What does an experienced chess player “see” when he looks at a chess

position? By analyzing an expert player’s eye movements, it has been

shown that, among other things, he is looking at how pieces attack and

defend each other (Simon & Barenfeld, 1969). But we know from other

considerations that he is seeing much more. Our work is concerned with

just what ahe expert chess pIayer perceives.

The most extensive work to date on perception in chess is that done

by de Groot and his colleagues (de Groot, 1965, 1966; Jongman, 1968).

In his search for differences between masters and weaker players, de

Groot was unable to find any gross differences in the statistics of their

thought processes: the number of moves considered, search heuristics,

depth of search, and so on. Masters search through about the same

number of possibilities as weaker players-perhaps even fewer, almost

certainly not more-but they are very good at coming up with the

“right” moves for further consideration, whereas weaker players spend

considerable time analyzing the consequences of bad moves.

De Groat did, however, find an intriguing difference between masters

and weaker players in his short-term memory experiments. Masters

showed a remarkable ability to reconstruct a chess position ahnost

perfectly after viewing it for only 5 sec. There was a sharp dropoff in

’ Correspondence should be addressed to: Dr. William G. Chase, Department of

Psychology, Carnegie-Mellon University, Pittsburgh, PA 15213. This work was

supported by Public Health Service Research Grant MH-07722 from the National

Institute of Mental Health. We are indebted to Hans Berliner for his masterful

performance as a subject.
55

Copyright @ 1973 by Academic Press, Inc.

All rights of reproduction in any form reserved.

56 CHASE AND SIMON

this ability for players below the master level. This result could not be

attributed to the masters’ generally superior memory ability, for when

chess positions were constructed by placing the same numbers of pieces

randomly on the board, the masters could then do no better in reconstructing them than weaker players,
Hence, the masters appear to be

constrained by the same severe short-term memory limits as everyone

else ( Miller, 1956), and their superior performance with “meaningful’

positions must lie in their ability to perceive structure in such positions

and encode them in chunks. Specifically, if a chess master can remember

the location of 20 or more pieces on the board, but has space for only

about five chunks in short-term memory, then each chunk must be

composed of four or five pieces, organized in a single relational structure.

One key to understanding chess mastery, then, seems to lie in the

immediate perceptual processing, for it is here that the game is

structured, and it is here in the static analysis that the good moves are

generated for subsequent processing. Behind this perceptual analysis,

as with all skills (cf., Fitts & Posner, 1967), lies an extensive cognitive

apparatus amassed through years of constant practice. What was once

accomplished by slow, conscious deductive reasoning is now arrived at

by fast, unconscious perceptual processing. It is no mistake of language

for the chess master to say that he “sees” the right move; and it is for

good reason that students of complex problem solving are interested in

here is to discover and characterize the structures, or chunks, that are

seen on the board and stored in short-term memory.

SCOPE OF THE STUDY

The previous studies of chess perception make highly plausible the

hypothesis that the chess master encodes information about a position

in chunks, but provides no direct methods for delimiting the chunk

boundaries or detecting the relations that hold among ‘the components

of a chunk. Evidence is needed on these points in order to discover how

many pieces typically constitute a chunk, what the relative sizes are of

the chunks of masters and weaker players, and how many chunks

players retain after a brief view of a position.

The player’s perceptual processing of the board is so rapid (and

probably unavailable to conscious introspection) that it is impossible

to obtain an accurate verbal description of the process from him. Although eye movements give us a
record of how the board is scanned

(de Groot, 1966; Simon & Barenfeld, 1966; Tichomirov & Poznyanskaya,

1966; Winikoff, 1967), they don’t tell us precisely which pieces are

observed (especially in peripheral vision) and in what order; they only

PERCEPTION IN CHESS 57

tell us the general area being aimed at by the fovea. And, of course, data

on eye movements can’t tell us what information is being abstracted

from the display.

There are, however, other techniques, which have been used with

verbal materials, that would appear promising for the problem at hand.

Tulving (1962) has looked at clusters in free recall protocols; and Bower
and Springston (1970) have looked at the timing relations and pauses

in the output. McLean and Gregg (1967) have used pauses to define

chunks in rote learning. Ein-Dor (1971) has studied chunking of visual

stimuli in the form of Chinese ideograms, using a method essentially

identical with our perception experiment.

The central objective of this study, then, is to isolate and define the

chunks into which information is hypothesized to be encoded in chess

perception tasks. We use two techniques. In the perception task, we ask

chess players to reconstruct a chess position while it remains in plain

view, and we use subjects’ successive glances at the board as an index

of chunking. The basic assumption is that, under the conditions of the

experiment, the subject will encode only one chunk per glance while

reconstructing the position.

In the memory task, which is very similar to de Groot’s task, we ask

chess players to reconstruct a position from memory after brief exposure

to it, and we use the timing or clustering in recall to segment the output

into chunks.

