Implementation and Evaluation of
Two Prediction Techniques
for the Lorenz Time Series
G r a n t E H u d d l e s t o n e
A s s i g n m e n t p r e s e n t e d i n p a r t i a l f u l f i l m e n t o f t h e r e q u i r e m e n t s f o r t h e
d e g r e e M a s t e r o f S c i e n c e a t t h e D e p a r t m e n t o f A p p l i e d M a t h e m a t i c s
o f t h e U n i v e r s i t y o f S t e l l e n b o s c h , S o u t h A f r i c a
Declaration:
I, the undersigned,
hereby declare that the work contained in this assignment
is my own original work,
and that I have not previously in its entirety, or in part, submitted
it at any university
for a degree.
Signatm~
.
A b s t r a c t
T h i s t h e s i s i m p l e m e n t s a n d e v a l u a t e s t w o p r e d i c t i o n t e c h n i q u e s u s e d t o f o r e c a s t d e t e r m i n i s t i c c h a o t i c t i m e s e r i e s . F o r a l a r g e n u m b e r o f s u c h t e c h n i q u e s , t h e r e c o n s t r u c t i o n o f t h e p h a s e s p a c e a t t r a c t o r a s s o c i a t e d w i t h t h e t i m e s e r i e s i s r e q u i r e d . E m b e d d i n g i s p r e s e n t e d a s t h e m e a n s o f r e c o n s t r u c t i n g t h e a t t r a c t o r f r o m l i m i t e d d a t a . M e t h o d s f o r o b t a i n i n g t h e m i n i m a l e m b e d d i n g d i m e n s i o n a n d o p t i m a l t i m e d e l a y f r o m t h e f a l s e n e i g h b o u r h e u r i s t i c a n d a v e r a g e m u t u a l i n f o r m a t i o n m e t h o d a r e d i s c u s s e d . T h e f i r s t p r e d i c t i o n a l g o r i t h m t h a t i s d i s c u s s e d i s b a s e d o n w o r k b y S a u e r , w h i c h i n c l u d e s t h e i m p l e -m e n t a t i o n o f t h e s i n g u l a r v a l u e d e c o m p o s i t i o n o n d a t a o b t a i n e d f r o m t h e e m b e d d i n g o f t h e t i m e s e r i e s b e i n g p r e d i c t e d . T h e s e c o n d p r e d i c t i o n a l g o r i t h m i s b a s e d o n n e u r a l n e t w o r k s . A s p e c i f i c a r c h i t e c t u r e , s u i t e d t o t h e p r e d i c t i o n o f d e t e r m i n i s t i c c h a o t i c t i m e s e r i e s , n a m e l y t h e t i m e d e p e n d e n t n e u r a l n e t w o r k a r c h i t e c t u r e i s d i s c u s s e d a n d i m p l e m e n t e d . A d a p t a t i o n s t o t h e b a c k p r o p a g a t i o n t r a i n i n g a l g o r i t h m f o r u s e w i t h t h e t i m e d e p e n d e n t n e u r a l n e t w o r k s a r e a l s o p r e s e n t e d . B o t h a l g o r i t h m s a r e e v a l u a t e d b y m e a n s o f p r e d i c t i o n s m a d e f o r t h e w e l l - k n o w n L o r e n z t i m e s e r i e s . D i f f e r e n t e m b e d d i n g a n d a l g o r i t h m - s p e c i f i c p a r a m e t e r s a r e u s e d t o o b t a i n p r e d i c t e d t i m e s e r i e s . A c t u a l v a l u e s c o r r e s p o n d i n g t o t h e p r e d i c t i o n s a r e o b t a i n e d f r o m L o r e n z t i m e s e r i e s , w h i c h a i d i n e v a l u a t i n g t h e p r e d i c t i o n a c c u r a c i e s . T h e p r e d i c t e d t i m e s e r i e s a r e e v a l u a t e d i n t e r m s o f t w o c r i t e r i a , p r e d i c t i o n a c c u r a c y a n d q u a l i t a t i v e b e h a v i o u r a l a c c u r a c y . B e h a v i o u r a l a c c u r a c y r e f e r s t o t h e a b i l i t y o f t h e a l g o r i t h m t o s i m u l a t e q u a l i t a t i v e f e a t u r e s o f t h e t i m e s e r i e s b e i n g p r e d i c t e d . I t i s s h o w n t h a t f o r b o t h a l g o r i t h m s t h e c h o i c e o f t h e e m b e d d i n g d i m e n s i o n g r e a t e r t h a n t h e m i n i m u m e m b e d d i n g d i m e n s i o n , o b t a i n e d f r o m t h e f a l s e n e i g h b o u r h e u r i s t i c , p r o d u c e s g r e a t e r p r e d i c t i o n a c c u r a c y . F o r t h e n e u r a l n e t w o r k a l g o r i t h m , v a l u e s o f t h e e m b e d d i n g d i m e n s i o n g r e a t e r t h a n t h e m i n i m u m e m -b e d d i n g d i m e n s i o n s a t i s f y t h e b e h a v i o u r a l c r i t e r i o n a d e q u a t e l y , a s e x p e c t e d . S a u e r ' s a l g o r i t h m h a s t h e g r e a t e s t b e h a v i o u r a l a c c u r a c y f o r e m b e d d i n g d i m e n s i o n s s m a l l e r t h a n t h e m i n i m a l e m b e d d i n g d i m e n s i o n . I n t e r m s o f t h e t i m e d e l a y , i t i s s h o w n t h a t b o t h a l g o r i t h m s h a v e t h e g r e a t e s t p r e d i c t i o n a c c u r a c y f o r v a l u e s o f t h e t i m e d e l a y i n a s m a l l i n t e r v a l a r o u n d t h e o p t i m a l t i m e d e l a y . T h e n e u r a l n e t w o r k a l g o r i t h m i s s h o w n t o h a v e t h e g r e a t e s t b e h a v i o u r a l a c c u r a c y f o r t i m e d e l a y c l o s e t o t h e o p t i m a l t i m e d e l a y a n d S a u e r ' s a l g o r i t h m h a s t h e b e s t b e h a v i o u r a l a c c u r a c y f o r s m a l l v a l u e s o f t h e t i m e d e l a y . M a t l a b c o d e i s p r e s e n t e d f o r b o t h a l g o r i t h m s .Opsomming
In hierdie
tesis word twee voorspellings-tegnieke
geskik vir voorspelling
van determ inistiese
chaotiese
tydreekse
ge"im plem enteer en geevalueer.
Vir sulke tegnieke word die rekonstruksie
van die aantrekker
in
fase-ruim te
geassosieer m et die tydreeks
gewoonlik vereis.
Inbedm etodes
word aangebied
as 'n m anier om die aantrekker
te rekonstrueer
uit beperkte
data.
M etodes
om die m inim um
inbed-dim ensie
te bereken uit gem iddelde
wedersydse
inligting
sowel as die optim ale
tydsvertraging
te bereken uit vals-buurpunt-heuristiek,
word bespreek.
Die eerste voorspellingsalgoritm e
wat bespreek
word is gebaseer op 'n tegniek
van Sauer.
Hierdie
algo-ritm e m aak gebruik van die im plem entering
van singulierwaarde-ontbinding
van die ingebedde
tydreeks
wat voorspel word.
Die tweede
voorspellingsalgoritm e
is gebaseer
op neurale
netwerke.
'n Spesifieke netwerkargitektuur
geskik vir determ inistiese
chaotiese
tydreekse,
naam lik
die tydafhanklike
neurale
netwerk
argitektuur
word bespreek
en ge"im plem enteer.
'n M odifikasie van die terugprapagerende
leer-algoritm e
vir gebruik
m et die tydafhanklike
neurale
netwerk word ook aangebied.
Albei algoritm es
word geevalueer deur voorspellings
te m aak vir die bekende
Lorenz tydreeks.
Verskeie
inbed param eters
en ander algoritm e-spesifieke
param eters
word gebruik
om die voorspelling
te m aak.
Die werklike waardes
vanuit die Lorentz tydreeks
word gebruik om die voorspellings
te evalueer en om
voorspellingsakkuraatheid
te bepaal.
Die voorspelde tydreekse
word geevalueer op grand van twee kriteria,
naam lik voorspellingsakkuraatheid,
en kwalitatiewe
gedragsakkuraatheid.
Gedragsakkuraatheid
verwys na die verm oe van die algoritm e
om
die kwalitatiewe
eienskappe
van die tydreeks
korrek te sim uleer.
Daar word aangetoon
dat vir beide algoritm es die keuse van inbed-dim ensie
grater as die m inim um
inbed-dim ensie soos bereken uit die vals-buurpunt-heuristiek,
grater akkuraatheid
gee. Vir die neurale
netwerk-algoritm e
gee 'n inbed-dim ensie
grater
as die m inim um
inbed-dim ensie
ook betel' gedragsakkuraatheid
so os verwag. Vir Sauer se algoritm e,
egter, word betel' gedragsakkuraatheid
gevind vir 'n inbed-dim ensie
kleiner as die m inim ale
inbed-dim ensie.
In term e van tydsvertraging
word dit aangetoon
dat vir beide algoritm es
die grootste
voorspellingsakku-raatheid
verkry word by tydvertragings
in 'n interval rondom
die optim ale
tydsvetraging.
Daar
word ook aangetoon
dat
die neurale
netwerk-algoritm e
die beste
gedragsakkuraatheid
gee vir
tydsvertragings
naby
aan
die optim ale
tydsvertraging,
terwyl
Sauer
se algoritm e
betel'
gedragsakku-raatheid
gee by kleineI' waardes van die tydsvertraging.
Contents
1
Introduction
2.2 Tim e D elay
2
Concepts in embedding theory
3.3 S election of the N earest N eighbours
3
Prediction with the Sauer algorithm
3.4 C enter of M ass of the N earest N eighbours
1.1
1.2
1.3
1.4
1.5
2.1
2.33.1
3.23.5
3.63.7
B ackground P roblem D escription .Lorenz system , general inform ation and equations
N otation conventions
C hapter P review . . .
