Professional Documents
Culture Documents
Carnegie-Melloft Vnioersity
lems for the programs (as defined by the number of variants of the
program that failed to solve them), Additional evidence is needed before
we can be satisfied that the theory gives even an approximate explica-
tion of the human processes.
The purpose of the present paper is to present and examine a larger
body of empirical evidence, from the behavior of 14 Ss during the prob-
lem solving process; the order in which they looked at letters in a
sequence, the hypotheses they formed about probable solutions, their
methods for recognizing periodicity in the sequence, the manner in
which they extrapolated it once they had discovered the concept, the
types of errors they made, and so on. On the basis of this evidence, we
can test the detailed hypotheses that are implicit in the organization and
behavior of the computer program.
Since the publication of our previous paper, our theory has been
examined and extended by several other investigators, including Pivar
and Finkelstein ( 1964), WiIIiams ( 1969 f , and KIahr and Wallace ( 1970).
In the following discussion, we will draw upon their investigations as
well as our own data. (See also, ,Restle, 1970; Simon & Sumner, 1968;
Gregg, 1967.)
HYPOTHESES
The basic components of the theory can be summarized in a set of
hypotheses about the characte~stics of the human behavior in this task.
Each hypothesis describes some important characteristic of the computer
program and its variants.
Processes
These hypotheses predict the manner in which Ss proceed through the
problem situation.
AL S first discovers the periodicity in the sequence, then induces the
pattern description (the rule) using this periodicity. (Example:
abmcdm . . . has the period 3-abm, cdm, . . . .)
A2. When extrapolating a sequence, S first initializes the rule (marks
his place on certain lists), then uses the relations of same (I, for
“identical”) next ( N) and buc?cwa7d next (BN), or predecessor,
to extrapolate.3
s We will later discuss a difference between the program and Ss behavior on this
point. The program executes its extrapolation by initiahzing at the bc&Gng of the
letter sequence as presented; Ss usually initialize at some point near the end of the
letter sequence, and extrapolate from there. Thus, the program, given abmcdm . . . ,
would initialize the alphabet at “a,” and produce the extrapolation, abmcdmefmg . . . .
The human S would more often initialize at “c,” say, and produce only cdmefmg . L. .
~EQUENTL~LPATTERN &QUISIT~~N 401
A3. S experiences difficulty when he must keep his place on more than
one list in extrapolating. (Information for place-keeping is as-
sumed to be held in short-term memory along with other dynamic
information during extrapolation.)
Mechanisms
These hypotheses are assertions about the mechanisms or abilities that
Ss have available (stored in long-term memory) for solving these kinds
of problems.
Individual Differences
While our main concern will be with the communalities among Ss,
we will also comment on individual differences, in order to indicate the
relative independence among the various submechanisms.
METHOD
Fourteen Ss were given 15 Thurstone-type letter series completion
problems. The Ss’ thinking-aloud verbalizations were recorded, and the
order in which they observed the letters in the sequence was auto-
matically recorded. The Ss only observed one letter of the sequence at a
time.
Subjects
The Ss were 14 Carnegie Institute of Technology (now Carnegie-
Mellon University) male and female freshmen and sophomore students
who were fulfilling an Introductory Psychology course requirement by
participating in psychology experiments.
Procedure
S was seated at a table containing the display device and pencils. A
microphone was placed about his neck and he then read a page of
standard thinking-aloud instructions. Next, he was given a series of prac-
tice problems and if he did not talk much while solving them, was en-
couraged to verbalize more. During the performance of the practice
problems, his questions about the procedures or the problems were an-
swered. At the end he was told the correct answers to the practice prob-
SEQUENTIAL PA’ITERS ACQUISITIOX 403
lems and asked if he had any questions. S then read the instructions for
use of the display device, after which the operation of the device was
demonstrated to him. The answer and information sheet was then
given to S.
He was told that while he should work through the problems as
rapidly as hc conveniently could, hc was to “work for accuracy and not
speed in performing the problems, since speed is not an important factor
in this experiment.” He was then told to begin the first problem. At the
conclusion of each problem, E a~~nounccd “OK,” to provide a clear indi-
cation on the tape of when the problem ended. The time for solution is
the time between the first punch (first letter looked at) and the “OK”
by E.
