Vous êtes sur la page 1sur 7

election of vapours and odours from

a multisensor array using pattern


recognition: self-organising daptive
esonance Techniques
by .J.W. Gardner, E.l. Hines and C. Pang
Department of Engineering, University of Warwick, Coventry

Introduction

,,

The human sense of smell is still the


primary 'instrument' used to assess the
odour quality of raw materials and end
products in many industries. Conventional analytical instruments, such as
gas chromatographs and mass spectrometers, have been unable to satisfy
the industrial need for a sensitive, reproducible and rapid odour measurement system. This inadequacy has
encouraged researchers to design an
electronic instrument that can mimic
the remarkable behaviour of the human olfactory system 1 The first international conference on Electronic
Noses was held in 1991 2 and since then
there has been an almost exponential
growth in research activity. A historical overview is given by Gardner and
Bartlett'.
Figure 1 shows the manner in which
signals are processed within an electronic nose. An odour (e.g. a floral
compound) comprises a set of molecular components of varying concentration and can be represented here by a
vector in sample (concentration) space
delivered by the 'odour delivery system'. Various active materials and sensor types can be employed as odour
detectors in the sensor array but the
most common ones are metal oxide
semiconducting (MOS) or conducting
polymer (CP) chemoresistors, coated
quartz crystal microbalances (QCMs)
and metal oxide semiconductor fieldeffect transistors (MOSFETS) with
catalytically active gates. Generally the
sensor output is non-specific so they
are suitable for application within an
electronic nose, but because they are
highly temperature and humidity dependent and tend to suffer from drift
or poisoning, it is desirable to apply

intelligent signal compensation routines and pattern recognition (PARC)


paradigms such as artificial neural networks to classify the output space 4
The current commercial instruments
may be described as first generation
because they possess limited capabilities to train sensor array signals and
compensate for undesirable characteristics. One possible way of improving
the instrument is to design signal
processing techniques that more closely
resemble our own well-proven olfactory system. Here we examine the implementation of one such technique,
called adaptive resonance theory, which
enables the electronic nose to learn
new odour patterns in an on-line manner.

Data preparation
An odour sensor has a response characteristic that approximates to that of a
first-order instrument, i.e. a simple RC network. It is common to ignore the
transitory response because it is determined by the physical time to deliver
the odour to the sensor (i.e. transfer
function of the odour delivery system)
and the dynamics of the odour source.
In some applications it is possible to
use an automated or robotic headspace
injection system and then the variations in the delivery system are suitably small to permit analysis of the
transient signals.
The choice of the data pre-processing algorithm has been shown to affect
significantly the performance ofPARC
methods. For example, in MOS odour
sensors, various sensor response parameters x 11.. have been used by workers,
see Table 1. It is our view that the
exact choice of parameter should
depend upon the underlying sensor

172 Measurement+ Control. Volume 29, July/August 1996

principle and the nature of the interfering signals. It has been shown that the
fractional change in conductance
( Godow_ Gai')/Gair helps to linearise the
sensor output (with concentration) and
to reduce its temperature sensitivity,
thus improving the performance of
chemometric 5 and neural networking
techniques 6 The concentration-dependence of the odour sensor can be
removed in linear sensors by normalising the sensor parameter according
to:
X

If

I)

~
I.x .

.I
"\J i==l

/}

where x ' is the normalised response of


v
sensor 1 to complex odour j over an
array of n sensors. In some respects
this simulates a characteristic of the
human nose namely that its perception
of the intensity of smells is rather poor.
One disadvantage is that, in the case
of weak smells the transform enhances
the noise.
The output from the sensor array
may be regarded as an n-dimensional
input vector S to the pattern recognition system where,
S1

= [x;; x r
2

... x,) and jE [1, m]

The purpose of the pattern recognition


system is then to discriminate between
a set of m classes of odour.

