STAHLp: Belief Revision in Scientific Discovery

From: AAAI-86 Proceedings. Copyright ©1986, AAAI (www.aaai.org). All rights reserved.
STAHLp:
Belief
1)epartmenI
Revision
of Inforrr~aLiori
IJniversity
RrpaNet:
in Scientific
and
of California,
Abstract
this
ring
paper
we describe
components
of chemical
ing cornponential
models.
S’I’AI-11, system
cal
reactions
and
known
rrlodels.
I lowever,
effective
strategy
covrbri rig from
upon
the
cacti
illfclrred
belief
tiple
methods
inferences
which
can
proc&
beliefs,
for detecting
been
comptating
revision
how
ttieories
issue
how
evolve
with
STAlI
recovering
premises,
they
after
frorn
method
are
them,
and
from
revisions.
the
further
belief
duction
any
which
have
in these
(Xytkow
systtlrn
undergone
areas
has
which
inferred
els are
from
ptllo-
f’rorll both
constructed
which
ciuri~~g
ory.
‘I’his theory
ing su Ipstances
swiii
lose
to
thclory
also
oxidizatiorl),
lost
prevailing
evolved
reduce
which
phlogiston
and
(2)
was
18th
f’rom
the
models.
to explain
believed
transformed
beliefs
its
periodically
/ SCIENCE
to
through
would
have
to
be
ence.
Next,
(4) the
work
then
carried
out.
I, sufficient
prt~riiises
Icad
n~odels,
then
of newly
to the system.
Then
(I)
based
on these
is detected.
belief
revision
proposing
ways
to
“best”
avoid
proposed
l~‘inally,
begins
as (3) hy-
in which
premises
erroneous
to iritc~rrricdiate
the
assigning
least
blame
modifications
(step
is itself
reactions,
to triort’ irltornlediate
infer-
having
again,
rllodcls.
a cycle
then
reactions,
Step
in which
to
and
to
are
I) starts
of more
are noted,
until
illfererlc-
premise
inferencing
prelnises
the
is chosen,
when
new mod-
Normal
hypothesis
models
if no tbrrorb
reactions
consists
begins
Ie&id ing to the construction
(possibly)
century
cycle
rllodified
its
pro-
of chemical
inference
and
generated,
18th
its output
inference
suspt~nded
are
models
of
and
inferred
a11 erroneous
tlre-
inferred
so
011.
wheu
S’L’A II I,, the
kinds
beliefs
i3ot
h
sys-
thus
represent
The
puts; a iriodel is represented
as a list contaiuing
ttle substumx being modelled,
and its cornyonents. tj’or example,
as
coniponential
S’I’A 11Lp deals
tenis
known
and
of
burn-
and
reactions
f,W(J
are
process.
associatoti
Like
with
a reaction
18th century
chemists
observed
“citlx”.
COiLI reacted
to
iron
this
form
rtbactiotl
as
models.
as a list
a Ineta
resent
528
consists
are input
premises;
system
that
(now
to occur
into
importantly
evolve
chemistry,
combustion
calcination
more
of STAHLp
S’l‘AllLp’s
certain
techniques
observation
iri the
its
S’I‘A t I Lp in part
theories
to construct
input
of Al
was phloyistm
framework
models,
tiowt>ver,
S’I’A tl I,p is a forward-chaining
models,
on existing
of chemical
century
but
scientific
designed
impact
production
combining
(phlogiston)
was
yet
S’I’AtIl,
models
towards
in size durirlg
something
seemed
The
cotnponential
step
Its domain
the
investigation,
a forward-chaining
1986)
areas.
are two areas
overlapped.
was a first
511 tJst;tIlcrJs
revision
considerable
rarely
& Simon,
belief
collft icts between
we created
Overview
known
ing is then
and
to
be-
revision.
Its
su bstar1ces.
and
We analyze
Introduction
discovery
in-
111 addition
inferences.
in scope;
for
and
S’l’A tll, employed
erronwus
how
system
potheses
Scitarltific
domain,
S’I’A I 11,‘s deficielbcies,
II
prernise
theory.
I
limited
investigate
determines
shift
were
explanations
chemists,
phenomena.
to resolve
certain
I,ike its predecessor,
Lp,
as interpretations
involving
techniques
to
made
repeated
this
from
erroneous
thus
within
lief revision
1,‘s mul-
hypotheses
chosen
both
rational
frustrated
long
relate
and
recover
provided
had
1984)
lead
in STAH
revise
The
models.
an example
to oxygeii
or model).
can be viewed
theories;
this
which
to
models
repeated
is based
(de Kleer,
to one
faulty
new
strategy
method
and
reduced
hypothesize
belief
giston
This
re-
which
seemed
to remedy
unified
and
or premises,
reaction
(i.e.,
a more
deed
methods
to construct
detecting,
inferences.