Results
A brief samhle of the. data obtained from the card punch records is
presented in Fig. 1. The tape-recorded verbalizations were transcribed,
but analysis had to involve the original tapes as well as the typewritten
versions in order to correlate S’s verbalizations with the particular letter
Subject l-Problem 13
Punching Sequence Verb& Protocol
WXAXYBYZCZADAB
Subject 8-Problem 12
Punching Sequence Verbal Protocol
NPAOQAPRAQSA
. . . . . .
he was observing. The tape was also used to identify voice pitch changes,
rhythmic punching, and voice cadences in naming letters-an important
source of information about the processes used by S-and to obtain so-
lution times.
In summary, the data we obtained were: what S said, his correct or
incorrect extrapolations, the order in which he looked at the letters, the
temporal patterns in his naming of letters, and the time it took him to
solve each problem.
The data were organized into six categories: measures of problem dif-
ficulty, the feature or features of each sequence first noticed by S, de-
scriptions of the concepts attained by S, the initialization of the concepts
prior to extrapolation, the method of extrapolation, and S’s errors and
sources of problem difficulty.
Dificulty
Numerous alternative measures of problem difficulty were computed
from the data: median and mean times and numbers of runs through
the sequence for all Ss, for Ss solving the problem, and for Ss not solving
it, and median and mean numbers of punches. The intercorelations
among these were so high (the lowest rank-order corelation was .81, and
most of the corelations were over .9) that they can all be represented
adequately by mean time for all S’s (Table 1) . Table 1 also shows the
number of Ss, from among our 14, solving each problem; the numbers
from the earlier groups of 12 and 67 Ss, respectively, reported in Simon
and Kotovsky (1963); the mean times for the earlier group of 12 Ss and
the list of problems solved and left unsolved by Variant C of the com-
puter program.
Table 2 gives the corelation coefficients (rank-order corelations, and
in one case the point biserial) among the difficulty measures of Table 1.
There are high corelations between the times required by the Ss in the
two previous studies, run under rather different experimental conditions.
The rather lower corelations between the items correct in the present
and previous studies is probably due mainly to the fact that the range
of the data was very small. A difference of four in the number of Ss
missing a problem in the present study meant a difference between first
and tenth rank in difficulty. Similarly, in the group of 12 Ss a difference
of four in the number of subjects obtaining a correct solution could de-
termine whether a problem ranked seventh or fourteenth in difficulty.
We shall see later that the specific conditions under which the Ss
worked affected the difficulty of solving particular problems. The I4 col-
lege Ss of the present study viewed the letters of each sequence singly,
as we have seen. The 67 high school Ss of the one earlier study were
40s KOTOVSKY AND SIMON
TABLE 2
Intercorrelations of Difficulty Measuresa
tested as a group, with all the letters and problems visible at once; the
12 SS of the other previous study were tested individually, with
all letters visible, and with the importance of speed emphasized in their
instructions.
Table 2 also shows high corelations among the difficulty measures of
all three studies (except the number-correct measure of the present
study), on the one hand, and the difficulties experienced with the differ-
ent problems by Variant C of the computer program, on the other.
Among the 14 Ss in the present study, the mean number of problems
missed (out of 15) was 3.9, and the range was 0 to 8.
TABLE 3
Priorities in Noticing Features”
Number of Ss
Problem Feat’ures present noticing first
cdcdcd (1) 2
N/ -
aaabbbcccdd 3
2 -
atbataatbat (1) 4
W) 1
abmcdmefmghm (1) 4
1
(2 -
5 defgefghfghi (1) 2
7
&( -
6 qxapxbqxa (1) 8
09 1
7 aduacuaeuabuaf (1) 7
(N -
WV -
8 mabmbcmcdm (1) 3
N/ 1
!I urtustuttu (1) 7
(N) -
10 abyabxabwab (1) 4
-
(fi& 2
11 rscdstdetuef (1) -
N/ 3
12 npaoqapraqsa (1) 11
1
(r; -
13 wxaxybyxczadab (1) 3
3
i% 2
14 jkqrklrslmst (1) -
N/ 3
15 pononmnmlmlk (1) 1
BN/ 5
(BN) -
Additional Features
Several features of the sequences were noticed and mentioned by the
SS that were not incorporated in our theory or programs. For example,
at least eight different Ss noticed in one or more problems, that some
items of the sequence were drawn from one end or another of the alpha-
bet. If we represent by B, M, and E letters belonging to the beginnhg,
middle, and end of the alphabet, respectively, then the sequence of Prob-
lem 4 might be abstracted as: BBMBBMBBMBBM, and the sequence of
P6 as MEBMEBMEB.