P hase space reconstruction
E m bedding D im ension.
P rem ise of S auer's A lgorithm ..
R econstruction of the A ttractor .
S ingular V alue D ecom position and P rojections
R egression of the Linear M odel
C alculation of the Target P oint
1
1
22
5
5
7
8
9
1115
16
16
17
18
18
20
20
4
Prediction with Time Delayed Neural Networks
4.1 A n Introduction to N eural N etw orks ...
23
II C O N T E N T S
4 .2 S tru c tu re s in N e u ra l N e tw o rk A rc h ite c tu re . . . .. 2 5
4 .3 T im e D e p e n d e n t N e u ra l N e tw o rk s . . . .. 2 9
4 .4 T ra in in g A lg o rith m fo r T im e -D e la y N e u ra l N e tw o rk s . . . .. 3 2
5
Evaluation
of the algorithms
37
5 .1 E v a lu a tio n o f th e T D N N a lg o rith m . . . .. 3 8
5 .1 .1 P e rfo rm a n c e C rite ria 3 9
5 .1 .2 E x p e rim e n ta l re s u lts o b ta in e d fo r th e T D N N a lg o rith m . . . .. 4 3
5 .1 .3 A n a ly s is o f re s u lts fo r th e T D N N a lg o rith m . . . .. 4 4
5 .2 E v a lu a tio n o f th e S a u e r a lg o rith m . . . .. 4 6
5 .2 .1 P e rfo rm a n c e C rite ria
. . . ..
47
5 .2 .2 E x p e rim e n ta l re s u lts o b ta in e d fo r th e S a u e r a lg o rith m . . . .. 4 7
5 .2 .3 A n a ly s is o f re s u lts fo r th e S a u e r a lg o rith m . . . .. 4 8
6
Conclusion
51
A
Time series scaling functions
53
B
Sauer algorithm
Matlab code
55
B .1 M a in p ro g ra m . . . .. 5 5 B .2 E m b e d d in g th e tim e s e rie s . . . .. 5 6 B .3 F in d in g n e a re s t n e ig h b o u rs . . . .. 5 6 B .4 C a lc u la tin g th e c e n tre o f m a s s o f th e n e a re s t n e ig h b o u rs . . . .. 5 6 . . . .. 5 7 B .5 S in g u la r v a lu e d e c o m p o s itio n a n d th e s p a n n in g m a trix B .6 P ro je c tin g th e d is p la c e m e n t m a trix B .7 P e rfo rm in g th e lin e a r re g re s s io n . B .8 E s tim a tin g th e ta rg e t p o in t ... . . .. 5 7 . . .. 5 7 5 7
C
TDNN
algorithm
Matlab code
59
C .1 T D N N tra in in g c o d e . . .. 5 9
CJ
n
0
'<
w
~
~
'<
I'D~
riG"
::r-...•
::J ;:;: iii"N"
OJ...•
0"
::J-.
C ::J n...•
0"
::JList of Figures
1.1 P hase space plot of the L orenz attractor obtained from the solution of the three L orenz equations. 3
1.2 P lot of the solution of 1J
(t)
for the L orenz system s of equations. . . .. 42.2 A m ap that is not one-to-one.
2.3 A n em bedding F of the sm ooth m ap A .
2.1
2.4
2.5
2.6
A one-to-one m ap that is not continuously differentiable. i.e. not an im m ersion.
D ifferent choices of tim e delays on a L orenz tim e series ..
A verage m utual inform ation versus tim e delay for the L orenz tim e series ..
A n illustration of false and true nearest neighbours. . .
9 9 9
10
1 112
3.1 Illustration of points on the attractor relevant to the S auer algorithm . . . .. 19
4.4 T he effect of the bias on activation function output.
4 .1
4.2
4.3
N eural netw ork architecture.
T he structure of a single neuron.
P opular activation functions. . .
24
25
27
28
4.5
Left:
F eedforw ard connections for a T D N N .Right:
F eedback connections for a T D N N . . .. 304.6 F eedback and F eedforw ard connection detail.
4.7
Left:
F eedforw ard connections for R N N .Right:
F eedback connections for R N N ..4.8
D ecision tree for adapted back-propagation neural netw ork training. . . .31
31
33
5.1 P lot illustrating a predicted tim e series versus the associated real tim e series obtained from a
prediction by the
TDNN
algorithm . . . .. 385.2 C haracteristic points for the L orenz tim e series. . . .. 40
V I
5 .3 C h a ra c te ris tic p o in ts o n th e L o re n z a ttra c to r. ..
LIST
O FFIGURES
41
5 .1 0 P lo t o f a q u a lita tiv e ly c o rre c t s im u la tio n b y th e S a u e r a lg o rith m . 5 .4 P lo t o f th e C -s p a c e p o in ts fo r re a l tim e s e rie s fro m tw o v ie w p o in ts .
5 .1 3 P lo t o f a p re d ic tio n b y th e S a u e r a lg o rith m th a t d is p la y s d iv e rg e n t b e h a v io u r. . 5 .1 1 P lo t o f a p re d ic tio n b y th e S a u e r a lg o rith m th a t is tra p p e d o n o n e w in g .
5 .1 2 P lo t o f a p re d ic tio n b y th e S a u e r a lg o rith m th a t d is p la y s o s c illa to ry b e h a v io u r ..
41
4 243
4 3 4 3 4 348
4 848
48
P lo t o f a p re d ic tio n th a t firs t c a p tu re s th e d y n a m ic s b u t is tra p p e d .
P lo t o f a p re d ic tio n th a t d is p la y s o s c illa to ry b e h a v io u r. .. . .. P lo t o f a p re d ic tio n th a t d e c a y s to w a rd s a fa ls e fix e d p o in t. P lo t o f a q u a lita tiv e ly c o rre c t s im u la tio n .
P lo t o f th e firs t d e riv a tiv e o f th e L o re n z tim e s e rie s v e rs u s th e L o re n z tim e s e rie s . 5 .5
5.6
5.7
5 .8
List of Tables
2 .1 P e rc e n ta g e o f fa ls e n e a re s t n e ig h b o u rs a t d iffe re n t e m b e d d in g d im e n s io n s fo r th e L o re n z tim e s e rie s . 1 35.1
5 .2 T e s tin g p a ra m e te rs fo r T D N N a lg o rith m .A v e ra g e M S E v a lu e s fo r T D N N p re d ic tio n s u s in g a c c u ra c y c rite rio n .
38
44
5 .3 A v e ra g e e u c lid e a n d is ta n c e b y T D N N p re d ic tio n fo r th e b e h a v io u r c rite rio n u s in g th e fx -s p a c e
d is c rim in a tio n . . . .. 4 4
5 .4 A v e ra g e e u c lid e a n d is ta n c e b y T D N N p re d ic tio n fo r th e b e h a v io u r c rite rio n u s in g th e (-s p a c e
d is c rim in a tio n . . . .. 4 5
5 .5 A v e ra g e M S E v a lu e s fo r T D N N p re d ic tio n s u s in g a c c u ra c y c rite rio n w ith h id d e n c o n te x t la y e rs . 4 5
5 .6 A v e ra g e e u c lid e a n d is ta n c e b y T D N N p re d ic tio n w ith h id d e n c o n te x t la y e rs fo r th e b e h a v io u r
c rite rio n u s in g th e (-s p a c e d is c rim in a tio n . . . .. 4 5
5 .7 A v e ra g e e u c lid e a n d is ta n c e b y T D N N p re d ic tio n w ith h id d e n c o n te x t la y e rs fo r th e b e h a v io u r
c rite rio n u s in g th e fx -s p a c e d is c rim in a tio n . . . .. 4 5
5 .9 A v e ra g e M S E v a lu e s fo r S a u e r a lg o rith m p re d ic tio n s u s in g th e b e h a v io u r c rite rio n .
5 .1 0 A v e ra g e M S E v a lu e s fo r S a u e r a lg o rith m p re d ic tio n s u s in g th e a c c u ra c y c rite rio n .
5.8
T e s tin g p a ra m e te rs fo r S a u e r a lg o rith m . .47
49
. . . 4 9 5 .1 1 A v e ra g e M S E v a lu e s fo r d im e n s io n v e rs u s d e la y . . . .. 4 9 5 .1 2 A v e ra g e M S E v a lu e s fo r d im e n s io n v e rs u s s p a n n in g d im e n s io n . m a x im u m v a lu e s fo r a ro w .) . VB (V a lu e s in ita lic s in d ic a te49
List of Symbols
T : T h e L o r e n z t i m e s e r i e s u s e d f o r a l g o r i t h m e v a l u a t i o n . N : T h e l e n g t h o f t h e t i m e s e r i e s T . X t : A d i s c r e t e v a r i a b l e d e s c r i b i n g t h e s t a t e o f t h e s y s t e m a t t i m et.
m:
T h e e m b e d d i n g d i m e n s i o n . l : T h e t i m e d e l a y . s : T h e s p a n n i n g d i m e n s i o n .b
t: T h e p r e d i c t i o n p o i n t o n t h e r e c o n s t r u c t e d L o r e n z a t t r a c t o r . l l i : T h e i - t h n e a r e s t n e i g h b o u r t o t h e p r e d i c t i o n p o i n t o n t h e L o r e n z a t t r a c t o r . ei: T h e t a r g e t p o i n t f o r t h e i - t h n e a r e s t n e i g h b o u r . c : T h e c e n t r e o f m a s s o f t h e n e a r e s t n e i g h b o u r s .d(t):
A t i m e d e l a y e d c o o r d i n a t e s e t a t t i m et.
k : R e g r e s s i o n c o e f f i c i e n t s . h ~ : T h e s u m m i n g s u b - u n i t v a l u e f o r n o d e i i n l a y e rn.
l ~ : T h e v a l u e f o r n o d e i i n l a y e rn.
wD:
T h e w e i g h t c o n n e c t i n g t h e j - t h n o d e i n l a y e rn -
1 w i t h t h e i - t h n o d e i n l a y e rn.
