Anda di halaman 1dari 27

Normative Systems of Discovery and Logic of Search

Author(s): Jan M. Zytkow and Herbert A. Simon


Reviewed work(s):
Source: Synthese, Vol. 74, No. 1, Knowledge-Seeking by Questioning, Part I (Jan., 1988), pp.
65-90
Published by: Springer
Stable URL: http://www.jstor.org/stable/20116486 .
Accessed: 19/11/2012 14:15
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Springer is collaborating with JSTOR to digitize, preserve and extend access to Synthese.

http://www.jstor.org

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

JAN

M.

ZYTKOW

NORMATIVE
AND

AND

HERBERT

SYSTEMS
LOGIC

A.

SIMON

OF DISCOVERY

OF SEARCH

for logic
create a research
New
program
computer
systems of discovery
consist
of
of science.
These
rules and control
inference
systems
philosophy
are influenced
Their paths of discovery
that guide
the discovery
process.
by
knowledge
with
of results. The
the justification
data and the discovery
the available
steps coincide
can be described
of artificial
in terms of fundamental
concepts
process
discovery

ABSTRACT.
and

intelligence
traditional

in terms of logic. The


and can also be interpreted
search,
of
the philosophy
that places
outside
studies of scientific
discovery
or history,
is no longer valid in view of the existence
sociology,

as heuristic

such

distinction

in psychology,
science,
It becomes
and attractive
of computer
of discovery.
both reasonable
systems
were
in the same way as the criteria of justification
of discovery
the schemes
norms.
as
as
and
facts,
empirically
logically

1.

NORMATIVE

SYSTEMS

OF

to study
studied:

DISCOVERY

In the last decade


several computer
have been constructed
programs
et
of discovery
in detail, various processes
that simulate,
(Feigenbaum
al. 1971; Lenat
1977,
1983; Buchanan
1978; Langley
1978,
1981;
et al. 1980; Michalski
et al. 1983,
and Stepp 1981; Langley
Bradshaw
either on
and Simon
1983a,
1987; Zytkow
1986). They
operate
or on quantitative
data. They
different know
incorporate
qualitative
about
the forms of the laws they
ledge and different
expectations
should
consider.
Some of these systems
very specific
incorporate
of a certain domain
in science, while others are based only
knowledge
on general-purpose
construct
heuristics.
Some programs
discovery
new concepts,
are
others
laws, and still others
capable of constructing
construct
in the form of descriptions
of the hidden struc
explanations
ture of things or processes.
from two discovery
In this paper we will draw examples
systems
that address a few facets of scientific
BACON
and GLAUBER
formation and generation
of empirical
concept
discovery:
regularities,

either for qualitative (GLAUBER) or quantitative (BACON) bodies

in
forms classes of objects based on regularities
states
in
terms
laws
of
and
these
classes.
data,
qualitative
qualitative
for finding numerical
includes heuristics
BACON
laws, for postulating
common
intrinsic
for
and
(theoretical
terms),
properties
noting
of data. GLAUBER

Synthese 74 (1988) 65-90


? 1988 byKluwer Academic Publishers

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

66

JAN

M.

ZYTKOW

AND

HERBERT

A.

SIMON

divisors.
Since both systems are described
in numerous
publications
and their full discussion
the limits of this paper, we
goes well beyond
to the minimum
to support
will condense
their descriptions
required
the ideas of this paper.
BACON:

A System

That Discovers

Numerical

Laws

et al. 1980; Langley


et al.
1978, 1981; Bradshaw
(Langley
on
bodies
of
data.
induction
1983, 1983a, 1987) makes discoveries
by
a
It deals primarily with numerical
set
Given
of
(variables).
concepts
a
one
and
variables
BACON
varies
variable,
independent
dependent
variable, while fixing the values of all other independent
independent
It registers
the corresponding
of the dependent
values
variables.
variable and it looks for a function that relates the varied independent
variable. Once
such a function
is found,
variable with the dependent

BACON

in that function the status of dependent


gives the parameters
at a higher
level of its search tree. At that higher
level the
next
to
the
and
tries
it to the
varies
variable
relate
system
independent
new dependent
terms. This process continues
until
all the
recursively
a
are
variables
into
relation
incorporated
independent
quantitative
an example,
two independent
the system may
consider
ship. As
variables
variable. For the fixed
x\ and x2, while
y is the dependent
=
the regularity
value of x2 = 1, it discovers
2jci + 3. This regularity
y
=
a and b are created.
has the form y
axx + b, and thus the concepts
For x2 = 1 their values are a = 2, b = 3. Then,
for another value of
=
is y = 3*i + 5, and for still another x2 = 3, the
x2
2, the regularity
the system confronts
is y = 4jCi + 7. Now
the variable
x2
regularity
that a = x2 + 1, while b = 2x2 + l.
with a and with b, and it discovers
=
law
these
form
the
Taken
y
regularities
together,
(jc2+1)jci +(2*2+1).
can also consider
nominal
BACON
that is,
variables,
independent
variables
that have symbols as their values. A nominal variable cannot
to numerical
be directly
variables
related
by the use of numerical
a new numerical
functions.
introduces
In this case BACON
variable
BACON
variables

it with the same record as


and associates
called an intrinsic property
can be related
to
Thus
the nominal
variable
the nominal
variable.
the
intrinsic
other numerical
variables
Intrinsic
property.
through
are defined by numerical
to
variables
available
dependent
properties
In this process
the numerical
values of the dependent
the system.

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

SYSTEMS

OF

DISCOVERY

67

are copied as numerical


variables
values of the independent
variable.
are defined,
the numerical
Once
values of the independent
variable
the same values will be used whenever
the same objects occur later in
BACON's
search.
a large number of laws in 18th and 19th
BACON
rediscovered
and chemistry
Law, Kepler's
(Ohm's Law, Boyle's
century physics
of momentum
the conservation
Laws, Coulomb's
Law,
law, and so
these laws BACON
intrinsic
introduced
forth). While
discovering
to mass,
terms corresponding
of a battery,
internal resistance
and
others.

GLAUBER:

A Qualitative

Discovery

System

Input data. The data utilized by GLAUBER


(Langley et al. 1983a,
of a set of qualitative
observational
facts, such as
1987) consist
tastes
sour"
and
acid
acid
combines with
"hydrochloric
"hydrochloric
to
to
form
Facts
sodium
sodium
chloride".
inputted
hydroxide
as lists beginning
are represented
with a predicate
fol
GLAUBER
lowed by a number of attribute-value
lists. Thus our second example
above
is represented
the
(reacts (inputs HClNaOH)
by
proposition
is
"reacts"
The
the attributes
here, while
NaCl)).
(outputs
predicate
are "inputs"
A set of values
and "outputs".
is allowed
for each
attribute.
In our example,
two
the
attribute
has
values,
inputs
the two substances
that combine
in the
{HO NaOH}, which represent
reaction.