Feed-forward multilayer
perceptron
An important class of neural network is
the multilayer feed-forward network
which typically consists of a set of
sensory units (note that in an electronic
nose there is a source or input node for
each odour sensor) that constitutes the

Intelligent Instruments - Gardner et al


ANALOGUE SENSING

DIGITAL PROCESSING

OdOUf

s1gnal

Xo

ODOUR
DELIVERY
SYSTEM

l ayer

Fig 1 Schematic representation of signal processing within a generic electronic


nose

input layer, one or more hidden layers


of computational nodes and an output
layer of computational nodes (one for
each classj of odour). The input signal
propagates through the network in a
forward direction as illustrated in Figure 2. These networks are commonly
referred to as 'multilayer perceptrons'
(MLPs) as a generalised form of the
single-layer perceptron and their widespread application has been promulgated by Rumelhart and others 4 .
Multilayer perceptrons have been
successfully applied to solve electronic
nose problems by training them in a
supervised manner with a popular algorithm called the error back-propagation algorithm. Basically, the sensor
parameters x,j are set to lie between [1, + 1] and the odour vector S.1 fed into
the input layer of the m ultilayer
perceptron. Each neurone computes a
weighted sum vjk of the input, i.e.

p n. Next the weighted sum is transformed by a non-linear activation function, usually a logistic or sigmoidal
function, where the output is _v,
v

l+exp(-l)

The output of the neurone is then feel


into the next layer. The output layer
compares the computed output J\ with
the desired output dk to define an error
signal,
5

Then the signals are passed back


through the network and used to adjust
the initially random weights wu in the
network. The error correction rule used
is usually the back-propagation and is
given by

vk

=L w
I= 0

ke X e

where in the case of the first input layer


of a fully connected network, I i and

Type

Log-fractional

Formulae

Iog[(x,/"""'- x,/") I x,/'']

Table 1. Definitions of several


common sensor response parameters'

L1w11

= Y] o v

I"" k

where the constantY] is called the learning rate parameter (generally set to
1.0) and 01 is the local gradient and
given by the product of , with the
gradient of the activation function (in
most implementations an additional
'momentum' term is added to improve
the learning process). The process is
repeated for a number of steps until the
network error has converged to a suitably small value. The network error is
commonly defined as the sum of
squared errors in the output layer, e.g.

Fwst

Si~iicond

hidden

layer

Outer
layer

layer

Fig 2. Architectural graph of a jitllyconnected feed-forward multilayer


perceptron that processes the sensor
array data

Adaptive resonance theory


Adaptive resonance theory (ART) was
introduced as a theory of human cognition in information processing~. It is
based on the fact that a human brain
can leam many new events without
necessarily forgetting events learnt in
the past. This characteristic is not observed in back-propagation algorithms
which have supervised learning, i.e.
designed using the data-set in an 'offline' sense. If we wish to have an intelligent system capable of autonomously
adapting in real time to changes in the
world, then we face the so-called 'Stability-Plasticity Dilemma'. In other
words,
(a) how can a learning system preserve
its previously learnt knowledge while
continuing to leam new things?
and
(b) what prevents the new knowledge
from destroying the memories of prior
learning?
ART networks have been designed
to resolve this stability-plasticity elilemma and so be stable enough to incorporate new information. The basic
principle is to add a feedback mechanism between the competitive layer
and the input (i.e. sensor) layer of the
network.
Three types of ART network have
been studied here, namely, ART 1 which
requires binary inputs, ART2 which
deals with binary and analogue signals
and SMART2 which is a supervised
variant of ART2R 10

ART1 algorithm

The basic features of an ARTI system


are shown in Figure 3. The two major

Measurement+ Control, Volume 29. July/August 1996

173

Intelligent Instruments - Gardner et al


ATTENTIONAL

ORIENTING

Initial top-down L TM traces are initialised according to:

- =

.....J i -

B -1
Dl

(0) >_I_

-,,

15

if iri-S

and

and bottom-up L TM traces are initialised according to:


L
O < :,, (O) < L- 1 + M

if iES

-~i-

L
L- 1 +lSI

if iES

if iri-S

16

Rest rule - An active F2 node is reset


if

Fig 3 Architecture (){ an ART1


network. The rectangular blocks
represent layers in which the STM
patterns are stored. The arrows
indicate paths that are not adaptive.
Solid circles represent gain control
nuclei which sum input signals

The STM activities x of all nodes v.