'! '( p ) : I n p u t p a t t e r n p f o r a n e u r a l n e t w o r k . Y ( p ) : O u t p u t p a t t e r n p f o r a n e u r a l n e t w o r k .!\.:
T h e n u m b e r o f l a y e r s i n a n e u r a l n e t w o r k . H : T h e n u m b e r o f c o n t e x t l a y e r s i n a n e u r a l n e t w o r k .J :
T h e n u m b e r o f s e t s o f i n p u t p a t t e r n s , '! '( p ) . a n d a s s o c i a t e d o u t p u t p a t t e r n s , Y ( p ) , f o r a n e u r a l n e t w o r k . MT l : T h e n u m b e r o f n o d e s i n l a y e rn.
I XChapter 1
Introduction
1.1
Background
W h e n a d isc u ssio n tu rn s to p re d ic tio n , a p e rso n w ill in v a ria b ly a sk w h e th e r p re d ic tin g th e fin a n c ia l m a rk e ts
is p o ssib le . T h is q u e stio n h ig h lig h ts th e c o m m o n d e sire to d iv in e th e fu tu re a n d h a s b e e n w ith h u m a n k in d
fo r m ille n n ia , a t first sim p ly w o n d e rin g w h a t th e w e a th e r m ig h t h o ld fo r c ro p s to th e m o d e rn d a y d e sire to
fo re c a st th e m a rk e t-d riv e n e c o n o m ie s o f th e w o rld .
In e x p la in in g th e p ro c e ss o f p re d ic tio n to a p e rso n , th e d istin c tio n b e tw e e n d iffe re n t fo rm s o f c h a o tic b e h a v io u r
h a s to b e h ig h lig h te d . P u re ly sto c h a stic b e h a v io u r is w h a t p e o p le im m e d ia te ly th in k a b o u t w h e n th e y h e a r
c h a o s. T h e te rm c h a o s is h o w e v e r a lso c o n sid e re d fo r sy ste m s th a t illu stra te irre g u la r b e h a v io u r b u t a t th e
sa m e tim e a re n o t c o m p le te ly ra n d o m .
T h is fo rm o f c h a o s is d e te rm in istic [1 ] [1 2 ] a n d h a s tw o k e y fe a tu re s; th e sy ste m is se n sitiv e to in itia l c o n d itio n s
a n d th e sy ste m , o v e r tim e , se ttle s d o w n to a fix e d re g io n o f th e sp a c e d e fin e d b y its v a ria b le s.
S e n sitiv ity to in itia l c o n d itio n s m e a n s th a t a v e ry sm a ll c h a n g e in th e in itia l, o r sta rtin g , c o n d itio n s o f a
sy ste m c a n le a d to v e ry d iffe re n t b e h a v io u r in a sh o rt p e rio d o f tim e . F o r th e p u rp o se s o f th is th e sis, o n ly a
d e te rm in istic c h a o tic sy ste m th a t p o sse sse s a n a ttra c to r, b e in g a lo c a liz e d a re a in th e v a ria b le sp a c e fo r th e
sy ste m to w h ic h th e so lu tio n w ill c o n v e rg e fro m in itia l c o n d itio n s, w ill b e c o n sid e re d .
A c c u ra te ly p re d ic tin g th e fu tu re fo r a sto c h a stic sy ste m m ig h t to a ll in te n ts a n d p u rp o se s b e im p o ssib le . T h is
h o w e v e r is n o t th e c a se fo r a sy ste m e x h ib itin g d e te rm in istic c h a o tic b e h a v io u r w h e re m e th o d s m a y b e so u g h t
to m a k e p re d ic tio n s [1 1 ] [2 7 ] a b o u t th e fu tu re b e h a v io u r o f th e sy ste m .
A s w ith a ll a n a ly tic a l m e th o d s th e d a ta u n d e r in v e stig a tio n , in th is se n se th e tim e se rie s c o n ta in in g o b se rv a tio n s
o f th e c h a o tic sy ste m , h a s to b e p ro c e sse d in to a fo rm th a t is c o n v e n ie n t fo r a n a ly sis.
A la rg e n u m b e r o f p re d ic tio n a lg o rith m s [3 ] [4 ] [8 ] [1 8 ] [2 1 ] re ly o n th e re c o n stru c tio n o f th e a ttra c to r fro m
th e o b se rv e d m e a su re m e n ts. D u e to th e p o te n tia lly la rg e n u m b e r o f v a ria b le s a t p la y in a c h a o tic sy ste m ,
2
CHAPTER
1. INTRODUCTION
r e c o n s tr u c tio n o f th e a ttr a c to r is u s u a lly n o t p o s s ib le a n d a m e th o d is s o u g h t to a p p r o x im a te th e a ttr a c to r .
S a u e r [ 2 0 ] ju s tif ie d th e u s e o f e m b e d d in g a tim e s e r ie s in r e c o n s tr u c tin g a n a ttr a c to r . S u c c e s s f u l e m b e d d in g
r e q u ir e s th e c h o ic e o f a m in im a l e m b e d d in g d im e n s io n [ 1 4 ] a n d a n o p tim a l tim e d e la y [ 9 ] . M e th o d s e x is t to
d e te r m in e th e s e o p tim a l p a r a m e te r s , a n d th e s e m e th o d s w ill b e d is c u s s e d in c h a p te r 2 .
1.2
Problem Description
T h is th e s is im p le m e n ts a n d e v a lu a te s th e p e r f o r m a n c e o f tw o p r e d ic tio n a lg o r ith m s w h e n a p p lie d to th e L o r e n z
tim e s e r ie s ( s e c tio n 1 .3 ) .
T h e f ir s t a lg o r ith m is b a s e d o n w o r k b y S a u e r [ 1 8 ] a n d u tiliz e s lo c a l lin e a r m o d e ls , in w h ic h th e s in g u la r v a lu e
d e c o m p o s itio n p la y s a p iv o ta l r o le , to p e r f o r m th e p r e d ic tio n s .
T h e s e c o n d a lg o r ith m e m p lo y s a n e u r a l n e tw o r k a p p r o a c h , w h ic h , b e c a u s e o f th e n o n - lin e a r n a tu r e o f th e
m e th o d h a s a p p e a r e d s u ita b le f o r th e a n a ly s is a n d p r e d ic tio n o f c h a o tic d a ta . Q u o tin g S w in g le r [ 2 5 ] :"
applications
which require pattern
recognition
or simulation
of a physical system
too complex to model with rules are
perfectly
suited to neural computing
techniques."
V a r io u s p r e d ic tio n a lg o r ith m s [ 2 ] [ 3 ] [ 4 ] [ 7 ] [ 1 7 ] b a s e d o nn e u r a l n e tw o r k s h a v e b e e n f o r m u la te d .
T h e p e r f o r m a n c e o f th e s e tw o a lg o r ith m s is e v a lu a te d o n th e b a s is o f ( 1 ) p r e d ic tio n a c c u r a c y a n d ( 2 ) w h e th e r
th e p r e d ic te d tim e s e r ie s c a p tu r e s th e d y n a m ic s o f th e tim e s e r ie s u n d e r in v e s tig a tio n .
T h e f ir s t c r ite r io n o f p r e d ic tio n a c c u r a c y a s a m e a s u r e o f p e r f o r m a n c e is th e p r im e p u r p o s e o f th e a lg o r ith m .
H o w e v e r , th e s e c o n d c r ite r io n a s k s a m o r e s tr in g e n t q u e s tio n : if th e a lg o r ith m p r e d ic ts a n u m b e r o f tim e
s e r ie s v a lu e s in to th e f u tu r e , d o th o s e v a lu e s im ita te th e b e h a v io u r o f th e tim e s e r ie s b e in g p r e d ic te d ?
T h e q u e s tio n th e n a r is e s a b o u t w h ic h c o m b in a tio n s o f e m b e d d in g d im e n s io n a n d tim e la g a r e th e b e s t f o r
p r e d ic tio n a c c u r a c y a n d f o r c a p tu r in g o f th e d y n a m ic s . F u r th e r m o r e , a r e th e tw o c r ite r ia s a tis f ie d b y th e
s a m e c o m b in a tio n o f te s t p a r a m e te r s ? F in a lly , a r e th e b e s t c o m b in a tio n s in d e p e n d e n t o f th e c h o ic e o f th e
a lg o r ith m ?
1.3
Lorenz
s y s t e m ,
general information and equations
T h e L o r e n z tim e s e r ie s [ 1 6 ] [ 2 3 ] is a s ta p le s o u r c e o f d a ta f o r th e e v a lu a tio n o f p r e d ic tio n a lg o r ith m s .
T h e L o r e n z e q u a tio n s d e s c r ib e a s im p lif ie d m o d e l o f f lu id m o tio n d r iv e n b y c o n v e c tio n in th e a tm o s p h e r e b y
in c o r p o r a tin g v a r ia b le s f o r th e r a te o f c o n v e c tio n o v e r tu r n in g ( x ) , th e h o r iz o n ta l te m p e r a tu r e v a r ia tio n ( 1 } ) ,
1.3. LORENZ SYSTEM,
GENERAL
INFORMATION
AND EQUATIONS
350
Phase space plot of the Lorenz attractory(t) N -20 -15 -10 -5 x (t) o 5 1 0 1 5 20 F ig u r e 1 .1 : P h a s e s p a c e p lo t o f th e L o r e n z a ttr a c to r o b ta in e d f r o m th e s o lu tio n o f th e th r e e L o r e n z e q u a tio n s . d x u ( y - x ) ,
=
( 1 .1 ) d t d y -=
p x - y - x z , ( 1 .2 ) d t d z=
- l3 z+
x y . ( 1 .3 ) d tT h e P r a n d tl n u m b e r ,
u,
is p r o p o r tio n a l to r a tio o f th e f lu id v is c o s ity o f a s u b s ta n c e to its th e r m a l c o n d u c tiv ity .T h e R a y le ig h n u m b e r , p , is p r o p o r tio n a l to d if f e r e n c e in te m p e r a tu r e b e tw e e n th e to p a n d b o tto m o f th e
s y s te m a n d 1 3 is th e r a tio o f th e w id th to th e h e ig h t o f th e R a y le ig h - B e n a r d c e ll e n c a s in g th e s y s te m . T h e s e
p a r a m e te r s a r e u s u a lly s e t to th e f o llo w in g v a lu e s ;
u
=
10,
p=
2 8 a n d 1 3=
l
T h e L o r e n z e q u a tio n s h a v e c e r ta in im p o r ta n t f e a tu r e s :
• T h e e q u a tio n s a r e a u to n o m o u s , i.e . th e r ig h t- h a n d s id e s a r e n o t f u n c tio n s o f tim e .