Classes.
GLAUBER
Noting Patterns and Defining
accepts such a set
of facts and searches for facts that have the same predicate,
the same
such a collection
attribute, and the same value for this attribute. When
creates a class (or classes) of values
of facts is discovered,
GLAUBER
that differ in these facts and a pattern which
is stated in the same
as the original facts, save that differing values are replaced by
on sodium
names. For instance,
that while
suppose
focusing
notes
that
this
combines
with
chemical
GLAUBER
hydroxide,
one
to
form
in
with
acid
sodium
chloride
nitric
acid
case,
hydrochloric

manner
class

to form sodium
Na2S04
would
contains

in another case, and with sulfuric acid to form


salt) in a third case. In this situation, GLAUBER
classes. The first (let us call it NaOH-reactors)

nitrate

(Glauber's
two
create
hydrochloric

acid,

nitric

acid,

and

sulfuric

acid, while

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

the

JAN

68

M.

ZYTKOW

AND

HERBERT

A.

SIMON

sodium nitrate, and


contains
sodium chloride,
(NaOH-results)
The pattern would be stated as (reacts (inputs NaOH NaOH
reactors)
(outputs NaOH-results)).
second

Na2S04.

Identical or very similar classes may be


Related Classes.
Combining
classes and com
created for different patterns. The system compares
a
in common. At
of elements
bines those that have
high percentage
the same time, both sets of patterns are associated with the combined
also generated
the class of sour-tasting
class. For example,
having
of the sour-tastipg
notes that every member
substances GLAUBER
class (and
class also fits the pattern associated with the NaOH-reacting
and NaOH
of sour-tasters
vice versa). As a result, the members
associated
into a new class. This class becomes
reactors are combined
taste and the other summarizing
the
with two patterns, one involving
class of NaOH

reactions.

Towards More General Patterns.


Since patterns are stated in
Recursing
can apply its abstraction
the same form as the initial facts, GLAUBER
In this way it
to
it
the
method
patterns
already obtained.
recursively
acid alkali)
arrive at a more
(reacts
may
pattern
(inputs
general
is com
next
of
round
After
the
salt)).
patterns
abstracting
(outputs
classes
and
and
combines
the
compares
patterns.
system again
pleted,

Discovery

Systems

Preliminary

Conclusions

our brief description


of both systems trivializes some of
Undoubtedly,
behind various aspects of
the motivation
their features and obscures
reader can find the details of their con
interested
their design. An
et al. (1980);
struction and testing in Langley
(1978, 1981); Bradshaw
several
et
al.
Nevertheless,
1983a,
1987).
important
(1983,
Langley
can
seen
from our
be
of
of computer
features
discovery
systems
of computer
First, requirements
bring dis
programming
examples.
construe and
to
of
the
level
concreteness,
detail,
covery
strategies
on the com
for running the programs
necessary
tivity of principles
terms
of
are
in
the broad
both
the
programs
general
puter. Second,
make.
and
the
discoveries
of
data
they
They are
range
they accept
same
sense
that
or
the
in
universal,
logic is: each
perhaps
general,
in a class of
to
in
a
data
that
formalism
any language
applies
provides
a
a
the
admissible
have
that
Third,
program.
by
syntax
languages

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

SYSTEMS

OF

DISCOVERY

69

into rules of inference


and the
may be divided
that guides its application.
The control mechanism
sail safely
lets the program
failures and successes
in in
through
and changes
does bookkeeping,
the tasks when
the ap
ferences,
time comes, thus letting the systems cumulate and generalize
propriate
their discoveries.
and GLAUBER
Fourth, BACON
apply the same
system
discovery
control mechanism

inference mechanism
both to the initial facts and to the intermediate
results on all levels of abstracting
their concepts
and laws. This
is
both
because
the same syntax, and
facts and laws have
possible
because
there is no formal difference
between
based on
concepts
direct observation
and the concepts
created in the abstraction
process.

Discovery

Systems

and Heuristic

Search

can be viewed
BACON
and other systems of discovery
out the search through a space that includes possible
laws,
and other entities
entertained
the
the
scientists.
From
concepts,
by
of
view
search
both
the
is
the
point
program
programming
guided
by
on. From the point of view of artificial
and by the data it operates
use heuristic
the programs
search. Their
intelligence
principles,

GLAUBER,
as carrying

heuristics
about the discovery
and in
process,
comprise
knowledge
clude both heuristics
that make
local moves
of
rules
of
by application
and heuristics
that affect global control strategies.
inference,
is search crucial in problem
Why
solving, and why may heuristics be
so helpful?
If we do not know a method
that allows us to solve the
The less we
problem
directly, we search, trying various possibilities.
know the more we must grope
in the dark. But even in blindfolded
there may be a method which may even guarantee
that the
wandering
solution can be found at last. We can at least be systematic,
not trying
the same place twice and not skipping others. Such a control strategy
is treated as a very weak method.
The
less we know,
the weaker
are then inapplicable
methods we use, for strong methods
(they would
are not satisfied).
lead us astray when
their presuppositions
The
we use,
the methods
even
For problems

the longer it generally


takes to find a
of moderate
size an exhaustive,
sys
tematic search through the space of possible
states may take weeks of
time, hence a need for cutting down search. This need is
computer
the program with heuristics.
satisfied by providing
Some heuristics
moves
most
current
the
from
the
propose
likely
place in the space of

weaker
solution.

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

JAN

70

M.

ZYTKOW

AND

HERBERT

some other moves.


while
disregarding
the most suitable search strategies.
and GLAUBER
BACON
apply a variety

states

A.

SIMON

Other

heuristics

may

select

of heuristics.
In parti
fails to find a
cular, they use data to direct their search. If BACON
one changes
two variables,
it tests whether
linear dependency
between
with
of
If
the
the
other.
is
the
this
case, the
growth
monotonously
a new concept,
in the
the ratio of both variables,
system considers
hope that the new concept will be linearly related to other concepts
to the system. For the inversely monotonous
relation
the
available
in this direction
is created. By continuing
product of both variables
for instance, Kepler's
third law, R3/T2 =
BACON
quickly discovers,
const. GLAUBER'S
search for regularities
is also guided by data. For
heuristic picks a fact and creates a schema
instance its FORM-LAW
the system searches
by fixing one attribute and one of its values. Then,
rest
of
for
this
the
data
facts
that
match
schema.
In this
the
through
that was not originated
is considered
the
initial
way no schema
by
data.

for cutting
down
search
used
is problem
technique
can
If
the
and
be decom
(divide
conquer).
problem
decomposition
a
a
of
then
for
the final
into
search
collection
simpler problems,
posed
solution reduces to a sequence of searches for partial solutions. Partial
are easier to find, and they are usually combined
until the
solutions
Another

and GLAUBER
final solution
is reached. Both BACON
apply this
the systems keep cumulating
their findings as long
method. Moreover,
as they can and the final solution each of them reaches
is the maximal
can
from
data
with
abstracted
the inference
that
be
the
given
theory
Both
be
called
of
the
may
systems
system.
knowledge
capabilities
are
of
whose
laws
and
solutions
concepts
systems,
partial
cumulating
for their lack of
which may be later abandoned
lesser generality,
of final laws and concepts. What
within
the collection
importance
are
most
the
clear at the
and
laws
valuable
may only become
concepts
end of the search process.
of
of search go beyond
the technical
Some properties
description
of
of
with
the
BACON
the program.
complies
principle
simplicity
it applies
its heuristics.
order in which
knowledge
by the particular
are tried first, and if they succeed, more complex
Simpler possibilities
are
at all. The principle
not
of simplicity
is
considered
competitors
of
the
in an implicit way, detectable
observed
system.
by meta-analysis
It is analogous

to similar properties

of logical

systems.