to construct
during
in order
preventing,
of the
use chemi-
theory
problems
inferring
construct-
both
S’I’A 11Lp employs
source
have
- i.e.,
1986);
the assumption-based
for infer-
is a descendant
models
for
errorleous
of rccordillg
system
substances
& Simon,
any
new
STAlILp
92717
phlogiston
‘I’hus,
S’l‘AIil,p
(Xytkow
and
partly
the
Science
lar~gl~~y~I~~S.lJ~~I.~~~I~IJ
two
III
Computer
Irvine
drose~lC:S.IJ(II.I~:1)1J,
Discovery
arid
(reacts
of its
that
inputs arld
calx-of-iron’
ash;
S’I’AtII,p
inputs
uut-
and charWould
{calx-of
rep-iron
charcoal) outputs {iron ash)). We abbreviate
this to
CI Ch + I Ash. lf S’l’A~il,p eventually
infers that charcoal
is composc~d
of p tilogistoll
this
c~ornponential
rrlodel
and
ash,
S’l’All I,p
as (components
of
rcprest~~~ts
(charcoal}
are {phlogiston ash}), or Ch = Ph Ash. We can
of tflt3t:
at~OVt~
I’ll
beliefs
(‘an
t
usirig
kJC? ViPWd
its
At this
ASll.”
Ilc’Irts of (.harcoat
I
I 1’11
algebraic
“(:I
}
point,
into
1 t Ash”,
nstl
“(:I
an
f hen
(:/I
wc
the
first
S’l‘AII I, alIt
c~orrt~spond
S’1‘AlII.p.
of
u
basic
S indicate
belie/,
ther1
has
S’I‘A II I,p ilifers
only
sucfi
as
I
irlg rrt>wly inferred
tul ion can
occur
rt>ti uctions,
and
S’1‘A II Lp’s
basic
rthactions
Its
tras
rule
was never
ash
C~LII
At this
reduced
list,
into
tll, would
S’l‘A
would
stitution,
which
is again
the
leading
) I Ash
H EI)l.J(?I~:
to it.
IIow<~vt~r,
t I {Ash}.
fire, and I
in CI Ph
t’ON tI:N’I’S would
cannot
reducctl
curlier
which
lists,
(dcsc.t~lldent.)
If A
hence
rule ensures
thus
of VA. In short,,
III
%y tkow
The
and
l3elief Revision
Simon
t:rroIic’ous
illfererlctts:
rule;
“error
arid
prc’triist:s.
tililig
iri
these
that
faulty
in
tile
S’l’All t,p’s
STAllI,
input
lo
type
tyl,t:,
IJ~II
all (erroneous
t)y f’itulty
I’auIIy
havtb
premises.
I)rtl1Iii bt5 are
the
typos
irlfererlccs
tri
t,his rr~c:tt~ocf f;)r recoverirq
there
Since
(: we will discuss
again
from
using
erronf2ous
the
faulty
its
use
lj we will
1~~3 viewed
a simpler
11k:l)11(~14~
is only
in S’I‘A I 11,~ must
section
revised,
call
allowing
irLjerences.
arid
show
one error
be caused
how
rc4uct4
in later
DELAYED-
and A ulso occurs
belief.
in
For any
previously
reduced
rtlrnoved
S’rAIIf,p
befrom
to construct
the
reduced
list enables
wmpwe~lts
of’
VC)
S’l’AIILp
incorrect,
models
to
like C
as new information
so011
becomes
known.
of
key role iI1 han-
In section
error
prevented,
i.e.
A we will
in S’I’A t 114p, thus
~rit~tl~od for detectr~ly erroneous
rutt: errors
S’l’.AtlL”
erro7~ewus I‘f~jer(:ticf3.
error
two sources
of thck ltl~~l)~JCl~~
list plays
111 see-1 iou
tlow S’L‘AI] 1,‘s ttircc~ rirairi
ah a sirlgtc
has
applic-ations
rr:cluccd
problerr15.
prei~erltiny
show
noted
Process
cor-
its reduced
immediately
Ibrevc’lll t’rroneous
iufertlnc~cs
(e.g.
= Ph VA CC) by rc‘vising beliefs as
tthe
the
et~lcmbers
rr~odel C = Ph CC, by removing the new occur-
tile correct
(e.g.