• T h e e q u a tio n s a r e n o n - lin e a r b e c a u s e o f th e x z a n d x y te r m s .
• T h e s o lu tio n s a r e b o u n d e d .
• T h e r e a r e n o u n s ta b le p e r io d ic o r b its o r u n s ta b le s ta tio n a r y p o in ts .
E v e n th o u g h th e e q u a tio n s c o n ta in n o r a n d o m , n o is y o r s to c h a s tic te r m s th e r e is a s u r p r is in g le v e l o f u n p r e
-d ic ta b ility in th e s o lu tio n , a s is e v id e n t f r o m f ig u r e
1.1.
4
CHAPTER
1 .INTRODUCTION
• T h e in fin ite n u m b e r o f d a ta p o in ts th a t c a n re a d ily b e g e n e ra te d c o m p u ta tio n a lly fro m th e e q u a tio n s .
• T h e s o lu tio n o f th e L o re n z e q u a tio n s d e s c rib e s a s tra n g e a ttra c to r in p h a s e s p a c e .
T h e s o lu tio n fo r e q u a tio n s (1 .1 ). (1 .2 ) a n d (1 .3 ) w e re o b ta in e d fro m th e u tilis a tio n o f a n e x p lic it R u n g e -K u tta
fo rm u la . In th is w o rk th e o d e 4 5 fu n c tio n o f M a tla b w a s u s e d .
S o lu tio n s c a n b e fo u n d fo r x . 1 ) a n d z b u t th is th e s is w ill o n ly u tilis e th e tim e s e rie s o b ta in e d fro m th e s o lu tio n
o f
1 ).L e t X i =1 ) (( i. - l) h ) . w h e re h is th e tim e s te p fo r th e s o lu tio n o b ta in e d fro m th e fu n c tio n s o lv e r a n d fo r th is p a p e r is s e t a t 0 .0 1 7 0 s . T h e re fo re , th e tim e s e rie s u s e d fo r a lg o rith m e v a lu a tio n in th is th e s is h a s th e fo rm ,
T
a n d is g ra p h ic a lly re p re s e n te d in fig u re 1 .2 . { X l,X l, ... ,Xn } , P lo t o f L o re n z tim e s e rie s (1 .4 ) 0 .9 0 .8 . " ~ ~ 0 .4 0 .3 0 .2 0 .1 0 2 4 6 8 1 \ 1 0 1 2 1 4 1 6 -F ig u re 1 .2 : P lo t o f th e s o lu tio n o f 1 )( t1
fo r th e L o re n z s y s te m s o f e q u a tio n s .1.4. NOTATION
CONVENTIONS
1.4
Notation
conventions
5 V e c t o r s a r e p r i n t e d i n b o l d f a c e a n d d e n o t e a c o l u m n v e c t o r . T h e t r a n s p o s e o f v e c t o r s w i l l b e i n d i c a t e d b y t h e s u p e r s c r i p t T . T h e E u c l i d e a n n o r m o f a v e c t o r a=
[a
1 ,a2, ... , an]T
w i l l b e r e p r e s e n t e d b yII
aII
w i t h1.5
Chapter Preview
IIall=
Ja:a
=
JL;-
af.
( 1 . 5 )C h a p t e r 2 e x p l a i n s t h e f u n d a m e n t a l i d e a s b e h i n d e m b e d d i n g t h e o r y a n d e l a b o r a t e s o n t h e p r o c e s s e s b e h i n d c a l c u l a t i n g t h e o p t i m a l e m b e d d i n g d i m e n s i o n a n d t i m e d e l a y f o r a g i v e n t i m e s e r i e s . H a v i n g l a i d t h e g r o u n d w o r k f o r p r e d i c t i o n a l g o r i t h m s t o c o m e , c h a p t e r 3 e x p l a i n s S a u e r 's a l g o r i t h m . C h a p t e r 4 b r i e f l y i n t r o d u c e s t h e r e a d e r t o m a i n c o n c e p t s i n n e u r a l n e t w o r k a r c h i t e c t u r e a s w e l l a s e x p l a i n i n g t h e t i m e d e p e n d e n t a r c h i t e c t u r e u t i l i z e d i n t h e a l g o r i t h m . T h e i m p l e m e n t a t i o n o f t h e a l g o r i t h m s i s e x p l a i n e d i n c h a p t e r 5 a n d r e s u l t s a n d a n a l y s i s o f b o t h a l g o r i t h m s a r e p r e s e n t e d .
Chapter 2
Concepts
In
•
embedding theory
A d y n a m ic a l s y s te m is a s e t o f e q u a tio n s th a t g iv e s th e tim e e v o lu tio n o f th e s y s te m 's s ta te fro m in itia l
c o n d itio n s .
T h e d y n a m ic a l s y s te m m a y e ith e r b e d e s c rib e d b y c o n tin u o u s v a ria b le s , w h e re a c o m m o n fo rm o f th e d y n a m ic a l
s y s te m c a n b e d e s c rib e d b y firs t-o rd e r a u to n o m o u s o rd in a ry d iffe re n tia l e q u a tio n s s u c h a s th o s e fo r th e L o re n z
e q u a tio n s (1 .1 , 1 .2 , 1 .3 ). T h e s o lu tio n fo r th e d iffe re n tia l e q u a tio n s d e s c rib e s a c o n tin u o u s e v o lu tio n o f th e
s ta te o f th e s y s te m fro m in itia l c o n d itio n s .
S im ila rly , fo r d is c re te v a ria b le s , th e d y n a m ic a l s y s te m m a y b e d e s c rib e d b y a m a p o f th e fo rm ,
X n + l
=
F (x n l.fo r w h ic h th e s o lu tio n d e s c rib e s a d is c re te e v o lu tio n o f th e s ta te o f th e s y s te m fro m in itia l c o n d itio n s . (2 .1 )
E a c h m u lti-d im e n s io n a l p o in t o b ta in e d fro m th e s o lu tio n o f th e d y n a m ic a l s y s te m d e s c rib in g th e s ta te o f th e
s y s te m is re p re s e n te d b y a c o rre s p o n d in g p o in t in p h a s e s p a c e w h o s e c o o rd in a te s a re th e v a ria b le s o f th e s ta te .
T h e re fo re a n y p o in t in p h a s e s p a c e c o m p le te ly d e s c rib e s th e s ta te o f th e s y s te m a n d c o n v e rs e ly a n y s ta te o f
th e s y s te m h a s a c o rre s p o n d in g p o in t in p h a s e s p a c e . T h e re la tio n s h ip (in te rc o n n e c tio n ) o f p o in ts in p h a s e
s p a c e d e s c rib in g th e tim e e v o lu tio n o f th e s y s te m is te rm e d a tra je c to ry .
If, fo r s o m e in itia l c o n d itio n s th e tra je c to rie s c o n v e rg e to s o m e b o u n d e d s u b s e t o f p h a s e s p a c e , th is s e t is
re fe rre d to a s a n a ttra c to r. T ra n s ie n t d a ta a re v a lu e s o b ta in e d fo r th e s y s te m th a t h a v e n o t y e t c o n v e rg e d
o n to th e a ttra c to r. D u e to fa c t th a t tra n s ie n t b e h a v io u r d o e s n o t c a p tu re th e s ta b le d y n a m ic s o f th e s y s te m ,
its e ffe c t o n p re d ic tio n a c c u ra c y is a n im p o rta n t c o n s id e ra tio n a n d th u s fo r th e re m a in d e r o f th is th e s is it is
a s s u m e d th a t a ll tra n s ie n t b e h a v io u r is o m itte d fro m th e re le v a n t d a ta s e ts .
If th e v a ria b le s o b s e rv e d fro m a s y s te m d o n o t fu lly d e s c rib e a p h a s e s p a c e p o in t th e n s o m e m e th o d is re q u ire d
to re c o n s tru c t p h a s e s p a c e fro m th e in c o m p le te s e t o f v a ria b le s .
8
CHAPTER
2.
CONCEPTS
IN EMBEDDING
THEORY
2 .1
P h a s e s p a c e re c o n s tru c tio n
R eco n stru ctio n o f th e o rig in al p h ase sp ace fro m o b serv ed d ata is u su ally n o t p o ssib le. M eth o d s are th u s so u g h t
w h ich allo w th e reco n stru ctio n o f th e attracto r in p h ase sp ace fro m th e av ailab le d ata (fo r earlier m eth o d s
see [1 0 ] an d [1 9 ]). T h e req u irem en t fo r reco n stru ctio n is th at lin earisatio n o f th e d y n am ics at an y p o in t in
th e p h ase sp ace is p reserv ed b y th e reco n stru ctio n p ro cess. T h is is d u e to th e n eed to ex p lo it d eterm in ism in
th e d ata an d th e p ro v isio n th at th e tim e ev o lu tio n o f a trajecto ry in reco n stru cted p h ase sp ace o n ly d ep en d s
o n its cu rren t p o sitio n .