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

OF

SYSTEMS

DISCOVERY

71

are criteria
for selection
and priority. They may be
Heuristics
in search programs
in a variety
of ways:
in control
incorporated
that select the node in the search tree from which
the search
processes
is to continue or that mark certain nodes as unpromising,
in processes
to be applied
to a node,
that select the operator
in restriction of the
a
to
to the
subset that promises
operator
especially
rapid progress
that evaluate progress.
goal, or in processes

Discovery

Systems:

Search

and Logic

and operators
heuristics
change one state into another,
they are
to inference
rules in logic. Deductive
rules in
inference
analogous
the deductive
and hence
that guarantee
logic are operators
validity
of transformations.
in a broad sense, do
the non-creativity
Heuristics,
or
not guarantee
but instead they
deductive
validity
completeness,

Since

increase

the efficiency

of search.

TABLE

I.
Search

Logic
Space of syntactically
correct
formulas

Search
rules

Inference
Logical

consequence

of

Decidability

of

(space

states)
heuristic

Operators,
Point reached
a formula

Set of consequences
a premise

space

knowledge

of possible
rules

in search

space

from

state

given
tree
Search

Computational

from

a given

state

complexity

formal logic and heuristic


The analogy between
search may help in
what
lies implicit
in the
type of logic of discovery
understanding
Table
the analogy,
I, which describes
computer
systems of discovery.
can obviously
to other concepts
to search and
be extended
relevant
a
a set of
and
of
The
axioms
inference
rules
define
logic.
logic
reachable consequences,
but do not determine
the order in which they
will be reached. A logic can therefore
as a non-deter
be viewed
a
ministic
non-deterministic
algorithm.
By
algorithm we understand

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

JAN

72

M.

ZYTKOW

AND

A.

HERBERT

SIMON

all possible
here one that makes
inferences
from a given
permitted
automata
state (formal
rather than
theory notion of indeterminism)
one that selects one of the inferences
and that can make
randomly
at different
different
inferences
times (physical
notion
of indeter
A system of heuristic
search is deterministic.
The order in
minism).
in the search space is determined
which
points will be reached
by
new points
and operators
that generate
heuristics
from the points
to which
that determines
already reached, and by a control structure
point in the space of states they will be supplied.
a logic, viewed
as a non-deterministic
In other words,
algorithm,
can be supplemented
to determine
rules of in
which
by heuristics
to determine
structure
ference will be applied and a control
where
they will be applied (i.e., which premises
already proven will be taken
an automatic
as the inputs to the inference
theorem
rules). Thus,
of a logic combined
with a superimposed
system consists
proving
structure. Of course
and a control
this superim
system of heuristics
search. Or, it
say, a breadth-first
posed structure may be very simple
in the interest of efficiency,
be very complex
may,
say a best-first
to determine
to
function
search using some kind of an evaluation
of

which

the already-proven

the next

theorems

operator

should

be

applied.
and the control structure may be desig
The ordering of operators
not only
to issues of efficiency,
but also to con
ned to respond
of decidability,
siderations
completeness,
consistency,
validity of con
can be applied
same criteria
to the
and so on. The
sequences,
and control structure of any heuristic
search system.
heuristics
and Inference

Heuristics
We

have

earlier
as

characterized

rules of

systems
terms, the inferences
BACON

Rules

inductive

some

of

inference.

the
Let

rules in the discovery


us examine,
in logical

they make.

s heuristics

are a collection
of FIND-LINEAR-REGULARITY
The premises
of
statements
that tell what values of dependent
variable
y go together
=
with what values of independent
variable
jci. For instance "if jci 2
=
=
=
9". The conclusion
of FIM>LIN
then y
7", "// jti 3 then y

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

(1)

= 2& b= 3&
AxxEy

Eab(a

DISCOVERY

is a statement

heuristic

EAR-REGULARITY

OF

SYSTEMS

NORMATIVE

73

of the form

=
axx + b
y).

since it is intensionally
conclusion
is an inductive generalization,
two new quantitative
to the input data.
In addition,
limited
a and b, with their values of 2 and 3 respec
are created,
concepts
and the conclusion
the premises
(1) are only a partial
tively. Both
state
that
BACON
associates with them,
of
the
knowledge
description
if
variables have their values fixed. Even
for several other independent
are not involved
in the inference
since
these variables
explicitly,
for
in another place, they are needed
BACON
keeps this information
com
in
for
of
BACON's
reconstruction
reasoning,
particular
logical

This
not

bining

different

(2)

Ex2,

inferences.

& xn
where

x3,...,

xn(x2

has

the full conclusion

Thus,

ml&...

= 2& b =
3) & AxxEy

mn)Eab(a

the form

=
axx + b
y

rrin are

numbers.
On
level of BACON's
the next
are
link
that
and
b
considered
and if the
x2,
a,
abstraction,
regularities
search is successful,
conclusions
similar to (2) are found.
While
the statement
for a and b on
(2) with regularities
combining
a statement
we
the next higher level of BACON's
obtain
abstraction,
mx,...,

...

Ex3

xn(x3

m3 &

... &

xn

m?

=
=
=
& Ecdef(c
l,d
l,e
2,f=l
=
& Ax2Eab(a
+d8c
b = ex2 +f
cx2
+ b = y))).
& AxlEyaxl
This
to

conclusion

the

can be transformed
of

sequence

steps:

discovery

& Ax2,
X\ Ey y
(jc2+1)xi
in a similar way,
combined,
...

X3y X4,

in his
as

AxEyPxy
decomposed
intermediate

tikka's

...

xn(x3

+ (2jc2 +1)). This


into a regularity

m3

omits
&

... &

reference
xn

rrVn

result may be further


that includes variables

5Xn.

Hintikka

AE-facts.
bottom

to the form which


Ex3

initial
into

(1988)
facts,
atomic

conclusions

of the form
A?-statements
that they need not be further
argues
to notice
facts. It is interesting
that the
are of the same form as such
of BACON
considers

and

uses the same strategy of reasoning


for the
Since BACON
facts and for the intermediate
"atomic"
results, Hin
too.
initial facts can be used as initial facts for BACON,
level

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

JAN

74

M.

ZYTKOW

AND

HERBERT

A.

SIMON

can

to Hintikka's
add a few words of justification
If
claim.
an
as
a
is
atomic
is
fact
agree,
"pluto
dog"
philosophers
generally
what we know about processing
in
then, considering
complexity
we
can
vision
and pattern
that
AE
computer
agree
recognition,
statements may be facts, too, if their scope is limited to the duration of
We

observation.

