• No results found

CRL 90 2 pdf

N/A
N/A
Protected

Academic year: 2020

Share "CRL 90 2 pdf"

Copied!
21
0
0

Loading.... (view fulltext now)

Full text

(1)

V

alidated

Retriev

al

in

Case{Based

Reasoning

Ev

angelos

Simoudis

James

Miller

Digital

Equipmen

tCorp

oration

Cam

bridge

Researc

hLab

CRL

90/2

Decem

ber

12,

1990

Abstract

W

ecom

bine

simple

retriev

al

with

domain-sp

ecic

validation

of

retriev

ed

cases

to

pro

duce

au

seful

practical

tool

for

case-based

reasoning.

Based

on

200

real-w

orld

cases,

we

retriev

eb

etw

een

three

and

six

cases

over

aw

ide

range

of

new

problems.

This

represen

ts

aselectivit

yranging

from

1.5%

to

3%,

com-pared

to

an

average

selectivit

yof

only

11%

from

simple

retriev

al

alone.

cDigital

Equipmen

tC

orp

oration

1990.

All

righ

ts

reserv

(2)
(3)
(4)

ase-2

2

RETRIEV

AL

IN

CBR

man

tic

net

work8

].

In

order

to

build

our

validation

mo

del

we

are

faced

with

aclassic

kno

wledge

acquisition

task.

By

perusing

existing

data

bases

used

by

sp

ecialists

we

are

able

to

acquire

this

kno

wledge

with

areasonable

amoun

to

f

eort

|

and

with

only

asmall

investmen

tof

specialists'

time.

2

Retriev

al

in

CBR

CBR

systems

rst

retriev

ea

set

of

cases

from

acase

base

and

then

reason

from

them

to

nd

asolution

to

an

ewly

posed

problem.

Existing

systems

(1],

2],

3],

4],

9]

and

10])

mak

et

wo

assumptions

ab

out

the

initial

retriev

al

of

cases

from

the

case

base:

1.

Very

few

cases

will

be

retriev

ed

from

the

case

library

.

2.

The

retriev

ed

cases

are

relev

an

tto

the

problem

being

solv

ed.

In

man

yp

ractical

applications,

retriev

al

alone

issucien

tt

oso

lve

the

dif-cult

part

of

atask.

For

example,

in

our

domain

of

diagnosis

of

computer

soft

ware

failures,

sp

ecialists

can

easily

resp

ond

to

customer

problems

if

they

can

quic

kly

locate

afew

similar

cases

from

their

collectiv

epast

exp

erience.

For

this

reason,

weh

ave

concen

trated

on

the

retriev

al

asp

ect

of

case-based

reason-ing.

In

MBR

Talk

10

],

also,

the

essen

tial

task

is

retriev

al

the

\reasoning"

comp

onen

tconsists

of

merely

passing

the

retriev

ed

information

directly

to

an

output

unit.

2.1

Related

W

ork

Closest

to

our

own

work

is

the

work

of

Koton

on

casey

5],

aCBR

system

whic

h

has

been

applied

in

the

domain

of

medical

diagnosis.

casey

has

a

justication

comp

onen

twhose

goal

is

to

determine

whether

the

causal

expla-nation

of

aretriev

ed

case

applies

to

anew

problem.

This

frequen

tly

allo

ws

casey

to

avoid

invoking

its

causal

mo

del

when

creating

an

explanation

for

a

new

case.

casey

's

justication

phase

is

similar

to

our

validation

phase.

But

there

isa

n

imp

ortan

tdierence

betw

een

these

two

systems

arising

from

dieren

tassumptions

abo

ut

tests.

casey

relies

on

precisely

two

tests

(EK

G

and

X-ra

ys),

both

of

whic

h

are

inexp

ensiv

eand

non-in

vasiv

e.

Both

of

these

tests

are

performed

prior

to

the

retriev

al

phase

and

the

results

are

used

to

pro

vide

surface

features

for

the

retriev

al

algorithm.

By

con

trast,

there

(5)

2.2

Tw

oPhases:

Retriev

al

and

Validation

3

literally

hundreds

of

tests

to

be

performed

in

our

domains

and

it

is

far

too

exp

ensiv

eto

perform

all

of

them

in

adv

ance

of

initial

case

retriev

al.

As

a

result,

our

systems

dev

ote

atten

tion

to

minimi

zing

the

num

ber

of

tests

that

are

performed.

W

enot

only

perform

tests

incremen

tally

and

cac

he

the

results,

but

also

emplo

yk

nowledge

ab

out

the

tests

themselv

es

to

reduce

the

num

be

r

of

tests

performed.

The

CBR

system

chef

2],

whose

domain

is

Chinese

co

oking,

also

has

a

justication

comp

onen

t.

In

order

to

justify

eac

h

retriev

ed

case,

chef

uses

bac

kw

ard

chaining

rules.

While

our

validation

mo

del

is

not

appropriate

to

this

domain,

weh

ave

examined

and

rejected

the

use

of

rules

for

our

valida-tion

mo

dels.

This

decision

isb

ased

on

the

dicult

yof

acquiring

the

exp

ert

kno

wledge

needed

to

create

large

rule

sets,

esp

ecially

in

comparison

to

the

simplicit

yof

constructing

our

validation

mo

dels.