I
I
on F 1 are initialised to zero, but the
STM activities x of all nodes 1'. on F 1
are initialised to!
'
x, (0)

10

Binary Input Patterns are initialised


to:
1 if i E 1
l =
II
0 otherwise

' l

,.

subsystems are the attentional subsystem and the orienting subsystem;


whereas F 1 and F2 are layers of network
nodes. Nodes on each layer are fully
connected to the nodes on the other
layer. Patterns of neural activities that
develop over the nodes in the two layers are called short-term memory
(STM) traces because they only exist
for a single application of an input
vector. F and F2 are linked both bottom-up and top-down by adaptive filters where the weights represent the
long-term memory (L TM) as they remain in the network for an extended
period. The orientating subsystem is
responsible for generating a reset signal to F, when the bottom up input
pattern and top-down template pattern
mismatch at F 1 according to a l'igilance
criterion. The mathematical model is
summarised below and interested readers are refened to reference 9 for more
details.
The global network parameter and
constants are defined as:

The bottom-up activation and 2/3


rule requires that:

M : number of nodes on F,
N :number of nodes on F,
A 1 : network parameter (A 1 > 0)
C 1 : network parameter (C 1 > 0)
D 1 : network parameter (0 1 > 0)
B 1 : network parameter
(max !D 1 , 1) <B 1 <D 1 )
L : network parameter (L > I J
p : attentional vigilance parameter
(O<p<l)

Bottom-up (F 1 ~ F):

174

ifF, is active

s- JI

12

l1 n V'J'

if the F2 node 1' is active


1

F, Choice and Search:


-If .J is the index setofF, nodes which
have not yet been reset on the present
learning triaL then:

J I T, = max l T,:

=l0

fC\)
where

kE I

It(\

V(J)

,,,

<p

=0

17

Once a node is reset, it remains inactive for the duration of the trial. The
ART! algorithm is described in Appendix A.

ART2 algorithm
ART2 is a class of neural networks that
can self-organise recognition categories for an arbitrary sequences of analogue or binary inputs. Figure 4 shows
the system studied here. ART2 contains all the features of the ART 1 network:
(i) an attentional subsystem which contains an input representation field F 1
and a category representation field F2

l
13

ORIENTING
SUBSYSTEM

ATfENTIONAL
SUBSYSTEM

T = '2,:
J

1'=.\

1/

LTM traces equations - Top-down


(F

~F ):

fK
.
-,,

[(I - :,,) L- :,, (IS[- I )j

if iES andjC1) =I

-K[:,,SI]

ifi!l.Sand/(.r,)=l

()

lf/(.1)

r -:,, + 1

:,, =

=()

14

if iES andf(.r) =I
ifiri.Sandf(x)= 1

if f(.r)I = 0

Fast learning- It is assumed that fast


learning occurs so that, when 1'1 in F2 is
active, all L TM traces approach the
asymptotes,

Measurement+ Control, Volume 29. July/August 1996

Fig 4 Architecture (){ ART2 neMork.


