov prmysl. Zkouman LAN je vybudovan vhradn na bzi produkt Cisco a zaruuje tm efektivn
sdlen a vyuvn vekerch slueb, kter hardware poskytuje.

Sestv ze t hlavnch server na kterch b v dob prvnho men Windows Server NT 4.0, piem v souasn
dob probh migrace na nov operan systm Microsoft Windows Server 2003 Standard Edition. Dle pak z jednoho
Modular Access routera Cisco 1700 (vylennho ke komunikaci s mateskou spolenost a s divizemi v Evrop),

133

jednoho switche tdy 4000, dalch 6-ti switch Catalyst tdy 3500, cca 200 pracovnch stanic a zazen (tiskrny,
CNC stroje, apod), kter vyaduj IP adresu.

Nakolik zkouman LAN byla vyprojektovan a uvedena do provozu v roce 2000, vyvstal problm zjistit, zda bude
vkonnostn stait novm potebm a poadavkm kladenm v souasnosti. V sti dochz ke zmn filosofie ukldn
dat a to od dvjho loklnho na centrln ukldn vech soubor a dat.V tto souvislosti je poteba zjistit propustnost
jednotlivch prvk st, zda vyhovuj narstajc zti souvisejc s novmi funkcemi danho aplikanho
prosted.Samotn men, jak ji bylo naznaeno dve se provd na dvou rovnch: na rovni monitorovn prvk st
(switch) a monitorovn jednotlivch vtv resp. segment st.

Men propustnosti switch
Men je zameno na switche z toho dvodu, e veker provoz na sti je zabezpeena prv tmito zazenmi, a to
nm poskytuje podrobn pehled o celkovm provozu st, jeliko data zskvme pmo z tchto zazen samotnch.

K men se vyuvaj dva nstroje. Hodnoty z centrlnho switche Catalyst 4006 se zskvaj pomoc MRTG (The
Multi Router Traffic Grapher) Jde o monitor uren ke sledovn penosov zte na spojch st. Generuje HTML
strnky obsahujcch PNG zobrazen, kter poskytuje iv vizuln reprezentaci tchto tok (Check
http://www.stat.ee.ethz.ch/mrtg/)

Men switche se m pomoc nstroje zvanho Cluster management, co je v principu podobn monitor jako
pedchoz, jene je vyvinut pmo Ciscem samotnm pro sv st, co je znt na podrobnosti mench dat
prezentujcch cca 30 hodnot.

Menm prvk (switch) jsou k disposici nejrozmanitj informace o stvajc innosti st.
A ji jde o informace o portech samotnch, podrobn informace tkajc se odesln paket, paket pijmanch, nebo
o procentuln vyjden hodnot, poet samotnch paket, kter proli zazenmi, vyuit informace o ce psma apod.
Vedle vkonnostnch parametr poskytuj rovn informace o poruchovosti st a tm i dokonal bezprostedn
pehled o stavu st.

Jeliko za pomoc Cluster managementu je neustle vidt co se dje v zazench st, je mon na vznikl situace
pimen a adekvtn reagovat.

Men probhalo ve stejnch asovch secch bhem pracovn doby a tm se zskal dokonal pehled o efektivnosti
vyuit jednotlivch sloek st. Ukzalo se, e pro zskn dokonalejho pehledu je teba asovou periodu snit; byla
zvolena na 30 minut.

Sbr informac pomoc MRTG probhal automaticky, men pomoc Cluster managementu byla provdna manuln,
kdy se informace pekoprovaly do excelovch list a nsledn se zpracovvaly a vyhodnocovaly.

Penosov aktivity na Catalystu 4006 ilustruje obrzek Obr. 2. Jde o port . 13 spolu s namenmi hodnotami, kter
ns nejvce zajmaj.

Max Average Current
In 4560.0 B/s (0.0%) 1513.0 B/s (0.0%) 283.0 B/s (0.0%)
Out 5063.0 B/s (0.0%) 1924.0 B/s (0.0%) 486.0 B/s (0.0%)
Obr.2. Namen penosov rychlosti u vybranho portu.

134

Prbh men na Catalystech 3500 mlo odlin charakter. Vzhledem k rozmanitosti a mnostv daj byl pro ilustraci
vybrn switch 129.115.4.38 a u nj vybran parametr Transit Rate.

0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
6
:
0
0
7
:
0
0
8
:
0
0
9
:
0
0
1
0
:
0
0
1
1
:
0
0
1
2
:
0
0
1
3
:
0
0
1
4
:
0
0
1
5
:
0
0
FastEthernet0/1
FastEthernet0/2
FastEthernet0/3
FastEthernet0/4
FastEthernet0/5
FastEthernet0/6
FastEthernet0/7
FastEthernet0/8
FastEthernet0/9
FastEthernet0/10
FastEthernet0/11
FastEthernet0/12
FastEthernet0/13
FastEthernet0/14
FastEthernet0/15
FastEthernet0/16
FastEthernet0/17
FastEthernet0/18
FastEthernet0/19
FastEthernet0/20
FastEthernet0/21
FastEthernet0/22
FastEthernet0/23
FastEthernet0/24
GigabitEthernet0/1
GigabitEthernet0/2

Obr.3 Aktuln penosov rychlosti v danch okamicch.

Na obr.3 jsou zachyceny aktuln penosov rychlosti ( Mbps) s jakou jsou data na danm switchi peposlan na dal
zazen. V naem ppad jde o penos z hlavnho switche pes tento switch ke koncovm zazenm.

Vsledky men a vyhodnocen statistickch dat ukzaly, e potaov s z velk sti pokryje narstajc
vkonnostn poadavky na nj kladen. Ukzalo se ale tak, e na oddlen strojn a elektrokostrukce vznik nutn
poteba pozen jednoho 24 portovho switchu pro vyselektovn pracovi pracujcch se zvenm objemem dat;
dosavadn podmnky ji nevyhovuj kapacitn jejich potebm. Navc bylo zjitno, e pro zajitn efektivn prce st
bude teba vytvoit VLAN pro separaci urit skupiny pracovnk spolu s centrlnm serverem na ukldn dat. [4].

Men propustnosti segment, resp. vtv LAN
Pro zskn ucelenho pehledu o vkonnosti cel st LAN bylo rozhodnuto doplnit zvry zskan menm prvk o
monitorovn vybranch segment st.Jde o ty vtve, na kter maj bt vlivem modernizac na LAN pokldny
nejnronj aplikan poadavky.

Jde pedn o adu upgrad CAD aplikac, jako je EPLAN, SolidEdge, a dalch.
Men na nejvytenjch vtvch m zajistit plynul pechod paket v cel sti LAN a tm na nov podmnky prce
st.

K sledovanm parametrm pat:
monitorovn toku paket (pracovn stanice a uzly LAN), rovn tak ztrty, smrovn, vmna a dal
identifikace nejintenzivnji vyslajcch a nejintenzivnji pijmacch stanic
monitorovn pli dlouhch a pli krtkch rmc,chyb CRC, chybov statistika
monitorovn rznch parametr na jednotlivch vrstvch OSI (provn MAC adres, sovch adres, IP adres, a
pod)

Monitorovn st bude provedeno pomoc osvdench sovch analyztor Sniffer Analyzer a NetXrey, spolu
s prostedky pro sprvu st obsaen v OS WIN2003 Server SE. (Bandwidth test).

135

Sniffer Analyzer je prostedek uren k analze aktivit v stch LAN a WAN, vude tam kde se vyaduje sofistikovan
analza paket. Zachytv anomlie na vech rovnch st, odhaluje problmy souvisejc s propustnost st a
pedkld nvrhy monch een.Krom toho je osvdenm prostedkem i k odhalovn poruch st. Bandwidth test
se vyuv k ovovn propustnosti vtve st. [5].

3 Sledovn vkonnosti server pomoc Microsoft Operations Manager 2005
O vznamu sledovn vkonnosti sloitch potaovch struktur, tedy i potaovch st jsou vedle uivatel
pesvdeni tak pedn potaov firmy. U pedchozho pkladu jsme upozornili na monitory firmy Cisco, kter jsou
specieln tvoen ke sledovn vkonnosti potaovch st tvoench z prvk tto firmy.

Ponkud s jinm eenm pichz na trh fy Microsoft , kter pro monitorovn vkonovch a kapacitnch parametr
server Windows 2000 a 2003 uvd na trh monitor

Microsoft Operations Manager 2005 (MOM 2005). Umouje celou klu men ponaje vyuitm CPU, diskovho
prostoru, parametr databz, vetn konektivity soubn pracujcch uivatel, vyuit prostoru databze, apod.

Za elem oven relnch monost monitoru a jeho pouitelnosti v konkrtnm nasazen, byla provedena jeho
instalace v potaov sti s 250 uivateli na OS MS Windows ve W2k3 domn. S File serverem, Domain
Controlerem, Exchange serverem W2k a Ebi serverem.

Po instalaci monitoru se na firemn sti neobjevil dn problm.Po dkladnjm prozkoumn options byla zapnuta
vt filtrace a u se objevily prvn vsledky monitorovn. Rozpoznal z eho se skld dan s, ukzal, e jsou tam
switche a huby, e jsou pouity prvky jak 1Gbit tak 100Mbit, tak 10Mbit a hned doporuil nahradit 10Mbit minimln
za 100Mbit.

Dalm pozitivnm rysem MOM je, e se potaov s d rzn rozgrupovat a podle danch grup pouvat urit typy
filtrace a zkoumn problm.

Nap. na zkouman sti je k disposici oddlen DEMO, kter slou k pedvdn produkt firmy (server a program
na nich vytvoench na mru). V tto skupin jsou statick adresy a vude jinde krom oddlen IT oddlen je DHCP
(automatick pidlovn IP adres ze serveru). V tto skupin dochz asto ke kolizi IP adres. Ukzalo se, e MOM
doke tuto situaci dobe mapovat a informovat o danm problmu sprvce st. Uivatel si nastav IP adresu a MOM
hls, e je uivatel v kolizi. V zpt uivatel vol sprvce, e m problm se dostat na s. Sprvce vak ji v kde je
problm.

Velice uiten funkce poskytovan MOM, kterou uvt nejeden uivatel je, e doke pro lep orientaci vytvoit
pedpokldan model potaov st, kterou lze pehledn znzornit, popsat s uritmi informacemi jednotliv skupiny
kde se pesn v, co se dje a co by se dt nemlo.

Monitor MOM poskytuje uivateli monost upravovat si chybov hlen (error messages), jako i styl jejich zobrazen.
Stupe dleitosti konkrtnch chybovch hlen si ur uivatel sm a speciln si nastav dleitost tch situac, ped
ktermi ns m monitor varovat.Samozejm, e je mono jednotliv rovn rzn vybarvovat, mnit barvu apod.
I kdy zatm probhla pouze prvn st experiment s monitorem MOM (podzim 2005), lze potvrdit jeho uitenost a
vznam pi sledovn innosti st. Na zklad dosud zskanch poznatk :
meme jednoznan doporuit pozen monitoru MOM vude tam, kde se jedn o rozshlej topologii
potaov st. Prokazateln poskytuje cenn informace o skuten innosti st a navc doke pedchzet
problmm.Investice do MOM se vyplat.
konstatujeme, e je uren pro sledovn serverovch systm a co jsme pedpokldali nen pli vhodn jako
monitorujc nstroj pro problmy souvisejc s infrastrukturou smrova i pepna. [2, 6].

Zvr
O vznamu monitorovn potaovch st ji dnes nen pochyb.Pedn o nj maj zjem uivatel st, kte menm
zskvaj aktuln informace o reln innosti st. Vedle zskvn vkonnostnch parametr jako jsou propustnost,
vyuit, apod., mohou rovn zskvat rzn hlen o bezporuchovm stavu st a dal potebn informace.

136

Renomovan firmy, zabvajc se vrobou rznch komponent, nebo vstavbou rozshlch potaovch st proto
nabz celou klu monitor, ktermi lze zajistit sledovn (men) a nsledn efektivn provozovn tchto sloitch
potaovch struktur.

Zvrem je si vak nutno uvdomit, e monitory (sov analyztory) nejsou univerzln a jsou zvisl na tom jak
parametry maj mit, kterou sloku st maj monitorovat (nap.servery, switche, apod), jsou zpravidla vyvinut pro
prvky urit firmy (Cisco, Microsoft).

LITERATURA
[1] RUKOVANSK, I. Sledovn vkonnosti potaovch st. Kunovice : Vzkumn zprva EPI, s.r.o. prosinec
2005.
[2] HANCE, B. Microsoft Operations Manager 2005. Computerworld .8, 2005, str. 21.
[3] CHMELA, T. Vyuit WiFi ke zven propustnosti st. Kunovice : EPI, s. r. o. Bakalsk prce EPI, 2005.
[4] UNK, M. Men sovch prvk potaov st a vyhodnocen men. Kunovice : Podklady k Bc prci
EPI, s.r.o., 2005.
[5] SLOBODA, P. Men propustnosti lokln potaov st. Kunovice : Podklady k Bc. prci, EPI, s.r.o., 2005.
[6] KADLEC, R. Oven funknch schopnost MOM 2005 na konkrtn sti. Kunovice : Prezentace vsledk
tmu 3-2-1, EPI, s.r.o., 2006.

Adresa:
Osvobozen 699
686 04 Kunovice
tel. / fax.: +420 572 549 018, +420 572 548 788
e-mail: rukovansky@vos.cz

137

ANALZA DAT S VYUITM NEURONOVCH ST A KONTINGENNCH TABULEK
Jindich Petrucha


Abstrakt: lnek popisuje monosti pouit externch zdroj pro zskn vstupnch dat o Evropsk unii a
pevod tchto dat do formtu vhodnho pro OLAP analzu. Je popisovn proces transformace a itn dat
tak aby bylo mon provst pevod do kontingenn tabulky, ze kter jsou data vkldna do externho
programu realizujc analzu vybran asov ady. Extern program simultoru neuronov st provede
proces uen z vybranch dat a dal etapu procesu analzy.