These

mo

dels

are

captured

by

aseman

tic

net

work

that

represen

ts

groups

of

tests

and

information

about

the

relationships

betw

een

the

groups.

Furthermore,

as

with

casey

,

chef

do

es

not

use

the

results

of

comparisons

with

earlier

cases

to

prune

its

searc

hf

or

relev

an

tcases.

The

sw

ale

6]

system

concen

trates

primarily

on

mo

difying

the

explanation

(con

tained

in

an

explanation

pack

et

or

XP)

ofa

retriev

ed

case

to

matc

ha

new

problem.

Nonetheless,

it

has

asub

comp

onen

t,

xp

accepter

,that

justies

the

application

of

a

retriev

ed

XP

to

a

curren

tsituation.

The

accepter

veries

an

XP

by

determining

ifit

can

believ

et

he

applicabilit

y

checks

that

are

pac

kaged

with

eac

hX

P.

Eac

hs

uch

test

is

similar

to

the

tests

asso

ciated

with

our

validation

mo

del.

Because

of

the

small

num

ber

of

cases

(eigh

t,

by

our

coun

t)

xp

accepter

nev

er

addressed

the

issues

of

scale

whic

h

are

our

ma

jor

concern.

Th

us,

sw

ale

nev

er

dev

elop

ed

ajustication

mo

del

to

relate

the

various

applicabilit

yc

hec

ks

to

one

another.

2.2

Tw

oPhases:

Retriev

al

and

Validation

Ou

rg

oali

stot

ake

asizable

pre-existing

case

base

along

with

an

ew

problem

and

pro

duce

asmall

num

ber

ofrelev

an

tcases.

Lik

eour

human

sp

ecialists,

our

systems

perform

diagnosis

in

two

phases:

Retriev

al:

it

poses

aq

uery

to

the

case

base

using

asubset

of

the

features

that

describ

ethe

new

(6)
(7)

3.1

What

is

avalidation

mo

del?

5

validation

mo

dels

:structures

that

capture

muc

hof

an

exp

ert's

kno

wledge

in

a

wayt

hatm

akes

it

easy

for

the

validation

phase

to

pro

cess

the

tests

itrequires.

3.1

What

is

avalidation

mo

del?

Rather

than

require

acomplete

and

accurate

description

ofe

acht

est

used

by

the

sp

ecialist,

we

capture

the

overall

structure

of

the

test

space

itself.

The

resulting

structure,

our

validation

mo

del

,consists

of

related

groups

oft

ests

and

information

ab

out

the

relationship

betw

een

the

groups.

For

example,

if

we

wan

tt

ok

noww

hy

ahouse

is

hot

(the

problem),

wem

ay

rst

wan

tto

see

ifthe

air

conditioner

isw

orking

but

this

requires

us

to

nd

out

ifthe

house

has

an

air

conditioner.

In

this

example,

the

desired

test

(is

the

air

conditioner

working?)

is

related

to

another

test

(is

there

an

air

conditioner?).

This

kno

wledge,

as

well

as

kno

wledge

ab

out

the

imp

ortan

toutcomes

and

implications

ofa

test,

is

captured

by

the

validation

mo

del.

W

eh

avec

hosen

to

represen

tthis

kno

wledge

in

the

form

of

aseman

tic

net

work

whose

no

des

corresp

ond

to

sets

of

tests

and

whose

arcs

indicate

relationships

betw

een

these

sets.

3.2

Creating

aV

alidation

M

odel

W

ebuild

our

validation

mo

dels

by

rst

examining

existing

data

bases

that

are

used

byh

uman

sp

ecialists.

These

data

bases

ma

yb

eeither

formalized

(as

in

the

case

of

our

W

PS-PLUS

1

system)

or

merely

informal

notes

prepared

by

the

sp

ecialists

for

their

own

perusal

(as

in

the

case

of

our

VAX/VMS

system).

In

our

two

case-based

systems,

the

existing

data

con

tains

atextual

description

of

the

steps

that

the

specialists

used

to

verify

ah

yp

othetical

explanation

of

the

problem.

In

constructing

the

validation

mo

del,

itis

our

goal

to

capture

the

interrelationships

betw

een

the

validation

tests.

As

aresult,

weh

aveb

uilt

validation

mo

dels

that

corresp

ond

to

ap

articular

case

base

by:

1.

Reading

the

validation

pro

cedures

of

eac

h

case

and

building

a

list

of

all

the

validation

steps

used

in

the

en

tire

data

base.

In

the

pro

cess

of

reading

the

data

base

and

preparing

this

list,

the

implem

en

tor

dev

elops

asense

of

the

underlying

(but

unstated)

relationships

betw

een

tests

that

are

men

tioned

in

the

data

base.

1

DEC,

VAX,

VMS

and

WPS-PLUS

are

registered

trademarks

of

Digital

Equipmen

t

Corp

(8)

6

4

AN

EXTENDED

EXAMPLE

2.

Examining

the

resulting

list,

looking

for

groups

oftests

that

app

ear

to

form

related

sets.