Another type of relation not hypothesized in the original model is
based on counting. Thus, P12 involves a “double next” relation (labeled
above N’), which can be derived simply by applying N twice. Similarly,
410 KOTOVSKY AND SIMON
Periodicity
After noticing some significant feature or features, most Ss find a
periodicity in the sequence that they then use in constructing a pattern
description. Where they do not find such a period or where they find
the wrong one they almost invariably fail to solve the problem. Explicit
verbalizations indicating the discovery of a periodicity before going on
to form a concept are frequent, occurring in 125 of the 210 problem
protocols (60%). This evidence supports hypothesis Al that “Ss first dis-
cover the periodicity in the sequence and then induce the rule using the
discovered periodicity.” The hypothesis is also supported by the eye
movement studies of Williams ( 1969).
The Ss’ behavior departs in one respect from the model. Periodicity
is determined by noticing I and N relations. In the computer program,
information about relations that are noticed at this stage is not retained,
but is regenerated during the second stage, when the pattern description
is being built up, Ss clearly retain much or all of this information, and
use it while building the pattern description. Thus, the current program
separates the two phases of problem-solving activity-detection of pe-
riod;city and pattern description-more sharply than do the Ss.
In coding the concepts, however, we did not force them into the Simon
and Kotovsky scheme, but introduced new elements and mechanisms as
dictated by the data, thus permitting an empirical test of how well the
original theory fit the new evidence.
Overall fit of data to model. The data on the concepts generated by
the Ss are summarized in Table 4. In the column headed “Program Con-
cept,” concepts that are identical to those proposed by Simon and Kotov-
sky ( 1963, Table 2, p. 538) are designated by + . For 12 of the 15 prob-
lems, the pattern description most frequently generated by Ss (including
one tie) is identical to that proposed in the earlier paper.
Some 177 concepts are described at a moderate (24) or high (153)
level of certainty (out of a possible total of 210). Of these, 103, or 58%,
were in complete detailed agreement with those proposed by Simon and
Kotovsky ( 1963)) while 74 were different from those proposed in 1963.
Of the latter, 16 were minor variants of those proposed. The remaining
33 concepts which failed to reach a high enough certainty level for in-
clusion in Table 4 do not noticeably differ, to the extent that partial
information is available, from those included in the table.
The general qualitative agreement between the pattern descriptions
attained by the S’s of this experiment and the pattern descriptions pro-
posed by the 1963 model supports hypothesis B3, that Ss possess a
pattern generator that solves patterns by generating and fixating a pat-
tern description associating the relations I, N, BN, with the several posi-
tions within the period of the sequence. This agreement, together with
the fact that Ss were able to use those concepts to extrapolate the se-
quences correctly in 80 of 97 cases, confirms hypothesis B4 that Ss possess
a sequence generator that can interpret such pattern descriptions.
Finally, the fact that most of the concepts (177 out of 210) were de-
scribable at moderate or high certainty levels in the pattern language,
supports hypothesis B5 that the language (in a somewhat expanded
form) proposed by Simon and Kotovsky gives an empirically valid expli-
cation of S’s internal representations of patterns and the behavior attend-
ant on their acquisition and use. Only one of the concepts (S2, P12) was
not easily describable in terms of the expanded language),
Concept 1 on P3, and Concept 2 on P6 confirm, for at least some Ss,
hypothesis B2 that Ss can induce a new alphabet from a given sequence
and apply the N relation to extrapolate it.
The circular next relation. Both model and most subjects treated al-
phabets in a circular fashion. That is to say, on the alphabet: a&, the N
operator applied to a yields b, and applied to .h yields a. The regular
English alphabet is also usually treated circularly, the N of; z being a.
Some Ss encountered difficulties with circular alphabets.