The open arrows indicate specific
patterned inputs to target nodes.
Solid arrows represent non-spec(fic
gain inpllt controls. The solid circles
represent the gain control nuclei that
carry out the normalisation process

Intelligent Instruments - Gardner et al


(ii) an orientating subsystem wh:ch interacts with the attentional subsystem
to cauy out an internally controlled
search process.
The two fields are again linked by
bottom-up and top-down adaptive filters.
One of the main differences between
the ART I and ART2 network is that F 1
is split into three layers in order to
match and learn sequences of analogue
input vectors in a stable fashion. The
top layer feeds the bottom-up L TM
adaptive filter, reads and normalises
input vectors; while the middle layer
matches patterns from the top and bottom layers before sending a composite
pattern through to the other two layers.
Positive feedback loops within F 1 enhance salient features and suppress
noise. The multiple F 1 layers buffer the
network against recording of the
category structure as new input vectors
are presented. F0 , F 1 and Fe are shunting,
competitive networks that contrastenhance and normalise their activation
patterns.
The F, field chooses the node which
responds maximally to an F 1 pattern
being applied to the bottom-up adaptive filter by means of competitive interaction of the F, nodes. The F, field
also suppresses- the F, nodes- providing poor matches based on the value
of the vigilance parameter, as directed
by the orienting subsystem for the duration of the trial. The L TM is made up
of the bottom-up filter (: ), which incorporates learning into ''the network
and the top-down adaptive filter z i
which enables attentional priming,
matching and a self-adjusting parallel
search which is similar to the one in
ARTl. The STM, and LTM and reset
equations may be found in reference
10. The ART2 algorithm is in Appendix B.

Sensor No

Type

SMART2
The semi-supervised adaptive resonance theory (SMART2J is a modified
version of ART2. It has all the properties of ART2 but it is also capable of
supervised learning. This is achieved
by including two design principles.
Firstly in design principle 1 (DPl ),
when a training pattern is applied to the
network, it only allows categories of
the same class to compete for it. This
significantly reduces the problem of
classifying similar trainillg patterns
from different classes to the same category. Thus, the network is much less
likely to develop multiple class categories. Secondly in design principle 2
(DP2). it uses high vigilance only for
patterns which are difficult to classify.
This overcomes a shortcoming of ART2
namely that the attentional vigilance
parameter is not dynamically adjusted.
In other words, if a high vigilance value
is needed to distinguish patterns of
certain classes it will be retained and
may result in an excessive sensitivity
and so unnecessary expansion of the
number of categories associated with
patterns of a particular class. Hence.
two new rules are added to the orienting subsystem. The first is to reset all
F, nodes with class tags other than the
cfass of the current training pattern.
The second is to track the misclassified
training patterns during the test cycle
and to then increase the vigilance level
for just these patterns in the next training cycle. The only change to the attentional subsystem is in the category representation layer F,. SMART2 uses
class information during its training to
inhibit nodes associated with classes
other than those of the training set. As
each new category is formed, it is
tagged with the class of training pattern that fonned it. When a new pattern

Sensor No

Type

TGS 812

TGS 815

TGS 711

TGS 882

TGS 813

TGS 816

TGS 813

10

TGS 816

TGS 814

11

TGS 817

TGS 824

12

TGS 825

Table 2. MOS Odour sensors used in electronic nose

is applied the only active nodes in F,


are the uncommitted ones and those
tagged with the same class as the input
pattern. The SMART2 algorithm is
given in Appendix C.

Experimental procedures
Data were collected from an electronic
nose which consisted of an an-ay of I 2
commercial metal oxide semi-conducting gas sensors (Figaro Engineering
Inc .. Japan). Table 2 is a list of the 12
sensors.
The response.\ from the sensors was
defined as the (steady-state) fractional
change in conductance as this was found
to reduce error; r'. The data was also
normalised according to equation I to
set the range of .r to [0. I].
Two electronic nose data-sets were
analysed. The first set consisted of simple odours: 5 PPM in air of methanoL
ethanol. butan-1-ol. propan-2-ol. and
2-methy-1-butanol. The process was
repeated to provide 8 identical samples
of each of the 5 classes making a set of
40 input vectors 11 The second set consisted of the headspace of three different roasted coffees. Two of the coffees
were of the same roasting level but
different blends while the third was a
different roast level but of the same
blend as one of the first two coffees.
This was a much more difficult problem to solve because the headspace of
coffee forms a complex odour. The
procedure was repeated to produce 30
samples of the first coffee and 29 samples for the second giving a total of 89
input vectors 12
The ART algorithms were written in
C++ so that the software simulator could
be run under UNIX or DOS.