Heuristics

GLAUBER'S

FORM-LAW
heuristic of GLAUBER
conclusions
very
produces
to BACON's
similar in their logical structure
FIND-LINEAR.
They
can be represented
in first order
logic in the form of a pair of
statements

The

Ax Ey Pxyc

and

Ay Ex Pxyc,

in the case of two or more classes of differing values, x and y, and one
c in common. They can be represented
in the form
object
Ax Pxa

in the case of one class of members


x, and one object a in common.
can
on
be
in a way similar to BACON's
These
combined
regularities
In this way a law (Ax)
the higher
levels of GLAUBER'S
recursion.
in two steps. An example of the last
Rxyz can be produced
(Ay)(Ez)
is: "For all acids and all alkalis there is a salt which
schema
is their
as
as
All
of
reaction".
these
inferences
in
well
BACON's
product
as inductive
can be classified
start
schemes.
with
ferences
They
facts and conclude with generalizations
that are applicable
particular
to other facts and for that reason are subject to falsification.
Heuristics

and Transfer

of Justification

far we

that the discoveries


and
demonstrated
of BACON
are the conclusions
of inductive schemes, while the initial
are the initial facts provided
to both systems. Justification
is
premises
to the conclusions.
from the premises
transfered across these schemes
In principle,
the same schema may be used to justify conclusions
given
or to justify
are justified prior to the inference
that the premises
So

GLAUBER

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

SYSTEMS

OF

DISCOVERY

75

For a deductive
scheme of inference
premises
given the conclusions.
the first situation
is called deduction,
the second, reduction.
Initial data are taken to be true by BACON
and GLAUBER.
we
can
of
conceive
that
may
systems
Although
reject some data
and Simon 1986, Langley et al. 1987), we do
(STAHL system, Zytkow
not consider
such a possibility
in this paper. Conclusions
of inductive
are less certain
as additional
inferences
than the premises,
facts can
contradict
them. For exactly the same reason, conclusions
have larger
contents.
This refers only to possible
to
facts as opposed
empirical
if the latter contradict
because
the conclusion
then an
schema
of
must
inductive
inference
have
been
used.
as implemented,
BACON
and GLAUBER
do not insist on
However,
exact fits to the laws they find. BACON
small deviations
of
accepts
observed
from theoretical
values
and GLAUBER
values,
permits
some exceptions
from its generalizations.
In both cases, the precision
in the programs.
required is controlled
by parameters
known

facts

incorrect

Bootstrap

Confirmation

inDiscovery

Systems

Scientific
role of a
laws, some of them at least, play the double
a
definition
and a genuine empirical
law
which
Can
is used
regularity.
as a definition
be also empirically
confirmed?
confirmation
Bootstrap
criterion
answer to this ques
1980) justifies an affirmative
(Glymour
are
tion. Bootstrap
confirmation
that
situations
requires
empirical
on
such
that
the
values
defined
based
the
tested
possible
(computed)
can eventually
contradict
the hypothesis.
The bootstrap
hypothesis
can be used
to justify some
criterion
confirmation
inferences
in
In
the
criterion
cari
the
for
schema
BACON,
systems.
discovery
justify
introduction
of intrinsic variables.
a simple version of Ohm's
Consider
law (we adapt
rediscovering
et al. 1987, chapter 4). Experiments
here an example
from Langley
a simple circuit
consist
in arranging
(a battery and a resistor) and
the current in the circuit (I, dependent
numerical
variable).
measuring
Let the two independent
variables
under the control of the experi
menter
be voltage
and the choice
of the
(V, numerical
variable)
resistor wire (W, nominal variable).
In the initial sequence
of experi
ments
let voltage
be fixed as vu and let the experimenter
vary the
resistor wire, each time measuring
the resultant current. Since W is
a new numerical
introduces
variable
that we will
nominal, BACON

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

76

JAN

M.

ZYTKOW

AND

HERBERT

A.

SIMON

and defines
the reciprocal of the resistance),
call C (the conductance,
currents.
its values for all the wires by the values of the corresponding
FIND-LINEAR-REGULARITY
its
this, BACON
applies
Following
that / is linearly related to C with the slope (S)
heuristic and discovers
in view of
of 1.0 and the intercept of zero. This result is tautological
were
determined.
the values of C
the way in which
vari
varies the next independent
the experimenter
But later, when
same
of
of
the
the
collection
able V, while using the
wires,
findings
same
uses
the
The system
justification.
system are subject to genuine
circuit. For the valuers
value of C for the same wire placed in whatever
the
Then,
v2 and v3 of V, the slopes will be s2 and s3, respectively.
to V and S gives in
of
FIND-LINEAR-REGULARITY
application
S = aV
that a = V\. Taken
result S = aV. We may notice
together,
=
some
to Ohm's
values
of I,
law. If
and I
SC are equivalent
law would not
from the actual values, Ohm's
however, were different
the example
and three resistors
With
three batteries
be discovered.
the values of
includes nine facts. Three of them are used for defining
the propor
is used to calculate
C. Another
s2, and thus to propose
for
facts
all
other
confirmation
V.
But
and
S
between
provide
tionality
the discovered
regularities.
in discovery
of justification
In our discussion
systems, we have seen
that the very process of discovery,
along the lines of the systems we
the
involves
have described,
against
testing tentative generalizations
it has also been
available data. By the time a law has been discovered,
a large part of
to be consistent with these data. Hence,
demonstrated
law takes place during the process of
of any scientific
the verification
new data that may be
involves
verification
it. Further
discovering
to
the level
Of course, whatever
the discovery.
obtained
subsequently
no
there is
attained at the time of discovery,
of verification
guarantee
that new data will

Norms

inDiscovery

not

invalidate

the law that has been

found.

Processes

can be studied as facts, although


is little doubt that discoveries
There
some philosophers
of science are sure (Popper 1959, p. 32) that such
an enterprise
But can discoveries
be studied as norms?
is hopeless.
do better work
Some scientists
is definitely
affirmative.
The answer
teach better methods
than others. Some courses, and some textbooks
can be studied
and the
evaluations
These
intuitive
than others.

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

SYSTEMS

OF DISCOVERY

as discovery
reconstructed
systems. For
we
of norms,
three
may
types
distinguish
analysis
norms.
and SHOULD
MUST, MAY,
methods

77

the purpose of our


which we will call

norms

specify things that have to be done in order


the
goals. For instance: "In order for a statement
accomplish
a theorem, a proof MUST
be produced".
be considered
1. MUST

to
to

for a variety of ways


in
for the proof
the goal. A procedure
that searches
reaching
a MAY
norm if either this or another procedure MAY
embodies
the goal.
be used in order to accomplish
2. MAY

norms

are weaker

and

allow

are largely of the MAY


almost any
Scientific procedures
type because
can be achieved
in many ways. Among
them there
scientific objective
so poor that they seem
are better and worse methods,
and methods
futile to pursue even if better ones are not known.
there is no basis for choosing
the optimal method;
are applied
are "good
that experience
has shown
are
The
called
methods
enough".
"good-enough"
satisficing
too,
methods,
(Simon 1956, p. 129). We may call them SHOULD
and fhey are primary targets for discovery
systems:

3. Usually
methods

as a branch of the theory of


can be viewed
normative
processes
theory of discovery
a class
we wish
to
Given
of computational
computational
complexity.
problems,
on average;
that are most
efficient
find the most
discover
or, if we cannot
algorithms
some
at least to discover
that make
in
available
efficient,
good use of the information

The

the problem

situation

(Simon

1973,

p. 477).

are improved,
Of course, when
methods
the improved
SHOULD
versions
the old ones. In terms of our earlier comparison
supersede
norms correspond
to logic as a
between
search and logic, MUST
norms
to various
non-deterministic
MAY
algorithm,
correspond
of heuristics
and control,
while
SHOULD
systems
superimposed
norms correspond
to satisficing
that are relatively
theorem-provers
efficient.
to
In characterizing
different
types of norms we referred norms
norms.
are
to
source
of
crucial
validation
Goals
the
is
the
What
goals.
on
concentrates
of goals
for discovery
One
systems?
approach
to reconstruct
and attempts
scientific practice
the goals of real science
(scientists)
concept

and the relations among them. Law and model


to this category
and instrument
creation,
belong

generation,
of activity.