G iv en an o b serv ed d iscrete tim e series, T , a d elay reco n stru ctio n , w ith d im en sio n m an d tim e d elay 1 , is as
fo llo w s,
b (m -l )l+ n
=
[X (m -l )l+ n ,'" , X 2 l+ n , X l+ n ,x
n ] T .(2 .2 )
E m b ed d in g th eo ry states th at fo r id eal n o ise-free d ata th ere ex ists a m in im al v alu e fo r th e d im en sio n , m , su ch
th at th e reco n stru cted v ecto rs
b
are eq u iv alen t to p h ase sp ace v ecto rs. E q u iv alen ce, in th is sen se, refers to th e fact th at th e m ap is a o n e-to -o n e im m ersio n (to b e ex p lain ed b elo w ).A co n stru ctio n is o n ly co n sid ered an em b ed d in g if th e reco n stru cted m ap d o es n o t co llap se p o in ts o r tan g en t
d irectio n s, in o th er w o rd s, th at th e m ap p in g is o n e-to -o n e an d th at d ifferen tial in fo rm atio n is p reserv ed .
W h en th e state o f th e d eterm in istic d y n am ical sy stem , an d th e fu tu re ev o lu tio n , is co m p letely sp ecified b y a
sin g le p o in t in p h ase sp ace th en th e m ap p in g is co n sid ered o n e-to -o n e. If at a g iv en state x th e v alu e f(x )
is o b serv ed in th e reco n stru cted sp ace an d th at after a fix ed tim e in terv al an o th er ev en t fo llo w s, th en if
f
is o n e-to -o n e each ap p earan ce o f f(x ) w ill b e fo llo w ed b y th e sam e ev en t.If a sm o o th m ap , F o n
A,
is a co n tin u o u sly d ifferen tiab le m an ifo ld an d th e d eriv ativ e m ap D F (x ) is o n e-to -o n e fo r ev ery p o in t x o nA
th en th e m ap is an im m ersio n . U n d er an im m ersio n n o d ifferen tial stru ctu re is lo st ing o in g fro m
A
toF(A).
T h erefo re a m ap , F , is an em b ed d in g if an d o n ly if it is a o n e-to -o n e im m ersio n . F ig u res 2 .1 , 2 .2 an d 2 .3
illu strate d ifferen t v alid an d in v alid em b ed d in g m ap s.
In o rd er to em b ed a scalar tim e series w e n eed to co n stru ct m in d ep en d en t v ariab les, th e n u m b er req u ired
to u n iq u ely ch aracterize th e sy stem . T h e v ariab le set (v ecto r) w ill th u s ex ist in so m e m -d im en sio n al sp ace.
T h erefo re, to reco n stru ct th e attracto r in th e n ew p h ase sp ace w e n eed to fin d th e em b ed d in g d im en sio n , m ,
an d an o p tim al tim e d elay , w h ich w ill allo w th e b est ch o ice o f v alu es fro m th e tim e series so th at th e v ariab les
2.2.
TIME DELAY
9F
F ig u re 2 .1 : A o n e-to -o n e m ap th at is n o t co n
-tin u o u sly d ifferen tiab le, i.e. n o t an im m ersio n .
F ig u re 2 .2 : A m ap th at is n o t o n e-to -o n e.
F
F (A )
F ig u re 2 .3 : A n em b ed d in g F o f th e sm o o th m ap A .
2.2
Time Delay
T he reconstruction of phase space is equivalent to the original space of variables and is a m ethod for providing
independent coordinates com posed of the present observation Xt and earlier observations of the system . T he
tim e delayed coordinate set at tim e
t,
w ith tim e delay 1, appears as follow sd(t)
[X t, X t~ l, X t-2 l, ... ,] . (2.3)If
the tim e delay is chosen too sm all then successive delayed coordinates w ill not be independent enoughand not enough tim e w ould have passed for the system to have explored the state space sufficiently. A s
an exam ple, consider a highly oversam pled data set w here given a sm all tim e delay successive tim e delayed
m easurem ents w ould be alm ost identical and provide no inform ation about the system .
C onversely, if the tim e delay w ere too large, tw o successive m easurem ents in equation 2.3 w ould be nearly
random w ith respect to each other.
1 0 xx -Time delay =2 Time delay =9 Time delay = 30
CHAPTER
2 .CONCEPTS
IN EMBEDDING
THEORY
'"
z x X l'"
+ x -X l o'"
xZ X l F i g u r e 2 . 4 : D i f f e r e n t c h o i c e s o f t i m e d e l a y s o n a L o r e n z t i m e s e r i e s . o f 2 v e r y l i t t l e i n f o r m a t i o n a b o u t t h e e v o l u t i o n o f t h e s y s t e m i s o b t a i n e d a n d t h a t f o r a t i m e d e l a y o f 3 0 t h e p o i n t s h a v e a n e a r r a n d o m d i s t r i b u t i o n a n d l i k e w i s e g i v e n o i n f o r m a t i o n o n t h e s t a t e o f t h e s y s t e m . F o r a t i m e d e l a y o f 9 i t c a n b e f e l t b y v i s u a l i n s p e c t i o n t h a t t h e d y n a m i c s o f t h e s y s t e m i s b e i n g c a p t u r e d . T h i s v a l u e f o r t h e t i m e d e l a y w i l l b e s h o w n t o b e t h e t h e o r e t i c a l c o r r e c t c h o i c e f o r t h e L o r e n z t i m e s e r i e s . F i g u r e 2 .4 a l s o s h o w s t h e L o r e n z a t t r a c t o r i n 2 d i m e n s i o n s f o r t h e d i f f e r e n t t i m e d e l a y s . I t c a n o n c e a g a i n b e s e e n t h a t i n c o r r e c t c h o i c e s o f t h e t i m e d e l a y l e a d t o b a d r e c o n s t r u c t i o n s o f t h e a t t r a c t o r . A v a l u e f o r t h e t i m e d e l a y i s t h u s s o u g h t , w h i c h w i l l e n s u r e t h a t s u c c e s s i v e t i m e d e l a y e d m e a s u r e m e n t s , a r e n o t c o m p l e t e l y i n d e p e n d e n t i n a s t a t i s t i c a l s e n s e . A m e t h o d i s s o u g h t t h a t q u a n t i f i e s t h e d e p e n d e n c e o fd(t -
1 ) o n v a l u e s o fd(t)
i n a t i m e s e r i e s . M u t u a l i n f o r m a t i o n i s d e f i n e d b y 1 ( 1 )=
-.L
P i j ( 1 ) I n P i j ( 1 ) i j P i P j , w h e r e P i i s t h e p r o b a b i l i t y t o f i n d a n o b s e r v a b l e i n t h e i .- t h i n t e r v a l o f a t i m e s e r i e s a n d P i j i s t h e p r o b a b i l i t y t h a t a n o b s e r v a t i o n f a l l s i n t o t h e i .- t h i n t e r v a l a n d a f t e r a t i m e 1 i s o b s e r v e d i n t h e j - t h i n t e r v a l . F r a s e r [ 9 ] d e m o n s t r a t e s t h a t t h e f i r s t m i n i m u m o f t h e m u t u a l i n f o r m a t i o n o f a t i m e s e r i e s c o r r e s p o n d s t o t h e b e s t c h o i c e o f t i m e d e l a y . I n f i g u r e 2 .5 t h e a v e r a g e m u t u a l i n f o r m a t i o n f o r t h e L o r e n z t i m e s e r i e s i s s h o w n w i t h t h e f i r s t m i n i m a i n d i c a t i n g2 .3 . E M B E D D I N G D I M E N S I O N
Average mutual information for the Lorenz time series
2 . 5 2
11
c o'iii
E.E
. S ~ 1 . 5 ::> E O J O l co Q ; >«
0 . 5 IS> 2 IS> 3 (S) 4 (S) 5 l" l s 6 7 Time delay 8 8 9 A 1 0 e 1 1 8 1 2 F i g u r e 2 . 5 : A v e r a g e m u t u a l i n f o r m a t i o n v e r s u s t i m e d e l a y f o r t h e L o r e n z t i m e s e r i e s . t h e o p t i m u m t i m e d e l a y a t 9 . E x p e r i m e n t a l r e s u l t s i n c h a p t e r 5 i l l u s t r a t e h o w t h e c h o i c e o f t i m e d e l a y a f f e c t s p r e d i c t i o n a c c u r a c y .2 . 3
E m b e d d i n g
D i m e n s i o n
A n a t t r a c t o r i s c o n s i d e r e d u n f o l d e d , a n d t h u s a n e m b e d d i n g , i f a l l n e i g h b o u r i n g p o i n t s i n t h e a t t r a c t o r a r e t r u e n e i g h b o u r s . K e n n e l [ 1 4 ] d e f i n e s a f a l s e n e a r e s t n e i g h b o u r a s f o l l o w s : " A fa ls e n e ig h b o u r is a p o in t in th e d a ta s e t th a t is a n e ig h b o u r s o le ly b e c a u s ewe
a r e v ie w in g th e o r b it ( th e a ttr a c to r ) in to o s m a ll a n e m b e d d in g s p a c e ." T h e r e f o r e a l l n e i g h b o u r i n g p o i n t s , i n a n i n c o m p l e t e u n f o l d e d a t t r a c t o r , w i l l n o t b e s o s o l e l y d u e t o t h e d y n a m i c s o f t h e s y s t e m b u t a l s o b e c a u s e t h e g e o m e t r i c s t r u c t u r e o f t h e a t t r a c t o r w a s p r o j e c t e d d o w n t o s m a l l e r s p a c e t h a n r e q u i r e d b y t h e a t t r a c t o r . I t i s o b v i o u s t h a t p r e d i c t i o n s m a d e w i t h i n f o r m a t i o n f r o m f a l s e n e i g h b o u r s w o u l d d e l i v e r u n r e l i a b l e r e s u l t s . T h e m e t h o d o f f a l s e n e a r e s t n e i g h b o u r s , i n t r o d u c e d b y K e n n e l [ 1 4 ] ' i s u s e d t o c a l c u l a t e a m i n i m u m e m b e d d i n g d i m e n s i o n f o r t h e a t t r a c t o r s o a s t o a v o i d p r o j e c t i n g p o i n t s i n t o n e i g h b o u r h o o d s o f t h e a t t r a c t o r t o w h i c h t h e y d o n o t b e l o n g . T h i s m e a n s t h a t f o r d i m e n s i o n s m l e s s t h a n t h e m i n i m u m e m b e d d i n g d i m e n s i o n m a t h e t o p o l o g i c a l s t r u c t u r e o f t h e a t t r a c t o r i s n o t p r e s e r v e d a n d t h e r e c o n s t r u c t i o n i s n o t a o n e - t o - o n e i m a g e . T h e r e f o r e , o n c e t h e n u m b e r o f f a l s e n e a r e s t n e i g h b o u r s i s z e r o t h e m i n i m u m e m b e d d i n g d i m e n s i o n h a s b e e n12
CHAPTER
2.