Results
The ART! simulator was applied to
the alcohol data-set using a one-fold
validation procedure 1'. The input vectors were converted into their binary
equivalent using a gray code technique
to ensure that the Hamming distance
between successive numbers was one.
In these tests the vigilance parameter p
was set to 0.65.
Table 3 summarises the performance
of the ART! network as applied to
classify the alcohol samples.
The classification-rate of 50% is
rather disappointing and is believed
to be due to the loss of crucial features in the conversion of the analogue data into its binary equivalent.

Measurement+ Control, Volume 29, July/August 1996

175

Intelligent Instruments - Gardner et al


Consequently, ART! was not explored further as it was felt that ART2 had
more to offer.
The ART2 simulator was applied to
both the alcohol data-set and the coffee
data-set, where one-fold validation was
used for the former and three-fold validation for the latter. The vigilance parameter was set to 0.999 and the noise
threshold parameter 8 to 0.01; after
systematic variation of the vigilance
parameter p and threshold parameters
had shown these values were optimal
values for our experiment. Table 4
summarises the performance of the
ART2 network as applied to classify
both the alcohol and coffee samples.
Thus, ART2 misclassified 3 alcohol
samples and created 12 categories even
though there were only 5 types of alcohol. This is because when an input
pattern does not sufficiently match the
established categories, a new category
is generated. The propensity of the network to do this depends upon the vigilance parameter and it should be set
high when the input patterns are
similar. Even poorer results were found

Sample No.
I
2

3
4

5
6
7

when analysing the coffee samples with


43 samples being misclassified and 31
to 36 categories generated. The capability of ART2 to learn on-line was
also tested. First the ART2 network
was trained with patterns from just 4 of
the 5 different alcohols. Then the network was tested with a data-set containing all 5 classes. The results are
summarised in Table 5 which shows
that the network cannot recognise the
patterns belonging to the unlearnt alcohol (i.e. ethanol). Next, the network
was retrained with the new ethanol
patterns but not the other patterns to
see whether the network could learn
new patterns without forgetting the old.
The network was tested again and Table 5 shows that the retrained network
can preserve its previously learnt patterns while continuing to learn its new
pattern (odour) in 'real-time'. An overall accuracy of 92.5% was found on
alcohols and 52% on the coffees.
A series of algorithms were then
tested to evaluate the performance of
SMART2 against ART2, where
SMART2 is the semi-supervised

No. of categories in
training

version of ART2. Three versions of


SMART2 were tested, namely
SMART2-I (ART2 + DPl), SMART2II (ART2 + DP2) and SMART2-III
(ART2 + DPl + DP2) where DP stands
for design principle. As in ART2, onefold validation was used for analysing
the alcohols and three-fold validation
for the coffees. The vigilance parameter
was set to 0.999 for the alcohol data-set
and 0.9999 for the coffee data-set. Table
6 shows the performance of the
SMART2 networks in analysing alcohol
and coffee input vectors.
It appears that adding design principle DP 1 to ART2 significantly improves
the accuracy of the network but can
also increase the number of categories
(i.e. in the coffee data with similar
input patterns). Adding design principle DP2 to ART2 slightly improves the
accuracy for coffee but dramatically
reduces the number of categories
formed for alcohol (not for coffee).
The combination of both principles
(SMART2-III) appears to give the best
results by substantially increasing the
network accuracy from 92.5% to 97.5%
for alcohols and 52% to 80% for coffees
while substantially reducing the number
of categories formed for alcohols.