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

M.

JAN

78

ZYTKOW

AND

HERBERT

A.

SIMON

starts from goals stated in Plato's


fashion. Those
approach
and so forth. They need to be
truth, simplicity,
justification,
in order to become
In this process a favorite
constructive.
interpreted
can
be used of truth, justification,
and the like.
logical definition
Normative
systems may use both approaches. We are more
discovery
inclined towards the first type of goals, but may admit the use of the
Another
be

may

second

type.
the comparison
between
logic and search,
goals delimit
are satisfied
of
search spaces, while additional
efficiency
requirements
search.
for
heuristic
selective
by arranging
Within

Normative

Discovery

Systems:

Research

Paradigm

a research paradigm
with discovery
for
systems provides
Experience
This paradigm
to
the normative
both
theory of discovery.
belongs
of science.
artificial intelligence
and to the philosophy
1. Take

one

or

computer
system
that will
program
make
Rather,
up
the goal
instance,
goal of finding a

a few scientific
a
goals at a time and construct
a
that reaches
these goals. Do not try to develop
circumstances.
satisfy the goals under all possible
a reasonable
substitute
for the general
goal. For
of finding a law is understood
as the
BACON
by
law that can be decomposed
into linear regularities.

for discovery
Schemes of inference and control strategies
be abstracted
from case studies of particular
discoveries
advice of particular
scientists,
including ourselves.

systems can
or from the

to data. The program


its path in response
2. Let the program modify
it can adapt its computation
becomes more efficient
this way, because
to the particular
situation described
by the data. This can be done by
an
on
of the data (GLAUBER
the
basis
does
operator
instantiating
or
of
the
the
choice
its FIND-LAW
this with
by making
operator),
does this when it creates
depend on the data (BACON
a
data
driven. A data-driven
is
called
Such
program
concept).
not
to
the
be
gram may
only
input data but also
responsive
data or theories it has inferred.
intermediate

operator

a new
pro
to the

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

SYSTEMS

OF

DISCOVERY

79

the normative
3. Validate
the range of goals it
system by investigating
can reach and its efficiency
in attaining
the goals. Validation
employs
a combination
of experimentation
and theoretical
analysis. Theoretical
forms that a law
analysis can tell, for instance, the range of functional
can

we are usually not sure


In complex
however,
situations,
whether
the given
of the actual
abstraction
theoretical
computer
to the system itself. It is usually much
system is adequate
discovery
to let the system run on different bodies
easier and very enlightening
of initial data and to watch
its performance.
One
of the major
of having a computer
system is the capability of examining
advantages
it experimentally.
We can let the system run on a variety of data while
same
the output
program. We can let the program "forget"
using the
of the former computations,
each time starting from a "tabula rasa"
take.

state of knowledge,
which
is impossible
in experiments
with humans.
can vary the program
and compare
under
the results obtained
to the evaluation
of
different versions. All these tests may contribute
the system.

We

Individual components
of the system may also be subject to validation.
A concrete
rule of inference
that generates
particular
periodic
pat
terns is justified by a requirement
for periodicity
heuristics. Periodicity
in turn, along with monotonicity,
conservation
heuristics,
symmetry,
and so forth, are justified by a requirement
for pattern
heuristics,
(regularity) detection.
Evaluation
criteria may vary from one system to another, as these
criteria we use for
systems differ in their tasks. Whatever
acceptance
the purpose
of evaluation
of a given
system, we admit that these
criteria are neither necessary
in
(the same goals may be achieved
different ways,
for many MAY
norms) nor sufficient
allowing
(the
realistic criteria are subject to exceptions).
What we are really trying
to evaluate
is whether
are "conducive"
the system's
to
processes
discovering
regularities.
The normative
approach

to discovery
of
consists,
therefore,
and sets of criteria
parts: systems of discovery
complementary
judging the efficacy and efficiency of these systems.
4. Cumulate
systems
them to interact with

of discovery
into a larger
each other. This procedure

two
for

system by allowing
is best illustrated by

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

M.

JAN

80

a number

ZYTKOW

AND

HERBERT

A.

SIMON

in (Langley et al. 1987, chapter 9). Another


a
to its
is by responding
way of developing
system of discovery
a
overcomes
some
it into
limitations and imbedding
larger system that
in
the FAHRENHEIT
of them. In this way BACON
is imbedded
with a search for the scope of laws.
system which augments BACON
uses BACON
both to find a law and to
In doing this FAHRENHEIT
find

of examples

the scope

of that law.

individual
the scientific
5. Reconstruct
system of goals. Each
system
of discovery
takes care of limited goals. If the individual
systems are
between
the need for understanding
interrelations
combined
together,
for
of
is
also
of
crucial
validation
becomes
crucial.
Clarity
goals
goals
In the descriptive
normative
clarity of goals
systems.
perspective,
in its variety.
of
science
crucial for a comprehensive
understanding

is

in philosophy
of science,
and other
Long experience
psychology,
that
the
of
reveals
disciplines
going beyond
analysis
single goals or
small systems of goals presents
formidable
research problems.
Here
simulation may be essential
the aid of computer
and by the experi
we may come
to understand
mental
study of systems of discovery
systems of goals better.
complex
in science may be generally
Parts of the goal/subgoal
hierarchy
science. For instance,
for the whole
the goal of extending
applicable
the scope of the laws to the new range of a variable may call for the
that work
in the new range. But
devices
construction
of measuring
too. Thus
there are domain-dependent
the
relations,
goal/subgoal
a particular
tissue-slice
method
biochemical
may be called within
to analytical
is relevant
the liquid electrode method
chem
context,
of domain
(maybe even prevalence)
istry, and so forth. The existence
of the scientific goal structure.
dependent
goals adds to the complexity
from
and
separate
discovery
justification.
Discovery
are
at
An
in
intertwined
science.
processes
attempt
justification
manner
of
would
in
the
them
hypothetico-deductivism
separating
of discovery
the performance
systems.
handicap
6. Do

not

can be separated
and discovery
course,
conceptually,
justification
the systems of discovery may be studied from both perspectives.
and GLAUBER,
Such an analysis, when applied to BACON
demon

Of

and

strates

that

regularities

justified.
Since a typical

discovery

are discovered
process

at

proceeds

the

same

time

in a sequence

they

are

of steps,

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

SYSTEMS

OF

DISCOVERY

81

which may be very long, and since each successive


step may be guided
accumulated
steps, by the time the
by the information
by the previous
is completed
will already
the law that has been discovered
process
have been tested against
the evidence
and found consistent
with
it.
a
is
there
considerable
and
between
Hence,
overlap
discovery
justification
law against

In fact, justification
processes.
new evidence
that had not been

is needed
considered

a
only to match
as a part of the

(Simon 1973, p. 475).


process
discovery
and justification may be described
The relation between
discovery
in Bayesian
since it
process,
step in the discovery
language: Each
of the emerging
involves confrontation
hypothesis with data, increases
at the moment
the probability
of the hypothesis.
Its probability
when
the discovery
process has been completed
may therefore be taken as
the prior probability
in a subsequent
confirmation
process.
Bayesian

Descriptive

and Normative

View

of Discovery

Systems

to codify
been many
the
attempts
history
throughout
are the well-known
accounts
method.
There
of Bacon,
and Mill, as well as hundreds of accounts
of scientific method
Newton,
0
in textbooks
of all times. The main problem with all these studies is
that they leave a gap between
the advice
and the
they provide
of this advice
in the practice
of making
A
discoveries.
application
a strategy of discovery
in a detailed,
program
computer
incorporates
it is executed,
it actually makes
dis
way, so that, when
procedural
There

have

scientific

and exact. This


efficient
make
the execution
Computers
new
to
the
from
the contemplative
distinguishes
approach
discovery
manner
in which we can use descriptive
of the traditional
principles
coveries.

advice.