I
reappear
the
enables
and
occurrence
it
rule,
substances
arc
cornpo-
infer
can;
new
A from
that
this
through
in u belief,
beliefs
reduction
a
tJ1e
it
been
reaction,
cannot
if they
using
he11
remove
reduced
rerrwue
11’s ancestral
sub-
because
model
above
S’SA tll,p
been
occurs
therl
was
VA CC after
w
the
substances
by
lief t$, this
rence
IrlerIlOry.
have
such
list,
fj. I)ttlayed
and
llowever,
rts reduced
f‘rorrl
chin.
into
reaction
reactions
REDIJ~1E:
if VC’s
model
what bus ulreudy
S’I‘A 11I, caltriot
removes
copper
infer the correct
Yrerr~emtxr”
substituted
substances
and
1 VA Ph VA
Llowever,
C = Ph
as
because
incorrect.
to
appears,
is an-
present
resulted,
of VA.
in an injerenm
of’ V(: arc
keen
C VA
inferred
knowl-
of copper
had
have
in the
rllissing
model
would
conclude
dwuys
rnectlunisrn
S’l’AHL
applying
accepted
ujter the incorrect
S’I’AtlI,
the
L1,
Irlodtlt
inferred,
rrlodtxl.
belief
this correct
known
r’cact copper
w hiclr led
of coppt’r
haves eventually
btbcarrte
rlents
after
C = Ph CC. The
CC, tlrrrr
reductd all occurrences
YMJ
Finally,
copper
SA
then
C = Ph VC. liowever,thiscon-
was
the correct
applica-
substitution,
VC = VA CC. If this model
of VA from
wry
reactions
iIlk
a prcnlise,
situa-
-* SA VC and
of VA.
for
model
to construct
nlodel,
are
conclusions”
standard
C VA
reduction
the
so on.
a reduced
t;or example”~
niodet
paradigrrl
needed
“there
+ VA Ph VC after
is incorrect;
VA again
S’l‘A 11I, in
with
for
UYM~
point,
C VA
t Ph VC after
STAti/,
substi-
CI Ph Ash
aSSeI’kd
edge
that
erroneous
transforms
lNI~1~~l~-~~OMI’ON~~N’l‘S:
other
out
produces
1986).
rules
to further
and
indicating
reslllting
tJe
rnto A
from
track,
far from
lo reactions
WOUtd
or
for that
pointed
a cotrlponcntial
clusion
for assert-
lcad
augmented
thus
IN t~I~:H-(:OM
= CI Ph (Ash}
inputs
iron,
inferred,
differs
are
be rrkducecl,
now
in turn
being
is to keep
applic4
point,
Ph
C
tms
cotltain
of 11. fq’or c~xarrrple,
e1npty
a11
may
represctrlt,atiorl
rc~tiuced
to t.llc: assertion
{I}
this
purpose
of’ all su bstanct3
the
oj A UYL~S. At this
which
models
models
a~~tl
main
& Simon,
model
When
model
I), or iJ /I decornposcs
again;
UYL
lNF~R,-C<)MPONENTS:
present
rnore
Irelief.
in either
Jj is cornposed
are
the
a corriponetitiat
is
I/ .I arid S react to form
rtxiLct,ions
replac:e A
theyi
CI Ph. 'I'his
thirdrule,
=
models,
S’, then ltljer thd
froYr1
one substance
or
is: IJ A occurs
rultl
A
of both
is: I/ A occurs
rule
rerr1ove
rules
a set of one
of I) u?d ,S,
REDIJ~X
ttic: outputs,
sut~sl,aricct,
= VA
into
to get
Simou
I< b;l) IJ(:lj:
of S’I‘AIll.‘s
asserts
) I’ll
sides
to the two main
S~JL3STITIJTE
an equation
list.
both
I,cttillg
I,he
with ti ~rtd S. The
that
ash
“(:I
tion
phlogiston
fnor(L sul>stallcc>s,
if ottitlr
‘Y:h
compo-
the
to get
frolri
belief;
and
substilde
equation
reduce
111(1 bulfej, und A is cornpmed
s1de.s
{ Ash”
1
c’a~i
t]le
and
in which
(Zytkow
think
I”.
‘I’tttw~ two steps
both
metaphor;
Xytkow
tiolls
such
%y tkow
involving
rllnirl
and
catclgoric>s.
as being
to this
exist
first
error
c~oIrcurrt:ritty
st,itutiorl
classified
Ij’irst,
composed
SII bstallctl
can
Simon
rrlisal)plical.iorl
assu~rws
(e.g.
the erroneoub
of the
a 5‘11b 5,t ante
of itself
typta
ltt4:DIJ(1tl:
(infinite
(e.g.
where
in rtienlory)
can
irlferences
become
three
defined
recursion).
We refer
A = B C and B = A D
as a circularity,
a circillar
not
rult> into
definition
because
af’ttbr’ applying
a
sub-
A = A D C or B = B C D). kcondly, there
be two rnodets
secorid
error
type
special
case
where
for the same
as [i~odet
one
su bstanc.e.
su bsuriiption,
model’s
componerlts
We refer
to this
f&~usirig
on
are
a
subset
the
of
list in
irijt:renc.es.
LEARNING
/ 529
the
other
model’s
components
(e.g.