Klov slova: OLAP, neuronov s, asov ady, kontingenn tabulka, itn dat, ETL.

1. vod
Monosti pouit externch zdroj pro analzu dat se v souasn dob neustle zvtuj, protoe mnoh data jsou
prezentovna na internetu a jsou neustle aktualizovna. Problm tedy nespov ani tak v zskan dat, ale spe v
problematice jejich analzy pomoc modernch prostedk informanch technologi. Mnoh specializovan nstroje
vyaduj urit standardizovan formt, kter slou jako import dat, nebo je nutno erpat tato data z datovch sklad a
pout techniku datov pumpy, kter dovoluje vybrat poadovan data. Pokud pracujeme s daty na internetu vtina
tchto dat je v textovm formtu s rznmi grafickmi pravami, kter vizuln zpehleduj zobrazen data, ale na
druhou stranu pro automatick zpracovn je tento formt naprosto nehodn. Tato nadbyten data je nutno odstranit
pomoc programovch nstroj nebo run pomoc rznch editor do poadovanho formtu. Takto upraven data je
mon pout pro analzu pomoc nstroj uml inteligence, kter zkvalituj rozhodovac proces.

2. Problematika extrakce dat
2.1 Obecn principy
Dleitm krokem je etapa popsna v [1] jako ETL (Extraction Transformation Loading) tedy extrakce dat z uritho
transaknho zdroje, transformace tchto dat do potebnch struktur a nsledn nahran tchto dat do datovho skladu
nebo pmo do programovho systmu. Clem tto etapy je centralizovat data tak by byla splnna podmnka rychlho
pstupu k tmto datm. Je vhodn na potku si definovat cle tto etapy tak aby byla splnna obecn kritria.
Pedpoklad pro tyto innosti je nae schopnost zpracovvat data na principech uvedench ve firemn literatue firmy
Oracle kde je definovn proces transformace dat na informace tehdy kdy:
mme daje,
vme, e mme daje,
vme, kde tyto daje mme,
mme k nim pstup,
zdroji daj meme dvovat.

Ne vdy se poda jednotliv podmnky tto definice splnit, protoe ve vech systmech, kde pracuje lidsk faktor,
dochz k chybm, a u zmrnm nebo nhodnou chybou lidskho initele, kter pracuje v transakn rovni
zpracovn. Mnostv dat je v mnoha ppadech tak velk, e jen men st je podrobovna analze, ze kter v procesu
business inteligence vznikaj znalosti.

V poten fzi ETL potebujeme zskat data uloen v uritm externm informanm zdroji, kterm me bt
informan podnikov systm, ve kterm se provd vtina transaknch operac, a u se jedn o etnictv, skladov
hospodstv, odbratelsko dodavatelsk vztahy nebo podobn systmy. Z tchto systm pokud obsahuj kompatibiln
databzov systmy, meme zskat data pmo z relanch tabulek nebo pes export dat do patinch formt, kter
podporuje vtina systm. Napklad CSV textov soubory a podobn. Dal monost je monitorovan trnho
prosted internetu a odtud zskat potebn data. Zde je mon pout modern nstroje, jako jsou teky RSS kanl a
sledovat vznik informanch zdroj zmnou URL pmo na strnkch urit firmy. Nebo meme vyuvat datov

138

zdroje informanch a vyhledvacch server. Tento systm se vyuv hlavn v oblasti finannch trh pi sledovn
index jednotlivch akciovch trh a nslednm vyhodnocen hodnoty akci.

2.2 Extrakce a itn dat na konkrtnm pkladu
Pro konkrtn ppad si vyberme aktuln statistick data o evropsk unii, kter popisuj stav inflace a vvoj prmyslov
vroby v jednotlivch sttech evropsk unie za urit obdob. Tato data je mon zskat dle odkazu [5] v pdf formtu v
pehledovch tabulkch tak jak je znzornno na obr. 1. V tabulce je na lev stran asov osa a na horn sti se
nachzej nzvy jednotlivch stt.

Obr. 1. Statistick data z informanho zdroje EUROSTAT

Pokud chceme extrahovat tato data je mon vloit data s vyuitm koprovn a oznaen sti data do textovho
souboru, kter dle opatme oddlovacmi stednky mezi jednotliv daje a provedeme rozdlen na jednotliv
zznamy. Sloitost tohoto kroku ukazuje textov st po vbru dat z pdf formtu. Data vytvej jeden zznam bez
oddlovacch informac. Tato mlo strukturovan forma zpisu se mus transformovat tak aby bylo mon separovat
potebn data do tabulkovho kalkultoru.

EUEMUAustriaBelgiumDenmarkFinlandFranceGermanyGreeceIrelandItalyLuxembourgetherlandsPortugalSpainSwed
enUKCyprusCREstoniaHungaryLatviaLithuaniaMaltaPolandSlovakiaSloveniaI.99-0,2%-0,1%-0,1%0,4%0,2%-0,2%-
0,4%-0,1%-1,3%-0,8%0,1%-1,7%0,0%-0,4%0,3%-0,4%-0,6%-
0,7%0,9%1,1%2,6%1,0%1,0%na1,5%3,0%1,0%II.990,3%0,3%0,2%0,2%0,5%0,4%0,4%0,2%-
0,7%0,7%0,2%1,9%0,7%0,0%0,1%0,1%0,2%-1,8%-0,1%0,2%1,3%0,2%0,0%na0,5%0,8%0,4%

Transformaci je vhodn provst do CSV formtu ktermu rozum tabulkov kalkultor Excel a dovoluje importovat
data z tohoto textovho formtu. Dalm dleitm krokem pro analzu je vytvoen datov struktury, kter je vhodn
pro kontingenn tabulku, kter pedstavuje datovou OLAP kostku, ve kter je mon provdt rzn pohledy na
analyzovan data. Na obr. 2 je znzornna kontingenn tabulka z vloenmi daty. V lev sti je zachovna asov
osa, kter je rozdlena jednotliv roky a dal lenn pedstavuj jednotliv msce podle selnho oznaen. V
kontingenn tabulce je mon provdt rzn agregace podle asov osy a vybrat si stty, kter chceme v danm
ukazateli sledovat. Pod kadm rokem je pouit dek s agregujc hodnotou prmru dat pro urit sloupec. Tento
agregan vzorec lze tak podle poadavk mnit. Nsledujc obrzek obr. 2 zobrazuje data vybran pro rok 2004 a
2005 pro stty EU se zvraznnm CR. Hodnota inflace pro CR je velmi nzk a je srovnateln se zemmi jako je
Rakousko, Nmecko nebo Francie.

139

Obr. 2. Kontingenn tabulka s daji inflace pro vybran zem EU .

3. Analza dat s vyuitm neuronovch st
3.1 Analza asovch ad
Pro dal analzu daj v kontingenn tabulce meme pout urit dal nstroje uml inteligence, kter dovoluj
zlepit analytick proces vyhlednm souvislost a vazeb, kter nejsou do dat pmo vloeny. Jeden z tchto nstroj
pedstavuj neuronov st realizovan pomoc programovch simultor, kter dovoluj zadat architekturu neuronov
st, provst etapu uen podle nastavenho kritria, simulovat proces zpracovn vstupnch a vstupnch vzor. Jako
hodnotc kritrium lze pouvat rzn pstupy hodnocen v uc mnoin, Pro n ppad pouijeme stedn
kvadratickou odchylku (MSE, mean squared error) pes cel zahrnovan obdob, zahrnujc N predikc, kdy hodnota
nesm pekroit uritou zadanou hodnotu.

MSE = 1/N (predikce skutenost)
2

Pi analze budeme pouvat jednokrokovou predikci, kter je zabudovna pmo do programovho simultoru. Proces
zpracovn dat z kontingenn tabulky bude znzornn v nsledujcch odstavcch.

3.2 Pouit simultoru neuronovch st pro analzu asovch ad
Jako vstupn datov soubor je pouvn ACII soubor, kter m urenou strukturu dovolujc zadat parametry neuronov
st a asovou adu podle struktury vstupnho vzoru.

Ukzka ASCII souboru:
Neuron - casova rada inflace CR od roku 2000
5 -pocet vstupu

140

1 -pocet vystupu 1-krok predikce
70 -pocet dat casove rady
10 -pocet neuronu ve skryte vrstve
**** data casove rady ****
0.0170
0.0020
0.0000
-0.0020
0.0030

Data jsou vloena do tohoto souboru bu pmo z kontingenn tabulky nebo z dat ze kterch erp kontingenn
tabulka. Tento zpsob dovoluje modifikovat tak parametr jako je poet neuron ve skryt vrstv velmi jednoduchm
zpsobem. Program simultoru je napsn v jazyce PASCAL DELHI a dovoluje realizovat uen na vybran sti asov
ady. Na obr. 3 je zkladn okno simultoru zobrazujc data inflace pro CR normovan na interval od nuly do jedn.
Z obrzku je zejm kolsn hodnot, kter se bude snait neuronov s analyzovat a nauit. Pro uen byla vybrna
oblast dat od 1 do 50 prvku a pro vyhodnocen chyby oblast od 10 do 55 prvku. Proces uen byl nastaven na 50000
cykl uen pro vhodn sledovn totln chyby ve spodnm pravm oknu simultoru.

Po ukonen procesu uen byl proveden test, kter je znzornn v hornm levm okn simultoru, kde zelen hodnoty
jsou data asov ady a modr hodnoty (aerkovan) jsou predikovan hodnoty pomoc simultoru uml neuronov
st. Na obr. 4 je detailn vidt oblast od 50 prvku po 70 prvek, kdy od 55 prvku se pracovalo s ji neznmmi daty.
Simultor zachycuje zmny a spe m snahu pedvdat prud kolsn, tak jak bylo zejm v pedchozch letech.
Nejvt rozdl je pro hodnotu 61 prvku, kdy simultor pedpokld rst inflace, kter se ale nekonal. Pro pesnj
hodnoty by bylo vhodn prodlouit poet cykl dvojnsobn a uit simultor na del asov obdob.

Obr. 3. Simultor uml neuronov st s daty vloenmi z kontingenn tabulky pro zemi CR.

141

Obr. 4. Detailn zobrazen dat ze simultoru uml neuronov st pro data inflace CR.

Vyuit tto metodiky dovoluje velkou flexibilitu pi analze dat, kter se nachzej v rznch informanch zdrojch na
internetu a provdt hledn souvislost mezi jednotlivmi daty. Kontingenn tabulky dovoluj velmi pehledn
analyzovat souhrnn za vybran asov obdob a jejich vazba na simultory neuronovch st dv kvalitn nstroj do
rukou samotnho uivatele. Tento postup je vhodn pro rzn systmy, kter maj urenou asovou osu ve kter se lze
dobe orientovat.

3. Zvr
Z pkladu, kter byl postupn prezentovn v pedchzejcch odstavcch je zejm, e pro kvalitn analzu je velmi
dleit mt pipravena data, kter jsou oitna od rznch vkyv a dovolujc provst na jejich zklad rozhodovac
proces. OLAP nstroje jsou pipraveny pro zpracovn dat, ale maj jen velmi mlo monost na jejich transformaci do
poadovanho formtu. Velmi asto jsou k dispozici data mlo strukturovan, kter je nutno zpracovat v etap ETL, pro
kterou je nutn pipravit programov nstroje podle charakteru vstupnch dat. Tato etapa je podle mho nzoru
nejsloitj a zabr nejvce asu z hlediska ppravy dat. Kontingennm tabulky jsou vhodnm nstrojem pro analzu
dat a dovoluj vbr asov ady s monost vloen do dalch programovch systm. Simultory neuronovch st je
vhodn pout tam kde lze sledovat urit trend v datech, kter analyzujeme.

Pouit literatura:
[1] LACKO, L. Datov sklady analza OLAP a dolovn dat s pklady v Microsoft SQL Serveru a Oracle. 1. vyd.
Brno : Computer Press, 2003. s. 486. ISBN 80-7226-969-0.
[2] DOSTL, P. Modern metody ekonomickch analz Finann kybernetika. Prvn vydn. Zln : Univerzita
Tome Bati, 2002, s.110. ISBN 80-7318-075-8.
[3] PETRUCHA, J.; MIKULA, V. Application of Neural Networks for Time Series Prediction and Making the
Adequate Program Simulator. In Proceedings First INTERNATIONAL CONFERENCE ON SOFT
COMPUTING APPLIED IN COMPUTER AND ECONOMIC ENVIROMENTS January 30 -31. Kunovice :
Evropsk polytechnick institut, 2003. s.165-171. ISBN 80-7314-017-9.
[4] PETRUCHA, J. Technologie analzy dat OLAP systmy v prosted DBPROVE. ACTA UNIVERSITATIS
AGRICULTURAE ET SILVICULTURAE MENDELIANAE BRUNENSIS, 2000, ronk XLVIII, slo 2, s.
149-155. ISSN 1211-8516.
[5] http://www.csas.cz/banka/application?pageid=downloads&dtree=cs&selnod=57/dataEU_public.pdf [online].
2005 [cit. 2006-01-18]. Dostupn z WWW: < http://www.csas.cz/banka>.

142

Adresa:
Ing. Jindich Petrucha, Ph.D.
Osvobozen 699,
686 04 Kunovice
te./fax.: +420 572 549 018, +420 572 548 788
e-mail: petrucha@vos.cz

143

DETECTION OF INITIAL DATA GENERATING BOUNDED SOLUTIONS OF LINEAR
DISCRETE EQUATIONS
1

Jaromr Batinec, Josef Diblk


Abstract: In the situation when graphs of solutions of discrete equations remain in a prescribed domain,
the problem concerning determination of their initial data is discussed. Special attention is paid to linear
discrete equations and initial data generating its solutions such that their graphs remain in a
prescribed domain are found. Illustrative examples are considered, too.