In addition to the normative


another paradigm
discovery
paradigm,
a paradigm
also stems from the systems of discovery,
that is useful for
and for sciences
of science. The
science
latter paradigm
cognitive
one because
differs from the normative
it aims at different goals. The
normative
efficient
that makes
deals with
software
dis
paradigm
one deals with software
while
the descriptive
that models
coveries,
or reasoning
of an epoch
of an in
reasoning
(history of science)
dividual

Both the descriptive


and the normative
(cognitive modeling).
can
same
to
lead
the
of
when we study
system
paradigms
discovery
scientists.
highly productive

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

JAN

82

M.

ZYTKOW

AND

HERBERT

A.

SIMON

in this way may reproduce


the scientific work,
The system created
The same system may also des
justified descriptively.
thereby being
therefore
cribe the best available way of making
discoveries,
being
Of course, we are very far
perspective.
justified from the normative
in the multitude
of
from creating a system that would model a scientist
same
and in the variety of means
he uses. The
tasks he performs
of
system may also play both roles in the early stage of development
The early systems of discovery
the two research programs.
neither
(and are not good norms for
grasp a large scope of scientific activities
nor
reconstruct
human scientific
in great detail.
this reason),
activity
- or
equally badly.
They perform both tasks equally well
be an adequate
of the
Can
the same method
(true) description
method
of a particular
scientist or a particular
epoch, and at the same
can be examined
the same method
time be a valid norm? Certainly
and normative. We can confront
descriptive
perspectives:
or about particular
facts about science
with historical
the method
to those facts. On the other
to see how closely
it conforms
scientists
it with
the scientific
and see what
hand, we can confront
goals
are
this
the
method
Both
towards
offers.
progress
goals
perspectives
a
one
that
not
and
the
other.
method
satisfies
may
independent
satisfy
the method
of a good scientist,
if we reconstruct
However,
adequately
from both

and descriptive
criteria will be satisfied.
both normative
to the dis
facts and values corresponds
The distinction
between
that study the facts of real science
tinction between
disciplines
(psy
and
that
economy,
chology,
sociology,
history),
disciplines
study the
of
of
the
scientific
method
science,
logic,
epis
validity
(philosophy
to
return
distinction
in
of
We
will
this
the
second
the
part
temology).
of the contexts of justification
and discovery.
paper in our discussion
deal more with norms than
In fact, the existing systems of discovery
how scientific discoveries
facts. They usually demonstrate
might
the
have been made or may be made, and rarely claim that they model
actual path of discovery.

with

Conclusions
an existence
have presented
systems of
proof for normative
and
the
of
such
the
construction
extrapolation
systems
by
discovery
by
for further research. Since the best existence
of them into a paradigm
do we need to further justify our claim? A
proofs are by construction,
a
to
is that they are merely
the
systems
existing
typical objection
So far we

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

SYSTEMS

OF

DISCOVERY

83

we
of real-life
imitation
Before
discovering
capabilities.
us
even
note
that
if self-contained
this claim
let
examine
systems
to the remote future,
belong
capable of making
impressive discoveries
can benefit
of science
and even
scientists
from the
philosophers
us
recall the analogy with the formal notion of
systems. Let
existing
proof. This notion is only remotely akin to real proofs in mathematics
are second-rate
and the existing
theorem provers
imitations of the
remote

mathematicians'
capability of proving. But the notion of formal proof
and related logical concepts
exerted a considerable
influence on the
on
of science
and
the foundations
of mathematics.
The
philosophy
are ready to play a similar role.
of discovery
normative
systems
if self-contained
the existing
Further,
systems are yet to be developed,
to claim that
support the normative
systems
paradigm well enough
are a legitimate program within
normative
the
systems of discovery
of science.
philosophy
Are our systems a remote imitation of discovering
Yes
capabilities?
a
and no. If we consider
of
actions
that
the whole
variety
capable
scientist may perform, our systems are far behind. But if we consider
tasks in their areas of strength,
limited scientific
they are not doing
worse
and are perhaps
than real-life
scientists. When
doing better
BACON
faces a particular
of data, it discovers
collection
Kepler's
laws or Black's
law (including
the notion of specific heat) in a more
or Black. GLAUBER
efficient and elegant way than Kepler
does not
in developing
the regularity
yield to Johann Glauber
involving acids,
alkali, and salts.
Additional
features can and are added to the existing
systems, and
we do not see limits to the range of discoveries
they will be able to
make. The method
used in the development
of discovery
is
systems
not a search for the philosophers'
stone but rather a mundane
work of
which
adds more
features
and increases
the
gradual
improvement
of
systems would satisfy the opponents
capability of the systems. What
to discovery
we cannot say. But refusal to
the normative
approach
accept

discovery

systems

that

actually

make

discoveries

is merely

dogmatism.

2.

EVALUATING

THE

EXISTING

STEREOTYPES

of science
in the philosophy
Finally, we wish to discuss two stereotypes
of
niche for normative
that claim that the ecological
systems
discovery

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

JAN

84

M.

ZYTKOW

AND

HERBERT

are the stereotypes


does not exist. These
and "the context of discovery".

of Discovery"

"Logic

vs. Discovery

A.

of "the

SIMON

logic of discovery"

Systems

In the narrow sense it may be represen


What
is a logic of discovery?
such that, given any data, itmechanically
ted as an inference machine
applies rules of inference and comes up with a theory to explain these
of in
data (Hempel
inductive
schemes
1966, p. 14). In addition,
ferences were required to be logically
1959,
28),
p.
justified (Popper
a unique, best theory. In the tradition of the twentieth
and to produce
was claimed
to be
of science
such a mechanism
century philosophy
non-existential

and not possible

(Popper

1959, p. 29; Hempel

1966, p.