Organizing

the

list

pro

vides

ab

asis

for

discussion

with

domain

exp

erts,

who

help

\debug"

the

prop

osed

organization.

3.

Rening

the

structure

of

the

list

through

kno

wledge

acquisition

sessions

with

domain

exp

erts.

During

these

sessions,

signican

tranges

of

test

results

are

iden

tied,

as

are

inferences

from

these

results

that

eliminate

the

need

to

perform

other

tests.

That

is,

ad

ep

endency

graph

based

on

test

results

is

dev

elop

ed.

4.

Iterating

the

ab

ovet

wo

steps

after

consulting

additional

information

suc

ha

sm

anuals

and

cod

ed

ocumen

tation.

The

structure

of

the

domain

becomes

clearer

at

eac

hiteration.

(W

eh

ave

found

that

three

iterations

are

sucien

tto

pro

duce

auseful

structuring.)

The

nal

validation

mo

del

consists

primarily

of

en

tries

corresp

onding

directly

to

information

that

app

ears

in

the

original

data

base.

5.

Integrating

the

test

sets

into

the

structure

deriv

ed

in

the

previous

step.

This

integration

mak

es

explicit

the

prerequisites

of

eac

htest,

as

well

as

pro

viding

alternativ

ew

ays

of

obtaining

information

ordinarily

pro

vided

by

aparticular

critical

test

in

cases

where

that

test

cannot

be

performed.

4

An

Extended

Example

In

order

to

understand

the

validated

retriev

al

pro

cess,

consider

the

follo

wing

example.

Our

domain

is

automobile

diagnosis

and

repair,

and

we

assume

an

existing

case

base

with

its

asso

ciated

validation

mo

del.

W

eare

giv

en

the

follo

wing

case:

NEW

CASE

mak

e

:

MAZD

A

mo

del

:

626

mo

del

year

:

1985

engine

typ

e

:

2.0L

EFI

miles

:

50,000

problem

:

engine

do

es

not

start.

The

retriev

al

phase

uses

the

mak

e,

mo

del,

problem,

and

appro

ximate

year

of

man

ufacture

to

searc

hthrough

acase

base

of

previous

automobile

problems.

Based

on

these

surface

features,

we

retriev

ethree

cases

to

be

validated

before

presen

tation

to

areasoning

comp

onen

(9)

7

CASE

1

mak

e

:

MAZD

A

mo

del

:

626

mo

del

year

:

1988

engine

typ

e

:

2.0L

EFI

miles

:

10,000

problem

:

engine

do

es

not

start.

validation

:

The

fuel

injector

was

clogged.

Fuel

was

not

deliv

ered

to

the

com

bustion

cham

ber

for

the

engine

to

ignite.

For

this

reason

the

engine

could

not

start.

solution

:

cleaned

the

fuel

injector.

CASE

2

mak

e

:

MAZD

A

mo

del

:

626

mo

del

year

:

1984

engine

typ

e

:

2.0L

miles

:

60,000

problem

:

engine

do

es

not

start.

validation

:

The

car

had

afault

y

gas

pump.

Fuel

could

not

be

de-liv

ered

to

the

com

bustion

cham

ber.

For

this

reason

the

engine

could

not

start.

solution

:

Replaced

the

gas

pump.

CASE

3

mak

e

:

MAZD

A

mo

del

:

626

mo

del

year

:

1987

engine

typ

e

:

1.8L

miles

:

20,000

problem

:

engine

do

es

not

start.

validation

:

A

leak

existed

in

the

gas

line.

Fuel

could

not

be

deliv

ered

through

the

fuel

line.

For

this

reason

the

engine

could

not

start.

solution

:

Fixed

the

leak.

The

validation

mo

del

con

tains

(at

least)

the

three

tests

that

are

referenced

by

these

cases:

\chec

kif

afuel

injector

isc

logged",

\chec

kif

the

gas

pump

is

working",

and

\chec

kif

there

is

aleak

in

the

fuel

line".

The

rst

of

these

is

actually

comp

osed

of

two

simpler

tests:

atest

for

fuel

presen

tin

the

reserv

oir

of

the

injector

and

atest

for

fuel

exiting

the

injector's

nozzle.

Ifthere

is

no

fuel

in

the

injector

then

we

can

deduce

that

the

injector

is

not

at

fault.

Rather,

the

problem

lies

earlier

in

the

fuel

system

|

either

in

the

pump

or

the

fuel

line.

The

system

rst

attempts

to

validate

Case

1b

yrep

eating

the

validation

steps

from

that

case.

That

is,

we

wish

to

test

if

the

fuel

injector

is

clogged.

In

the

pro

cess

ofp

erforming

this

two-step

test

we

actually

acquire

kno

wledge

that

is

relev

an

tto

Cases

2a

nd

3:

if

the

fuel

reserv

oir

is

not

empt

yw

ec

an

eliminate

both

cases

ifit

is

empt

y,w

ecan

eliminate

Case

1.

This

(10)

8

5

RECENT

RESUL

TS

is

enco

ded

in

the

seman

tic

net

work

that

represen

ts

our

validation

mo

del

and

is

used

in

the

validation

phase.

