Professional Documents
Culture Documents
Cognitive Complexity in Matrix Reasoning Tasks: Marco Ragni (Ragni@cognition - Uni-Freiburg - De)
Cognitive Complexity in Matrix Reasoning Tasks: Marco Ragni (Ragni@cognition - Uni-Freiburg - De)
a formal characterization may elucidate future test development and can then form a formal foundation of reasoning
complexity. An analysis of the IQ-Test problems of Raven
has been conducted by Carpenter, Just, and Shell (1990). Figure 1 is an example, variations of which can be found in popular literature (e.g., Eysenck, 1962; Russell & Carter, 1994).
Introduction
For the past hundred year human intelligence has mostly been
tested by use of IQ-tests (Binet & Simon, 1905). Geometrical analogy problems (cf. Fig. 1) are part of a number
of IQ-tests, for example the Hamburg-Wechsler-IntelligenceTest (Wechsler, Hardesty, Lauber, & Bondy, 1961). A significant number of IQ-tests even consist exclusively of such geometrical reasoning problems, e.g., Cattells Culture Fair Test
(K. Cattell & Cattell, 1959) or Ravens Standard Progressive
Matrices (Raven, Raven, & Court, 2000) and Advanced Progressive Matrices (Raven, 1962). Such problems are sometimes classified as culture fair (R. Cattell, 1968) as they require less declarative knowledge than for instance word analogy problems. While the success in solving word analogy
problems can depend on additional knowledge, geometrical
reasoning problems can be modeled using mathematical functions exclusively. For this reason these problems are more
accessible in formal terms than other analogy problems. An
individuals intelligence is always measured by determining
the deviation of his or her performance on a given set of reasoning problems, from a particular group (specific age and
educational status, etc.). Problems in turn are classified empirically as simple or challenging, based on whether a given
population is able to solve most or only a limited number of
similar problems. While it is possible to empirically capture
the human reasoning difficulty it seems more desirable to
identify the characteristics of such problems formally. Such
this idea to the test for one of the best analyzed sample sets,
namely for Cattells Culture Fair Test (CFT). This test, along
with the aforementioned Ravens SPM and APM, is given an
empirical difficulty rating in the manual, i.e. the percentage
of persons who have correctly solved each problem. Our formal complexity measure is evaluated against these empirical
difficulty measures. Another empirical investigation, which
was recently performed for Evans Analogy problems (Evans,
1964; Lovett, Tomai, Forbus, & Usher, 2009), serves as an
additional benchmark. This article is structured as follows:
In the next section, we briefly review the literature on cognitive complexity, especially regarding explanations of human
reasoning difficulty. In Section 3, we analyze typical reasoning problems and develop a classification. Thereafter, we
introduce a functional complexity measure and compare it, in
Section 5, to empirical findings.
State-of-the-Art
Solving matrix problems requires the recognition and computation of similarities between the presented matrix objects
and their attributes. According to Representational Distortion (RD) theory, object representation similarity is defined
by the number of basic transformations necessary to transform one representation into another (Chater & Hahn, 1997).
Hahn and colleagues defined the complexity of the similarity computation as a special case of Kolmogorov complexity,
(e.g. Chater, 2000; Li & Vitanyi, 1997). Instead of defining
the length of the shortest program as a complexity measure,
they used the transformational similarity, i.e., the length of the
sequence of basic transformations (Hahn, Chater, & Richardson, 2003).
The problem of such a formal measure is the tractability
constraint (van Rooij, 2008). A transformation between two
object representations can be represented as a binary string.
The similarity can then be computed by a Boolean circuit, but
the outcome of this is super-polynomial complexity (Muller,
van Rooij, & Wareham, 2009). Consequently, it would be intractable and RD models would be psychologically implausible. This motivated them to argue for an analysis of restricted
problem parameters to avoid problems with intractability in
classical RD models.
Studied since the 1960s, the question of determining factors of the subjective difficulty of concepts, (Feldman, 2000)
has not found a sufficient answer. Feldman undertook a series
of experiments to test a wide range of Boolean functions with
respect to the question: Why are some concepts psychologically simple and easy to learn, while others seem difficult,
complex or incoherent? (Feldman, 2000). The data revealed
a surprisingly simple empirical law: The subjective difficulty
of a concept is directly proportional to its Boolean complexity. The influence of Boolean Complexity has, however, recently been questioned by Vigo (2006).
Evans (1964) wrote a program called ANALOGY to solve
geometric analogy problems frequently encountered in intelligence tests. The program can solve problems that can be
Classification of Problems
We will now characterize the problems by various properties.
The first and most simple distinction is made according to
the characteristic changes of functions. It is possible to extract functions for object changes by a horizontal and vertical
analysis of this sort of problem. Fig. 2 depicts an example of
a horizontal rotation transformation function of a rectangle.
Objects can be decomposed to simple squares, triangles, dots
or lines. As shown in Fig. 2, all objects consist of triangles,
rectangles or hexagons.
The second category, topological characteristic changes,
specifies the relationships between certain shapes or objects.
Topological characteristics are mathematically classified as
determining if two object are: (i) distinct, (ii) overlapping, or
(iii) contained within one another.
Category three specifies the changes of pattern. It is determined if there are any transformation changes concerning the
shapes of objects; that is, if an object (e.g. circle) is transformed into another object e.g., in Fig. 6 the square in the left
cell above can be associated with the triangle in the left cell
below.