A = B C). 1“
1Jr1’
n IIy, a reaction
its inputs
or outputs
We have
found
as one type;
Thus,
(but
be restated
alter
are
inputs
the
fuJldameJJtal
error
null
null
to nil
type
null
a sirrJplttr,
unified
tnethod
two error
note
types
the
to A = A D C after
(Conflict-
a complete
arid
begiJis
such
that
of all reduced
erroncow
reacllotl.
system
its
I1tllet.e 0 I’ronl the
revision.
II frorri
had
an erroneous
its rriairi revision
This
process
process
decides
iJift:reJJccA, revises
ory
(i.e.
ttitl original
hypothesized
in observed
rcactiorls,
c~XporirrJental
‘l‘henard
tlie
reaction
pOtaSSiu[Jl
that
I)avy
anti
oXygt:Jl.
and
JJot pure
lY8ci).
to exhibit
sucti
such
ln order
actually
to support
coJJt,aiJled
sotne
As we shall
see later,
hypothetical
c-oriflictand
and
their
claim,
l)avy’s
potastJ
water
S’I’A ttl,p
hyinto
(%ytkow
is also
able
ttit~ .w(lrce
tays
of’ the
lar to Jnechanisms
19X4), these
iJJg to each
new
each
belief
ultirriately
associated
premises
detected
tags
belief
store
the
revision
erroneous
underlying
in memory;
inference.
is associated
support
For each
source
it.
tag
contains
to belief
the
absent
in a belief
itself
ultimately
contributed
of’ rule
little
stanc-c
finds
tlith ways
plete
rlufl reaction
hypothetically
530
/ SCIENCE
in which
null
it could
(having
altering
reaction
empty
its l,IlS
have
if certain
of P
are
erro-
premises
reaction
P
f P H 0 if’ the
(2).
tiypothcsis
(4) arid
H 0, a
(fl 2
tag
riot
the
ItIfS
S’l’AIll,p
post4
instead
and
or IttlS.
That
the
114
fl, while
(0
3 should
rc3ult
premise
is
eff’eclt P
P
tt ftS of’ premise
3 r) would
not
is
sub-
given
rt,actioJJ
the
11ave
iJidicate
Ilad
ccluse-hypothesis
coJJsl,ruct,
would
which
erroneous
of premise
tiave
there
“l)elete”
For exdJtJplt&,
modified
efrect
rea(.tioll),
indicates
side).
as re-
t.l~cxdesired
for t*ach
that
revisioJJs
While
easy
13, its
cult.
plus
the
thut
riot
such
have
0.
Thus,
(4), whose
in effect-tlyI>ottJcsis
bud
pro-
(4):
The
of
0.
have
beeli
decide
“smaller”
the
side
that
aJJced
side
reactioJJ.
Our
of’ the
Jnodifitbd
Agai11
usirlg
tag
to work
to the
that
erroneous
Jnust
the
be add4
above
should
from,
to a prerrlisc
back
511 t,star1ces
substaJJccs
difii-
substance
is to IJse the
plugged
reaction
relatively
more
efl’cbct . ‘J’he prob-
source
substance
solution
were
this
dt3iruti
is no obvious
in before:.
wtJere
the
this
art:
in ari effect-hypothesis,
premise
which
iJJ to get
there
errotieous
outputs),
S’f’AklLp
substaIJce
“Add”
preserJt
art:
cause-hypotliescls
S’I‘AtI Lp will add
led to a comby
cause-~Jypol,lieses
omission
lj‘or each
S’I’AII t,p must
the
corrirnission
lo create,
ItlrrJ is that
STA11Lp
is,
lag
effectprclrriise.
be hypoltlesized
t’rrolieous
source
should
tile
iJi sorile
wheti
indicate
would
of’ substarlces
inputs
rJlust
(i.e.
modific~d
would
r)
in
It’1I.S O/ prerrllst: 2 ?rlust rwt huue hud 11, urltl tht: ti!iS
that
applicatiorls.
is inferred,
which
(as well as which
premises
substance
premise
the
which
substance
in ari effect-hypothesis
prerrrisc: 3 niust
Effect-Hypotheses
W heJJ an unbalanced
each
f’rorri the
irlvolved
riot exist
Generating
decide
matchirig
corJlplic;itioiJ;
sirice
1.
aJJd
I)elete
1111s. ‘J’hese
balancctl
is hypothesis
r~ow is to
I’rorrb soJne
is delelioJJ
way,
preJniscs
tfJe substance
a series
LlfS
and
to iJJfer a
to 132. In this
substance
which
1) after
Simi-
t:orrcspoJJd-
131 is used
with
on
(dtl K f(>(lr,
premises
propagated
in JnerJJory
based
systems
as a belief
are
of the premise
szlbslartce
for
in assuJrJZ)tioJl-bas<~c~
112, 131’s tags
belief
JJuJrJber
certain
ways
f’our opt,ioJis
LtlS
the
illstead
to a corresporrdiJig
2 should
reasoning.
by
In ltitl t’abt3 of’subslar~cc:s
ally
that
S’I’A I I I,p selects
many
to the modified
the
the
Causo-Hypot~leses
problem
revised,
hypottic3c5
deco~r~posed
tlJat
‘l’he
as water,
of potash
concluded
into
rnodijied
the
are
resulted
For example,
prt:Jrrises
back
tl to the
0 to the
changes
on the
reduction”
how
There
11 0 from
have
premises
nthw substaoccs)
(2) Add
(3) Add
would
Generating
2.
be
(iay-Lussac
consists
finds
using
l<llS.