In

the

best

case,

this

validation

mo

del

allo

ws

us

to

reduce

the

work

required

to

validate

cases

from

four

tests

to

two

tests

and

sim

ultaneously

reduces

the

num

ber

ofcases

to

be

considered

by

the

reasoner

from

three

to

one

(selectivit

y

of

33.3%).

The

rst

test

is

for

an

empt

yfuel

reserv

oir

if

the

reserv

oir

isfull

then

Cases

2and

3are

eliminated.

W

ethen

test

the

nozzle

for

fuel

exiting.

If

no

fuel

lea

ves

the

nozzle,

then

Case

1is

presen

ted

to

the

reasoner

but

if

fuel

is

lea

ving

the

nozzle

we,

unfortunately

,eliminate

Case

1as

well

and

lea

ve

the

reasoning

comp

onen

tto

its

own

resources.

The

worst

case

requires

all

four

tests

and

pro

vides

either

zero

or

one

case

to

the

reasoner.

5

Recen

tR

esults

5.1

An

Op

erating

System:

VMS

The

rst

system

we

dev

elop

ed

isused

for

the

diagnosis

ofd

evice

driv

er

induced

crashes

ofDigital's

VMS

op

erating

system.

The

kno

wledge

ab

out

surface

fea-tures

was

obtained

primarily

from

DEC

internal

publications

and

was

com-plemen

ted

by

an

exp

ert

from

the

VMS

supp

ort

team

during

three

kno

wledge

acquisition

sessions.

It

took

atotal

of

84

hours

to

acquire

the

domain

sp

e-cic

kno

wledge

about

surface

features.

Based

on

this

information,

the

domain

kno

wledge

used

by

unimem

in

order

to

organize

the

cases

into

ageneralization

hierarc

hyw

as

impleme

nted

in

v

ed

ays.

It

took

an

additional

four

days

of

reading

validation

pro

cedures

in

the

data

base

to

dev

elop

avalidation

mo

del

for

device

driv

ers.

In

addition,

four

more

kno

wledge

acquisition

sessions,

lasting

40

hours,

were

needed

to

rene

and

impro

vet

he

validation

mo

del.

Enco

ding

the

actual

validation

mo

del

too

k

ab

out

80

additional

days.

The

total

num

be

rof

days

sp

en

to

n

kno

wledge

acquisition

and

dev

elopmen

ti

ssh

own

belo

w:

Activity

Person

Days

Kno

wledge

acquisition

20

Dev

elopmen

t

85

Since

this

was

our

rst

attempt

to

build

acase

base

and

validation

mo

del,

these

num

bers

are

muc

hlarger

than

we

exp

ect

for

subsequen

tsystems.

(11)

5.2

AW

ord

Pro

cessing

System:

WPS-PLUS

9

work

to

date

on

the

system

describ

ed

in

Section

5.2

app

ears

to

conrm

this

exp

ectation.

The

system

was

evaluated

using

acase

base

of

200

cases

that

were

obtained

from

notes

written

by

specialists.

The

surface

feature

retriev

al

phase

of

the

system

was

evaluated

by

presen

ting

eac

ho

fthe

200

cases

to

the

retriev

er

(as

new

problems)

and

preparing

ah

istogram

of

the

num

ber

of

cases

retriev

ed.

unimem

pro

vides

am

echanism,

kno

wn

as

retrieval

weights

,for

tuning

its

re-triev

al

capabilities.

After

some

exp

erimen

tation,

wed

isco

vered

that

the

use

of

larger

retriev

al

weigh

ts

(i.e.

more

stringen

tm

atching

criteria)

caused

the

re-triev

er

to

miss

man

yrelev

an

tcases

and,

in

man

yoccasions,

to

fail

to

retriev

e

an

ycases

at

all.

With

less

stringen

tcriteria

this

problem

was

rectied.

Ho

w-ev

er,

man

yof

the

retriev

ed

cases

were

not

relev

an

tto

the

problem.

W

ith

the

optimal

weigh

ting,

wew

ere

able

to

retriev

eo

n

average

22

cases

per

retriev

al

(11%).

The

validation

phase,

ho

wev

er,

was

able

to

reduce

this

num

ber

ofc

ases

to

an

average

of

4.5

cases

out

of

200

(2.25%).

In

addition,

we

presen

ted

three

new

cases

to

the

system.

Based

on

surface

features

alone,

we

retriev

ed

20,

25,

and

16

cases

(10,

12.5,

and

8%

selectivit

y).

The

validation

phase

reduced

this

to

3,

5a

nd

3cases,

resp

ectiv

ely

(1.5,

2.5,

and

1.5%

selectivit

y).

Our

exp

erts

conrm

that

these

validated

cases

are

the

only

ones

relev

an

tto

the

problems

presen

ted.

5.2

AW

ord

Pro

cessing

System:

W

PS-PLUS

The

second

system

performs

diagnosis

of

customer

problems

with

the

word

pro

cessing

comp

onen

tof

an

oce

automation

pro

duct.

During

15

hours

of

kno

wledge

acquisition

sessions,

the

kno

wledge

ab

out

surface

features

was

ob-tained

from

asupp

ort

engineer

for

the

pro

duct.

It

then

too

k

an

additional

v

ed

ays

to

enco

de

the

domain

kno

wledge

for

use

by

unimem

.