TABLE 4
Concepts
No. of Ss
1 9 1 [C, Ul +
2 7 1 1 [3M,, NO%)1 + M, = A, D 8
4 1 2 [MI, M,, MI, N(W,)l + 341 = c, D 4
3 2 1 [A, T, MI, NC.%)1 + M, = B on A, B 9
2 2 2 IA, T, B, A, T, Al 5
1 3 [A, A, T, B, A, Tl *
4 13 1 [MI, NOW, MI, NM), Ml + M, = G, H, E, A
k3
5 8 1 [N(M,), M, = M,, MI, N(M,), + M2 = F, D, E
MI, N(M,), MI, NOWWI B
4 Other
g
6 7 2 1 [Q, X, -4, P, X, Bl +
2 1 2 [MI, NWL), X, Mz, N(MdI + M, = Q on P, Q;
M2= AonB,A
7 5 1 1 [A, Mk, N(Mk), ul Mk = M, = D, F
(MI, N(MdDL, BN(Md)
I I 2 [A, Mt + 1 -+ M,, Mk, N(Mk), N(Md, ul M, = D, C
Mk = (Mt, NUW, MI)
(Mt, BNDW, MI)
Other
8 11 [MI, NW), MI, Ml + Ml = A, C
9 11 (U, MI, N(W), I-1 + M, = It, S, T
10 13 [A, B, MI, BNUWI M, = W, Y
11 8 WL, NM), MI, Mz, + Ml = T, S, R, E
NCWW M, = E, C, U
2 Other
12 6 1 WI, NWI), Ma N(K), Al + MI = N, P, Q
Mn = P, Q, It, S
4 2 [Ml = M,, M,, N(M,), Mz = MI, nh = Q
NW), MI, Al
Other
13 5 [MI, N(Md, MI, Mt, NW)1 + M, = W, A; Mz = A, D
1 Other
14 10 [Ml, NM), MI, MB, + M, = J, L, T, S
NUW WI M, = Q, S, M, L
2 Other
15 4 1 1 [M,, BN(M,)M,, BN(Mz), M, = P Mz = 0
Ma, BN(&)l MS = N
1 3 2 [Ml, BN(M& M, = MI, Ml,
BN(M,), Ml, Ml = %I
3 Other
a Indicates agreement with the rule proposed by Simon and Kotovsky, 1963.
* Example: M, = A, D; M, = E, T reads: some subjects set memory cell M, at letter A on the Alphabet, some set it at D; somesubjects
set memory cell M2 at E and some set it at T when beginning their ext,rapolation. Information in t.his column is occasionally less certain
than the other information on the Table.
414 KOTOVSKY AND SJMON
The backward alphabet. The first change in the model that is neces-
sary to fit the data is to eliminate the backward alphabet as a separate
entity to which the NEXT operator applies. The behavior of even those
SS who could work backward through the alphabet is more accurately
described by encoding the process by a backward next (BN), or prede-
cessor, operator that is applied to the ordinary alphabet, The possible
need for this alternative method of characterizing backward movement
in the alphabet was mentioned by Simon and Kotovsky (1963, footnote
3). The BN operator appears mainly in P7, PlO, and P15. With this
change in the model, 13 more Ss on PI0 and two more on P15 would be
brought into agreement with it, reducing the number of discrepant cases
from 74 to 59, and increasing the number of positive agreements from
103 to 118 (67%).
The BN operator is not used by all Ss, and even where used, is usually
not used with the same facility as the N operator. In fact, some Ss only
perform “quasi-BN” operations. They pick a point in the alphabet some-
what ahead of the letter they want to perform a BN upon; then moving
down the alphabet, using N, they remember the immediately preceding
letter as well as the current letter; so that when they reach the target let-
ter, they also retain in memory the one before it, i.e., the BN of the target
letter.
Hierarchic patterns. A second change in the original model is necessary
to describe more accurately the organization of the pattern descriptions
used by many Ss on P7, Pll, and P14. A means must be provided for
describing the hierarchic organization of these Ss’ pattern descriptions.
In the pair of similar problems, Pll and P14, the second concept in each
case involves a memory chunk we have labeled Mk, which points to a
position on a two-item list, One item on this list names the subpattern
(Ml, N(M2), M2). Th us, the pattern description for these problems
(rscdstde . . . . . and jkqrklrs . . . . . ) involves a higher level switching be-
tween the two subpatterns-( M,N( Mk))-and the production of dif-
ferent alphabetic sequences by each subpattern-( Ml, N( Ml), Ml)
and (M2, N(M2), M2). Th e introduction of such hierarchic elements
in the descriptions of concepts constituted the only difference between
eight pattern descriptions in this study and those predicted by Simon
and Kotovsky ( 1963).
Counting procedures. In addition to the two changes in the language
for describing pattern descriptions discussed above, a third change was
introduced to describe the pattern descriptions some Ss generated on P7
as well as on P2, P3, and Pl5. On these problems they used counting pro-
cedures of various sorts, Thus in problem two the most frequent rule
was (3M1, N(M1)) w h ic h seemed more appropriate for those seven SS
SEQUENTIAL PATTERN ACQUISITION 415
than the related (Ml, Ml, Ml, N( Ml) ) used by four Ss and predicted
by Simon and Kotovsky. Other counting procedures involve keeping
tallies in memory (designated M, and M,).