Key words: Linear discrete equation, bounded solutions, initial data.
AMS Subject Classification: 39A10, 39A11

1 Introduction and the problem considered
1.1 General suppositions
We consider a scalar discrete equation
( ) ( ) ( ) k u k f k u , (1.1)
( ) ( ) { } { }. ,... 1 , 0 , and ... , 1 , where : + N N a a a a R R a N f Together with discrete equation (1.1) we consider an
initial problem. It is posed as follows: for a given N s we are seeking the solution ( ) k u u of (1.1) satisfying
the initial condition
( ) R u s a u
s
+ (1.2)
with a prescribed constant u
s
. Let us recall that the solution of initial problem (1.1), (1.2) is defined as an infinite
sequence of numbers { } ( ) i.e. , with
0
k s a u u u
k
k
k
+ +

( ) ( ) ( ),... ,..., 1 ,
1 0
n s a u u s a u u s a u u u
n
s
+ + + + +

such that for any ( ) s a N k + the equality (1.1) holds. Let us note that the existence and uniqueness of the solution of
the initial problem (1.1), (1.2) is a consequence of properties of the function f. If function f depends continuously on
second argument then the initial problem (1.1), (1.2) depends continuously on its initial data. Let b(k),c(k) be real
functions defined on N(a) such that b(k) < c(k) for every ( ) a N k . We define a set ( ) as R a N
( ) ( ) ( ) { } k u a N k u k , : , :
with
( ) ( ) ( ) { } k c u k b u k : :

and a closure of the set u; as
( ) ( ) ( ) { } k u a N k u k , : , :
with
( ) ( ) ( ) { } k c u k b u k : :

Obviously it holds:
( ) ( ) ( ) { }
( )
( ) ( ) { } k u u k k u a N k u k
a N k

, , , , ,
U

1
Preliminary version.

144

Let us involve set
2 1
B B B with
( ) ( ) ( ) { } ( )
( ) ( ) ( ) { } ( ) R , : , :
R, , : , :
2
1

a N k c u a N k u k B
a N k b u a N k u k B

and define the boundary of as
( ) ( ) ( ) ( ) ( ) ( ) { } B k c u k b u a N k u k 0 , : , :
Define, moreover,
( ) ( ) ( ) { } k c k b k , :

and ( ) ( ) ( ) R , a N u k for auxiliary functions
( ) ( )
( ) ( ). : ,
, : ,
2
1
k v u u k U
k b u u k U

Definition 1.1. The full difference ( )
( )
1
, 1
| ,
B u k
u k U

of the function ( ) u k U ,
1
for a given ( )
1
, B u k with respect to the
discrete equation (1.1) and the set B
1
is defined as
( )
( )
( ) ( ) ( ) ( ). 1 , : | ,
1
, 1
k b k b k b k f u k U
B u k
+ +

The full difference ( )
( )
2
, 2
| ,
B u k
u k U

of the function ( ) u k U ,
2
for a given ( )
2
, B u k with respect to the discrete
equation (1.1) and the set B
2
is defined as
( )
( )
( ) ( ) ( ) ( ). 1 , : | ,
2
, 2
k c k c k v k f u k U
B u k
+ +

Definition 1.2. A point ( ) B u k , with ( ) a N k is called the point of the type of strict egress for the set with
respect to the discrete equation (1.1) if
( )
( )
0 | ,
, 1

B u k
u k U
in the case when ( )
1
, B u k , and
( )
( )
0 | ,
, 2

B u k
u k U
2
, B u k .
The affirmation of following lemma is based on above Definitions 1.1, 1.2 and is an easy consequence of the
formulated notions.

Lemma 1.3. The point ( ) B u k , with ( ) a N k is a point of the type of strict egress for the set LJ with respect to the
discrete equation (1.1) if and only if
( ) ( ) ( ) ( ) 0 1 , : + + k b k b k b k f
1
, B u k , and
( ) ( ) ( ) ( ) 0 1 , : + + k c k c k b k f
2
, B u k .

1.2 Nonlinear case and description of problem considered
The following theorem, concerning asymptotic behavior of solutions of the equation (1.1) is a particular case of a more
general result in [3, Theorem 2, p. 520] (see [4] also).

Theorem 1.4. Let us suppose that f is defined on with values in R and is continuous with respect to the
second argument. If, moreover, each point ( ) B u k , is the point of the type of strict egress for the set with respect to
the discrete equation (1.1), then there exists an initial problem
( ) ( ) a u a u

(1.3)

145

such that the corresponding solution ( ) k u u

satisfies the relation
( ) ( ) k k u
(1.4)
for every ( ) a N k .
Now we are able to describe the general version of the problem considered. Analysing result given by Theorem
1.4 we conclude that the existence of (at least one) solution of the problem (1.1), (1.3) having the indicated
(asymptotic) behavior characterized by relations (1.4) is stated without concrete determination of the corresponding
initial data u* itself. In this contribution we try particularly to fill this gap in the linear case. Note that the
questions concerning behavior of solutions of discrete equations are considered e.g. in [1, 2], [5] [9].
Unfortunately, problem concerning the determination of corresponding initial data was not considered there.

1.3 Linear case and the problem considered
Let us put
( ) ( ) ( ) ( ) ( ) k k u k k u k f + : ,
in (1.1) with ( ) ( ) ( ) R : , a N k k and consider the corresponding linear equation
( ) ( ) ( ) ( ) k k u k k u + (1.5)
together with an initial problem
( )

u a u (1.6)
It is easy to verify that in the linear case Theorem 1.4 takes the form:
Theorem 1.5. Let the inequalities
( ) ( ) ( ) ( ) ( ) 0 1 1 + + + + k b k k b k (1.7)
( ) ( ) ( ) ( ) ( ) 0 1 1 + + + + k c k k c k (1.8)

hold for every ( ) a N k . Then there exists an initial problem
( ) ( ), a u a u

(1.9)
such that the corresponding solution ( ) k u u

of equation (1.5) satisfies for every ( ) a N k the inequalities
( ) ( ) ( ). k c k u k b

(1.10)
Now, let us formulate problem under consideration.

Problem 1.6. Determine at least one value u* such that the corresponding solution ( ) k u u

of the linear problem
(1.5), (1.6) satisfies the relations ( ) ( )
k u k, for every ( ) a N k , i.e. satisfies inequalities (1.10) for every

( ) a N k .
We will show in the sequel that conditions of Theorem 1.5 together with the condition ( ) 1 k for every ( ) a N k are
sufficient for determining at least one initial value u* .

2 Main Results
In the following we put ( ) ( ) ( )

2
1
2
1
2 1 2 1
, , if 0 and 1
k
k i
k
k i
k k a N k k i G i G and G is a function well-defined on
N(a).

2.1 Auxiliary Lemma
Lemma 2.1. Let for every ( ) ( ) 1 : k a N k and the inequalities (1.7), (1.8) hold. Then the sequence { }
0 s cs
u with
( ) ( ) ( ) ( )
( ) ( )
N ,
1
1
:
1
0
1
1
1
0
+ +
+ + +
+
+ +
s
i a
j i a s a c
u
s
i
s a
i a j
s
i
cs

(2.1)

146

is a decreasing convergent sequence and the sequence { }
0 s bs
u
( ) ( ) ( ) ( )
( ) ( )
, N ,
1
1
:
1
0
1
1
1
0
+ +
+ + +
+
+ +
s
i a
j i a s a b
u
s
i
s a
i a j
s
i
bs

(2.2)
is an increasing convergent sequence. Moreover
bs cs
u u holds for every N s and for limits , ,

b c where
, lim , lim

s
cs
s
b u c (2.3)

the inequality

b c holds.

Proof. Let us divide the proof into several steps.
a) Property
bs cs
u u , ( ) a N s .
Let us show that
bs cs
u u for every ( ) a N s . This follows from (2.1), (2.2), since ( ) ( ) s a b s a c + + for every
N s .

b) Sequence { }
0 s cs
u is a decreasing sequence.
Let us verify that
1
,
+
s c cs
u u for N s every . If 0 s and 1 s then (2.1) gives
( )
( ) ( )
( ) a
a a c
a c u
c
+
+

1
1
u and
c1 0

The inequality
1 0 b c
u u is due to the property ( ) ( ) 0 1 + a a consequence of (1.8) with a k since
( ) ( ) ( ) ( ) ( ) 0 1 1 + + + a c a a c a
Let us consider the general case. For N , + s s a k , the inequality (1.8) gives
( ) ( ) ( ) ( ) ( ) 0 1 1 + + + + + + + s a c s a s a c s a
or ( ) ( ) 0 1 since + + s a
( )
( ) ( )
( ) s a
s a s a c
s a c
+ +
+ + +
+
1
1
(2.4)

Then using (2.1) we estimate the general term N , s u
cs
of the sequence { } :
0
s cs
u
( ) ( ) ( ) ( )
( ) ( )
[ ]
( ) ( )
( )
( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( )
( ) ( )
1 ,
0
1 0
0
1
1
0
0
1
1
1
0
1
0
1
1
1
0
1
0
1
1
1
0
1
1 1
1
1 1
1
1 1 1
1
1
1
1
(2.4) to due
1
1
+
+
+ +
+
+ +
+
+ +
+
+ +
+
+ +
+ +
+ + + +
+ +
+ + + + +
+ +
+ +
,
_
+ + + + +
+ +
+ +
+ +
+ + +

+ +
+ + +

s c
s
i
s a
i a j
s
i
s
i
s a
i a j
s
i
s
i
s a
i a j
s
i
s
i
s a
i a j
s
i
s
i
s a
i a j
s
i
cs
u
i a
j i a s a c
i a
j i a s a s a c
i a
s a j i a s a s a c
i a
j i a
s a
s a s a c
i a
j i a s a c
u

147

So, the inequality
1
,
+
s c cs
u u holds for every N s .

c) Sequence { } :
0
s bs
u is an increasing sequence.
Let us show that
1
,
+
s b bs
u u for N s . If 0 s and 1 s then (2.2) gives
( )
( ) ( )
( ) a
a a b
u a b u
b b
+
+

1
1
and
1 0

The inequality
1 0 b b
u u is due to the property ( ) ( ) 0 1 + a a consequence of (1.7) with a k since
( ) ( ) ( ) ( ) ( ) 0 1 1 + + + a b a a b a

Let us consider the general case. For N , + s s a k , the inequality (1.7) gives
( ) ( ) ( ) ( ) ( ) 0 1 1 + + + + + + + s a b s a s a b s a

or ( ) ) 0 1 (since + + s a
( )
( ) ( )
( ) s a
s a s a b
s a b
+ +
+ + +
+
1
1
(2.5)

Then using (2.2) we estimate the general term N , s u
bs
of the sequence { } :
0
s bs
u

( ) ( ) ( ) ( )
( ) ( )
( ) [ ]
( ) ( )
( )
( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( )
( ) ( )
1 ,
0
1 0
0
1
1
0
0
1
1
1
0
1
0
1
1
1
0
1
0
1
1
1
0
1
1 1
1
1 1
1
1 1 1
1
1
1
1
2.5 to due
1
1
+
+
+ +
+
+ +
+
+ +
+
+ +
+
+ +
+ +
+ + + +
+ +
+ + + + +
+ +
+ +
,
_
+ + + + +
+ +
+ +
+ +
+ + +

+ +
+ + +

s b
s
i
s a
i a j
s
i
s
i
s a
i a j
s
i
s
i
s a
i a j
s
i
a
i
s a
i a j
s
i
s
i
s a
i a j
s
i
bs
u
i a
j i a s a b
i a
j i a s a s a b
i a
s a j i a s a s a b
i a
j i a
s a
s a s a b
i a
j i a s a b
u

Consequently, the inequality
1
,
+
s b bs
u u is verified for every N s .

The lemma is proved since all remaining affirmations are elementary consequences of the theory of number sequences.

2.2 Main Results
Lemma 2.2. Let for every ( ) ( ) 1 : k a N k . Then solutions ( ) ( ) ( ) a N k k U k u , , of two problems for linear
equation (1.5):
( ) ( ) a U and a u ,

with satisfy the inequalities
( ) ( ) k U k u (2.6)
for every ( ) a N k .

148

Proof. For 1 + a k we have
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) a a a a U a a U
a a a a u a a u

+ + + + +
+ + + + +
1 1 1
, 1 1 1

and ( ) ( ) 1 1 + + a U a u . Let inequality (2.6) holds for p a a a k + + ,..., 1 , . Then
( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ), 1 1
, 1 1
p a p a U p a p a U
p a p a u p a p a u
+ + + + + + +
+ + + + + + +

and obviously ( ) ( ) 1 1 + + + + p a U p a u . The proof is complete.