The

validation

mo

del

was

obtained

from

the

validation

pro

cedures

of

the

cases

in

the

data

base,

an

internal

publication,

and

10

hours

of

kno

wledge

acquisition

with

the

same

engineer.

While

the

work

is

not

yet

complete

(only

50

out

of

340

cases

ha

ve

been

enco

ded),

ith

as

tak

en

only

10

da

ys

to

implem

ent

the

validation

mo

del.

This

system

is

still

under

dev

elopmen

t.

Ho

wev

er,

the

time

wes

pen

to

n

kno

wledge

acquisition

and

dev

elopmen

tis

sho

wn

belo

(12)

10

6

SCOPE

OF

W

ORK

Activity

Person

Days

Kno

wledge

acquisition

5

Dev

elopmen

t

10

This

system

was

evaluated

using

acase

base

of

340

cases.

Rep

eating

the

same

exp

erimen

tperformed

with

the

VMS

case

base

led

to

an

average

of

26

cases

per

retriev

al,

or

7.6%

selectivit

y.T

he

validation

phase

reduced

this

to

two

cases,

or

0.58%

selectivit

y.

Since

the

validation

mo

del

for

this

case

base

is

not

yet

fully

enco

ded,

weh

ave

not

presen

ted

new

problems

to

the

system.

6

Scop

eof

W

ork

Our

validated

retriev

al

metho

dcan

be

applied

in

man

yt

yp

es

of

tasks.

The

basic

requiremen

ts

are:

an

existing

data

base

of

previous

practical

exp

eri-ence

aset

of

quic

ktests

that

serv

eto

reduce

the

searc

hspace

at

low

cost

a

set

of

more

exp

ensiv

etests

that

can

further

reduce

the

searc

hspace

and

an

understanding

of

the

relationships

bet

ween

the

exp

ensiv

etests.

W

eh

ave

iden

tied

four

areas

of

poten

tial

interest,

but

weh

ave

limited

our

implem

en

tation

work

to

the

rst

of

these:

Diagnostic

tasks

.A

ssh

own

in

the

example,

we

use

the

symptoms

of

ap

roblem

as

the

surface

features

for

the

retriev

al

phase.

The

validation

pro

cedure

describ

es

whic

htests

to

perform

in

order

to

determine

ifthe

case

is

relev

an

tto

the

new

problem.

Design

tasks

.The

surface

features

are

sp

ecications

that

adesign

must

satisfy

.The

validation

pro

cedures

verify

that

ap

rop

osed

design

meets

the

specication.

Sales

tasks

.The

tec

hnique

can

be

used

to

help

iden

tify

sales

prosp

ects

for

a

new

pro

duct.

The

surface

features

are

the

characteristics

of

a

customer

suc

has:

size

of

business,

typ

eof

business,

location

of

business.

Eac

hv

alidation

pro

cedure

describ

es

the

customer's

requiremen

ts

that

were

satised

in

a

previous

sale.

The

validation

mo

del

includes

tests

that

determine

whether

or

not

a

customer

needs

a

particular

typ

eo

f

pro

(13)

11

Managemen

ttasks

.T

hese

tasks

include

accoun

ting,

credit

analysis,

investmen

tdecisions,

and

insurance

underwriting.

In

eac

ho

fthese

ar-eas,

sp

ecialists

can

iden

tify

easily

recognized

features

in

their

problem

domain

(typ

eof

compan

y,size,

etc.)

that

allo

w

rapid

retriev

al

of

similar

situations

encoun

tered

in

the

past.

They

then

ha

ve

more

detailed

tests

that

can

be

applied

(debt/equit

yr

atio

,pa

ymen

thistory

,t

ypeo

fc

lien

t,

balance

sheets,

etc.).

7

Conclusions

Our

work

has

concen

trated

exclusiv

ely

on

the

issue

of

case

retriev

al.

A

care-ful

study

of

two

applications

in

whic

h

people

consciously

use

case

retriev

al

has

sho

wn

that

retriev

al

based

solely

on

surface

features

is

not

sucien

tly

discriminating

for

use

with

large

case

bases.

It

results

in

large

num

bers

of

cases

returned

to

the

reasoner,

eac

ho

fwhic

hm

ust

then

be

further

examined

at

great

exp

ense.

Adding

avalidation

phase

that

uses

kno

wledge

of

domain-sp

ecic

tests

to

prune

the

retriev

ed

cases

dramatically

reduces

the

num

be

rof

cases

that

must

be

examined

by

the

reasoner.

W

eh

ave

found

that

acquiring

kno

wledge

ab

out

domain-sp

ecic

tests

is

aided

by

an

initial

perusal

ofthe

existing

data

base

used

by

spe

cialists.

With

areasonable

amoun

tof

eort,

and

with

only

asmall

investmen

tof

special-ists'

time,

this

information

can

be

captured

in

avalidation

mo

del

represen

ted

as

aseman

tic

net

work.

W

eh

ave

used

this

metho

dology

to

pro

duce

twos

ys-tems.

One

of

these

systems

has

been

successful

in

practice,

and

the

other

(incomplete)

system

is

lik

ely

to

be

equally

useful.