The fourth distinction concerns the alteration of the number of elements in a problem. Take, for example, Fig. 3, there
the number of objects increases. In other words, if there is an
implicit sequence of numbers it must be determined whether
the number of objects increases or decreases from cell to cell
and whether this change happens horizontally, vertically, or
in both directions.
A further distinction concerns the characteristics of different shapes. To be analyzed is whether lines, dots, dashes,
or other characteristics act in compliance with Boolean functions such as, AND, OR, and XOR. Category number 6
number of parts characterizes how many parts the objects
are composed of.
The following categories horizontal function, vertical function and horizontal-vertical function classify the problems according to the objects of the horizontal/vertical succession of
changes. This refers to whether the objects in the cells are
dependant upon addition and/or subtraction or only succession. One aspect is rotation, indicating if at least one of the
objects is rotated during the changes of the cells. Another aspect takes moving objects into account, this is true if at least
one object is moving (e.g., clockwise or counter-clockwise,
up, down, or left or right movements of an object). Then,
the overall dimension characterizes whether the underlying
General Categories
The classification items for matrix reasoning problems, as
discussed above, can be grouped into three main types, which
we denote as (i) geometric operation problems, (ii) Boolean
operation problems, and (iii) grouping problems.
Mathematical Foundation
In theoretical computer science the difficulty of a problem is
determined with respect to the resources necessary to solve
The costs of a basic transformation can be weighted according to the different types of transformation. For example:
A translation of a stimuli might be easier to transform as a rotation. So the translated transformation gets a basic cost of 1
unit and the rotation a cost of 2.
For the problems listed in Table 1 nine basic transformations were used: identity function (with a cost of 0), translation (1), scaling (1), change in fill-in/background of a stimulus (1), addition and removal of a stimulus (1 each), rotation
(2), rotation of the fill-in/background of a stimulus (3) and a
change of the type of a stimulus (for example: transforming
a triangle into a square, 4). Without weights, all transformations receive a basic cost of 1.
Figure 6: An example of a more complex transformation
problem. The reasoner has to keep track of changes both horizontally and vertically.
ing the horizontal way, we have to decide weather we associate it with the black square or the shaded rectangle. Therefor the attributes of both stimuli have to be compared with the
ones of the white square. Again, the hierarchy of attributes
plays an important role, because these stimuli are associated
with each other. It depends on which share the highest ranked
attribute and not the most attributes. The number of comparisons which have to be performed to identify an association
are added to the transformation costs of an association.
Once the stimuli are associated to each other, the transformation between them have to be applied on the unassociated
stimuli (In this case the white triangle and the shaded rectangle) to determine the stimuli in the answer field. There it is
necessary to find the right pairs of stimuli, so that the white
triangle changes its color to black and not the shaded circle.
Overall it becomes obvious that the difficulty in this problem is raised by the complex relations between the stimuli and
this relations can be described as transformations and comparisons on the attributes of the stimuli.
ai
(1)
pi
Index P
76
93
78
96
95
90
94
90
47
85
Cw
8
13
18
5
6
4
10
4
26
21
C0
7
10
10
3
3
4
6
4
18
17
Q
0
1
0
1
P XOR Q
0
1
1
0
The solution for the problem in Fig. 7 requires the identification of a Boolean XOR function. If the middle cell in the
leftmost column (identified by the matrix coordinates (1, 2))
is compared with the middle cell of the same row (2, 2) and
both are compared with the middle cell in the rightmost column (3, 2) once can deduce that if the same object (triangle)
can be found in the left and middle cell it is not found in the
rightmost cell (so it will be deleted). And, whenever there is
an object only in the leftmost or middle column (e.g., the star
in the same row) it is found in the rightmost cell. These kinds
of functions are certainly not as easily recognizable as rotations or scaling and must be defined as such. Feldman (2000)
argues that the number of combinations of Boolean functions
determines the reasoning difficulty. A position which has recently been questioned (Vigo, 2006; Goodwin, 2006).
Correct
100
100
100
76
100
97
100
100
100
82
Time
10.8
7.3
6.7
8.7
13.4
8.5
6.1
6.1
6.0
23.6
C0
2
1
1
1
1
1
1
1
1
2
Nr.
11
12
13
14
15
16
17
18
19
20
Correct
100
97
91
94
97
100
56
91
82
97
Time
5.8
4.5
5.7
12.3
6.4
5.4
26.7
10.5
6.1
10.8
C0
1
1
2
1
1
1
3
1
1
2
General Discussion
The presented cognitive complexity measure for matrix problems allows the classification of problems with respect to the
number of basic functions needed to solve them. This measure is based on abstract units and might have a cognitive
counterpart: Nevertheless, it can predict problem difficulty
for Type 1 problems and pose a general enough framework
for Type 2 and Type 3 problems. What are the differences
to existing measures? Our approach is a combination of the-
Acknowledgement
This research was partially supported by the DFG (German National Research Foundation) in the Transregional Collaborative Research Center SFB/TR 8 within project R8[CSPACE] and by a Steinbuch-Stipend to the second author.
References
Binet, A., & Simon, T. (1905). The development of intelligence in children. L Annee psychologique, 12, 191244.
Carpenter, P. A., Just, M. A., & Shell, P. (1990). What one
intelligence test measures: A theoretical account of the processing in the raven progressive matrices test. Psychological Review, 97(3), 404431.