No. misclassified in
testing

Conclusions

12
12
11
13
10
12
13
12

Adaptive resonance theory neural networks have been applied to alcohol


and coffee input patterns generated by
a 12-element electronic nose instrument. ART networks are attractive compared with a MLP approach because
they are self-organising and have an
ability to learn new input patterns (i.e.
odours) without forgetting older ones.
It was found that ART2 out performed

3
3
1

2
1
1
2

Table 3. Poformance of"ARTI network in classifying alcohol samples

Sample No.

2
3
4

5
6
7
8

No. of categories in training

No. of misclassifications in testing

ALCOHOL

ALCOHOL

12
12
12
12
12
12
12
12

9
10

COFFEE
35
35
36
35
31
34
34
34
35
32

0
0

0
0
0

Table 4. Results of ART2 classificmion of alcohol and coffee data-sets


176

Measurement+ Control, Volume 29, July/August 1996

COFFEE
6
4
5
3
6
4
4
3
4
4

Intelligent Instruments - Gardner et al


Alcohol

Categories established
in training
Before

Methanol
Butan-1-ol
Propan-2-ol
1-methyl-2 butanol
Ethanol

Categories classified
in testing

After

0,6
!A
2,5,7
3,8
none

Before

0,6
1,4
2,5,7
3,8
9,10,11

After
0

2
3
None

2
3
9

Table 5. Pe1jormance of ART2 network on alcohol data-set hefou' and after the
ethanol input rectors are introduced

Algorithm

Accuracy%
Alcohol

ART2
SMART-I
SMART2-II
SMART2-III

92.5
97.5
97.5
97.5

No. of categories established

Coffee

Alcohol

Coffee

52
75
60
80

12
12
6
6

35
55
38
44

3. Gardner, J. W. and Bartlett, P. N.


(1994) Sensors and Actuators B, 1819. 211-220.
4. Hecht-Nielson, R. (1990) Neurocomputing, Addison-Wesley, USA.
5. Gardner, J. W. (1991) Sensors and
Actuators B. 4, 109-115.
6. Gardner. J. W., Hines. E. L. and
Tang, H. C. ( 1992) Sensors and Actuators B. 9, 9-15.
7. Hines. E. L. and Gardner. J. W.
(1994) Sensors and Actuators B. 19,
661-664.
8. Grossberg, S. (1976) Biological Cybernetics, 23. 187-202.

Table 6. Pe1jormance of SMART2 networks on alco!zol and coffee data

ART1 because of a need to handle


analogue data. SMART2-III achieved
the highest accuracy (97.5% on alcohol and 80% on coffee). It should be
noted that this was slightly outperformed by the popular back propagation algorithm (100% on alcohols 6 and
81% on coffees), but the training time
of ART2/SMART2 is typically an order of magnitude faster than BP on
small data-sets.
Users do not want to retrain their
entire smell libraries every time new
odour patterns are introduced as this
would be very time consuming and
labour-intensive. An alternative approach would be for users to employ a SMART2-III network that
permits the learning of new odours
and gradually forgets older patterns

2. Gardner. J. W., and Bartlett, P. N.


(1992) Sensors and Sensory Svstems
for an Electronic Nose, Kluwer Academic Publishers, Dordrecht, 327.

over an extended period of time. Periodic calibration of the most important


odours will help restore the long-term
loss of old patterns. It could be argued
that the duration of the network's longterm memory should match long-term
sensor drift thereby both increasing the life of the sensor array and
minimising the network error (as
patterns that would have been misclassified clue to sensor drift are forgotten). In conclusion, there is a need
to develop more intelligent neural
networks for electronic nose instrumentation.

References

9. Carpenter, G. A. and Grossberg, S.


( 1987) Computer Vision, Graphics and
Image Processing. 37, 54-115.