LliS,
be inferred
of revisiJig
chemists
to explain
I,tiat potash
Thenard
but
& Simon,
substances,
potassiurrl
observed
that
difI’ert?JJt.
“inverse
them
system
its
J.evise
t P H 0 { -},
(without
f
is nil
the-
is historical
century
lj’or example,
effect
the
(4) I)clete
t P H would
not include
there
18th
in arJ attempt
a new
does
Iri fact,
missitlg
this error.
Now
II 0 I,O the
ItIfS,
is to
an
to
it f’rorn Jnemory
beer1 left isolated
pluggiJlg
eflect-hypotheses
h.~~l
H
Its goal
is P
RIIS,
reaction
deletes
ver-
lead
the errorleous
c~onstructs
which
method.
results.
claimed
while
(;a.~-Lussac
was
revision
frorri
caused
and
iliference.
sorrlt~times
drogen,
premises,
erroneous
for this
S’I‘AII I,p invokes
to recover
pretniscs
a new set of beliefs)
motivation
irlg
in order
which
these
reinfers
inference,
erroneous
here
equals
to use
(1) Add
I~WIJS
detecting
LIIS
here:
reduclioris.
JJot have
reaction
detect
S’l‘AIll,p’s
IJporr
this
or rrJore
is to perform
result
supporting
would
S’I’Alll,p
list substances,
the
so that
the
step
to
a Lalutlced
011~ which
one
the
0 would
‘I‘he first
reaction;
af’tcr
revisioJJ.
11 and
charlges
of inferring
reaction
detected,
belief
how
eff’ect
siippose
Once
two
reductions.
the
JJ1Jll reactiori
For example,
IllIS.
about
have
erroneous
it carI change
S’I‘A 11Lp must
enabling
of this
to B C = B C D
= D after
of belief
as
how
could
H 0 {P).
in S’l’AlI Lp;
reduction.
inf’erences
react,ioJJs,
outputs
(or simply
reactions,
= D C after
to nil
be viewed
hypotheses
premises
as tlJe third.
reaction
B = A D leads
erroneous
uribalaJiccd
can
To see how the first
therJ
only
constructs
either
or empty
A = B C a11d A = B C D lead
substitution,
‘I’hus,
types
be restated
empty
as unbalanced
then
models
error
can
un6alunced
A = B C and
sJJbst,itution,
where
sion
three
either
reaction).
circJJlarity
ing
these
is the
to it as an
A = B C D subsumes
be irlterred
empty.
two types
with
JJot both)
an erroneous
can
that
the first
a reaction
we refer
are
can
it, did
source
tags
er/Jplty side
arc’
IJOW
reaction.
of
the
‘I‘his is
to tlffect
exanrple,
OJl
a bal-
STAlll,p
into
rrierriory.
fj‘irially,
in response
to the
the result
‘1‘1it> qtlc:st,iori
is uhrc~h I,IIS rrirlst be revised
now
whit Ii premise’s
1’ is to
answer-.
(I’ 3 I), the cltatailed
If’ it was
woc~ld
0.
the
r.csuIt:
(Nottt
that
t’ 01~
such
a
rriadc
iri 011e of’ its ruris,
claiIii
that
the
otJst:rvc~d
sirriply
had
to
>ource
of leliej
PIredly
coIl(.IuskJrl,
wtlich
S’I’AlII,p
(I’)
sotr~e
in the
wakr
and
are
Wllile
we
aid
have
tly[“Jtht?S
shown
(1)
alld
t,ti iihc~i
to construct
laitiirig
both
Ily~JothtWs
how
tt ltb reasoIking
(‘$
hybrid
omission
(2)
S’I’A I I Lp
and
null
will
back
irlto
f’orrri
&scribed
rrlodols),
ori(~ can
evolvirlg
iri response
reactions
and
(e.g.
new
irlf’erencing
repeated
as the
theory
incrementally
detections
revisions
con-
set of coniponential
an initial
to
subsequent
IV
of erroneous
of’ selected
Phlogiston
‘f‘llus f’ar we llave discussed
operation.
causealso
thosta
errors
a new
design
a complete
ot’ reinferencing
(i.e.
visualize
the
the orig-
the
that
tjurirlg
result
of’ d n~‘w ttieory
structiori
pur-
would
sausc~-IiyIJottieses
irlstedci
the
Ieast
be reinferred;
guarantees
occur
premises.
vs.