While

the

burden

of

kno

wledge

acquisition

in

our

metho

dology

is

small

compared

with

other

metho

ds,

itis

not

negligible.

Automating

this

work

by

com

bining

anatural

language

system

to

analyze

existing

data

bases

with

AI-assisted

statistical

comparison

of

surface

features

pro

vides

a

fertile

area

for

further

investigation.

Furthermore,

we

susp

ect

that

acareful

study

of

suc

h

asystem

in

practice

will

rev

eal

validation

tests

that

are

sucien

tly

common

that

it

ma

yb

ereasonable

to

promote

them

to

surface

features,

leading

to

a

system

with

better

retriev

al

capabilities.

This

analysis,

whic

hm

ust

rst

be

performed

man

ually

to

validate

our

assumption,

isitself

an

excellen

tarea

for

the

application

AI

metho

ds.

Ac

kno

wledgemen

(14)

12

7

CONCLUSIONS

The

authors

wish

to

thank

Prof.

David

W

altz

for

his

help

with

this

researc

h.

In

addition,

weh

ave

receiv

ed

helpful

commen

ts

from

Mark

Adler,

Andrew

Blac

k,

Da

ve

Hanssen,

Rose

Horner,

and

Candy

Sidner.

W

eare

also

indebted

to

the

engineers

who

ha

veg

iven

their

time

and

kno

wledge

to

help

us

understand

the

two

domains

weh

ave

studied:

the

Digital

Supp

ort

Engineers

at

Colorado

Springs,

Colorado

and

Spitbro

ok,

New

(15)

REFERENCES

13

References

1]

Ra

yB

areiss,

Karl

Bran

ting,

and

Bruce

Porter.

The

role

ofe

xplanation

in

exemplar-based

classication

and

learning.

In

Pr

ocee

dings

of

Case-Base

d

Reasoning

Workshop

,1

988

.

2]

Kristian

Hammond.

Case-Base

dP

lanning:

AnI

ntegr

ate

dThe

ory

ofP

lan-ning,

Learning,

and

Memory

.P

hD

thesis,

Yale

Univ

ersit

y,1

986

.

3]

Janet

L.Kolo

dner.

Reconstructiv

em

emory:

A

computer

mo

del.

Co

gnitive

Scienc

eJournal

,7:281{328,

1983.

4]

Janet

L.K

olo

dner,

Jr.

Rob

ert

L.

Simpson,

and

Katia

Sycara-Cyranski.

A

pro

cess

mo

del

ofcase-based

reasoning

in

problem

solving.

In

Pr

ocee

dings

of

the

International

Joint

Confer

enc

eo

nA

rticial

Intel

ligenc

e

,1

985

.

5]

Ph

yllis

Koton.

Using

exp

erienc

ei

n

learning

and

pr

oblem

solving

.

PhD

thesis,

Massac

hussetts

Institute

of

Tec

hnology

,1988.

6]

Da

vid

Leak

e.

Ev

aluating

explanations.

In

Pr

ocee

dings

of

the

Seventh

National

Confer

enc

eo

nA

rticial

Intel

ligenc

e

,1

988

.

7]

Mic

hael

Leb

owitz.

Exp

erimen

ts

with

incremen

tal

concept

formation:

Unimem.

Machine

Learning

,2:103{138,

1987.

8]

M.R.

Quillian.

Seman

tic

memory

.In

Marvin

Minsky

,editor,

Semantic

Information

Pro

cessing

,pages

227{270.

MIT

Press,

1968.

9]

Edwina

Rissland

and

Kenneth

Ashley

.H

yp

otheticals

as

heuristic

device.

In

Pr

ocee

dings

of

the

Fifth

National

Confer

enc

eo

nA

rticial

Intel

ligenc

e

,

1986.

10]

Craig

Stanll

and

Da

vid

W

altz.

Tow

ard

memory-based

reasoning.

CA

CM

,

29:1213{122

(16)

14

A

CREA

TING

THE

VMS

VALID

ATION

MODEL

A

Creating

the

VMS

Validation

Mo

del

This

app

endix

describ

es,

in

detail,

the

metho

dused

to

construct

the

valida-tion

mo

del

for

the

VAX/VMS

device

driv

er

diagnosis

system.

Eac

hsubsection

corresp

onds

to

one

of

the

steps

describ

ed

in

Section

3.2.

For

exp

ository

pur-poses,

of

course,

weh

ave

remo

ved

an

um

be

rof

intermediate

steps

and

sho

w

only

selected

results.

A.1

Reading

the

Initial

Data

Base

W

eb

egin

with

the

data

base

in

curren

tuse

by

Customer

Supp

ort

Sp

ecialists.

This

data

base

(actually

,an

informal

textual

\notes

le")

consists

of

ab

out

150

en

tries.

A

typical

en

try

is

sho

wn

in

Figure

1.

At

this

stage

we

not

only

read

these

en

tries,

but

we

tidy

them

up

a

bit.

For

example,

some

of

these

en

tries

were

incomplete

and

were

man

ually

expanded

into

multiple

cases

for

our

case

base,

resulting

in

200

entries

to

be

studied.