Extensions of the model: summary. The use of a tally combined with
hierarchic pattern descriptions in P7 accounted for the differences be-
tween the concepts of four Ss and those predicted by Simon and Kotov-
sky. These together with the seven concepts using the counting pro-
cedure in P2, the eight concepts using hierarchic pattern descriptions,
and the 15 concepts using BN instead of N on the backward alphabet
account for 34 of the 74 concepts identified in the current study that
differ in any respect from those predicted by Simon and Kotovsky. Thus
the introduction of these three changes substantially improves the pre-
dictive capacity of the model without a great proliferation of new
mechanisms.
Short-term memory. A few of the concepts used by Ss require more
than two chunks, or place-keepers, to be held in short-term memory. We
will have more to say below about the burden this places on short-term
memory.
A different point concerning memory allocation that is only implicit
in the 1963 paper is that a recurring letter (the I relation determining M
in abmcdmefm, for example) is assumed to be held in long-term memory,
hence not to add a chunk to the short-term memory load. Short-term
memory is used only for those portions of the pattern description that
involve N and BN operators associated with continually changing posi-
tions on alphabetical lists. In the computer model also, application of the
I relation involves a simple set of computer operations compared to the
relatively involved operations for applying N or BN. A related distinc-
tion was made by Gregg ( 1967), who differentiated between the direct
and indirect production of items from memory (operations M and
D( M,) ) in the meta-language he used for his sequential patterns.
Whether the M,< chunks postulated for hierarchic patterns impose a
short-term memory load is not easy to determine. The Ss tend to use
auxiliary procedures that have the effect of reducing the load of this in-
formation on short-term memory. Thus rhythmical shifts in voice pitch,
or separate and independent extrapolation of the two sequences (even
though the S has previously achieved a unified pattern description) re-
duce the simultaneous demands on short-term memory. Further experi-
ments will be needed to confirm this conjecture, although in the present
experiment Ss using hierarchic procedures made fewer errors on prob-
lems involving two (P7) or three (Pll and P14) chunks in short-term
memory than did the other Ss. We will look later at the exact process
they used for extrapolating.
416 KOTOVSKY AND SIMON
Initialization
Once S has attained a pattern description from the sequence, he then
uses that pattern description to extrapolate the sequence. Occasionally
this temporal order is not followed strictly. Instead S obtains part of a
pattern description-say, for one position of the period-and extrapolates
it; then obtains another part and extrapolates it; then obtains another
part and extrapolates it.
Generally, in order to use a pattern description to produce an extrap-
olation, S first has to initialize-that is, set “pointers” in short-term mem-
ory to specific letters in alphabets associated with the pattern. Table 5
describes the 147 (of a possible 210) initializations about which informa-
tion was obtained at a moderate (17) or high (130) level of certainty.
The table indicates whether S initialized his memory cells at the be-
ginning (B) of the presented sequence (thus reproducing the sequence
as well as extrapolating it), at the end (E) of the sequence (thus pro-
ducing only the extrapolation), or at some point in the middle ( M) re-
gion of the sequence. Again, the initializations not included on the table
(not discernible at a 3, 4, or 5 certainty level) do not, at rough inspec-
tion, appear different in this respect from those that do appear.
Ss tend to use B or M initializations on harder problems, and E initial-
izations on easier problems. A majority of Ss used E or in one case E
TABLE 5
InitializationU
Problem B M E Other
Pl cdcdcd 4
P2 aaabbbcccdd 1 1 5
P3 atbataatbat 3 4
P4 abmcdmefmghm 1 1 7 1
P5 defgefghfghi 2 1 7
P6 qxapxbqxa 3 4 2
P7 aduacuaeuabuaf 4 1 4 3
P8 mabmbcmcdm 3 2 5
P9 urtustuttu 7 3 1
PI0 abyabxabwab 3 7
PI1 rscdstdetuef 3 7 2
P12 npaoqapraqsa 5 1 4 1
P13 wxaxbybzczadab 5 3 4 2
P14 jkqrklrslmst 7 3 1
P15 pononmnmlmlk 6 3
- - - -
53 14 69 11
Extrapolation
Table 6 contains information about the last operation performed by
the Ss: the actual extrapolation from the initialized pattern description.