CONCEPTS
IN EMBEDDING
THEORY
re a c h e d a n d th e a ttra c to r h a s b e e n u n fo ld e d in p h a s e s p a c e . F ig u re 2 .6 illu s tra te s th e e ffe c t o H a ls e n e ig h b o u rs .
False and true nearest neighbours on the Lorenz attractor
0 -.
+
+'><
X t F ig u r e 2 .6 : A n illu s tr a tio n o f f a ls e a n d tr u e n e a r e s t n e ig h b o u r s . F o r th e p o in t P th e n e a re s t n e ig h b o u r is th e fa ls e n e ig h b o u r F n (t*)
a s c a n b e s e e n b y th e d iv e rg e n c e o f th e fa ls e n e ig h b o u r's tra je c to ry . T h e re a l n e ig h b o u r is p o in tNn(t)
w h o s e tra je c to ry is s im ila r to th a t o f p o in t P .T h e s q u a re o f E u c lid e a n d is ta n c e b e tw e e n a p o in t
b
n a n d a n e a re s t n e ig h b o u rb
r in d -d im e n s io n a l s p a c e is [R d (n , r)]2 d-lL.
(X n -k l - X r_ k l)2 ,k=O
w h e re 1 is th e c h o s e n tim e d e la y . If th e e m b e d d in g d im e n s io n is in c re a s e d b y o n e to d+
1 , th e n [R d+
1 (n , r)]2 - [R d (n , r)]2 [X n -d l - X r-d tl 2 .A c c o rd in g to K e n n e l [1 4 ]' a firs t o f tw o c rite ria fo r a p o in t to b e d e fin e d a s a fa ls e n e ig h b o u r is
IX n d l - X r-d ll
>
Rto 1, R d (n ,r)
w h e re Rto l is a c h o s e n to le ra n c e .
A s e c o n d c rite rio n is re la te d to th e ra d iu s o f th e a ttra c to r. If th e s u p p lie d tim e s e rie s is s m a ll a p ro b le m
m ig h t a ris e b e c a u s e a s th e e m b e d d in g d im e n s io n s in c re a s e s , s o th e p o in ts in th e a ttra c to r s p re a d o u t to fill th e p h a s e s p a c e . T h u s in s m a ll d a ta s e ts n e a re s t n e ig h b o u rs m ig h t n o t b e c lo s e , i.e . th e E u c lid e a n d is ta n c e
2.3. EMBEDDING
DIMENSION
D i m e n s i o n 1 2 3 4 P e r c e n t a g e o f f a l s e n e a r e s t n e i g h b o u r s 4 2 .6 1 .6 0 013
T a b l e 2 . 1 : P e r c e n t a g e o f f a l s e n e a r e s t n e i g h b o u r s a t d i f f e r e n t e m b e d d i n g d i m e n s i o n s f o r t h e L o r e n z t i m e s e r i e s . b e t w e e n n e i g h b o u r i n g p o i n t s i s o f t h e s a m e o r d e r a s t h a t o f t h e 'r a d i u s ' o f t h e a t t r a c t o r , R A, w h e r e[R
A ] 21
N (
1
N
)2
N';
xn-N,;
xn R A i s t h e n t h e r o o t m e a n s q u a r e v a l u e o f t h e d a t a a b o u t i t s m e a n . T h e r e f o r e , i f a n e i g h b o u r i n g p o i n t m e e t s t h e c r i t e r i o n i n e q u a t i o n 2 .4 i t i s d e e m e d f a l s e , w h e r e A t o l i s a c h o s e n t o l e r a n c e . Rd+
1 ( n , r )>
A t o l .RA
( 2 .4 ) A n i m p o r t a n t c o n s i d e r a t i o n n e e d s t o b e t a k e n i n t o a c c o u n t i f t h e d a t a i s o v e r s a m p l e d . T h e n , w h e n e x a m i n i n g a n e a r e s t n e i g h b o u r t h e p o i n t c o u l d e x i s t o n t h e s a m e t r a j e c t o r y a n d j u s t b e d i s p l a c e d b y t h e t i m e e q u i v a l e n t o f t h e s a m p l i n g r a t e . T h i s p o i n t i s a l w a y s a t r u e n e i g h b o u r a n d m i g h t i n f l u e n c e t h e f a l s e n e i g h b o u r s t a t i s t i c . A s i m p l e s o l u t i o n i s t o d o w n s a m p l e b u t t h i s d i s c a r d s d a t a a n d i s u s u a l l y u n a c c e p t a b l e . A n o t h e r s o l u t i o n i s t o d i s c a r d p o i n t s t h a t h a v e a t i m e i n d e x w i t h i n a b o u n d f r o m t h e i n d e x o f t h e p o i n t u n d e r i n v e s t i g a t i o n . A m o r e e l e g a n t s o l u t i o n i s t h a t o f f a l s e s t r a n d s , a s i n t r o d u c e d b y K e n n e l [ 1 3 ] . T h e c o n c e p t i s s i m i l a r t o t h a t o f t h e f a l s e n e a r e s t n e i g h b o u r a p p r o a c h e x c e p t t h a t a g r o u p o f n e a r e s t n e i g h b o u r i n g p o i n t s o n a t r a j e c t o r y , o r s t r a n d , a r e c h o s e n . I f i n i n c r e a s i n g t h e d i m e n s i o n a n y p o i n t o n a s t r a n d i s f a l s e , t h e e n t i r e s t r a n d i s d e e m e d f a l s e . T h e m e t h o d p r o c e e d s i n a s i m i l a r f a s h i o n t o t h a t o f t h e f a l s e n e a r e s t n e i g h b o u r s m e t h o d . E x p e r i m e n t a l r e s u l t s i n c h a p t e r 5 i l l u s t r a t e h o w t h e c h o i c e o f a n e m b e d d i n g d i m e n s i o n s a f f e c t s t h e p r e d i c t i o n a c c u r a c y .Chapter 3
Prediction with the Sauer algorithm
T o m a k e a s e r ie s o f p r e d ic tio n s f o r a tim e s e r ie s , a m o d e l o f th e g iv e n tim e s e r ie s h a s to b e c o n s tr u c te d .
T h e m e th o d u s e d f o r p r e d ic tio n th e n e m p lo y s a c o n tin u a tio n o f th e m o d e l s o th a t f u tu r e v a lu e s o f th e tim e
s e r ie s m a y b e c a lc u la te d b y m e a n s o f s o m e f o r m o f e x tr a p o la tio n . T h e r e f o r e , p r e d ic tio n m o d e ls illu s tr a te h o w
th e tim e s e r ie s d a ta m u s t b e v ie w e d o r in te r p r e te d a n d p r e d ic tio n m e th o d s p r o c e s s e s th e d a ta to m a k e th e
p r e d ic tio n .
M o d e ls m a y b e c la s s if ie d a s e ith e r lo c a l o r g lo b a l m o d e ls . G lo b a l m o d e ls m a k e p r e d ic tio n s o f f u tu r e v a lu e s o f
th e tim e s e r ie s p o s s ib le b y u s in g a ll k n o w n p o in ts in th e tim e s e r ie s . N e u r a l n e tw o r k s ( c h a p te r 4 ) a r e g lo b a l
m o d e ls a s th e y a r e tr a in e d o n a ll a v a ila b le tim e s e r ie s d a ta to m a k e p r e d ic tio n s . L o c a l m o d e ls in tu r n o n ly
u tiliz e a p o r tio n o f th e a ttr a c to r in a n e ig h b o u r h o o d a r o u n d th e d a ta p o in t f o r w h ic h th e p r e d ic tio n is b e in g
m a d e . T h e a lg o r ith m d e v is e d b y S a u e r [ 1 8 ] a n d s tu d ie d in th is c h a p te r is a n e x a m p le o f a lo c a l m o d e l.
T h e s im p le s t m o d e ls u s e d f o r p r e d ic tio n o f a tim e s e r ie s a r e b a s e d o n lin e a r te c h n iq u e s . L in e a r m o d e ls o n ly
p r e d ic t p e r io d ic o s c illa tin g , e x p o n e n tia l s o lu tio n s o r c o m b in a tio n s th e r e o f .
I n tu r n , n o n - lin e a r p r e d ic tio n te c h n iq u e s a tte m p t to a p p r o x im a te m o r e o f th e d y n a m ic s o f a s y s te m a n d in
s o d o in g a p p r o x im a te b o th th e lin e a r a n d n o n - lin e a r b e h a v io u r o f th e tim e s e r ie s . D u e to th e f le x ib ility o f
n o n - lin e a r m o d e ls o v e r f ittin g o n d a ta c a n b e c o m e a r e a l c o n s id e r a tio n b e c a u s e th e m o d e l f its e q u a lly w e ll o n
tr u e tim e s e r ie s f e a tu r e s a n d o n n o is e in th e tim e s e r ie s .
16
CHAPTER
3.
PREDICTION
WITH
THE SAUER
ALGORITHM
3 .1
P re m is e o f S a u e r's A lg o rith m
T h e S a u e r a lg o rith m to b e d isc u sse d in th is c h a p te r h a s its o rig in s in a v e ry sim p le b u t p o o r p re d ic tio n
te c h n iq u e u tiliz in g e m b e d d e d a ttra c to rs th a t w ill b e te rm e d th e c e n te r o f m a ss (C O M ) p re d ic tio n te c h n iq u e .