,.,.
l

10. Pang. C. (1993) BEng dissertation.


University of Warwick, UK.
II. Gardner, J. W .. Hines, E. L. and
Wilkinson. M. (1990) Meas. Sci.
Techno!., 1. 446-451.

12. Gardner, J. W., Shurmer, H. V. and


Tan, T. T. ( 1992) Sensors and Actuators B, 6, 71-75.
13. Weiss, S.M. and Kulikowski, C. A.
(1991) Computer Systems That Learn,
Morgan Kaufmann Publishers, San
Fransisco.

1. Persaud, K. C., and Dodd, G. H.


(1982) Nature, 299, 352-355.

See over for appendices

1997

Forthcoming Feature Issues

of
Measurement + Control
1996-97

DECEMBER/JANUARY

Building Management
FEBRUARY

Open Feature
MARCH

& APRIL

Development and Application


of the Artificial Nose

Measurement + Control, Volume 29. July/August 1996

177

Intelligent Instruments - Gardner et al


Appendices
START
1. Apply an input pattern I to F, and calculate F, activities.
2. Calculate the output vector, S, for F,.
3. Propagate S forward to F2 and then calculate F2 activities.
4. Find the F? node with largest activation, and set the output of this node to 1 and the outputs of the rest of the nodes to 0.
5. Propagate the output from F2 back to F, and calculate the net input, V, from F2 to the F, nodes.
6. Calculate the new F, activities.
7. Determine the new output vector, S, for F,
8. Determine the degree of match between the input pattern and the top-down expectation.
9. IF (I S I I I I I < p )mark v, as inactive, zero the outputs of F2 , and return to step 1 using the original input pattern.
10. Update only the bottom-up weight and top-down weight connected to the winner node on F2 .
11. Remove the input pattern, restore all inactive F? nodes and return to step 1 with a new input pattern.
END

Appendix A ART 1 algorithm

START
Set vigilance p;
FOR (each training pattern)
WHILE (a training pattern is applied to network) DO
F0 and F, processing;
Apply outputs ofF, to bottom-up adaptive filter;
F, processing (choose the winning node} or add a new node);
Send top-down expectation from node } through the top-down adaptive filter;
IF (degree of match between bottom-up inputs and top-down expectation > p)
THEN Adjust the LTM trace to make the bottom-up and top-down adaptive filters look more like the training pattern;
IF (the number of weight adjustments is exceeded) remove the pattern from the network;
ELSE Activate the orienting subsystem to reset the fth node in F, (i.e. inhibit it from competing again while the current training
pattern remains on the network);
END

Appendix B ART!. algorithm

START
1. Set p == low vigilance (usually p = 0 0);
2. FOR (each training pattern)
WHILE (a training pattern is applied to network) DO
IF (the network misclassified the pattern on the last testing cycle) THEN increase vigilance (usually top = 0.9999);
F. and F. processing;
Apply output ofF. to bottom-up adaptive filter; Reset all F, nodes that are associated with a class that differs from the class
of the current training pattern;
F.. processing (choose the w1nn1ng node i or add a new node);
Send top-down expectation from node j through the top-down adaptive filter;
IF (degree of match between bottom-up inputs and top-down expectation > p)
THEN
Adjust the LTM trace to make the bottom-up and top-down filters look more like the training pattern;
IF (fth node is an uncommitted node) tag it with the class of the current training pattern;
IF (the number of weight adjustment is exceeded) remove the pattern from the network;
ELSE
Activate the orienting subsystem to reset the fth node in F,. (i.e inhibit it from competing again while the current training
pattern remains on the network);
3. FOR (each training pattern)
Apply each training pattern to the new version of network allowing any category to win. See if the class of pattern is the same
as the tag of the category it mapped to.
4. Repeat Step 1 to Step 3 until all patterns are classified correctly in Step 3.
END

AJlpcndix C SMART2 algorithm


178

Measurement + Control. Volume 29. July/August 1996

Vous aimerez peut-être aussi