15).
Our

to heuristic
of the concepts
search with
related
comparison
can
that
be con
demonstrates
systems
logical concepts
discovery
in logical terms and that they can be parts of the logic of
sidered
if only we relax the criteria for such a logic. Three criteria
discovery,
to be lowered:
of inferences,
need
(logical validity)
non-fallibility
of the unique, best solution, and applicability
of one system
existence
to all data.
In the logical tradition, the label of "validity"
is often limited to an
infallible rule of inference or to a system of such rules. We do not
to
intend to seek non-fallible
results as it is unreasonable
inductive
treat validity as a relative notion:
seek the impossible. We
there are
more and less valid systems. Even
if it is not clear how a given system
can be improved,
it worse. Thus we see room for
it is easy to make
we
even
not
if
Bacon's
do
share Francis
inductive
overly
logic
a
is
that
method
that
leads
belief
there
infallibly to scientific
optimistic
errors. In a similar way we reject the task of
truth, without producing
In doing
best
this we follow
the
and
conclusions.
seeking
unique
that dominates
of science.
solves
Science
the actual practice
of uniqueness
the problems
and non-fallibility
them.
by ignoring
on
some
of
will
work
fine
of
data
and
types
discovery
Finally, systems
learn only very slowly,
will fail on some others. Scientists
and science
over centuries,
how to use new forms of data and how to seek for new
of
that summarize
them.
types
regularities
of discovery
Our
of
demonstrates
the
systems
existing
analysis
norms
can
in
that
used
between
be
discoveries
relationship
making
attitude

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

SYSTEMS

OF DISCOVERY

85

and rules of inference. These norms do not guarantee


(what
discovery
in
efficiency
logic could do that?), nor do they guarantee maximum
a
a
to
of
It would be unreasonable
expect
logic
discovery.
making
a logic of discovery
to meet either of these criteria. What
discovery
has been done by several systems of
should do, and what we believe
that would con
is to provide
rules of procedure
heuristic
discovery,
a
to make
stitute good advice for someone desiring
discovery. A logic
of a reasonable
scientist
hence a set of
of discovery
is a description
norms for doing science.

Context

of Discovery

vs. Context

of Justification

was for a long time exorcised


from the
of discovery
on
to
and
science
passed
psychologists,
"generously"
philosophy
as a nonscientific
task entirely
when not considered
(Popper 1959, p.
in
becomes
if talking about discovery
32). Even
popular
increasingly
The

context

of

of science,
it is rather talking about discoveries
the philosophy
actually
of historical
reconstructions
made
cases), not about the nor
(usually
It was the normative
mative
processes.
aspect that
aspect of discovery
was denied
its place under the sun by a great number of philosophers
if the existing
in the tradition of logical positivism.
Even
systems of
a
create
the
and
make
strong enough
discovery
paradigm
they
for themselves,
the context
of dis
the vigor with which
argument
an
was
to
in
the
first
is
covery
place
rejected
interesting phenomenon
so
were
context
the
the
of
arguments
against
discovery
explain. Why
to these arguments
breach
is caused by
agreed on? What
widely
the existing discovery
systems? In order to address these questions we
will first recall

The Distinction,

the problem

and the argument.

the Thesis,

the Argument,

and

the Reasons

occurs
in the posi
simple and tempting distinction
repeatedly
us
Let
it in a quo
recall
literature.
tivistically oriented
philosophical
tation from Feigl:

The

It is one

at our

thing

and what

thing to ask how we arrive


to ask what
sort of evidence

the testing,
the confirmation
govern
of knowledge
claims of science
(Feigl

scientific

knowledge

general,
or disconfirmation
1965,

claims...

it is another

rules and standards


objective
or rejection
and the acceptance

p. 472).

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

JAN

86

in the passage

and
...

M.

AND

ZYTKOW

A.

HERBERT

SIMON

from Reichenbach:

to remain
in correspondence
tendency
to obtain
valid
the tendency
thinking;
task. (Reichenbach
and the critical
descriptive
the

with

from

and

actual

so we
1938,

have

thinking must
to distinguish

be

separated
between
the

p. 7)

were applied
in both quotations:
Different
distinctions
the
conceptual
to
first opposes
while
the
second
confronts
justification
discovery,
norms with descriptions.
The two pairs of concepts
used at the same
the borderline
time determine
between
the context of discovery
and
the context of justification.
both contexts
The distinction
between
is
is of interest
followed by the claim that only the context of justification
to philosophy.
For Reichenbach,
and for many
others
"... episte
in
context
is
of
the
mology
only occupied
constructing
justification"
in
claim can be illustrated
(Reichenbach
1938, p. 7). This dominant
II.

Table

TABLE II.
Discovered
How

are

laws

Justified

(A) Psychology
context

(descriptions)

(B) Psychology

of

discovery
How

should

laws be

(C) Logic

of discovery

non-existent,
not possible

(norms)

(D) Epistemology
context

of

justification

the
of
the emptiness
(C) are essentially
a previous
we
in
discussed
that
of
discovery
logic
arguments
against
to be empty both actually
and
section. The area (C) was believed
Another
of
seemed
the
because
impossible.
discovery
logic
potentially
refers to (D). In order to keep the intended
important observation
to distinguish
it is important
the two contexts,
between
distinction
a traditional
In
context
of
from
the
justification.
justification processes
context
to
while
a
the
of
discovery
finding
proof belongs
example,
even
context
But
to
of
a
the
justification.
proof belongs
checking
in
Actual
work.
creative
a
involve
proofs
proof may
checking
of
never
standard
the
mathematics
proof proclaimed
satisfy
virtually
This
thinking to be understood.
by formal logic, and require creative
The

arguments

for

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

SYSTEMS

OF

DISCOVERY

87

to the
is not an isolated example. Most
processes
justification
belong
context of discovery.
Take, for instance, Popper's urge for severe tests
severe tests is nothing
of hypotheses.
less than discovering
Designing
new experimental
to the
and although
this belongs
arrangements,
to the context of
it certainly does not belong
process of justification,
in the justification
is the
processes
Similarly
important
justification.
new methods
of preparation,
etc.
of new instruments,
discovery
In conclusion,
if the area (D) is determined
"how
by the question
to the context
should laws be justified?",
then it only partially belongs
of justification,
to the context
because
it also partially
of
belongs
are intertwined
contexts
Both
and the scope
of the
discovery.
to tests that answer
context
is very limited. It is reduced
justification
the question
"Is such and such justification
valid?", and is destitute of
constructive
if so limited,
of
has the context
then,
steps. Why
justification
the context

of science? And why was


dominated
logic and philosophy
of discovery
Let us address both questions
in
banned?

sequence.

reasons made
Several
the context of
on an attractive
formal
domain,

It was
attractive.
justification
This
formal,
logic.
promised
- all
and general
results
abstract,
difficult,
analytical,
precise,
on logical foundations
of
scientifically
By concentrating
prestigious.
could
abstract
from
of
the
processes
justification
philosophers
scientific
and avoid criticism of their results by scientists.
justification
of the new field. Since
This gave them an exclusive
the
possession
of mathematics
in the foundations
that were
exciting
developments
based

to research
in the
logic could be also applied
yielded
by formal
some philosophers
of science,
that they might
foundations
believed
or right
even
whether
tell scientists
in their
wrong
they were
the domain of expertise was com
justification
procedures.
Finally,
was supposed
to enter the stage only
fortably limited. The philosopher
of the solution were
after all pieces
for his inspection.
arranged
all pieces
of the puzzle
else had to organize
Someone
and the
was supposed
to check whether
the proposed
solution was
philosopher
to carry out this job, the philosopher
really a solution. To be prepared
has had to carry out a prior task: finding the examination
criteria and
the way to apply them.
In contrast,
the context
reconstructions

Attempted
ingly far from satisfying

of discovery
any payoffs.
hardly offered
of the creative
process were embarrass
the scientific standards of formalism, general

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

88

JAN

M.

ZYTKOW

AND

HERBERT

A.

