Biomedical Fuzzy Systems Association
NII-Electronic Library Service
BiomedicalFuzzy Systems Association
Biomed, FuzzySys.Bulr,Vol.1,No,2,1991
CONNECTIONISM,
ARTIFICIAL
INTELLIGENCE
AND
FUZZY
CONTROL
Elie
Sanchez
Laboratoire
d'Infbrmatique
Medicale,
Faculty
ofMedicine,
University
ofMarscille
and
Neural
&
Fuzzy
Systems
Institute,
Neurinfo
Research
Department
lnstitut
Mediterran6en
de
Technologie,
Marseille,
France
After
recalling somefundamentals
ofknowledge-based
systemsand of
fuzzy
controlthrough
approximate
reasoning concepts,it
is
introduced
artificial neural networks(connectionism)
andit
is
shown
how
these
three
scientificfields,
that
are under activedevelopment,
share common
features
andtechniques.
Modelling
brain
functions,
particular[[y
those
involved
in
perception,
leaming,
memory and movement,has
long
challenged many researchers.ln
the
late
60's,
M.
Minsky
andS.
Papert
discussed
the
theoretical
limitations
ofpercepmons,the
first
precisely
spechied, computationally oriented neuralnetwotks.
As
a consequenceto
these
criticisms, expert systems soonemerged as a major
branch
ofArtificial
intelligence
providing
a strongimpetus
to
the
development
of approximate reasoning methodologies,following
L.A.Zadeh's
theory
offuzzy
sets and associatedpossibMty
distributions.
But,
expert systemshave
shown some weaknesses,fbr
examplein
the
process
of eiicitingknowtedge
from
experts,in
leaming
capahilities orin
producing
poor
resultsat
the
limits
ofthe
system'sdomain
of expertise.Neural
networks are offering noticeable contributionsto
expert systems such as :training
by
example,dynamical
adjustment of changesin
the
environment, ability
to
generalize,
tolerance
to
noise,gracefu1
degradation
atthe
border
ofthe
domain
of expenise, abilityto
discover
new relationsbetween
vaiiables.Fuzzy
logic,
supportinginterpolative
reasoning,is
playing
akey
rolein
human
cognitive systems.The
standards of accuracy andprecision
prevailing
in
traditional
computers arepresently
questioned
ordiscarded,
especially while narrowing
the
gap
between
human
reasoning and machinereasonmg.
-55-Biomedical Fuzzy Systems Association
NII-Electronic Library Service
BiomedicalFuzzySystems Association
l
Connectionist
netwotks(or
anificialneural netwoiks)tools
are nowused
for
learning
controlproblems
like
the
cart-polebalancing
system.Combination
offuzzy
logic
with neural netwotkstheory
is
enhancingthe
capability of
intelligent
systems tolearn
from
experience and adaptto
changes
in
an environment withquaiitative,
imprecise,
uncertain orincornplete
inforrnation.
In
ahybrid
architecture associating neura1 netwotksand expert systems
techniques,
neuial netwoiks can act aspreprocessors
for
the
treatment
oflow
level
irifbrmatiog
oras
internal
subsystemsfor
learning
tasks,
generalization
or classification.A
connectionist expert systemis
anexpert system
that
uses connectionist netwotksfor
constructingthe
knowledge
base
from
training
examples,giving
a wayto
automatethe
generation
of expert systems.Fuzzy
logic
has
been
used recentlyin
conjunction with anificialneura1 netwotks.
We
propose
now an expert classMcation systemin
which aconnectionist
inodel
is
usedto
extract or totune
the
knowledge
from
atraining
set of examples.An
impertant
feature
ofthis
modelis
its
fuzzy
nature with an
intrinsic
treatment
offuzzines.
inputs
to
the
neurai system areweighted,
but
we assumethat
weights are eftwo
types
:primary
weights,in
general
followed
by
secondary weights.Primary
weights expressthe
main
information
onknowledge,
they
have
a1inguistic
form
andthey
areinterpreted
aslabels
offuzzy
sets.Depending
on applications,these
fuzzy
setsaiedefined
on universesof
discourses
relatedto
the
nature ofthe
input
celis or,1ike
in
fuzzy
control,they
canbe
members of agiven
partition
ofthe
interval
[-1,+1],
withtriangular
shaped membershipfunctions
typically
meaning '7VegativeLarge
(N]L],
Negatiye
Medium
iIVM),
Negative
Smagt
(NS),
2`tppreximatelyZero
(ZR),
Positive
Smali
(PS),
Positive
Medizam
(Pl{di,
Positive
Lai:ge
(RL)",
ormore simply, 'tDecreased,
Nbrmal,
Increased."
Secondary
weights are numbersin
[O,1],
they
refiectthe
grade
of weakness ofthe
correspondingconnection
(t!ie
weakerthe
connection,the
closerto
1
the
weight) andthey
do
not necessarily act on cormectionsbut
whenthey
do
so,they
follow
aprimary
weightthey
are combined wnh.The
fuzzy
neural netwotkis
afeedfbrward
system. 'Ihere are nodlrected
cycles and nofeedback,
oneiteration
is
sufficientfor
inferencing.
The
training
phase
is
notperformed
frdm
methodsinvelving
weighted sums efinputs,
but
from
intrinsically
fuzzy
equations, using min and max operators.-Biomedical Fuzzy Systems Association
NII-Electronic Library Service
BiomedicalFuzzySystems Association
It
is
not askedto
find
the
membershipfunctions
ofthe
primary
weightsin
the
general
case,for
any universe ofdiscourse.
It
is
assumedthat
a
human
experthas
a roughidea
ofthe
shapes,the
task
is
to
tune
the
curvesaccording
to
the
information
provided
from
input-output
examples.Learning
consists
in
finding
the
numerical weights andthe
netwotktopology.
Here
are someissues
that
willbe
illustrated,
gommented
andaddressed
for
discusssion.
i)
It
has
been
saidthat
"neura1netwotks and
fuzzy
legic
systems are notopposed, on
the
contrarythey
go
hand
in
hand,"
but
how
?
ii)
What
is
afuzzy
neuron?
What
arethe
possible
approaches todefine
it
?
iii)
What
canbe
afuzzy
artificialneura1 network?
iv)
Should
one consider conical colums asunits,instead
of neurons?
v)wrat
canfuzzy
logic
(or
fuzzy
aritirmetic)bring
to
connectionism?
vi)How
canfuzzy
neural networkstechniques
be
usedin
medical systems,for
examplein
knowledge
elicitation orin
diagnostic
classification?
vii)