Lemma 2.3. Let for every ( ) ( ) 1 : k a N k and the inequalities (1.7), (1.8) hold. Then
a) The solution ( ) ( ) a N k k u u
cs

, of the problem
( ) N ,
s u a u
cs cs
(2.7)
for the linear equation (1.5) satisfies the relations
( ) ( ) 1 ,..., 1 , , + +
s a a a k k k u
cs
(2.8)
and
( ) ( ) s a c s a u
cs
+ +
(2.9)
Moreover,
( ) ( ) s a a a k k u k u
cs s c
+ +

+
,..., 1 , ,
1 ,
(2.10)

b) The solution ( ) ( ) a N k k u
bs

, of the problem
( ) N ,
s u a u
bs bs
(2.11)
for the linear equation (1.5) satisfies the relations
( ) ( ) 1 ,..., 1 , , + +
s a a a k k k u
bs
(2.12)
and
( ) ( ) s a b s a u
bs
+ +
(2.13)
Moreover,
( ) ( ) s a a a k k u k u
bs s b
+ +

+
,..., 1 , ,
1 ,
(2.14)

Proof.
) Consider the initial problem (2.7). Then
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) a u a a a u a a u
cs cs cs
+ + + + +

1 1 1 (2.15)
Suppose that the formula
( ) ( ) ( ) ( ) ( ) ( )

+
+ + + + + + +
1
1
1
0
1
0
1 1
k
i j
k
i
k
i
cs cs
j a i a i a u k a u (2.16)

holds for 1 ,..., 1 , 0 s k . For 0 k and 1 k it holds since we consequently get the relations (2.7) and (2.15).
Moreover,
( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( )

+ + + + + +
+ + + + + + + +
+ +
1
1
]
1
+ + + + + + + +
+ + + + + +
1
0
1
1
1
0
1
1
2
0
1
0
2
0
2
1
2
0
1 1
1 1 1
1 1 1 1 1
1 1 1 1
s
i
s
i j
s
i
cs
s
i j
s
i
s
i
cs
s
i
s
i j
s
i
cs
cs cs
j a i a i a u
s a j a i a i a u
s a j a i a i a u s a
s a s a u s a s a u

149

Comparing the last expression with (2.16), we conclude that (2.16) holds for s k and, moreover, for every N k .
Using the representation (2.1), we get
( )
( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( ) ( ) ( ) s a c j a i a i a
i a
j a i a s a c
s a u
s
i j
s
i
s
i
s
i
s
i j
s
i
cs
+ + + + + + +
1
1
1
1
1
1
]
1
+ +
+ + + +
+

1
1
1
0
1
0
1
0
1
1
1
0
1 1
1
1

So, the formula (2.9) holds.
) Consider the initial problem (2.11). Then
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) a u a a a u a a u
bs bs bs
+ + + + +

1 1 1 (2.17)

Suppose that the formula
( ) ( ) ( ) ( ) ( ) ( )

+
+ + + + + + +
1
1
1
0
1
0
1 1
k
i j
k
i
k
i
bs bs
j a i a i a u k a u

holds for 1 ,..., 1 , 0 s k (for 0 k and 1 k we consequently get relations (2.11) and (2.17)). Moreover,
( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( )

+ + + + + +
+ + + + + + + +
+ +
1
1
]
1
+ + + + + + +
+ + + + + +
1
0
1
1
1
0
1
0
1
1
2
0
2
0
2
1
2
0
1 1
1 1 1
1 1 1 1 1
1 1 1 1
s
i
s
i j
s
i
bs
s
i
s
i j
s
i
bs
s
i
s
i j
s
i
bs
bs bs
j a i a i a u
s a j a i a i a u
s a j a i a i s u s a
s a s a u s a s a u

Comparing the last expression with (2.17), we conclude that (2.17) holds for s k and, consequently, for every N k .
Using the representation (2.2) we get

( )
( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( ) ( ) ( ) s a b j a i a i a
i a
j a i a s a b
s a u
s
i j
s
i
s
i
s
i
s
i j
s
i
bs
+ + + + + + +
1
1
1
1
1
1
]
1
+ +
+ + + +
+

1
1
1
0
1
0
1
0
1
1
1
0
1 1
1
1

So the formula (2.13) is proved.
) By Lemma 2.1 we have
bs cs
u u , N s .. Then, by Lemma 2.2, for every, ( ) a N k .inequalities
( ) ( ) k u k u
bs cs

holds. Since, by Lemma 2.1,
1
,
+
s c cs
u u , N s and
1
,
+
s b bs
u u , N s , the properties (2.10) and
(2.14) are a consequence of Lemma 2.1 and Lemma 2.2. Let us prove relations (2.8), (2.12). Since
( ) ( ) a u a u
cs bs

and ( ) ( ) ( ) ( ), ,
1 , 1 ,
a u a u a u a u
cs s c bs s b

+

+
Lemma 2.2 gives
( ) ( ) ( ) ( ) k u k u k u k u
cs s c s b bs

+

1 , 1 ,

For s a k + we get
( ) ( ) ( ) ( ) ( ) ( ) s a c s a u s a u s a u s a u s a b
cs s c s b bs
+ + + + + +

+
1 , 1 ,

The last inequalities can be rewritten as
( ) ( ) ( ) ( ) ( ) ( ) s a c s a u s a u s a u s a u s a b
s c s c s b s b
~ ~ ~ ~ ~ ~
1 , ,
+ + + + + +

+

150

Putting here consequently 1 ..., 2 , 1 , 0
~
s s we see that (2.8) and (2.12) hold, too.

Theorem 2.4. [Main Result] Let for every ( ) ( ) 1 : k a N k and the inequalities (1.7), (1.8) hold. Then
every initial problem (1.9) with [ ], ,

c b u where b* and c* are defined by (2.3), determines a solution satisfying
inequalities (1.10).

Proof. The proof is a straightforward consequence of Lemmas 2.1, 2.2, 2.3. For solutions of the problems
( ) ( )

c a U b a u , we have ( ) ( ) ( ) ( ) k c k U k u k b for every ,... 1 , + a a k and ( ) ( ) ( ) k U k u k u
~

for every ,... 1 , + a a k if ( ) [ ]. ,
~
c b a u

Consequence 1. If Theorem 2.4 holds, then the expression
( ) ( ) ( ) ( ) ( ) ( )

+
+ + + +
,
_
+ + +
1
1
1
0
1
0
1 1
s
p
s
i
s
i
p a i a i a u s a u

with [ ]

c b u , is a solution of the problem (1.5), (1.9) satisfying inequalities (1.10).

Proof. The proof follows immediately from the statement of Theorem 2.4 and from the explicit form of the
problem (1.5), (1.9) which can be derived from (1.5) directly.

Theorem 2.5. Let for every ( ) ( ) 1 : k a N k and the inequalities (1.7), (1.8) hold. The initial problem
( )

u a u

with ( ) ( ) [ ] [ ]

c b a c a b u , \ , generate a solution ( ) k u u

of equation (1.5) not satisfying inequalities (1.10) for all
( ) a N k

Proof. Let us suppose that ( ) ( ) a c c u ,

. Then there exists a number N

s s such that
cs
u u . By Lemma
2.2, the inequalities
( ) ( ) k u k u
cs

where ( ) k u
cs
is solution with ( )

cs cs
u a u , hold for very ( ) a N k . In accordance with Lemma 2.3,
( ) ( ) 1 ,..., 1 , , + +

s a a a k k k u
cs

and
( ) ( )

+ + s a c s a u
cs

Moreover, due to (1.5) and (1.8),
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) 1 1 1 1 + + + + + + + + + + + + + +

s a c s a s a c s a s a s a u s a s a u
cs cs

Consequently,
( ) ( ) 1 1 + + + +

s a c s a u

So the inequalities (1.10) do not hold for 1 + +

s a k . The case ( ) [ ]

b a b u , can be considered similarly. The
following two corollaries follow obviously from Theorem 2.4 and Theorem 2.5.

Corollary 1. Let for every ( ) ( ) 1 : k a N k and the inequalities (1.7), (1.8) hold. Then a solution
( ) k u u of equation (1.5) satisfies inequalities (1.10) for every ( ) a N k if and only if ( ) [ ]

c b a u , .

151

Corollary 2. Let for every ( ) ( ) 1 : k a N k and the inequalities (1.7), (1.8) hold. Let, moreover, b* = c*. Then
the equation (1.5) has a unique solution u = u*(k) satisfying for every ( ) a N k inequalities (1.10). This solution is
determined by initial data
u*(a) = u* = b*.

It is interesting to find sufficient conditions for the case b* = c*. Let for every ( ) ( ) 1 : k a N k and the
inequalities (1.7), (1.8) hold. Let us denote
( ) ,... 1 , 0 , s u u s
bs cs

Then the length of the interval [b*,c*] can be estimated (due to the monotonicity of sequences { } { }
0 0
,
s bs s cs
u u as
( ) ,... 1 , 0 , 0

s s b c

From the definition of the expressions
bs cs
u u , we see that ( )
bs cs
u u s [due to (2.1) and (2.2)] =
( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( )
( ) ( )
( ) ( )
( ) ( )
,
1 1
1
1
1
1
0
1
0
1
1
1
0
1
0
1
1
1
0

+
+ +
+
+ +
+ +
+ +
+ +
+ + +
+ +
+ + +
s
i
s
i
s a
i a j
s
i
s
i
s a
i a j
s
i
i a
s a b s a c
i a
j i a s a b
i a
j i a s a c

i.e.
( )
( ) ( )
( ) ( )
+ +
+ +

1
0
1
s
i
i a
s a b s a c
s

The following corollary is obvious.

Corollary 3. Let for every ( ) ( ) 1 : k a N k and the inequalities (1.7), (1.8) hold. Then the following inequalities
hold obviously:
( )
( ) N , b 0
N , 0

s s u
s s c u
bs
cs

Theorem 2.6. Let for every ( ) ( ) 1 : k a N k and the inequalities (1.7), (1.8) hold. Then b* = c* if
( ) 0 lim

s
s
(2.18)

Proof. From Theorem 1.5, the existence a solution of problem (1.5), (1.9) follows. Then
( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( )
( ) ( )
( ) ( )
( ) 0 lim
1
lim
1
1 1
lim
1
1
lim
1
1
lim
1
0
1
0
1
1
1
0
1
1
1
0
1
0
1
1
1
0
1
0
1
1
1
0

+ +
+ +
+ +
+ + + + + + +
+ +
+ + +
+ +
+ + +

+
+ +
+
+ +
+
+ +
+
+ +

s
i a
s a b s a c
i a
j i a j i a s a b s a c
i a
j i a s a b
i a
j i a s a c
b c
s
s
i
s
s
i
s a
i a j
s
i
s a
i a j
s
i
s
s
i
s a
i a j
s
i
s
s
i
s a
i a j
s
i
s

Then c* = b*.

152

Remark 2.7. The condition (2.18) is valid e.g. in the case when
( ) ( ) [ ] ( ) ( ) 0 1 lim 0 lim
1
0
+ + + +

s
i
s s
i a s a b s a c
or in the case when functions c(k), b(k) are bounded on N(a) and ( ) , 0 k const, for every ( ) a N k .

2.3 Concluding remarks
Let us consider a partial case of equation (1.5) with ( ) 1 R, , 1 + A A A k and with a function ( ) k ,
( ) ( ) ( ) k k Au k u + +1 (2.19)

The following theorem is a consequence of the previous results:
Theorem 2.8. Let A > 1 and ( ) ( ) a on M k . Then initial problem
( )
( )
+

+

0
1
i
i
A
i a
u a u

(2.20)

generates a unique bounded solution of equation (2.19) on N(a).

Proof. The series (2.20) is obviously convergent since it can be majorized by a convergent series
0
1
1
i
i
A
M
A
A
M

Put ( ) ( ) M k b M k c : , : with ( ) , 1 / 1 A =const. Then inequalities (1.7) and (1.8) hold since
( ) ( ) ( ) ( ) ( ) ( ) ( ) 0 1 1 1 1 + + + + A M k b k k b k
and
( ) ( ) ( ) ( ) ( ) ( ) ( ) 0 1 1 1 1 + + + A M k c k k c k

for every ( ) a N k . All assumptions of Theorem 2.4 hold. Let us compute the limits c*, b* . In accordance with (2.3),
(2.1), we get

( ) ( ) ( ) ( )
( ) ( )
( )
( )
+
+ +
+
+ +
+

+
+ +
+ + +
0
1 1
0
1
1
1
0
1
0
1
1
1
0
lim
1
1
lim
i
i s
i
s a
i a j
s
i
s
s
i
s a
i a j
s
i
s
u
A
i a
A
A i a M
i a
j i a s a c
c

Since
( )
( ) ( )
( ) ( )
0
2
lim
1
lim lim
1
0

+ +
+ +

s
s
s
i
s s
A
M
i a
s a b s a c
s

then in accordance with Theorem 2.6 we conclude
w* = c* =b*.

This is in accordance with (2.20). Then by Corollary 2 there exists a unique solution u = u*(k) (generated just by
relation (2.20)) satisfying for every ( ) a N k inequalities
( ) M k u M

These inequalities express the boundedness of u*(k) on N(a). Since inequalities (1.7), (1.8) are valid for arbitrary
positive M we can conclude that bounded solution of equation (2.19) is really only one.

153

Theorem 2.8 serves only as an illustration of results obtained. Let us note that this result coincides with result
described in the book [5, p. 77, Exercise 20, part (c)].

2.4 Examples
In this section, two illustrative examples are considered. In the first one we determine initial data generating unique
bounded solution of equation of the type (1.5). In the second one every solution is unbounded, but initial data
determine a particular solution with slightly different asymptotic behaviour.