SIMON

a
and testability. No
offered
ity of principles,
respected
discipline
source of transfer of ideas and solutions.
The
idea of a logic of
was dismissed.
was con
While
the scope of attention
discovery
narrow
context
well
for
work
within
of
and
determined
the
veniently
or
area
to
to
reconstruct
of
focus
the
any
required
explain
justification,
to be overwhelmingly
broad. By narrowing
it
seemed
given discovery
we
ad
since
consider
the
down we risk the explanation
hoc,
being
only
is relevant
relevant
facts. But how do we know what
before
the
has been made?
discovery
But an even more principled
that
problem was seen. Big discoveries
of
attracted
and
attention
influenced
the
image
particularly
discovery
were said to go beyond
structures
the limits of the existing conceptual
and beyond
the existing ways of thinking. Thus no matter how big the
was
in which
the discovery
framework
the discovery
considered,
consists

Reasons

in going

beyond

that framework.

for Rehabilitation

the context of discovery


have lost
The arguments
against
apparently
their power, while at the same time attractive
features of the context
to the context of discovery.
In
of justification
have become
applicable
in
form
of
be
the
fact, discovery
programs may
systems
computer
to people who
like detail, precision,
and formalism
the
exciting
in the
that made
the context of justification
attractive. Work
features
area of discovery
of research as the areas of
systems offers autonomy
to which
and
machine
artificial
learning
they belong
intelligence
are
own
methods.
The
their
systems of discovery
existing
develop
on
it
of
has
limits
the
will
for
them
clear
data
"closed
each
systems"
it can make.
and the discoveries
these systems
handle
However,
some discoveries,
to simulate
and
how little is required
us
a
matter
to
that
it
is
of
construction
convince
practical
they
how
of
how much
in
cir
demonstrate
little)
(or perhaps
complexity
to create an effective
cumstances
is enough
and associations
discovery
demonstrate

system.

of
in attitude was brought
about by the development
on
and
heuristic
search
fast
systems
computers.
symbolic processing
we can
of systems of discovery
Instead of arguing for the possibility
now present a constructive
in
form
of
the
running systems. The
proof
are
of the field.
the
for
the
systems
plausibility
existing
key argument
The

change

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

NORMATIVE

SYSTEMS

OF

DISCOVERY

89

an early stage of development


if they represent
of discovery
can
structure
from
their
and
behavior
be
systems, enough
extrapolated
to create the vision of a research program.
The new approach
is strengthened
that
by the empirical possibilities
of programs
it offers, and the cumulation
that may someday go beyond
are very patient
and
human
programs
comprehension.
Computer
in what
in humans.
virtues
uncharacteristic
they do
systematic
can test ideas that we humans accept
Therefore
programs
computer
but are not able to follow in accurate detail.
Even

REFERENCES
of
4: The Discovery
'BACON
and H. A. Simon:
1980,
of the Canadian
Conference
Proceedings
of the Third National
Studies of Intelligence,
pp. 19-25.
for Computational
of Production
'Model-Directed
B. G. and T. M. Mitchell:
1978,
Learning
and F. Hayes-Roth
in A. Waterman
(eds.), Pattern-Directed
Inference Systems,

Bradshaw,
Intrinsic

P. Langley,

G.,

Properties',

Society
Buchanan,
Rules',
Academic

Press, New York.


and J. Lederberg:
E. A., B. G. Buchanan,
The DENDRAL
A Case
Study Using
Solving:

Feigenbaum,
Problem

Press,
University
telligence 6, Edinburgh
of Science',
1965,
'Philosophy
Feigl, H.:
Cliffs.
Prentice-Hall,
Englewood
Glymour,

C:

Hempel,
Hintikka,

C. G.:
J.:

1980,

Edinburgh.
in R. M. Chisholm

and

'On Generality
Machine

1971,

In

Program',
et al.

(eds.),

Philosophy,

Princeton
and Evidence,
Press, Princeton.
University
Cliffs.
Prentice-Hall,
Science,
of Natural
Englewood
Philosophy
of Experimental
'What is the Logic
74, next
Synthese
Inquiry?',

Theory

1966,
1988,

issue.
Langley,
Second

P. W.:
National

1978,

'BACON

Conference

of

1: A General

Discovery

the Canadian

Society

System',

Proceedings

for Computational

of

Studies,

the
pp.

173-80.
P. W.:

Langley,
31-54.

1981,

'Data-Driven

Discovery

of Physical

Laws',

Cognitive

Science

5,

and H. A. Simon:
with
P., G. Bradshaw,
1983,
Langley,
'Rediscovering
Chemistry
J. G. Carbonell,
and T. M. Mitchell
BACON
4', in R. S. Michalski,
(eds.), Machine
An Artificial
Press, Palo Alto.
Learning:
Intelligence
Approach,
Tioga
M.
J.
G.
and
H.
A.
Simon:
'Three Facets
of
P.,
Bradshaw,
1983a,
Langley,
Zytkow,
on
Joint Conference
Scientific
International
Proceedings
of the Eighth
Discovery',
Artificial

Intelligence.
H. A.
P. W.,

G. Bradshaw,
and J. M.
Simon,
1987,
Scientific
Zytkow:
MIT Press, Boston.
Processes,
Discovery:
Computer
of the Creative
Explorations
in Mathematics',
'Automated
Formation
Lenat, D. B.: 1977,
of the
Proceedings
Theory
on Artificial
Joint Conference
pp. 833-42.
Fifth International
Intelligence,
of Heuristics
D. B.:
Three
Case
'The Role
in Learning
Lenat,
1983,
by Discovery:

Langley,

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

90

M.

JAN

in R.

Studies',

AND

ZYTKOW

J. G.

S. Michalski,

HERBERT

Carbonell,

A.

and T. M.

SIMON

Mitchell

(eds.), Machine

An

Press, Palo Alto.


Artificial
Inellligence
Approach,
Tioga
'An Application
of AI Techniques
R. S. and R. E. Stepp:
1981,
Into an Optimal
Objects
Conceptual
of
Structuring
Proceedings
Hierarchy',
on Artificial
Joint Conference
Seventh
International
pp. 460-65.
Intelligence,

Learning:
Michalski,

K.:

Popper,

1959,
H.:

Reichenbach,

The

Logic

1938,

of Scientific Discovery,
and Prediction,

Experience

Basic

Books,

University

New

to
the

York.

of Chicago

Press,

Chi

cago.
Choice
of the Environment',
'Rational
H. A.:
and the Structure
1956,
Psy
Review
63, 129-38.
chological
a Logic?',
'Does Scientific
Have
Simon, H. A.:
1973,
Discovery
of Science
Philosophy
40, 471-80.
Simon,

J. M.
Zytkow,
Construction

and

H.

A.

Simon:

of Componential

1986,

Models',

'A Theory
of Historical
Discovery:
Machine
1, 107-36.
Learning

Science
Department
Computer
Wichita
State University
67208
Kansas
Wichita,
U.S.A.
and
of Psychology

Department
Carnegie-Mellon
Pittsburgh,
U.S.A.

University

Pennsylvania

15213

This content downloaded by the authorized user from 192.168.82.206 on Mon, 19 Nov 2012 14:16:00 PM
All use subject to JSTOR Terms and Conditions

The

Anda mungkin juga menyukai