Information is presented in the table for 174 extrapolations (out of a
possible 210), 137 at a high certainty level, and 37 at a moderate cer-
tainty level. Table 6 distinguishes extrapolations by positions, from those
by period, and from mixtures of the two.
In extrapolating by position the letters occupying a specific position
in the period of the answer are initialized and extrapolated separately
from the letters occupying other positions. When a problem is classified
as extrapolated by position, every position for that problem was initial-
ized and extrapolated separately. In contrast other Ss extrapolated the
sequence by period, that is, by iterating through the complete pattern
TABLE 6
Extrapolation
Extrapolation method
Percent
BY Within within-
Problem STM chunks position Mixed period period
TABLP: 7
Analysis of Diagnosable I<rrors
(by Problem)
Error in
Pl cdcdcd one
P2 aaabbbcccdd one 3
P3 atbataatbat one 1
P4 abmcdmefmghm one 1
P5 defgefghfghi several 1 1
P6 qxapxbqxa one 2 2
P7 aduacuaeuabuaf several 1 3 1
P8 mabmbcmcdm one 1 2 1
P9 urtustuttu one 7
PlO abyabxabwab one 1
- Pll rscdstdetuef several 2 1
P12 npaoqapraqsa several 1
P13 wxaxybyzczadab several 2 3
P14 jkqrklrslmst several 1 1
P15 pononmnmlmlk several 3 1
11 57 5
220 KOTOVSKY AND SIMON
CONCLUSION
The current study attempts to record details of human behavior in the
problem solving process so as to test the hypotheses generated by the
theory and computer model presented in our 1963 paper.
The data on Ss’ behavior were obtained in two main ways: from
verbalizations and from records of a presentation device that required
422 KOTOVSKY AND SIMON
APPENDIX I
Problems Missed by ljifferent Variants of the Computer Model
A* 4 x x x x x x x x x x x x
B* 4 x x X xxxxxx
C* 4 x X x x x x x x
I)* 4 X x
1s: 4 x x x x x
F 4 x x X x x
(: 4 x x
H 4 x
I 4 X x x
J 4 X
K 4
I, 3 x x x
M 3 x x X x
N 3 x x
0 3 x x
2 s x X
i 2 x x X
1: 1 x x x x x x x x x
s 1 X x x x x xxxxxxx
T 1 x x x x x x x x x x x x x
U 1 xxxxxxx xxxxxxxx
REFERENCES
GREGG, L. W. Internal representations of sequential concepts. In B. Kleinmuntz (Ed. ),
Concepts and the structure of memory. New York, Wiley, 1967.
KLAHR, D., & WALLACE, J. G. An information processing analysis of some Piagetian
experimental tasks. Cognitive Psychology, 1970, 1, 358387.
KOTOVSKY, K. An empirical test of the Simon & Kotovsky “concept former” model
of human letter series sequence extrapolation behavior. Unpublished M. S. thesis,
Carnegie-Mellon University, 1970.
NEWELL, A., & SIMON, H. A. Human problem solving. Englewood Cliffs, N. J.: Pren-
tice-Hall, 1972.
PIVAR, M., & FINKELSTEIN, M. Automation using LISP, of inductive inference on
sequences, 125-216. In E. C. Berkeley & D. Bobrow (Eds.), The programming
language LISP. Cambridge, Mass.: Information International, 1964.
&STLE, F. Theory of serial pattern learning. Psychological Review, 1970 (Novem-
ber), 77, 481495.
SIMON, H. A., & KOTOVSKY, K. Human acquisition of concepts for sequential pat-
terns. Psychological Review, 1963, 70, 534-546.
SIMON, H. A., & NEWELL, A. Information processing in computer and man. American
Scientist, 1964, 52, 281300.
SIMON, H. A., & SUMNER, R. K. Pattern in music. In B. Kleinmuntz (Ed.), Formal
representation of human judgment. New York: Wiley, 1968.
THURSTONE, L. L., & THURSTONE, T. G. Factorial studies of intelligence. Chicago:
University of Chicago Press, 1941.
VAN DE GEER, H. P., & JASPERS, J. F. M. Cognitive function. In P. R. Farnsworth,
0. McNemar & Q. McNemar (Eds.), Ann& Review of Psychology, 1966, 17,
145.
WILLIAMS, D. S. Computer program organization induced by problem examples.
Unpublished Doctoral dissertation, Carnegie-Mellon University, 1969.