Example 2.9. Let us consider the equation (1.5) with ( ) ( ) ( ) 2 / 1 , / 1 + k k k k and a = 1, i.e. the equation
( ) ( ) ( ) 1 ,
2
1 1
N k
k
k u
k
k u
+
(2.21)

Define ( ) ( ) ( ) 1 , 1 , 0 N k k c k b . Then
( ) ( ) ( ) ( ) ( ) ( ) 0
2
1
1 1
+
+ + +
k
k k b k k b k

for all ( ) 1 N k and the inequalities (1.7) hold. Moreover,
( ) ( ) ( ) ( ) ( )
( )
0
2
2
2
1 1
1 1
+
+
+ + +
k k k k
k c k k c k
and the inequalities (1.8) hold for all ( ) 1 N k . All conditions of Theorem 2.4 hold and therefore there exists a solution
( ) k u
~
of equation (2.21) such that
( ) 1
~
0 k u (2.22)

for every ( ) 1 N k . Moreover, using (2.3) and (2.1) we get

( ) ( ) ( ) ( )
( ) ( )
( )
( ) ( )
( )

+ + + +
+ +
+
+
+
+ +
+ +
1
]
1
,
_
+
+ +
,
_
+
,
_

,
_
+

+
+
+
,
_
+ +
+
+
+ + + + +
+
+ +
,
_
+
+
,
_
+
+
,
_
+
+
,
_
+
+ +
+
+ +
+ + + +

1
2
1
2
1
0
1
0 2
1
0
1
1
1
0
2
1
2
1
1
1
...
4
1
3
1
3
1
2
1
lim
1
1 1
lim
1
1
lim
1
1
lim
1
1
1
.
2
1 1
.
1
1
...
3
1
.
4
1
2
1
.
3
1
1 1
lim
1
.
1
.....
4
5
.
3
4
.
2
3
. 2
1 1
2
1
...
1
.....
4
5
.
3
4
4
1 1
.....
3
4
.
2
3
3
1
1
lim
1
1
1
1
1
2 1
1
1
lim
1
1
lim lim
s
i
s
i
s s s s
s s
s
i
s
i
s
i j
s
s
i
s a
i a j
s
i
s
cs
s
s s i i i i s
s
s s s s
s
s
s
s
s
s s
s
s
s
i
j i
i a
i a i a s a c
u c

Since
( )
( ) ( )
( ) ( )
0
1
1
lim
1
2
1
lim
1
lim lim
1
0
1
0
+
+
+ +
+ +

s
i
i
i a
s a b s a c
s
s
s
i
s
s
i
s s

then, in accordance with Theorem 2.6, b* = c* = 1/2. Therefore the equation (2.21) has a unique solution
( ) ( ) 1 ,
~
N k k u u satisfying inequalities (2.22), determined by the initial data ( ) 2 / 1 1 u . Indeed, it is easy to verify,
that the function
( ) ( ) 1 ,
1
~
N k
k
k
k u
+

154

is such a solution of equation (2.21). Moreover the general solution of (2.21) is given by formula
( ) ( ) , .
~
k C k u k u +
where C is any constant. It means that ( ) k u
~
is a unique bounded solution of equation (2.21).

Example 2.10. Let us consider the equation (1.5) with ( ) ( ) ( ) 2 , 1 / 2 + k k k and a = 2, i.e. the equation
( ) ( ) ( ) 2 , 2
1
2
N k k u
k
k u +
+
(2.23)

Define b(k) = k
2
, c(k) = (k +1I)
2
, ( ) 1 N k . Then
( ) ( ) ( ) ( ) ( ) ( ) 0
1
1
1 2
1
2
1 1 1
2 2
+ +
,
_
+
+ + + +
k
k
k k
k
k b k k b k

for all ( ) 2 N k and the inequalities (1.7) hold. Moreover,
( ) ( ) ( ) ( ) ( ) ( ) ( ) 0 1 2 2 1
1
2
1 1 1
2 2
+ + +
,
_
+
+ + + + k k
k
k c k k c k

and the inequalities (1.8) hold for all ( ) 2 N k . All conditions of Theorem 2.4 hold and therefore there exists a solution
( ) k u
~
of equation (2.23) such that
( ) ( )
2 2
1
~
+ k k u k (2.24)

for every ( ) 2 N k . Using (2.3) and (2.1) we get
( ) ( ) ( ) ( )
( ) ( )
( )
( )
( )
( ) ( )( )
( )( )
( ) ( )( )
( )
( )( )
( ) ( )( )
( )
( )( )
6
4
24
12
4
1
3
1
3
1
2
1
...
6
1
.
5
1
5
1
.
4
1
. 24
4
3
. 12 lim
4 3 .
12
1
1
1 1
4 3 2 3
lim
4 3 .
12
1
1
1
4 3 2 3
lim
4 . 3
4 3
4
1
.
3
1
.
3
1
.
2
1
...
6
1
.
5
1
5
1
.
4
1
4 3 2 3
lim
2
4
.
1
3
.
2
.
1
1
.
2
.....
6
8
.
5
7
.
4
6
.
3
5
1
2
4
...
2
4
.....
6
8
.
5
7
2
4
.....
5
7
.
4
6
2 3
lim
3
5
1
3
2 3
lim
1
2
1
1
2
1 2 1
lim
1
1
lim lim
3
4
2
3
4
2
2
2
1
0
1
3
1
0
2
1
0
1
1
1
0
2
1
0
1
1
1
0

1
]
1
,
_
+
+
+
+
+ + +
+
+
+ +
,
_
+
+ + +
+ +
+
+ + +
+ +
1
]
1
+ + + +
+ + + + + +
+
+
+
+ +
1
]
1
+
,
_
+
+
+ +
,
_
+
+
+
,
_
+
+
+
+
+
+
+
+
,
_
+ +
+
,
_
+
+ + +
+ +
+ + +

+
+
+
+
+ +
+
+
+
+
+ +
+
+ +
+ +

s s s s s
s
s s
i i
s s s
s s
i i
s s s
s s
s s s s
s s s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
i
i
j
j
s
i a
j
s a
i a
j i a s a c
u c
s
s
i
s
s
i
s s
s
s
i
s
i j
s
i
s
s
i
s
i a j
s
i
s
s
i
s a
i a j
s
i
s
cs
s

155

Since
( )
( ) ( )
( ) ( )
( ) ( )
( )( )
( )
( )( )
0
4 3
5 2 12
lim
4 3 .
12
1
2 3
lim
1
lim lim
2 2
1
0
+ +
+
+ +
+ +
+ +
+ +

s s
s
s s
s s
i a
s a b s a c
s
s s
s
i
s s

then, in accordance with Theorem 2. 6, b* = c* = 6. Therefore the equation (2. 23) has a unique solution
( ) ( ) 2 ,
~
N k k u u satisfying inequalities (2.24), determined by the initial data ( ) 6 2
~
u . Indeed, it is easy to verify that
the function
( ) ( ) ( ) 2 , 1
~
N k k k k u +

is such a solution of equation (2.23). It is easy to see that the general solution of (2.23) is given by formula
( ) ( ) ( )( ) 2 1 .
~
+ + + k k C k u k u
where C is any constant.

Acknowledgment
The first author was supported by the Grant 201/04/0580 of the Czech Grant Agency (Prague), the second author was
supported by the Council of Czech Government MSM 00216 30503.

References
[1] AGAWAL, R.P. Differential Equations and Inequalities, Marcel Dekker, Inc., 2nd ed., 2000.
[2] AGAWAL, R.P.; POPENDA, J. Periodic solutions of first order linear difference equations, Mathl. Comput.
Modelling 22, 11-19, 1995.
[3] DIBLK, J. Discrete retract principle for systems of discrete equations, Comput. Math. Appl 42 (2001), 515-
528.
[4] DIBLK, J. Asymptotic behaviour of solutions of discrete equations, Fund. Differ. Equ., 11 (2004), 37-48. J.
Diblik, Retract principle for difference equations Proceedings of the Fourth International Conference on
Difference Equations, Poznan, Poland, August 27-31, 1998. Eds.: S.Elaydi, G. Ladas, J. Popenda and J.
Rakowski, Gordon and Breach Science Publ., 107-114, 2000.
[5] ELAYDI, S. N. An Introduction to Difference Equations, Springer, 1999. Second Edition.
[6] GOLDA, W.; WERBOWSKI, G. Oscillation of linear functional equations of the second order, Funkc. Ekvac.
37 (1994), 221-227.
[7] GYORI, I.; PITUK, M. Asympotic formulae for the solutions of a linear delay difference equation, J. Math.
Anal. Appl. 195 (1995), 376-392.
[8] GYORI, J.; PITUK, M. Comparison theorems and asymptotic equilibrium for delay differential and difference
equations, Dyn. Systems and Appl. 5 (1996), 277302.
[9] MIGDA, M.; MIGDA, J. Asympotic behaviour of solutions of difference equations of second order, Demonstr.
Math. XXXII (1999), 767-773.

Adresa:
Doc. RNDr. Jaromr Batinec, CSc.
Technick 8,
616 00 Brno,
bastinec@feec.vutbr.cz

Adresa:
Doc. RNDr. Josef Diblk, DrSc.
Technick 8,
616 00 Brno,

156

157

ON SOME PROPERTIES OF FRACTIONAL CALCULUS
Vlasta Krupkov, Zdenk marda

Abstract: This paper is devoted to some properties of the fractional integral and derivative. There are
shown applications of this calculus in cases of some classes of fractional integral equations.

Key words: Fractional integral, fractional derivative, Laplace transform.

1. Introduction
The fractional calculus is a generalization of integration and derivation to non-integer order operators [4,7]. The idea of
fractional calculus has been known since the development of the normal calculus, with the first reference probably being
associated with Leibniz and L'Hospital in 1695.

Fourier, Euler, Laplace are among the many that dabbled with fractional calculus and the mathematical consequences [6].
The most famous of these definitions that have been popularized in the world of fractional calculus are the Riemann-
Liouville and Grunwald-Letnikov defintion [1]. From the view of requirements of physical reality Caputo reformulated the
more "classic" definition of the Riemann-Liouville fractional derivative in order to use integer order initial conditions to
solve fractional order differential equations [7]. As recently as 1996, Kolowankar reformulated again, the Riemann-
Liouville fractional derivative in order to differentiate no-where differentiable fractal functions [2].

In this paper we will also be devoted to other constructions of the fractional derivative especially, and we show some
particularities of these ones occuring at solving of certain classes of integral equations.

2. The fractional integral
Understanding of definitions and use of fractional calculus will made necessary some but relatively simple
mathematical definitions that will arise in the study of these concepts.

Euler's Gamma function:
( )

0
1
, dx e x t
x t
(1)
special case when n x :
( ) ( )! 1 n n (2)

Mittag-Leffler function in two parameters:
( )
( )

+

k
t
t
k
k 0
,
0 , 0 (3)

It is a generalization of exponential function
( )
( )
t
k
k
k
k
e
k
t
k
t
t
+

! 1
0 0
1 , 1

More particular cases
( ) ( ) ( ) ( ) t erfc e t t t
t

2
, cosh
1 , 2 1 1 , 2
(4)

158

The Laplace transform:
( ) } { ( ) ( ) p F dt t f e t f L
pt

0
: (5)

Also commonly used is a convolution of two function
( ) ( ) ( ) ( ) ( ) ( ) t f t g d g t f t g t f
t

0
:
( ) ( ) } ( ) ( ) { p G p F t g t f L (6)

One final important property of the Laplace transform that should be addressed is the Laplace transform of a derivative
of integer order n of the function ( ) t f
( )
( ) } ( )
( )
( ) ( )
( )
( )

'

1
0
1
1
0
1
0 0
n
k
k n k n
n
k
k k n n n
f p p F p f p p F p t F L (7)

Cauchy formula for evaluating the n
th
integration of the function:
( )
( )
( )
( )
( )

t t
n
d f t
n
d f
0 0
1
! 1
1
... (8)
For the abbreviated of this formula, we introduce the operator
n
I
( ) ( )
( )
( )
( )
( )

t
n
n
n
d f t
n
t f t f I
0
1
! 1
1
: (9)

For direct use in (8), n is restricted to be an integer. The primary restriction is the use of the factorial which in essence
has no meaning for non-integer values. The Gamma function is however an analytic expansion of the factorial for all
reals, and thus can be used in place of the factorial as in (2). Hence, by replacing the factorial expression for its Gamma
function equivalent, we can generalize (9) for all
+
R

( ) ( )
( )
( ) ( )

t
d f t t f t f I
0
1
1
:
(10)
This approach is commonly referred to as the Riemann-Liouville approach. This formulation of the fractional integral
carries it some very important properties, that will later show importance when solving equations involving integrals
and derivatives of fractional order. First, we consider integrations of order
0
to be an identity operator, i.e.
( ) ( ) t f t f I
0

Also, given the nature of the integral's definition, and based on the principle from which it came (Cauchy repeated
integral equation), we can see that just as
N n m I I I I I
n m n m m n

+
, ,
(11)
so to,
R I I I J I
+

, ,

The one presupposed condition placed upon a function
( ) t f
that needs to be satisfied for these and other similar
properties to remain true, is that
( ) t f
be a causal function, i.e. that it is vanishing for
0 t
. Although this is a
consequence of convection, the convenience of this condition is especially clear in the context of the property
demonstrated in (11). The effect is such that
( ) ( ) ( ) 0 0 0 0

f f f
n
. Using the Gamma function we define a
function
( ) t

( )
( )

1
:
t
t

From this we obtain

159

( ) ( )
( )
( )
( )
t
d f
t
t f t
0
1

+
t
denotes the function vanishes for
0 t
. Now the formula of the fractional integral (10) can be written in the form
( ) ( ) ( )
( )
( ) ( )

t
d f t t f t t f I
0
1
1

(12)

We will find the Laplace transform of the Riemann-Liouville fractional integral. In (12) we showed that the fractional
integral could be expressed as the convolution of two functions,
( ) t
and
( ) t f
. The Laplace transform of
1
t
is
given by
} ( ) {

p t L
1

Thus, the Laplace transform of the fractional integral is found to be
( ) } ( ) { p F p t f I L

(13)

3. The fractional derivative
Because the Riemann-Liouville approach to the fractional integral began with an expression for repeated integration of
a function, one's first instinct may be to imitate a similar approach for the derivative. In the following we introduce two
definitions fractional derivative in the sense of the Riemann-Liouville approach. Consider a differentiation of
order
+
R
. Now, we select an integer m such that
m m 1
. Given these numbers, we have two possible
ways to define the derivative. Having found the integer m, the first step of the process is to integrate our function
( ) t f
by order
m
and second, we differentiate the resulting function
( ) t f
m
by order m. This method we will call
Left Hand Definition (LHD) of the fractional derivative and there is given
( )
( )
( )
( )
( )
'
1
1
]
1
+
m m t f D
t
m m
m
m
m
d
t
f
m dt
d
m t f
dt
d
L