Oxygen
the
for this
would
result
l3y viewilig
may
Hopefully,
at the very
not
strategy
that
is performed.
but
will
revision
rcac.tiorl
irlf’erencing
side
cause-tly[>ot,tieses.
i~~ld coInmission
and
S’l’AtiLp
reaction
of’ S’l’Alil,p’s
cycle.
new
will be new models,
erroneous
In general,
art’ used
orrlission
iri constrllctitlg
allegedly
promise
pIugg:cA
c,rr~pl> bide of an er~roric~o~~s reaction
to
intleed
irl it.)
0)
(1)
‘I’heriard’s
substances.
that
the
had 11 und
I)avy
relevarit
orr~ittod
hOl(lS
anti
reactiou
(II
the
hypothrsize
all rc~ci~lc~c:d-list substa~lct~
[JOSC
tag
causal-hyI)ottiesis
I,ifS
the
uscxtf I’ to pillpoint
in which
I”5
r~~otl~~ls (lay-l,ussac
potash
really
blat~~.
i.e.
inal
ariy
II~W premise(s)
Let
walkirlg
us riow
through
assertion
a detailed
three
of
tht: fundamentals
synthesize
the
of’STAlll,p’s
previous
example,
sections
by
with
the
chains
(4
2 into
3,
beginning
prerriises:
(‘ori-
cause(1
(2
(3
(3)).
(M --> CM Ph i-1))
(CM = M 0 i-1))
(M CI --> I CM t-1))
3.
gc:rieraled
tlavirig
to
detailed
rc>vistt prerrrises
(~1lct~, S’I’AIlLp
to CXtY II te.
art’
This
step
the
tJy
cortairl
to
begills
[Jrerriises
erroneous
sets
wouId
applkd.
supports
1 rnod<tl,
+
he
ttitln
to premises
of
cost
each
wit Ii ttlcl lowest
the
2 arid
1~:
cost
cost
orl
i.e.,
arid
8.
After
tJle
011~
wtlose
belief>
lesl
3
trlodifi-
corrlputing
the one
revisiorls
as t,he
2
[)rerriise
S’I’A 111,p sekcts
existiIlg
let
3 7IlUSf not
(of’ rrlaking
3) is 7 1 I
~ausc-llyI)ottlesis,
1h(> IC’ilbl. irripact
ctlariged
example,
Vor
7 models
total
flmlt:ls
‘I’h.5 /l//S o/ premise
(4):
tiutt 0. If prerriisr~ 2 supports
cations
the
cause-
existing
must IlOl have bud II, UPLdthe IIlI,S of yremzse
tlnoe
ilif’er-
These
premises
through
will
have
tly]Jottlesrs.
(4
(M Cl
(5
(6
(CI
(Cl
-->
(7
(M -->
(8
(nil
M 0 Ph {-))I
At lhis
0 Ph {M)))
point,
rrlovclci aud
atirlg
hypot
have
been
striic
tirig
a hew
lhcory
01‘( ortlporleritial
lorigclr
present,
potht3i5,
alid
that
sorml
Second,
ill
the
hy[)OthcSiS
a possibly
Iriay
will 1)c charigecl
the
that
hystcrrl
allc~
asserts
COII-
tlifl~ert~rit set
be view, s0111e may
0t~ir~r.s trl;iy have
‘5 IJdWc
all t Jt’,I’I(’ 1;
,’ 1 011
rrlc~rllory.“’
IJcJSc‘(1
c~olitairlirig
rr~odels;
for cbacti premise
S’l‘A 1I I,p starlb
tm3ri rnodiliecl.
no
Iqirst,
rt>action
belief
two
inference
hescs
atJout
avoided
sclve~,
~llus
prevcrlt,irlg
bclirig
asserted.
The
irl
eH~c1, turning
analyzes
(wit,tlout
H could
a corrll>lcl,e
balanced
tlt’w
is re-
been
the
reduced
non-
them-
11ull reactiorl
null
first
could
in
by recognizing
reaction
constructs
M back
reaction
reactions
of r’oaction
;rny
have
7.
would
both
‘I’hen,
relight
had
frorrl
the
the
into
which
7 if premises
1,
gener-
reaction
substances
comes
to
inference)
errorleous
a11 uIlbalanced
system
2 into
S’I‘AII Lp starts
the
by plugging
the four
using
the
fromg
7.
erroneous
how
its attention
irll’errt~~l irlskaci
substituting
reducing
how
The
reaction
after
after
answer
wllicll
irlskaci.
substituting
reducing
4,
infer
components
bctgi1ls.
i.e.
of’ reaction
after
after
after
8 (an
revision
side
empty
errorlt’ous
hyl~othesis,
to
8):
I M 0 {-I>>
-->
resulkd
4.