N o te th a t fo r th is c h a p te r a tra je c to ry re fe rs to a n o rd e re d se t o f se q u e n tia l p o in ts in a lo c a liz e d a re a o f
th e a ttra c to r a ro u n d th e p re d ic tio n p o in t. It th e re fo re m a k e s se n se to re fe r to d iffe re n t tra je c to rie s o n o n e
a ttra c to r. T h is is in c o n tra st to th e p re v io u sly d e fin e d v ie w o f a sin g le tra je c to ry b e in g d e fin e d fo r th e e n tire
a ttra c to r.
In sh o rt, th e C O M te c h n iq u e fin d s a n u m b e r o f n e a re st n e ig h b o u rs a ro u n d th e la st p o in t o n th e a ttra c to r,
sh ifts th e n e a re st n e ig h b o u rs o n e p o in t fo rw a rd o n th e re sp e c tiv e tra je c to rie s a n d p re d ic ts th e n e x t p o in t o n
th e a ttra c to r a s b e in g th e c e n te r o f m a ss o f th e se n e ig h b o u rs. T h is te c h n iq u e is v e ry se n sitiv e to p re d ic tio n
e rro rs a n d m u ltiste p p re d ic tio n s q u ic k ly fa il.
S a u e r ta k e s th is sim p le a p p ro a c h a n d im p ro v e s o n it b y in tro d u c in g th e c o n c e p t o f p rin c ip a l c o m p o n e n t
a n a ly sis th ro u g h th e u se o f sin g u la r v a lu e d e c o m p o sitio n (S V D ).
G iv e n a n e m b e d d e d tim e se rie s (e m b e d d in g d im e n sio n
m),
th e n e ig h b o u rin g p o in ts to th e fin a l, o r p re d ic tio n , p o in t o n th e a ttra c to r a re fo u n d . T h e S V D o f th e m a trix fo rm e d b y th e v e c to rs d e sc rib in g th e p o sitio n o fn e ig h b o u rin g p o in ts re la tiv e to th e c e n tre o f m a ss o f th e se n e ig h b o u rin g p o in ts is o b ta in e d .
A se t o f o rth o g o n a l v e c to rs sp a n n in g th e ro w sp a c e o f a m a trix is o b ta in e d fro m th e S V D o f th e m a trix .
B y c h o o sin g th e first s o rth o g o n a l v e c to rs, a su b sp a c e is fo u n d w h ic h is th e b e st s-d im e n sio n a l lin e a r sp a c e
(s ::;
m )
th ro u g h th e p o in ts in a le a st-sq u a re s se n se .P ro je c tio n s o f th e m a trix o n to th e c o lu m n sp a c e p ro v id e d b y th e S V D a re c a lc u la te d . L in e a r m o d e ls a re b u ilt
w h ic h m a p th e p ro je c tio n s to k n o w n fu tu re tim e se rie s v a lu e s o f th e n e ig h b o u rin g p o in ts. A sim ila r m a p p in g
is th e n a p p lie d to th e p re d ic tio n p o in t to p re d ic t th e n e x t v a lu e in th e tim e se rie s.
3 .2
R e c o n s tru c tio n
o f th e A ttra c to r
T h e re c o n stru c tio n sta rts w ith c h o o sin g a d im e n sio n ,
m,
a n d tim e d e la y , 1 . If th e tim e se rie s h a s th e sa m e fo rm a s p re v io u sly d e fin e d in c h a p te r 2 th e n th e e m b e d d in g m a trix , B , a p p e a rs a s fo llo w s,B
X ( m - l) l+ l X ( m - l ) l+ 2 X l+ l X l+ 2 X l X2 b ( m - l) l+ l b ( m - lll+ 2 (3 .1 ) X t X t- ( m - 2 ) l X t- ( m - l) lb
TtIt c a n b e n o te d fro m e q u a tio n 3 .1 th a t th e tim e se rie s is e m b e d d e d in 're v e rse ' a c ro ss th e ro w s o f th e m a trix .
T h is is u se fu l in sim p lify in g so m e in te rp re ta tio n a sp e c ts a n d d o e s n o t c h a n g e th e fu n c tio n in g o f th e a lg o rith m
3.3.
SELECTION
O FTHE NEAREST
NEIGHBOURS
17
T h e p re d ic tio n p o in t,
h
t , in te rm s o f th e e m b e d d e d tim e s e rie s is th e la s t ro w , o r c o o rd in a te p o in t, in B . T h eta rg e t p o in t is h t + 1 a n d is to c o n ta in th e firs t p re d ic te d v a lu e .
T a rg e t v a lu e s , W i = X ( m - l ) l + ( i + 1), a re d e fin e d a s th e s c a la r v a lu e s s u c h th a t h T m - l ) l + i m a p s to W i . It is th u s im m e d ia te ly a p p a re n t th a t th e p re d ic tio n p o in t, h t, m a p s to th e ta rg e t v a lu e W t = X t +1 w h ic h is n o t a n e m b e d d e d v a lu e in (3 .1 ).
3 .3
S e l e c t i o n o f t h e
N e a r e s t
N e i g h b o u r s
T h e firs t ta s k is lo c a te th e 'V n e a re s t n e ig h b o u rs to th e e m b e d d e d p re d ic tio n p o in t,
h
t , fro m th e tra in in gs e t (d a ta p o in ts g iv e n a t th e s ta rt o f th e p re d ic tio n ). In e x te n d e d p re d ic tio n ru n s o f m o re th a n o n e p o in t it
is im p o rta n t to d iffe re n tia te b e tw e e n th e tra in in g s e t a n d p re d ic tio n s e t. T h e tra in in g s e t is th e e m b e d d in g
m a trix w ith w h ic h th e in itia l p re d ic tio n ru n b e g a n . N e ig h b o u rs a re o n ly c h o s e n fro m th e tra in in g s e t s o th a t
e rro rs in p re d ic tio n d o n o t in flu e n c e th e p re d ic tio n a c c u ra c y o f la te r p o in ts .
T h e E u c lid e a n d is ta n c e fro m h t to a n y p o in t h i in th e a ttra c to r is c a lc u la te d a s fo llo w s
. 6 t i = l l h t - h i
II.
T h e v e c to rs h i c o rre s p o n d in g to th e s m a lle s t . 6 t i 'S a re c a n d id a te s fo r n e a re s t n e ig h b o u rs . H o w e v e r, o n ly o n e p o in t o n e a c h tra je c to ry s h o u ld b e ta k e n a s a n e a re s t n e ig h b o u r to e n s u re th a t e n o u g h d iv e rs e in fo rm a tio n
c a n b e fo u n d a b o u t th e b e h a v io u r o f th e tra je c to rie s in th e a p p lic a b le re g io n o f th e a ttra c to r. T h e re fo re th e 'V
c lo s e s t n e ig h b o u rs o n s e p a ra te tra je c to rie s a re c h o s e n a s th e 'V n e a re s t n e ig h b o u rs , a n d a re re la b e lle d ,
nj,
fo rj
=
1 , ... ,'V. T h e m a trix N is c o n s tru c te d w ith th e 'Vv e c to rsnT
a s ro w s a n d th e ir a s s o c ia te d ta rg e t v a lu e s a re re la b e lle d C j. T h e ta rg e t v e c to r is d e fin e d a sN=
e =nT
nI
T noy C lCz
C o y(3.2)
( 3 .3 )A m e th o d o f e n s u rin g a c h o ic e o f 'V n e ig h b o u rs o n 'V d iffe re n t tra je c to rie s is : if a n y c a n d id a te n e ig h b o u r, h i
h a s a n in d e x w h ic h d iffe rs a c h o s e n
b
o r le s s fro m a p re v io u s ly s e le c te d n e ig h b o u r's in d e x th e n th e c a n d id a te1 8
CHAPTER
3.
PREDICTION
WITH
THE SAUER ALGORITHM
A s tr a ig h tf o r w a r d w a y to p r e d ic t w o u ld b e to f in d th e b e s t r e g r e s s io n o f e j o n l l j a n d to u s e it to p r e
-d ic t W t f r o m
b
t . H o w e v e r , th is m e th o d is s u s c e p tib le to n o is e in th e d a ta . I t is th e r e f o r e m o r e a p p r o p r ia teto d r a w th e m o r e d o m in a n t tr e n d s f r o m th e m a tr ix N a n d u s e o n ly th a t f o r th e r e g r e s s io n . T h e u s e o f th e
S V D th u s b e c o m e s p r u d e n t a n d th e f o llo w in g s e c tio n s d e s c r ib e th e p r o c e s s in v o lv e d .
3.4
Center of Mass of the Nearest Neighbours
H a v in g o b ta in e d th e "V n e a r e s t n e ig h b o u r s to
b
t th e c e n te r o f m a s s o f th e s e n e ig h b o u r s h a s to b e f o u n d . T h e n o r m a l m e th o d f o r c a lc u la tin g th e c e n te r o f m a s s v e c to r , c o f a n u m b e r o f p o in ts is a s f o llo w s , "V . L l l j j = l C=--"V i = 1 , ... ,m . ( 3 .4 )A w e ig h te d c e n tr e o f m a s s is u s e d in th is th e s is , b e c a u s e o f s lig h tly im p r o v e d p r e d ic tio n a c c u r a c y , w ith ,
"V . L W j l l j j = l
c=-
"V . L W j j = l i = 1 , ... ,m , ( 3 .5 ) w h e r e Wj is th e w e ig h t c o n s tr u c te d in s u c h a w a y th a t th e w e ig h tin g in c r e a s e s w ith a d e c r e a s e in th e d is ta n c eo f a n e a r e s t n e ig h b o u r to th e p r e d ic tio n p o in t. T h e c h o s e n w e ig h tin g e x p r e s s io n , Wj, is a s f o llo w s ,
W j
=
( 1 _ ~ (II
l l j - b tII)
2) 3
2II
ll" V - btII
A m a tr ix C is c o n s tr u c te d w ith "V r o w s o f c a s f o llo w s , cT
c
T
C=
cT
3.5
Singular Value Decomposition
and Projections
( 3 .6 )
H a v in g o b ta in e d th e m a tr ix C th e s in g u la r v a lu e d e c o m p o s itio n , S V D [ 2 4 ] ' o f th e r e s u ltin g d is p la c e m e n t
m a tr ix ,
D
=
N - C c a n n o w b e c a lc u la te d . T h e d is p la c e m e n t m a tr ix ,D,
d e s c r ib e s p o s itio n v e c to r s f o r e a c h n e ig h b o u r r e la tiv e to th e c e n tr e o f m a s s . B y c a lc u la tin g th e S V D f o r th e m a tr ix , v e c to r s a r e o b ta in e d th a ts p a n th e r o w s p a c e o f th e m a tr ix .