1 , :
0
1
1
,
(14)
The Right Hand Definition (RHD) attempts to arrive at the same result using the same operations, but in the reverse
order. The mathematical results of this is the form

( )
( )
( )
( )
( )
( )
'
m m t f D
t
m
m
m
m
d
t
f
m
m t f
dt
d
R

1 , :
0
1
1
,
(15)

This second definition, although referred to here as the Right Hand Definition, was originally formulated by Caputo,
and is therefore, commonly referred to as the Caputo fractional derivative. Demonstrating the practicality of the RHD
over the LHD is conveniently simple. For example, the fractional derivative of a constant using the LHD is not zero,
and in fact there is valid
( )
1
Ct
C D

which is a substantial problem in the physical world. In the section 2. we demonstrated the Laplace transform of the
fractional integral (13). Using this definition, we may find similarly the Laplace transform of LHD fractional derivative.
The fractional derivative of this one may be written in the form
( )
( )
( ) ( ) ( ) m m t f I t g where t g t f D
m m
L

1 , ,

Using the formula (7) and the definition of the fractional integral Laplace transform, we obtain
( ) } ( )
( )
( ) ( )
( )
( )

'

1
0
1
1
0
1
0 0
m
k
k
L
k
m
k
k m k m
L
f D p p F p g p p G p t f D L

(16)

160

It is obvious that the required initial conditions are for all k to n 1 terms, fractional order derivatives of
( ) t f
.
For the RHD, we will write the derivative in the form
( ) ( ) ( )
( )
( ) m m t F t g where t g I t f D
m m
R

1 , ,

Using the formula (13), we get
( ) } { ( ) ( ) ( )
( )
( )

1
0
1
0
m
k
k k
R
f p p F p p G m p t f D L

(17)
In this formulation, the order a does not appear in the derivatives of
( ) t f
. So, quite conveniently , integer order
derivatives
( ) t f
are used as the initial conditions, and therefore easily interpreted from physical data and observations.
Consider the fractional integral equation of the first kind
( )
( )
( )
( )
( )

t
t f d
t
u
0
1
1 0 ,
1

(18)
This equation can be written in the form
( ) ( ) t f t u I

There is valid
( ) ( ) ( ) ( ) } {
( )

p
p U
t L t u t t u I
(19)
We can reorder the result (19) into one of two forms.
( ) ( )
( )
,
_
1
p
p F
p p F p p U

or
( ) ( ) ( ) ( ) ( )
( )

+
1 1
0
0
1
p
f
f p pF
p
p F p p U

Inverting the first form into the time domain, we get

( )
( )
( )
( )
( )
t
t f d
t
f
dt
d
t u
0
1
1

which is equivalent to solution of (18) with the LHD. The second form can be similarly inverted to yield
( )
( )
( )
( )
( ) ( )
( )
t
t
f t f d
t
f
t u
0
1
0
1
1

which is equivalent to solution of (18) with the RHD.

Now we consider the fractional integral equation of the second kind in the form
( )
( )
( )
( )
( ) ( ) ( ) ( )
+

+
t
t f t u I t f d
t
t u
t u
0
1
1

(20)
Applying the Laplace transform to (20) we get
( ) ( ) } ( ) } ( ) ( )
'
'
1
]
1
+ + p F p U
p
t f L t u I L

1 1
(21)
The equation (21) can be rearranged in many ways, but we will solve this one using of LHD. We can rewrite the
equation (21) as follows
( ) ( ) ( ) p F p F
p
p
p p U +
1
]
1
1
1
(22)

161

The equation (22) is next inverted back into the normal function domain. In order to do this one must address
comprehend the Laplace transform of the Mittag-Leffler function
( ) }
'
1
1
1 ,

p
p
t E L
(23)

By the relationship given in (7), it is clear that between the brackets in (22) is the Laplace transform of the first
derivative of the Mittag- Leffler function in (23), i.e.
( ) }
( )
( ) }
'
'
1
1
1
1 , 1 ,
p
p
p t E L t E
dt
d
L

From this the inverse (22) we obtain
( ) ( )
( )
( ) ( ) t f t E t f t u +

1
1 ,

Acknowledgement
This research has been supported by the Czech Ministry of Education in the frame of MSM002160503 Research
Intention MIKROSYN New Trends in Microelectronic Systems and Nanotechnologies.

Reference
[1] CHEN, Y.Q. Fractional-order calculus in Signal processing and Control, CSOIS, ECE Dept. of Utah State
University, 2003, 1-83.
[2] KOWANKAR, K. M.; GANGAL, A. D. Fractional Differentiability of nowhere differentiable functions and
dimensions, Chaos, Vol.6, No 4, 1996, Amer. Inst. of Physics.
[3] NISHIMOTO, K. An essence of Nishimoto's Fractional Calculus, Descartes Press Corp., 1991.
[4] OLDHAM, K. B.; SPANIER, J. The Fractional Calculus, Acad. Press, New York, 1974.
[5] PIRES, E. J. S. Fractional Order Dynamics in a GA planar, Signal Processing 83, 2003, 2377-2386.
[6] PODLUBNY, I. Fractional Differential Equations, Mathematics in Science and Engineering Vol. 198, Acad.
Press 1999.
[7] PODLUBNY, I. Fractional Differential Equations, Acad. Press, San Diego 1999.
[8] RAYNAUD, H. F.; ZERGALNOH, A. State-space Representation for Fractional order controllers,
Automatica 36, 2000, 1017-1021.

Address:
RNDr. Vlasta Krupkov, CSc.
University of Technology
Technick 8
616 00 Brno
e-mail: krupkova@feec.vutbr.cz

Address:
Doc. RNDr. Zdenk marda, CSc.
Technick 8
616 00 Brno
e-mail: smarda@feec.vutbr.cz

162

163

ON THE STABILITY OF LINEAR INTEGRODIFFERENTIAL EQUATIONS
Zdenk marda
University of Technology, Brno

Abstract: The stability solutions of linear integrodifferential equations with respect to an asymptotic
behaviour of base and regular ordinary differential equations is investigated.

Key words: Stability solutions, integrodifferential equations, base and regular system.

1. Introduction
While studying integro-differential equations (IDE) we often meet two basic problems, which do not have an analogy in
the theory of ordinary differential equations:
The system of IDE has singular points of the first and second order , i.e. points that more one solution
passes through or no solution exists there at all, in simple cases, presuming continuity (see [1,2]).
An unknown function is under the integration sign, so the right-hand side of the IDE cannot be evaluated
in given point. This means that we are actually unable to determine the direction field of the IDE, although
the direction field exists. Therefore known qualitative methods of an investigation of ordinary differential
equations, e.g. Wazewki's topological method, cannot be applied to IDE (see [4,5,6,7]).

In this paper, we investigate stability of solutions of linear systems of IDE and appropriating base systems. There is
introduced some examples base systems and linear systems of IDE in which integral terms will change asymptotic behaviour
of base systems. Character of asymptotic behaviour of linear systems of IDE with respect to base systems depends on
kernels of integral terms, especially. There are given sufficient conditions under which linear systems of IDE is stable or
asymptotic stable with assumption that base systems is unstable.

2. Examples and basic notions
Consider the linear system of IDE
( ) ( )
+
t
ds s u s t K t u t A t u
0
, , ) ( ) ( ) ( ' (1)

where u
(i)
(t) = (u
1
(i)
(t),.,u
n
(i)
(t))
T
, i = 0,1, A(t) is n x n- the matrix function, A(t) C
1
(J), detA(t) 0, for t J, K(t,
s) is n n - the matrix function, K(t, s) C
1
(J x J), J = [0, ). The function K(t, s) will be called the kernel of (1).

We also consider the base system of (1)
u'(t) = A(t)u(t). (2)

Let V(t), V(0) = E, be the fundamental matrix of (2) then the Cauchy problem of (2) with u(0) = b, b = (b1,..., b
n
)
T
is a
constant vector, has the unique solution
u(t) = V(t)b. (3)

The zero solution of (2) will be called stable if for any constant vector c and t there is valid
, ) ( ) ( c M c t V t u R M
+

If as t , ||u(t)|| 0 then the zero solution of (2) will be called asymptotic stable.
If as t , ||u(t)|| then the zero solution of (2) will be called unstable, ||.|| is a usual norm in R
n
or R
2n
.

164

The system (2) is the linear system. Thus, there is alwayes valid one of these cases:
(I) All solutions of (2) are stable.
(II) All solutions of (2) are asymptotic stable.
(III) All solutions of (2) are unstable.

Consequently, we will say that the system (2) is stable, asymptotic stable, unstable. There is also valid in the case of (1).

Example 1.
Consider the system of IDE
dr u
t
t u
t
t
t u
t
) (
0 0
1
1
0
) (
1
1
0
0
1
1
) ( '
0
,
_
+
+
+
,
_
(3)

with the initial conditions u1(0) = b1, u
2
(0) = b
2
. The particular solution of (3)
.
1
1
0
) 1 ( 2 1
1
) (
2
b
t
t
t
t
t u
,
_
+
+ +

Thus, the system (3) is unstable. Appropriating the base system has the form
u'
k
(t)= ) (
1
1
t u
t
k
+
2 , 1 , ) 0 ( k b u
k k
(4)

with the particular solution
u
k
(t) =
k
b
t 1
1
+

From here the system (4) is asymptotic stable so that the integral term in (3) changed the asymptotic stable system (4) in
the unstable system (3).

Example 2. Consider the system of IDE

d u
t
t t
t u
t
t t
t u
t
) (
) 1 (
) 1 ( 2
0
) 2 ( ) 1 (
1
0
) (
1
1
0
0
) 2 )( 1 (
1
) ( '
0
3
3 4
,
_
+
+
+ +
+
+
,
_
+
+ +
(5)

with the initial conditions u1(0) = b1, u
2
(0) = b
2
. The particular solution of (5)

1
]
1
,
_
+
+
+
+
+
+
+
+
+
+
3 2
2 1 1
) 1 (
2
) 1 (
3
1
2
2
1
ln
2
) 1 ( 2
) (
t t t t
t
b b
t
t
t u

. 2 2
1
1
2 ) ( b
t
t u
,
_
+

165

From here the system (5) is only stable. Appropriating the base system has the form

) (
1
1
0
0
) 2 )( 1 (
1
) ( ' t u
t
t t
t u
,
_
+
+ +
(6)

with the particular solution
,
2
) 1 ( 2
) (
1 1
b
t
t
t u
+
+

2 2
) 1 ( ) ( b t t u +

Thus, the system (6) is unstable. Now it is obvious that the integral term in (5) changed the unstable system (6) in the
stable system (5).

3. Main results
Let R(t, s) C
1
(JX J) be a resolvent the matrix kernel A
-1
(t)K(t, s).
Put
+
t
ds s z s t R t z t u
0
. ) ( ) , ( ) ( ) ( (7)

Substituting (7) into (1) we obtain

t
t
ds s z s t R t z t B t z
0
'
, ) ( ) , ( ) ( ) ( ) ( '

where B(t) = A(t) + A
-1
(t)K(t,t). Consider the system
) ( ) ( ) ( ' t z t B t z (9)
which will be called the regular system with respect to (1) and let W(t) be the fundamental matrix of (9).

Theorem. Let following assumptions hold:
(i) The base system (2) is unstable.
(ii) The regular system (9) is either stable or asymptotic stable.
(iii)

< <
t
t
N dsdt s W s t R s W
0
' 1
0
) ( ) , ( ) (

. ) ( ) , ( ) (
0
< <
L ds s W s t R t F
t

F(t) > 0
ast
in the case of the asymptotic stability.
Then the system (1) is either stable or asymptotic stable.

Proof. The resolvent R(t, s) of (7) satisfies the equation (see [3])
( ) ( ) ( )
+
t
d s R t Q s t Q s t R
0
, , , ) , (
(10)

where Q(t, s) = -A
-1
(t)K(t, s). From (8) we get

+
s
t
t
ds d z s R s W t W b t W t z
0
'
0
1
. ) ( ) , ( ) ( ) ( ) ( ) (
(11)

166

Multiplying by
( ) t W
1
the both sides of (11) and putting
( ) ( ) ( ) t z t W t
1

we obtain
. ) ( N b t
Thus
bN t W t z ) ( ) (

Now, from (7) it follows
+

t
ds s z s W s W s t R t z t W t W t u
0
1 1
) ( ) ( ) ( ) , ( ) ( ) ( ) ( ) (
+

ds s z s W s W s t R t z t W t W
t
0
1 1
) ( ) ( ) ( ) , ( ) ( ) ( ) (
. ) ( ( ) ( ) , ( ) (
0
b N L t W b N ds s W s t R t W
t
+
1
1
]
1
(12)

From here it is obvious that the system (1) is stable . In the case F(t) > 0 and ||W(t)|| > 0 as t , we obtain from
(12) that the system (1) is asymptotic stable. The proof is complete. Results of this Theorem we apply to the system of
IDE in the example 2, i.e.
dr u
t
t t
t u
t
t t
t u
t
) (
) 1 (
) 1 ( 2
0
) 2 ( ) 1 (
1
0
) (
1
1
0
0
) 2 )( 1 (
1
) ( '
0
3
3 4
,
_
+
+
+ +
+
+
,
_
+
+ +

with the initial conditions ( ) ( )
2 2 1 1
0 , 0 b u b u . We already know that the base system
( )
( )( )
( ) t u
t
t t
t u
,
_
+
+ +
1
1
0
0
2 1
1

is unstable. The matrix B(t) of the regular system z'(i) = B(t)z(t) has the form
( ) ( ) ( ) ( )
( )( )
,
_
+ + +
+
+

1 0
2 1
1
2
1
1
1
,
1
t t t
t
t t K t A t A t B
and the resolvent
( )
( )( )
,
_
+ +
2 0
2 1
1
0
1
1
,
t t
t R

The fundamental matrix of the regular system
( )
( )
( )
( )
.
1
1
0
1
1
1
2 3
1
2
1 2
3
,
_
+
1
1
]
1
+
+
+
+
t
t
t
t
t
t
t W

Thus, the regular system is stable. Remain to verify the conditions (iii),(iv). In this case there is valid
( ) ( )
( )
( )( )
,
_
+ +
2 0
2 1
2
0
1
1
,
2
t t
W t R

( ) ( )
N d W t R
t
2 ln
6
5
,
0

167

( ) ( ) ( )
( )( ) ( ) ( )
.
0 0
1 0
2
1
1
1
1 1 2
2
,
2 2 2
1
,
_
,
_
+ + +
+
t t t
t
W t R t W
t

( ) ( ) ( )

L dt d W t R t W
t
t
3 ,
0
1
0

Now, it is obvious that all assumptions of Theorem are fulfilled and withought computting of the particular solution of
the system of IDE (5) we get that this one is also stable. The example is complete.

