CONNECTIONISM, ARTIFICIAL INTELLIGENCE AND FUZZY CONTROL(Fuzzy Logics in Biomedical Systems)

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

of

Medicine,

University

of

Marscille

and

Neural

&

Fuzzy

Systems

Institute,

Neurinfo

Research

Department

lnstitut

Mediterran6en

de

Technologie,

Marseille,

France

After

recalling some

fundamentals

of

knowledge-based

systems

and of

fuzzy

control

through

_approximate

reasoning concepts,

it

is

introduced

artificial neural networks

(connectionism)

and

it

is

shown

how

these

three

scientific

fields,

that

are under active

development,

share common

features

and

techniques.

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

and

S.

Papert

discussed

the

theoretical

limitations

ofpercepmons,

the

first

_precisely

spechied, computationally oriented neural

netwotks.

As

a consequence

to

these

criticisms, expert systems soon

emerged as a major

branch

of

Artificial

intelligence

_providing

a strong

impetus

to

the

development

of approximate reasoning methodologies,

following

L.A.Zadeh's

theory

of

fuzzy

sets and associated

possibMty

distributions.

But,

expert systems

have

shown some weaknesses,

fbr

example

in

the

_process

of eiiciting

knowtedge

from

experts,

in

leaming

capahilities or

in

producing

poor

results

at

the

limits

of

the

system's

domain

of expertise.

Neural

networks are offering noticeable contributions

to

expert systems such as :

training

by

example,

dynamical

adjustment of changes

in

the

environment, ability

to

_generalize,

tolerance

to

noise,

gracefu1

degradation

at

the

border

of

the

domain

of expenise, ability

to

discover

new relations

between

vaiiables.

Fuzzy

logic,

supporting

interpolative

reasoning,

is

_playing

a

key

role

in

human

cognitive systems.

The

standards of accuracy and

_precision

prevailing

in

traditional

computers are

_presently

_questioned

or

discarded,

especially while narrowing

the

_gap

between

human

reasoning and machine

reasonmg.

-55-Biomedical Fuzzy Systems Association

NII-Electronic Library Service

BiomedicalFuzzySystems Association

l

Connectionist

netwotks

(or

anificialneural netwoiks)

tools

are now

used

for

learning

control

_problems

like

the

cart-pole

balancing

system.

Combination

of

fuzzy

logic

with neural netwotks

theory

is

enhancing

the

capability of

intelligent

systems to

learn

from

experience and adapt

to

changes

in

an environment with

quaiitative,

imprecise,

uncertain or

incornplete

inforrnation.

In

a

hybrid

architecture associating neura1 netwotks

and expert systems

techniques,

neuial netwoiks can act as

_{preprocessors}

for

the

treatment

of

low

level

irifbrmatiog

or

as

internal

subsystems

for

learning

tasks,

_{generalization}

or classification.

A

connectionist expert system

is

an

expert system

that

uses connectionist netwotks

for

constructing

the

knowledge

base

from

training

examples,

giving

a way

to

automate

the

generation

of expert systems.

Fuzzy

logic

has

been

used recently

in

conjunction with anificial

neura1 netwotks.

We

propose

now an expert classMcation system

in

which a

connectionist

inodel

is

used

to

extract or to

tune

the

knowledge

from

a

training

set of examples.

An

impertant

feature

of

this

model

is

its

fuzzy

nature with an

intrinsic

treatment

of

fuzzines.

inputs

to

the

neurai system are

weighted,

but

we assume

that

weights are ef

two

types

:

_primary

weights,

in

_general

followed

by

secondary weights.

Primary

weights express

the

main

information

on

knowledge,

they

have

a

1inguistic

form

and

they

are

interpreted

as

labels

of

fuzzy

sets.

Depending

on applications,

these

fuzzy

setsaie

defined

on universes

of

discourses

related

to

the

nature of

the

input

celis or,

1ike

in

fuzzy

control,

they

can

be

members of a

_given

_partition

of

the

interval

[-1,+1],

with

triangular

shaped membership

functions

typically

meaning '7Vegative

Large

(N]L],

Negatiye

Medium

_iIVM),

Negative

Smagt

_(NS),

2`tppreximately

Zero

(ZR),

Positive

Smali

_(PS),

Positive

Medizam

_(Pl{di,

Positive

Lai:ge

_(RL)",

or

more simply, 'tDecreased,

Nbrmal,

Increased."

Secondary

weights are numbers

in

[O,1],

they

refiect

the

_grade

of weakness of

the

corresponding

connection

(t!ie

weaker

the

connection,

the

closer

to

1

the

weight) and

they

do

not necessarily act on cormections

but

when

they

do

so,

they

follow

a

primary

weight

they

are combined wnh.

The

fuzzy

neural netwotk

is

a

feedfbrward

system. 'Ihere are no

dlrected

cycles and no

feedback,

one

iteration

is

sufficient

for

inferencing.

The

training

_phase

is

not

_performed

frdm

methods

invelving

weighted sums ef

inputs,

but

from

intrinsically

fuzzy

equations, using min and max operators.

-Biomedical Fuzzy Systems Association

NII-Electronic Library Service

BiomedicalFuzzySystems Association

It

is

not asked

to

find

the

membership

functions

of

the

_primary

weights

in

the

general

case,

for

any universe of

discourse.

It

is

assumed

that

a

human

expert

has

a rough

idea

of

the

shapes,

the

task

is

to

tune

the

curves

according

to

the

information

_provided

from

input-output

examples.

Learning

consists

in

finding

the

numerical weights and

the

netwotk

topology.

Here

are some

issues

that

will

be

illustrated,

gommented

and

addressed

for

discusssion.

i)

It

has

_been

said

that

"neura1

netwotks and

fuzzy

legic

systems are not

opposed, on

the

contrary

they

_go

hand

in

hand,"

but

how

?

ii)

_What

is

a

fuzzy

neuron

?

What

are

the

_possible

approaches to

define

it

?

iii)

What

can

be

a

fuzzy

artificialneura1 network

?

iv)

Should

one consider conical colums asunits,

instead

of neurons

?

v)

wrat

can

fuzzy

logic

(or

fuzzy

aritirmetic)

bring

to

connectionism

?

vi)

How

can

fuzzy

neural networks

techniques

be

used

in

medical systems,

for

example

in

knowledge

elicitation or

in

diagnostic

classification

?

vii)

How

can connectionist

(expert)

systems

be

applied

to

Medicine

and

Biology

?

and

for

what

kmd

of applications

?