S in g u la r v a lu e d e c o m p o s itio n a llo w s th e f a c to r is a tio n o f th e "V x m m a tr ix ,
D,
a s f o llo w s ,3.5.
SINGULAR
VALUE DECOMPOSITION
AND PROJECTIONS
o
P r e d ic tio n
P o in t
•
N e ig h b o u r s
E 9 C O M
o f N e ig h b o u r s
•
o th e r p o in ts
F ig u r e 3 .1 : I llu s tr a tio n o f p o in ts o n th e a ttr a c to r r e le v a n t to th e S a u e r a lg o r ith m .
s o th a t
r
0"1 liL VT ~ [u, ... u,
]~m
[ v , ... vm
r
0
1 9 w h e re U ( v X 'V ) a n d V ( m x m ) a re b o th o rth o g o n a l m a tric e s a n dL
( 'V X m ) is d ia g o n a l. T h e v e c to rs , V 1 , ... , Vm, a re o rth o n o rm a l b a s e s fo r th e ro w s p a c e o f 0 a n d U 1 , ... ,u " a re lik e w is e o rth o n o rm a l b a s e s fo r th e c o lu m n s p a c e o f D . H a v in g p re v io u s ly c h o s e n a s p a n n in g d im e n s io n s , s o th a t s :::;m,
th e firs t s c o lu m n s in V a re s e le c te d . T h e s e le c te d v e c to rs a re k n o w n a s th e d o m in a n t rig h t s in g u la r v e c to rs a n d s p a n th e lin e a r s p a c e , RS . RS is a s u b s p a c e in th e p h a s e s p a c e Rm. T h e s p a n n in g v e c to rs , V 1 , V 2 ," " Vs , a re p a c k e d in to th e ( m x s ) m a trix S a s fo llo w s ,(3.7)
S= [
V1 Vs ] , ( 3 .8 )2 0
CHAPTER
3. PREDICTION
WITH
THE SAUER ALGORITHM
T h e s p a n n in g v e c to r s , S , th u s d e s c r ib e a lo w - d im e n s io n a l lin e a r s p a c e p a s s in g th r o u g h th e c e n te r o f m a s s , c , o f th e n e ig h b o u r s . T h e a d v a n ta g e o f u s in g th e S V D b e c o m e s a p p a r e n t b e c a u s e th e d y n a m ic in f o r m a tio n o f th e d is p la c e m e n t v e c to r s a r e c o m p r e s s e d in to th e f ir s t f e w d o m in a n t v e c to r s w h e r e a s th e n o n - d o m in a n t v e c to r s m a in ly s to r e le s s im p o r ta n t in f o r m a tio n s u c h a s n o is e . T h e r e f o r e , b y p r o je c tin g th e d is p la c e m e n t v e c to r s o n to th e s u b s p a c e RS th e v e c to r s a r e f o r c e d o n to th e s u b s p a c e p a s s in g th r o u g h th e c e n tr e o f m a s s w h ils t d e c r e a s in g th e e f f e c t o f n o is e in th e d a ta . T h e p r o je c tio n , P , is c a lc u la te d a s f o llo w s pT
=eDT,
w h e r ee
is th e p r o je c tio n m a tr ix .e
=
S ( S T S ) - ' S T .3.6
Regression of the Linear Model
I n th e f o llo w in g s e c tio n it is a s s u m e d th a t th e d is c u s s io n r e v o lv e s a r o u n d a a ttr a c to r f o r w h ic h th e r o w c o u n t
o f th e e m b e d d e d m a tr ix is f a r g r e a te r th a n th e d im e n s io n
m.
O n c e th e c o lu m n s o f 0 h a v e b e e n p r o je c te d d o w n to th e s u b s p a c e , RS
, th e lin e a r r e g r e s s io n m o d e l c a n b e
f o u n d w h ic h b e s t f its th e p r o je c te d v e c to r s .
L e t
Q
b e th e p r o je c tio n m a tr ix P a u g m e n te d w ith a c o lu m n o f o n e s , v iz .Q
= [P u ] w ith u = [ 1 ,1 , ... , l] T . B y a u g m e n tin g th e p r o je c tio n m a tr ix w ith o n e s th e r e g r e s s io n c o e f f ic ie n ts a s s o c ia te d w ith th a t c o lu m n p la y sth e r o le o f a c o n s ta n t. I t is th u s r e q u ir e d to s o lv e th e r e g r e s s io n c o e f f ic ie n ts , k , in th e f o llo w in g e q u a tio n , Q k e . E x p a n d in g ( 3 .9 ) g iv e s , r P ','
P',l
...
P ',m 1k,
Pl,'
Pl,l
...
Pl,m
1kl
P " , ,P",l
...
P " ,m 1k
m+,
( 3 .9 )C,
Cl ( 3 .1 0 ) C"3.7
Calculation of the Target Point
H a v in g o b ta in e d a lin e a r e q u a tio n f o r th e p r e d ic tio n p o in t b t th e ta r g e t p o in t b t + ' c a n b e c a lc u la te d .
3.7.
CALCULATION
O F
THE TARGET
POINT
21
T h e so lu tio n fo r
k
le tse ~
Qk
su c h th a tIIQk -
ell
is a s sm a ll a s p o ssib le . T h e re fo re th e e x p re ssio n qTk
a p p ro x im a te s W t =b t+ l (').T h e c o m p le te e x p re ssio n fo r th e p re d ic te d e m b e d d e d p o in t b t+ l is
b t+ l
=
[q T k b t-l+ l(l) b t-1 + 1 (1 )b
t-1+
1(m - 1) ] .
(3 .1 1 )N o te th a t th e
1
th ro u g h m -th e n trie s in b t+ l (3 .1 1 ) a re p re v io u sly e m b e d d e d p o in ts in B .T h e ta rg e t p o in t th e n b e c o m e s th e p re d ic tio n p o in t in th e n e x t p re d ic tio n a n d th e a lg o rith m p ro c e e d s in th e
Chapter
4
Prediction with Time Delayed
Neural Networks
4.1
An Introduction to Neural Networks
A n e u ra l n e tw o rk [5 ] [2 2 ] [2 5 ] [2 8 ] is a s y s te m o f d a ta s to ra g e p o in ts , c a lle d n o d e s . T h e s e n o d e s a re s e t u p
in s u c h a w a y th a t th e v a lu e s in s o m e o f th e n o d e s d e te rm in e th e v a lu e s in o th e r n o d e s .
If
th e v a lu e in o n en o d e d e p e n d s o n th e v a lu e in a n o th e r n o d e th e n th e s e n o d e s a re s a id to b e
connected.
A w e ig h t is a s s o c ia te dw ith e a c h c o n n e c tio n th a t d e te rm in e s to w h ic h d e g re e th e v a lu e in th e n e x t n o d e is in flu e n c e d b y th e v a lu e
o f th e p re v io u s n o d e .
A n e u ra l n e tw o rk a lw a y s h a s a s e t o f in p u t n o d e s a n d a s e t o f o u tp u t n o d e s . It a ls o h a s s o m e in te rm e d ia te
n o d e s c a lle d
hidden
n o d e s . T h e s im p le s t d e s c rip tio n o f th e fu n c tio n in g o f a n e u ra l n e tw o rk (N N ) is th a t it isa n o n lin e a r o p e ra to r th a t m a p s a v e c to r, d im e n s io n
n,
in th e in p u t s p a c e o n to a v e c to r, d im e n s io nm,
in th e o u tp u t s p a c e , w h e re th e re a ren
in p u t n o d e s a n dm
o u tp u t n o d e s .W h e th e r th e N N is u s e d fo r c la s s ific a tio n o r fo r c o m p u ta tio n a l p u rp o s e s , th e u s e r o f th e N N s h o u ld h a v e a
te s t s p a c e o f in p u t v e c to rs w ith th e ir a s s o c ia te d d e s ire d o u tp u t v e c to rs . T h e te s t s p a c e , o r tra in in g s e t, is
th e n u s e d to tra in th e n e tw o rk in th e fo llo w in g w a y : fo r e a c h in p u t a p p lie d to th e n e tw o rk th e w e ig h ts a re
a d ju s te d u n til th e e rro r b e tw e e n th e a c tu a l o u tp u t a n d th e c o rre c t o u tp u t is m in im is e d . O n c e a s a tis fa c to rily
lo w e rro r is o b ta in e d th e n e tw o rk is c o n s id e re d
trained
a n d th e n e tw o rk c a n th e n b e u s e d to o u tp u t v a lu e s fo rin p u t v e c to rs w h ic h h a v e n o t p re v io u s ly b e e n p re s e n te d to th e m o d e l.
T h e fu n c tio n in g o f a n e u ra l n e tw o rk is a s fo llo w s . A n in p u t v e c to r in tro d u c e d to a n e tw o rk is s to re d in th e
in p u t la y e r, (s e e fig u re 4 .1 ). T h e n o d e s o f th e in p u t la y e r a re c o n n e c te d to e v e ry n o d e in th e n e x t la y e r, th e
firs t h id d e n la y e r. If a ll p o s s ib le c o n n e c tio n s a re m a d e b e tw e e n th e n o d e s in tw o a d ja c e n t la y e rs th e n th e
la y e rs a re c o n s id e re d fu lly c o n n e c te d (in g ra p h th e o ry th is c o n fig u ra tio n is re fe rre d to a s a c o m p le te b ip a rtite