Acknowledgement

Reference
[1] BYKOV, J. V. Theory of integrodifferential equations , Kirg. Univ. Frunze 1957 (in Russian).
[2] IMANALIEV, M. Oscillation and stability of solutions of singular-p erturbation integro differential equation,
Akad. nauk , ILIM Frunze, 1974 (in Russian).
[3] KREK, J. Zklady aplikovan matematiky III., Praha : SNTL, 1993.
[4] MARDA, Z. On some particularities of integro-differential equations. Proceedings of APLIMAT 2003,
p.253-257.
[5] MARDA, Z. On solutions of an implicit singular system of integro differential equations depending on a
parameter, Demonstratio Mathematica, Vol.XXXI, No 1, (1998), 125-130.
[6] MARDA, Z. On an initial value problem for singular integrodifferential equations, Demonstratio Mathe-
matica, Vol. XXXV, No 4, (2002), 803-811.
[7] YANG, G. Minimal positive solutions to some singular second-order differential equations, J. Math. Anal.
Appl. 266 (2002), 479-491.

Address:
Doc. RNDr. Zdenk marda, CSc.
Technick 8
616 00 Brno
e-mail: smarda@feec.vutbr.cz

168

169

EXISTENCE OF POSITIVE SOLUTIONS FOR RETARDED FUNCTIONAL DIFFERENTIAL
EQUATIONS WITH UNBOUNDED DELAY AND FINITE MEMORY
Josef Diblk, Zdenk Svoboda
University of Technology, Brno

Abstrakt: For systems of retarded functional differential equations with unbounded delay and with finite
memory sufficient and necessary conditions of existence of positive solutions on an interval of the form
[ )
, t are derived. A general criterion is given together with corresponding applications (including a linear
case, too). Examples are inserted to illustrate the results.

Key words and phrases: Positive solution, delayed equation, p- function.

1 Introduction
One of the basic projects in the theory of regulation is to find the control of some process such that parameters of this stay
in the required area. The most simply type of this area is often described by inequalities. Description of continuous
processes is usually realized by using the differential equation. In this paper is given a criterion for the existence of positive
solutions (i.e. a solution with positive coordinates on a considered interval) for systems of retarded functional
differential equations (RFDE's) with unbounded delay and with finite memory. At first let us give short explanation of
emphasized above terms. Let us recall basic notions of RFDE's with unbounded delay but with finite memory. A
function p C[R x [-1,0], R] is called a p -function if it has the following properties [12, p. 8]:

(i) p(t,0)=t.
(ii) p(t, -1) is a nondecreasing function of t.
(iii) there exists a - such that p( t, ) is an increasing function for for each t ( , ). (Throughout the
following text we suppose t ( , ).)

In the theory of RFDE's the symbol y
t
, which expresses "taking into account", the history of the process y(t) considered,
is used. With the aid of p - functions the symbol y
t
is defined as follows:

Definition 1 ( [12, p. 8] ) Let t
0
R, A > 0 and y C ( [p(t
0
, -l), t
0
+ A), R
n
). For any t [t
0
,t
0
+ A), we define
( ) ( ) ( ) 0 1 , , : t p y y
t

and write

y
t
C := C[[-l,0],R
n
].

Note that the frequently used symbol y
t
(e.g., y
t
(s) := y(t + s), where , 0 , 0 s = const) in the theory of
delayed functional differential equations for equations with bounded delays is a partial case of the above definition. Indeed,
in this case we can put p (t, ) := t + T , [-1, 0].
In this paper we investigate existence of positive solutions of the system

(t) = f(t,y
t
) (1)
where f C([t
0
,t
0
+A) x C,R
n
), A > 0, and y
t
is defined in accordance with Definition 1. This system is called the system of
p-type retarded functional differential equations (p-RFDE's) or a system with unbounded delay with finite memory.

Definition 2 The function y C ([p(t
0
, -l),t
0
+A),R
n
) C
1
([t
0
,t
0
+ A),R
n
) satisfying (1) on [t
0
,t
0
+A) is called a solution
of (1) on [p(t
0
, - 1),t
0
+ A).

Suppose that . is an open subset of R C and the function f : . > R
n
is continuous. If (t
0
, ) , then there

170

exists a solution y = y(t
0
, ) of the system p-RFDE's (1) through (t
0
, ) (see [12, p. 25]). Moreover this solution is
unique if f(t, ) is locally Lipschitzian with respect to second argument ([12, p. 30]) and is continuable in the usual
sense of extended existence if f is quasibounded ([12, p. 41]). Suppose that the solution y = y(t
0
, ) of p-RFDE's (1) through (t
0
,
) , defined on [t
0
, A], is unique. Then the property of the continuous dependence holds too (see [12, p. 33]), i.e. for
every > 0, there exists a ( ) > 0 such that (s, ) , |s - t
0
| < and || - || < imply

|| yt (s, ) - (to, )|| < , for all t [ ,A]

where y(s, ) is the solution of the system p- RFDE's (1) through (s, ), = max {s,t
0
} and || || is the
supremum norm in R
n
. Note that these results can be adapted easily for the case (which will be used in the sequel) when
has the form = [t*, ) x C where t* R.

1.1 Problem of existence of positive solutions
In this paper we are concerned with the problem of existence of positive solutions (i.e. problem of existence of
solutions having all its coordinates positive on considered intervals) for nonlinear systems of RFDE's with unbounded
delay but with finite memory. Let us cite some known results for retarded functional differential equations. For the
scalar equation
( ) ( ) ( ) ( ) 0 + t t x t p t x & (2)

with p, C ( [t
0
, ), R+), (t) t,
t
lim (t - (T)) = and R
+
= [0, ) a criterion for existence of a positiv solution
is given in the book [10]. Namely, (2) has a positive solution with respect to t1 if and only if there exists a continuous
function (t) on [T1, ) with T
1
=
1
inf
t t
{t - (t)}, such that (t) > 0 for t t
1
and

(t) p(t)e
t
t t ) (
(s)ds
, t t1. (3)

(A function x is called a solution of (2) with respect to an initial point t
1
t
0
if x is defined and is continuous on [T
1
, ),
differentiable on [t
1
, ), and satisfies (2) for t t
1
.) Results in this direction are formulated in the book [11] and in the
papers [1, 2], too. Positive solutions of (2) in the critical case were studied e.g. in [4]-[10]. The cited criterion was
generalized for nonlinear systems of RFDE's with bounded retardation in [3] and for nonlinear systems of RFDE's with
unbounded delay and with finite memory in [6]. These generalizations are in a sense "direct" generalizations since in their
formulations existence of a positive (vector) functions playing a similar role as in (3) is supposed.

2 Sufficient conditions
Let a constant vector k >> 0 and a vector (t) defined and locally integrable on [p*, ) are given. Then the operator T
is well defined by
T(k, ) (t) := ke
t
p*
(s)ds
= (k
1
e
t
p*
(s)ds
,..., k
n
e
t
p*
(s)ds
).

Define for every i {1, 2,..., n} two types of subsets of the set C:

i
: = { C: 0 ( ) T (k, )t ( ), [-1,0] except for
i
(0) = k
1
e
t
p*
(s)ds
}

and
i
: = { C: 0 ( ) T (k, )t ( ), [-1,0] except for
i
(0) = 0 }.

Theorem 1 Suppose f C ( ,R
n
) is locally Lipschitzian with respect to the second argument and quasibounded. Let a
constant vector k >> 0 and a vector (t) defined and locally integrable on [p*, ) are given. If, moreover, inequalities

171

i
ki i i
t

> ) ( e
t
p*
(s)ds
) ( t f
i
(4)
hold for every i { 1 , 2 , . . . , n}, (t, ) [t*, ) x
i
and inequalities
0 ) , ( > t f
i i
(5)

hold for every i { 1 , 2 , . . . , n}, (t, ) [t*, ) x
i
, where
i
= - 1 for i = 1,... ,p and
i
= 1 for i = p+1,... ,n,
then there exists a positive solution y = y(t) on [p*, ) of the system p -RFDE's (1).
The proofs of this and next theorems are based on the retract method and on the Lyapunoff method. Analogous
consideration can be found in [6].

3 Scalar linear application
Let us consider the scalar linear equation with delay
( ) ( )
( )
( )ds s y s t K t y
t
t
, & (6)

where [ ) [ )
+
R p x t K , , : is a continuous function, and [ ) [ )

, , : p t is a nondecreasing function with
( ) t t .

Theorem 2 The equation (6) has a positive solution ( ) [ )

, p on t y y if and only if there exists a
function [ ) ( ) R p C , ,

, such that ( )

t t for t 0 and
( )
( )
( )

,
_
t
t
du u
ds e s t K t
t
s
, (7)
on the interval [ )
, t .

Inequality (7) can be used for finding sufficient conditions for the existence of a positive solution of Eq. (6). Let us give
two of them.

In the case when ( )

t p t and ( ) ( ) t c s t K , for every [ )

, t t , Eq. (6) takes the form
( ) ( ) ( )

t
p
ds s y t c t y& (8)

Theorem 3 For the existence of a solution of Eq. (8), positive on [ )
, p , the inequality
( )
( )
[ )

, ,
1
2
t t
e
t c
p t
(9)
with a positive constant S is a sufficient condition.

In the case when ( ) ( ) ( ) t c s t K and R t t
+
, 1 , 1 for every [ )

, t t , Eq. (6) takes the form
( ) ( ) ( )

t
t
ds s y t c t y
1
& (10)

Theorem 4 For the existence of a solution of Eq. (10), positive on [ )
, 1 , t , the inequality
( ) [ )

, , t t M t c (11)
is sufficient for ( )
2
/ 2 l M const with a constant a being the positive root of the equation

e 2 2 . (The
approximate values are a = 1., 5936 and M = 0., 6476/l
2
.)

172

Acknowledgment

Reference
[1] BEREZANSKI, L.; BRAVERMAN, E. On oscillation of a logistic equation with several delays, J. Comput.
and Appl. Mathem. 113 (2000), 255-265.
[2] ERMK, J. A change of variables in the asymptotic theory of differential equations with unbounded delay,
J. Comput. Appl. Mathem. 143 (2002), 81-93.
[3] DIBLK, J. A criterion for existence of positive solutions of systems of retarded functional differential
equations. Nonl. Anal., TMA 38 (1999), 327-339.
[4] DIBLK, J. Positive and oscillating solutions of differential equations with delay in critical case, J. Comput.
Appl. Mathem. 88 (1998), 185-202.
[5] DIBLK, J.; KOKSCH, N. Positive solutions of the equation x(t) = c(t)x(t r) in the critical case. J. Math.
Anal. Appl. 250 (2000), 635-659.
[6] DIBLK, J.; SVOBODA, Z. An existence criterion of positive solutions of p-type retarded functional differential
equations. J. Comput. Appl. Mathem. 147 (2002), 315-331.
[7] DOMSHLAK, Y. On oscillation properties of delay differential equations with oscillating coefficients, Funct.
Diff. Equat., Israel Seminar 2 (1996), 59-68.
[8] DOMSHLAK, Y.; STAVROULAKIS, I. P. Oscillation of first-order delay differential equations in a critical
case, Appl. Anal. 61 (1996), 359-371.
[9] ELBERT, .; STAVROULAKIS, I. P. Oscillation and non-oscillation criteria for delay differential equations,
Proc. Amer. Math. Soc. 123 (1995), 1503-1510.
[10] ERBE, L. H.; KONG, Q.; ZHANG, B. G. Oscillation Theory for Functional Differential Equations. New York :
Marcel Dekker, 1995.
[11] GYRI, I.; LADAS, G. Oscillation Theory of Delay Differential Equations. Oxford : Clarendon Press, 1991.
[12] LAKSHMIKAMTHAN, V.; WEN, L.; ZHANG, B. Theory of Differential Equations with Unbounded Delay,
Kluwer Academic Publishers, 1994.
[13] HALE, J.K.; LUNEL, S. M. V. Introduction to Functional Differential Equations, New York : Springer-Verlag,
Inc., 1993.