The memory task permits us to replicate the basic findings of de Groot

and Jongman. These results are so important that it is essential to have

an independent replication; moreover, the empirical results for the case

of the random boards have never been reported in detail in the literature.

By using two different tasks, we obtain some protection against artifacts that might compromise the
interpretation of our findings. One important question we shall investigate is whether the chunks
defined by

the data from the perception task are essentially of the same size and

character as the chunks defined by the data from the memory task.
In the following sections of this paper, we will report and analyze the

main body of data obtained by presenting the two tasks to a chess master

and to weaker players. Then we will investigate in somewhat greater detail the data for the chess master
in middle game positions. In a final

section, we will summarize our findings and our interpretation of them.

METHOD

Three chess players, a master (M), a Class A player (A), and a beginner ( B ), were used as subjects.
Twenty games were selected from

chess books and magazines to generate the stimuli. These were games

58 CHASE AND SIMON

between advanced players (masters, experts, and perhaps a few Class A

players). Ten were middle game positions, at about White’s 21st move,

with 24-26 pieces remaining on the board. Ten were end-game positions,

at about the 41st move, with 12-15 pieces remaining on the board. Not

all the positions were “quiet,” i.e., some of them caught games at a point

where an exchange of pieces was in progress.

In addition to the positions from actual games, eight random positions

were generated, four from middle games and four from end games, by

taking actual positions and replacing the pieces randomly on the board.

Perception Task

In this task, two chess boards were placed side by side, separated by

about 6 in.: One of the 28 chess positions was set up on the subject’s

left, and the other board, free of pieces, was placed directly in front of

him. A full set of pieces was placed to the right of the blank board. A

partition between the two boards prevented the subject from seeing the

position on the left. When the partition was removed, the subject’s task
was to reconstruct the position on the board in front of him as quickly

and accurately as possible, glancing at the position on the left as often as

he wished. His behavior was recorded on videotape.

Memory Task

The procedure in the memory task was similar to that used by de

Groot ( 1965), except that the subject was given multiple trials in each

position. The boards were set up exactly as in the perceptual task. When

the partition was removed, the subject was allowed to view the position

on the left for 5 set, and the partition was then placed in position again.

The subject then recalled, by placing pieces on the board in front of

him, what he could remember of the position on the left, being allowed

as much time as he wished (subjects rarely took more than 1 min). If

the position was not reconstructed perfectly, the board in front of the

subject was cleared and a second trial was conducted in the same way:

5 set of viewing, followed by free recall of the position. Additional

trials followed until the subject recalled the position perfectly, except

for the random positions, which were too difficult to continue to

criterion.

In the perception task, each subject processed five middle-game

positions, five end-game positions, two randomized middle-game positions, and two randomized end-
game positions. He also processed the

same number of each kind of position in the memory task.

PERCJ3PTION IN CHESS 59

RESULTS

The videotape records for both tasks were analyzed by recording each

piece as it was placed on the board, and by recording the time, within
r/l0 set, between the placing of that piece and the next one.

The time intervals were used to segment the protocols, in order to test

the hypothesis that long pauses would correspond to boundaries between

successive chunks, while short time intervals between pieces would

indicate that the pieces belonged to the same chunk in memory.

The nature of the chess relations between successive pieces, separated

by long and brief pauses, respectively, were analyzed for information

that would reveal how pieces are chunked perceptually. The occurrence

of each of five chess relations between successively placed pieces was

recorded: (1) attack: either one of the two pieces attacks the other; (2)

defense: either one of the two pieces defends the other; (3) proximity:

each piece stands on one of the eight squares adjacent to the other; (4)

common color: both pieces are of the same color; and (5) common type:

both pieces are of the same type (e.g., both are pawns, rooks, etc.).

Accuracy of Reconstruction

The accuracy with which the subjects reconstructed positions on the

first trial in the memory task was analyzed for comparison with the

previous findings of de Groot and Jongman. Accuracy was measured

by the number of pieces placed on the correct squares of the board on

the first trial after a 5-set view of the board. The number of pieces

correct on subsequent trials was also computed, but chief interest for

our purposes centers on the first-trial results.

Figure 1 shows the results for the middle-game positions, actual and

random. Figure 2 shows the results for the end-game positions, actual

and random. The figures show the average number of pieces placed
correctly by each subject on successive trials for all positions of the type

in question. The standard errors, based on five scores, are shown for the

first trial of the middle- and end-game positions.

In the actual middle game positions, M was able to place an average

of about 16 pieces correctly on the first trial, while A and B placed about

eight and four, respectively. M was able to reproduce the board perfectly

in three or four trials, while A typically required about one or two more

trials than M, but B took considerably more trials (as many as I4 in

one case). M showed no such superiority in additional pieces placed in

successive trials. In trials just beyond the first, M typically added about

four more pieces to his previous reconstruction, while the gains for A

and B averaged five or six pieces per trial. Of course, A and B, because