111th best
lead
7 and
--> I 0 {M3))
= I 0 {Ml))
I’c%rerlt ways
AlI’tt~r choosirlg
t heri
6, then
ot’ revisions
by each
rriariy
wllich
is
Iiow
S’I’A tl I,p computes
l10w
cause-tiypot,tittsis
about
suggested
rtbftt:cts
those
the
of these
when
causc~-IlylJot~lc~sis
II:, c~xarriirie
avoid
which
rrlodific;tt.iorls
‘I‘tlt: cost
supportecl
if’ a
in order
riow selects
cost of’ rrlaking
Ily~Joth~sis.
cause-hy[~otheses
been
difhave
modified
iI1
sides
S’I‘,~kll,p
have
been
different
sutJstaIlces):+
due to the chosc>Ii hy-
~>r~~lrliSt!
~Jerforrrls
the
arc:
c~eh!kd
f’rCJll1
tflc
changes
pro-
nlodified
[)rt!IrliSC’(S)
(EHl)
(EH2)
M [O Ph] --> M 0 Ph
M I.01 --> M 0 (Ph)
needed
needed
0 and Ph on LHS;
0 on LHS, no RHS Ph;
(EH3)
M [Ph]
-->
(EH4)
M -->
M (0)
needed
needed
Ph on LHS, no RHS 0;
no RHS 0, no RHS Ph.
M (0)
(Ph)
Ph
LEARNING
/
53 1
‘I’0 determine
which
ill rcactiofi
7,
its source
tag
1 r>
(Ph
ing ri~‘w I,llS
su hstanccs
used
it
1 1)
(M
su t,stmce
cornplcte
+ (M 2 r)
corresponding
‘I‘he tag
t’h are
each
analyze
(M
N ow the
c‘au be gerleratcd.
contributed
must
irlf’orrnaliou:
{-}.
for 0 and
premises
S’I‘Afll~p
t)esl
lions
(Cl 2 r)
aud
for hypot
errors),
for hypothesizing
while
The
hesix-
liever
1.
(CIf2)
Belief
Belief
1,
1,
(CH3)
(CH4)
LHS:
should
force
1,he tags
commission
tlieory
(i.e.
p”riHiigUl
errors:
have
had
had
0 and
Ph;
had
Ph,
Ph;
should
not
had
0;
01’ I)avy’s
Ph.
0.
l,avoisier-era
RHS:
Belief
1.
Belief
2,
RHS:
RHS:
should
should
have
not
not
have
have
IIavy’s
tile
cost
hypottleses
of carrying
recommend
011e rr~oclt~l (CI
= I
(Ill3
have
arid
I)elielh
(:114
1 arid
thtl lowest)
had
had
cost
point,,
sirice
rrlodific4)
would
1 is the
betiet’
premise
vt%rsioIl of premise
of’ (:fi’L:
‘I’wo
rtlacl iou,
whittb
null
irrg existing
its
10:
M 0
a c.orIlpleLt:
upon
resulted
not irt the
+
uull
p;LradigIir
can
by fi)Ilowers
propcjse
of’ bvoisier
would
(: t I’L, sirice
that
if
frorri
of reaction
rto belief’s
(: 112’s rtbvisio1is
cxecu
told phfoyzslon
were
reac tiolls
rrlotlels
then
S’I‘A 111,p to revise
phlogistori
very
least
theory
shiR
dofirtcs
oxygerl
w t1ic.h 00
over
represent
not
a period
to orit:
t~mhodyirig
tkic views
small,
910
but
of the
rsisted
of
t ht! existeuc-e
it is possible
from
one
oxygen
the
alaf’ter
as a systcnl
inc. I udt:
for
errltmdyiug
theory.
hypotheses
competing
SIIct1
at tile
tlieories;
r Note how IIO hylmtl~eses
use I,elId 3 ‘1s it did IIUL cotItCbute
to the erI,ike all ~LsaIIIIIl)tIUII-t)‘Lse(l systems,
S’l’AltIdp ih rl
I’~,I,t’~>lls IIIfereIlce.
d~l~~~tlc~tcy-clurc ted I‘~LL~OIICI~LLUS~ it 011ly IIypuLlIesixes Ixvisims
1~ ,,I (‘i~~Iht:h uI1 w Ibis II t he er1~0lIeolI~
iIIfcI e~lc e ulC iIIIdt,t’ly tlepiIcIS.