Address:
Doc. RNDr. Josef Diblk, DrSc.
Technick 8,
616 00 Brno,

Address:
RNDr. Zdenk Svoboda, CSc.
Technick 8,
616 00 Brno,
svobodaz@feec.vutbr.cz

173

APPLICATION OF NON SIMPLEX METHOD FOR LINEAR PROGRAMMING
Marie Tomov
University of Technology Brno

Abstract: This paper showes a method for the solution of optimizing problems which is different from the
usually used Simplex method. The simplex method is considered one of the basic models fom which many
linear programming techniques are directly and indirectly derived. The simplex method is an iterative
process which approaches, step by step, an optimum solution in such a way that an objective function of
maximization or minimization is fully reached. Each iteration in this process consists of shortening the
distance (mathematically and also graphically) from the objective function to the intercepted vertex of a
convex set determined by the inequalitis which describe the problem. The simplex method is not the only
technique known and used for solving linear programming problems. For the pedagogical expendiency are
more useful also other methods, see for example R. Dorfman, P.A.Samuelson, and R.M.Solov, Linear
Programming and Economic Analysis, New. York: McGraw-Hill Book Comp. Inc,1958. I interduce an
other method from the simplex method. This method will be based on the princip of the graphical method
of optimization of linear problems for two variables, but my method will be generalized for n variables and
an arbitrary finite number of inequalities descibing the problem.

Key words. Liner programming, system of inequalities, disposal and slack variable, dummy variable,
objective function, iteration, key row, key column, key element, polyhedron, hyper-plane.

1. The leading article
Remark I reported this problem on the fourth Mathematical Workshop in Brno 2005. This contribution
aims at spreading knowledge of the described method - I prepared a programme for solving of concrete problems.
This programme is given as an appendix.

Introduction The generall problem of linear programming is usually formulated as follows:
Let a
ij
, b
i
, c
j
(i = 1,2, ... , m ; j = 1,2, ... ,n) be given real numbers and let us denote I1 C I = {i = 1,2, ... , m} and J
1
C
J = j = 1, 2, ... ,n}. The problem of maximizing of the function
1
n
j j
i
c x
(1)
on the set of
1
n
ij j j
j
a x b

( )
1
i I
(2)
1
n
ij j j
j
a x b

( )
1
i I I
(3)
0
j
x

( )
1
j J
(4)
is called maximizing problem of linear programming in mixed form if
1 1 1
0, , . I I I or J J

The problem of linear programming given by (1) till (4) where I
1
= I and J
1
= J that is the problem of maximizing of the
function
1
n
j j
i
c x
(5)

174

on the set of linear independent system of linear unequations
1
n
ij j j
j
a x b

( )
1
i I
(6)
0
j
x

( )
1
j J
(7)
is called maximizing problem of linear programming in the form of unequations.

With respect to the fact that for arbitrary set
n
R M where R
n
is n dimensional vectorial space and for arbitrary linear
function
n
R M z :
min ( ) max( ( )),
n
z x z x where x R

holds then if one of extrems exists we can transform also minimizing problem on the problem with linear equations or
or linear unequations. We do the rearrangement by multiplication by number -1.

2. The solution of the general problem
We desist from the condition (4) and hence also from (7) in the following considerations. We rewrite the system (6) and
add the objective function as the last row into the form:

11 1 12 2 1 1
... 0
n n
a x a x a x b + + + +

21 1 22 2 2 2
... 0
n n
a x a x a x b + + + +

...
(8)

1 1 2 2
... 0
m m mn n m
a x a x a x b + + + +

1 1 2 2
... 0
n n
c x c x c x + + +

We call the set x = {x
1
,x
2
, ... ,x
n
} C R
n
of elements as polyhedron. We call the polyhedron opened if n m . We know
that the objective function ( )
n
R x x z , receives its optimal values at vertixes or at all areas of polyhedron. In the first
part of computational procedure there we find one from vertixes of polyhedron and we transform it into the coordinate
origin simultaneously. Simultaneously we transform all hyperplanes of polyhedron and the objective function with
respect to the given transformation. We continue as folows:

We select arbitrary hyperplane and we denote it by the index { } m i ,..., 2 , 1 . We divide the whole column at the variable
x
1
by coefficient a
i1
and we put simultaneously
1 1 2 2 3 3
...
i i in n i
x x a x a x a x b
(9)

We introduce the transformation relation (9) into the system (8) and the mathematical representation of the problem
transforms for m > 1 on to the following form:
11 2 11 3 11 11 11
1 12 2 13 3 1 1
1 1 1 1 1
... 0
i i in
n n i
i i i i i
a a a a a a a a
x a x a x a x b b
a a a a a
_ _ _
+ + + + +

, , ,

21 2 21 3 21 21 21
1 22 2 23 3 2 2
1 1 1 1 1
... 0
i i in
n n i
i i i i i
a a a a a a a a
x a x a x a x b b
a a a a a
_ _ _
+ + + + +

, , ,

1,1 1,1 2 1,1 3 1,1 1,1
1 1,2 2 1,3 2 1, 1
1 1 1 1 1
... 0
i i i i i i in i
i i i n n i i
i i i i i
a a a a a a a a
x a x a x a x b b
a a a a a

_ _ _
+ + + + +

, , ,

1
0 0 ... 0 0 0 x + + + + +
1,1 1,1 2 1,1 3 1,1 1,1
1 1,2 2 1,3 2 1, 1
1 1 1 1 1
... 0
i i i i i i in i
i i i n n i i
i i i i i
a a a a a a a a
x a x a x a x b b
a a a a a
+ + + + +
+ + + +
_ _ _
+ + + + +

, , ,

175

21 2 21 3 21 21 21
1 22 2 23 3 2 2
1 1 1 1 1
... 0
i i in
n n i
i i i i i
a a a a a a a a
x a x a x a x b b
a a a a a
_ _ _
+ + + + +

, , ,

In the following step there we choose some of rows where the coefficient at the variable x2 is different from zero
arbitrary. The existence of such road follows from the assumption that m linear rows is independent. In the next we
suppose hat this assumption is satisfied by the row m s s , . It is obvious that i s . We continue such that we divide
the whole second column by the expression
2
2
1
si i
s
i
a a
a
a

and then we introduce the following transformation:

1
1 ,
1
1
1 ,
, 3
1
3 1
3
2
1
1
2
...
i
s
s n
i
in s
n s
i
i s
s
i
s
a
a
b b x
a
a a
a x
a
a a
a x
a
a
x +
,
_
,
_
+

After this transformation the s-th row will be of the form:

We continue till we do the all m < n transformations by analogy. Thus we calculate one point of one edge of polyhedron
which transformed into the coordinate origin.

In the account that n m there we find after n transformations one vertix of polyhedron which transformed into the
coordinate origin. The whole calculus is done on computer therefore we calculate only with the matrix of coefficients of
polyhedron. We apply all the steps of transformation to the objective function
0 0 ...
2 2 1 1
+ + + +
n n
x c x c x c

and we receive after the first transformation
0 0 ...
1
1
1
1
3
1
3 1
3 2
1
2 1
2
1
1
1
+
,
_
+ +
,
_
,
_
+
i
i n
i
in
n
i
i
i
i
i
a
c
b x
a
a c
c x
a
a c
c x
a
a c
c x
a
c

Further adaptations of coefficients of the objective function run over simultaneously with the adaptations of coeficients
of polyhedron such as it was given in the previous description of the hash algorithm applied onto polyhedron.

As the next step we extend the matrix of coefficients descibing the system (8) with the objective function which is of
the type (m +1) x (n +1) such that we add the matrix of the type n x (n +1) which consists of unit matrix type n x n with
added column vector of zeros of the length n.This step is necessary for explicit expression of the poin of edge optionally
vertex of polyhedron and the optimal value of the objective function. Onto such expanded matrix are applied all before
described affinite transformations. We obtain after the making described transformtion algorithm the original
coordinates of the point of edge or the vertex of polyhedron which is transformed into the coordinate origin in the last
column of the matrix n x (n + 1). We show the expanded matrix of coefficients of the type (m + 1 + n) x (n + 1) before
the transformation algorithm.
2
0 0 0 ... 0 0 0 x + + + + +
1 1 1 1 1
2 2 2 2 1
1 2 3
1 2 3 ...
1 2 3 ...
. . . ... . .
1 2 3 ...
... 0
1 0 0 ... 0 0
0 1 0 ... 0 0
. . . ... . .
0 0 0 ... 1 0
n
n
m m m m m
n
a a a a b
a a a a b
a a a a n b
c c c c
_

,

176

3. Optimizing and decision making process small We suppose that m>n and that the transformation
algorithm transformed one of vertices of polyhedron was transformed into the coordinate origin. If after this
transformation all the coefficients in the first m rows of the last (n + 1)-the column are nonnegative numbers and
simultaneously all transformed coefficiens of the objective function it is c
,
k
in the (m + 1-the row negativw the
maximizing process of the objective function z(x) is finished. In the last n rows of the (n + 1)-the column there are
original coordinates of the vertex of polyhedron in which the objective function acquires its maximum and the value of
this maximum is at the position [m + 1, n + 1] of transformed matrix.
If previous situation does not occure then it is necessary to do the following analysis. We suppose for coefficients in the
first m rows of the (n + 1)th column nonnegative again but some of transformed coefficiens of the objective function in
the (m + 1)-th row is positive. Let this situation in the j0-th column occure. We look at all transformed coeficients in the
j
0
-th column. If all transformed coefficients of the polyhedron
0 ij
a . are nonnegative then the problem has not any
solution. It is possible to get along this edge incident to polyhedron to infinity. Polyhedron is not bounded ant the
solution does not exist.

4. Example
4.1 Remark I drafted a program for explanation of the given method which adress is on the server Pal of
the Technical University in Brno: Q: \vyuka\ matematm\Tomsova\Polyhedron\ matice. exe. We do an applivcation of
this program. Data are denoted as CONCRETE

4.2 Example Maximize the objective function z = 3x+5y in the area bounded by the following restrictions:
1. 0 x
2. 0 y
3. 5 x
4. 12 2 + y x
5. 19 3 2 + y x

Solution: We line a figure for the better graphical preview where the area of polyhedron will be bounded with bisectors
suitabled to the restrictions with the vertixes 0 = [0, 0], A = [5, 0], B = [5, 3], C = [2, 5], D = [0, 6]

177

After startup of program we browse the vertixes in the sentence 0, A, B, C, D. We can monitor the coefficients and
values of the objective function in the green coloured row. The coordinates of the vertices are in the last column in the
last two rows. The program finishes at that moment when all coefficients of the objective function are nonpositive and
the result is: The problem has just one solution at the point C and the value of the objective function is 31.

References
[1] CHURCHMAN, CH. W.; ACKOFF, R. L.; ARNOFF, L. Introduction to Operation Research. New York :
John Wiley & Sons. Inc. 1957.
[2] KLAPKA, J.; DVOK, J.; POPELA, P. Metody operanho vzkumu. Brno : VUTIUM, 2001
[3] RAIS, K. Vybran kapitoly z operan analzy. Brno : PGS, 1985.
[4] VACULK, J.; ZAPLETAL, J. Podprn metody rozhodovacch procesu. Brno : Masarykova univerzita,
1998.
[5] WALTER J. a kol. Operan vzkum. Praha : SNTL, 1973.
[6] ZAPLETAL, J. Operan analza. Kunovice : Skriptorium VOS, 1995.

Address:
Mgr. Marie Tomov
Technick 8,
616 00 Brno,
e-mail: tomsova@feec.vutbr.cz

178

179

AUTHOR INDEX

A
ABDURRZZAG, T....................27
AMALKA AL K. ......................81

B
BATINEC, J. ................... 35, 143

D
DIBLK J. ................. 35, 143, 169
DOSTL, P. .............................21

F
FAJMON B. .............................51

K
KLIETIK T. ..........................105
KOSTIHA J. .............................65
KRUPKOV V. .....................157

L
LACKO, B. ..............................73
LAK, V. .............................89

M
MSZROS, I. .........................57
MIKULA, V. ............................45
MINAK, M. ..........................99

N
NOVK, M. .......................43, 51

O
OMERA, P. ............ 89, 109, 123

P
PETRUCHA, J. ................. 45, 137
POPELKA, O. ..........................89

R
RUKOVANSK, I. .......... 89, 131

S,
SVOBODA, Z. .......................169
MARDA, Z. .................. 157, 163
ASTN, J. ............................99

T
TOM, I..................................57
TOMOV, M. ......................173

V
VRTESY, G. ..........................57

Z
ZAPLETAL, J. .......................... 9

180

SPONSORS

Sverepec 365
Povask Bystrica
PS: 01701
tel.: 00421-42-4321110
fax: 00421-42-4379930
majitelia:
Milan Richtrik
Jozef urajka

VS-mont, s. r. o.,
Lazy pod Makytou, SLOVENSKO
email: vsmont@vsmont.sk
tel./fax: 00421 42 4681 965
tel.: 00421 42 4681 952

Udia 366,
018 01 Povask Bystrica
telefn: 042/4260768042/4260769
fax:042/4340269
e-mail:objednavky@febro.sk - objednvky tovaru
ekonomika@febro.sk - ekonomick oddelenie
sekretariat@febro.sk - sekretarit

Nzev:
ICSC 2006 Fourth International Conference on Soft Computing Applied in
Computer and Economic Environment

Autor:
Kolektiv autor

Vydavatel, nositel autorskch prv, vyrobil:
Osvobozen 699, 686 04 Kunovice
Nklad: 200 ks
Poet stran: 182
Vydn: prvn
Rok vydn: nor 2006

ISBN 80-7314-084-5

9 7 8 8 0 7 3 1 4 0 8 4 7
I SBN 80 - 7314 - 084 - 5

Pavel Ošmera - Vortex-Fractal Physics: Page 123-129

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Pavel Ošmera - Vortex-Fractal Physics: Page 123-129

Uploaded by

Copyright:

Available Formats

EUROPEAN POLYTECHNICAL INSTITUTE KUNOVICE

+ < < X P , which

is the learning rate

is the derivative update

k u k, for every ( ) a N k , i.e. satisfies inequalities (1.10) for every

You might also like