The

primary

purp

ose

of

this

initial

reading

isto

create

alist

ofall

of

the

tests

that

are

explicitly

men

tioned

by

the

sp

ecialists

in

their

explanations

of

the

resolv

ed

cases.

From

the

case

presen

ted

in

Figure

1w

eextract

three

tests:

1.

Retriev

eP

rocess

PCB

2.

Chec

kJ

IB

Address

3.

Chec

kCoun

t

For

this

particular

data

base,

we

deriv

ed

an

initial

set

of

ab

out

100

tests.

A.2

Grouping

the

Tests

Reading

through

this

data

base

led

to

anatural

decomp

osition

of

the

tests

by

grouping

them

according

to

data

structures

men

tioned

in

the

cases.

Th

us,

we

group

ed

tests

related

to

the

JIB

(job

information

blo

ck)

together,

joining

the

tests

\Chec

k

JIB

Address"

and

\Chec

kC

oun

t"

from

the

case

sho

wn

in

Figure

1.

In

this

particular

data

base,

we

group

ed

the

tests

into

roughly

sev

en

sets

of

tests.

By

con

trast,

ino

ur

work

with

the

WPS-PLUS

data

base,

the

natural

group-ing

was

along

functional

lines.

While

weh

ave

no

rm

evidence,

it

seems

lik

ely

that

suc

h\natural"

groupings

will

occur

in

most

sizable

data

(17)

A.2

Grouping

the

Tests

15

Version

ofs

ystem:

VAX/VMS

V4.2

Reason

for

bugc

hec

kexception:

ACCESS

VIOLA

TION

Pro

cess

curren

tly

executing:

SP=>

8018E7AC

00000004

8018E7B0

7FFE7DE4

8018E7B4

FFFFFFFD

8018E7B8

0000026A

8018E7BC

00000000

!THIS

S

HOULD

H

AVE

B

EEN

O

N

PREVIOUS

LINE

!AND

VISA

VERSA

8018E7C0

00000001

8018E7C4

00000005

!BEGINNIN

GO

F

S

I

G

N

A

LA

R

R

A

Y

8018E7C8

0000000C

!ACCES

VIOLATION

8018E7CC

00000004

!WRITE

ACCESS

PROTECTIO

N

8018E7D0

00000020

!VA

8018E7D4

80004A4E

IOC$BUFPO

ST+

017

8018E7D8

04040000

8018E7DC

80004A6D

IOC$BUFPO

ST+

036

8018E7E0

800CD108

DZDRIVER+

118

the

reason

for

this

bugc

hec

kis

because

the

pid

eld

ofthe

IRP

has

been

zero

ed.

IN

post

pro

cessing

the

system

tries

to

giv

ethe

buered

byte

coun

tb

ackt

o

the

pro

cess

that

did

the

io.

HE

doe

sthis

by

using

the

PID

eld

oft

he

irp

to

index

into

the

pid

table

then

gets

the

PCB(n

ull

pro

cess

in

this

case)

and

gets

the

JIB

address.

THE

JIB

address

for

the

null

pro

cess

is0

.T

HE

system

then

go

es

to

the

JIB$L

BYTCNT

oset

(20)

from

the

base

of

the

JIB

(0

in

this

case)

and

tries

to

add

bac

ki

nt

he

byte

coun

t.

IN

this

case

we

get

an

access

violation.

ha

ve

resp

onse

from

VMS

dev

elop

emen

tthat

there

is/w

as

ab

ug

xed

in

4.2

it

has

been

men

tioned

to

try

and

low

er

the

baud

rate

on

terminals

as

a

work

around.

Figure

1:

An

Initial

Data

Base

En

(18)

16

A

CREA

TING

THE

VMS

VALID

ATION

MODEL

A.3

Rening

the

Test-Group

Structure

Armed

with

this

group

oftests,

we

returned

tot

he

sp

ecialists

who

had

prepared

the

data

base.

With

their

help,

we

learned

the

meaning

ofeac

hof

the

tests

we

had

iden

tied.

The

sp

ecialists

poin

ted

out

that

our

natural

groupings

didn't

corresp

ond

to

groupings

that

they

would

ha

ve

made

|

primarily

because

they

had

a(non-articulated)

set

of

abstractions

that

our

groupings

did

not

mak

eclear.

By

reviewing

our

groupings

they

were

forced

to

verbalize

the

abstractions

that

they

took

for

gran

ted,

and

wew

ere

able

to

sort

our

tests

into

15

groupings

that

reected

the

specialists'

notion

of

the

correct

abstractions

for

their

domain.

In

the

case

of

the

VAX/VMS

data

base,

the

specialists

typically

sub

divided

our

groupings

into

three

distinct

categories:

those

that

test

for

the

existence

of

the

exp

ected

location

in

memory

,those

that

tested

the

internal

consistency

of

the

data

structure,

and

those

that

prob

ethe

data

structures

for

spe

cic

value.

In

the

case

ofthe

tests

extracted

from

Figure

1,

\Chec

kJIB

Address"

is

in

the

rst

category

,the

sp

ecialists

poin

ted

out

for

the

need

for

an

additional

test

we

had

not

iden

tied

that

performed

the

consistency

test,

and

\Chec

kC

oun

t"

is

in

the

third

category

.