532
/ SCIENCE
null
reduced
it assigns
are
tectkrriqilt5
iltto
it,5 su hst
lo arc: rttvised,
iril’cbrred.
work
for rrlotit~llirig
011 sc.ierttific
11ow t,tieorirts
References
for
cartle
Once
t)lc> error,
arlti
S’I’A tll,p’s
of’ more
re-
itllCCS
premises.
blarrle
be
Lfte reduced
a11 erroneous
to accoiirit
lies ill its iIlc.or-pc,l.aliori
iI1 its potential
l~‘iually,
of ils
is
t,yI)tls can
is choseri
t~\~t~lll~lliilly
intor-
iIll’ert+rictts
error
t’roIl1 such
where
ct-
thari
i1lf’ertbric.m caused
erroIit>ous
to sortie
modand
list. coIlGus
rc‘actiorls.
ioll atmut
revisions
unified
irrfcrt~Iict3
erroIlt>ous
I)t!tc~ctilig
hypothesis
prerrlises
~orrlI)ollent,ial
a more
S’I‘A t 11,‘s Z) mdin
irif’orrriat
to propose
revision
‘I‘he
but
exarriptt!,
lo7iyer
it to
new ones
discovery.
erroueous
S’I’Alf I,p to recover
using
Lributiori
with
hyI>othesis
was that,
theory
of tirtic
it,s set. of’ blief’s
is not guaraliteeti,
ol)sthrvetf
a fOl lower
of oxygc~~,
of ttiis
was kept
If one
of‘ pit togisI,ori,
point
of beliefs
111 (~71~ belie/s.
alld
to believe
in the c:xistcllce
‘I’he important
the number
a Ihtlory
be prone
believe
nontzx-
for revising
enables
of scierllific
eIttploys
reduction.
a pl;tusit)te
the
rc:port,edly
a beliclf’
Summary
S’f’A t II,.
as unbalarlct:d
shows
I~or t~xarnplc,
iriferences
of 5
process
riot t)xist
rrlec*llallisrrl
for c’onstructirlg
iIi S’l‘A f1 I,p;
actiori.
how fol towckrs of’ one
of data
paradigm.
probably
he would
not. ptltogistori.
t,hough
rttvlsions
of mother
dornaiI1
1,115
a set
revision
as well as propose
for preverlt.iIlg
by ulisapplictd
simpler
viewed
Thus,
he inferred.
loosc~ly model
giveri
does
to erroneous
for dealing
rleedeti
set, of’ bclicls
S’I‘A II I,p’s hypotht~ses
0 on the
just. as phlogis~orl’s
assurriptioris,
substaricths,
predecessor
rcinf’erencing
outputs,
caloric
be-
revised
reillterpretatiorl
belief
S’l’ALf I,p’s
in any
strategy
lisl erlables
in t hta irtputs
will
ltle
f’rorri rriernory.
that
modt:l:,
fectivtf
rrlatiorr
tjelief
that
in response
its basic
I1 and
example,
by chemists
S’I’A I I I,p, a system
begirls;
to
the dispute
S’TAlll~p
STAlll,p’s
that
observations.
Thenard’s
Iri another
IN short,
els of chemical
already
the cttartgcs
to
in rnodclling
and
a bck-
ariolher
hypothesize
obtained
V
support,ed
has
and
from
to be
incorporat,es
presertt
was actually
Lelicf
to how
betiels
theory),
reactions,
ability
corresponds
cou td arlalyxe
9 K 0 to include
At
prerllise
reinferencing
discovered
reactioir
place.
1. I,et
9, a rnodifitbd
lead
P
was.++
a vital
example
Gay-l,ussac
accepted
queslion
belief
dtlleted
were
was actually
phlogiston
corilradic1
then
is harrnlcssty
ttith coniplele
1, and
Now
reacpremises
hence
belief
which
asserts
2 ir1t.o 9 leads
models
borrtt’ 0xygtsIi
the only
I wtlicti
reductions
which
110 uew
ttcrt’,
belief
of’ belief
M 0 { -}.
uow
(aud
best, hypothesis.
reactiou,
+ CM { -}.
M 0
subslitulion
here
from
C:ay-I,ussac/‘l’~ie~iard.
resulI,s.
premises
Thus
chaIigirig
zero
1 (the
ist,crlc.e
:! supports
changirig
as the
cause-
none.
propose
have
011 htllief
erroIitf0us
S’I‘A tll,p
bolh
propose
habed
to take
rnisint,erpreted
ItAt1 to l,ht3 hypothesis
t,hese
IJetief
CliZ
chosc~n
be deleted;
t)tlcaIl titbteted.
changes
I supports
I, since
both
all bczliefs
t)y
belief
Cl1 1 arid
us say (:11X is arbitrarily
this
the
is computed.*
0), alid
2, while
cos;,
out.
and
replicating
t~ventually
Now
of erroneous
actually
were
I)avy
have
1,
0,
were
results
tween
not have
have had
2,
shift
phlogiston
beliefs
should
Belief
presence
of phlogistorl
theory
in this
should
should
Belief
the
revisiorl
LHS:
RHS:
LHS:
repeated
to removal
in oxygen
Sirliilar
Belief
the
lead
helice
those
(CHl)
case,
would
cause-hypottlc~ses
I I) is used
(omission
iri the
with
a rlcw
IItain
<‘On-
powc>rlul
Ixtirbl’
discovery,
alld
evolve.