Th

us,

our

initial

mo

del

of

sev

en

structure-based

test

groups

was

rened,

through

interaction

with

specialists,

into

15

groups

with

ala

yered

structure.

Testing

alw

ays

pro

ceeds

from

one

layer

to

another,

starting

with

tests

for

exis-tence

ofthe

data

structure,

pro

ceeding

to

tests

on

the

consistency

ofthe

data

structure,

and

nally

on

to

probing

specic

values

within

the

data

structure.

A.4

Acquiring

Additional

Kno

wledge

This

step

is

primarily

an

iteration

of

the

earlier

one,

but

twou

nusual

items

arose

during

this

phase

for

the

VAX/VMS

system.

The

rst

was

the

realiza-tion

that

certain

terms

used

in

the

descriptions

were

not

reected

in

anyo

f

the

tests

our

groupings

we

had

already

deriv

ed.

In

this

particular

case,

there

was

con

tin

ual

use

of

phrases

lik

e\during

AST

deliv

ery

,"

and

\during

post-pro

cessing."

Examination

of

additional

material

(material

from

an

internal

DEC

course

suggested

by

the

sp

ecialists

as

a

goo

d

starting

poin

t)

rev

ealed

that

these

reected

the

pro

cessing

stages

of

the

system

wew

ere

diagnosing.

Our

initial

breakdo

wn

had

been

along

data

structure

boundaries,

but

an

or-thogonal,

equally

imp

ortan

tstructuring

is

possible

along

temp

oral

lines

based

on

these

stages.

It

is

these

two

alternativ

edecomp

ositions

that

we

referred

(19)

A.5

Integrating

Tests

into

the

Mo

del

17

in

Section

3.2

as

the

\structure

oft

he

domain."

Further

iteration

with

the

course

materials

and

the

specialists

rev

eals

a

muc

hric

her

structure,

consisting

of

multiple

lev

els

of

abstraction.

With

the

temp

oral

decomp

osition

tak

en

as

primary

,F

igures

2and

3sho

wt

wo

dieren

t

lev

els

of

abstraction.

Figure

2is

tak

en

at

avery

abstract

lev

el,

sho

wing

the

four

ma

jor

pro

cessing

stages

and

the

data

structures

used

at

eac

hstage.

Figure

3

is

a\close-up"

of

the

single

pro

cessing

stage

kno

wn

as

\AST

deliv

ery

."

A.5

Integrating

Tests

into

the

M

odel

Armed

with

am

uc

h-im

pro

ved

mo

del

of

the

domain,

we

built

aseman

tic

net-work

capturing

as

muc

hof

this

mo

del

as

bears

directly

on

the

abilit

yto

select

and

apply

the

tests

and

test

groups.

In

the

VAX/VMS

system,

our

validation

mo

del

reects

both

decomp

ositions

along

data

structures

and

along

pro

cess-ing

stages.

The

seman

tic

net

work

has

15

nod

es,

for

eac

htest

group,

77

no

des

that

represen

tk

nowledge

ab

out

the

pro

cessing

stages

of

device

driv

ers,

and

22

no

des

that

represen

tk

nowledge

ab

out

data

(20)

18

A

CREA

TING

THE

VMS

VALID

ATION

MODEL

LEGEND:

data structure

test

test groups

IRP

UCB

PCB

valid

PCB?

processing

device

driver

postprocessing

AST Delivery

preprocessing

starting

address

starting

address

valid

IRP?

starting

address

valid

UCB?

Figure

2:

Device

Driv

er

Validation

Mo

del:

Abstract

(21)

A.5

Integrating

Tests

into

the

Mo

del

19

AST Delivery

Break Internal PID

retrieve process PCB

null process PCB?

PCB$PID PCB Vector

first ACB ACB$ASTQBL JIB

PCB$ASTQBL PCB$JIB JIB$BYTCNT Check ACB

type

get ASTQBL Check JIB Valid

Return byte count Get JIB

address

JSB to AST routine REMQ AST

Enqueue ACB in AST queue

Table

Check

address JIB? byte count

Figure

3:

Device

Driv

er

Validation

Mo

del:

AST

Deliv

References

Related documents

Using a spatial working memory task, Azuma and colleagues found lower activation in a parietal region in patients with 22q11 deletion syndrome compared with healthy controls,

17% of MIT undergraduate students patent an invention within 15 years of graduation; On average, MIT alumni inventors produce six patents each. MIT engineers are 5x

clinical faculty, the authors designed and implemented a Clinical Nurse Educator Academy to prepare experienced clinicians for new roles as part-time or full-time clinical

Vital Museum/Fire Hall 600

We support the idea of integrated STEM education in a Turkish context in ways that students spend efforts to solve a real-world problem, which requires content knowledge and skills

matrices of the multivariate time series data of solar events as adjacency matrices of labeled graphs, and applying thresholds on edge weights can model the solar flare

The tense morphology is interpreted as temporal anteriority: the eventuality described in the antecedent is localised in the past with respect to the utterance time.. Compare this

Using the same sample of students, the second empirical study investigates: (1) the relative contribution of vocabulary knowledge, grammatical knowledge, and idea- generating