Analogical Reasoning: A Core of Cognition

Helmar Gust, Ulf Krumnack, Kai-Uwe Kühnberger, Angela Schwering

Analogies have always been considered a central part of human intelligence and cognition. This survey offers an overview
of analogical reasoning and its applications, showing that analogies are an important element of various cognitive abilities
like memory access, adaptation, learning, reasoning, and creativity. Therefore, analogies can provide a basis for integrated
large-scale cognitive systems.

1 Introduction Proportional analogies have a nice counterpart in metaphoric

expressions. For example, a statement like Gills are the lungs of
Analogy research has a long history and was subject of inves- fish can be analyzed as the solution to the following analogy:
tigation on various levels: Ancient Greek philosophy examined mammals : lungs :: fish : X. Here the word gills refers to what-
analogy as a proportional relation between objects. In belletris- ever the analogical relation constructs for X. What is mentioned
tic texts, metaphors use analogy implicitly to describe objects in in the surface structure of the metaphoric expression is typically
a figurative way. In natural sciences, analogy is an indispensable a cross domain role relation meaning X plays the role of lungs
tool for scientific progress [18]. to fish, while the reference to the counterpart of fish (mammals)
While classical research on analogy mainly deals with higher- is assumed to be derivable by background knowledge.
level intelligent behavior, the focus in cognitive science has A second type of analogies are predictive analogies. They
shifted to the examination of more basic cognitive abilities, rel- describe a new domain (target) not only by specifying struc-
evant for agents interacting with their environment. Many of tural commonalities with a given domain (source), but also by
these abilities seem to rely on analogy as well and thereby give transferring information and explanations from the source to
an explanation for the analogical capacity on the higher-level. the target [5, 13]. Besides the classical domain of metaphoric
Therefore, it is not astonishing that analogy plays a central role expressions, complex conceptualizations of correlations between
in cognitive science research. seemingly different physical domains are good examples for such
analogies. They relate physical domains which are hardly acces-
sible by our direct experience to domains which have perceiv-
2 What is Analogy? able properties. The importance of analogies becomes evident
in cases where such working analogies are not available. Why is
Analogy making is a highly sophisticated cognitive process in
relativity theory or quantum mechanics so hard to conceive by
which two conceptualizations - a source and a target - are an-
humans? The main reason is, because the analogical transfer to
alyzed for common structural patterns [6]. In analogies, source
a domain which is accessible to our direct experience essentially
and target are typically of different domains (for metaphors this
fails. There are many other examples in modern physics where
is even essential). The purpose of analogies is to adapt knowl-
no good analogies are available or strange metaphors are used
edge available about the source conceptualization such that it
which do not explain very much: consider for example the wave-
can be applied to the target in a way that new analogous infer-
particle dualism, exchange forces (based on photons, gravitons
ences can be drawn. Analogy making requires intelligence since
etc.), or string theory.
analogous patterns and transfers often are not obvious and de-
The third type of analogy, analogical problem solving, can be
pend on a certain conceptualization of the domains.
used to solve a problem by transferring a solution from a well-
Metaphors and analogies occur in a large variety of domains,
known domain to an unknown domain. Case-based reasoning
as well as in quite different forms. In order to classify certain
[16] formalizes approaches of this type. Other examples are the
aspects and properties of analogies, three types of analogies are
famous tumor-fortress problem [8] or the usage of program code
often distinguished1 [14, 22]: First, proportional analogies have
for developing new programs by abstraction [23].
the general form (A : B) :: (C : X). Proportional analogies
Another aspect connected to analogies is the re-
can be in domain which means that A, B, C are expressions from
conceptualization of known domains (called rendition). This
the same domain. This type is characteristic for intelligence test
may occur if an obvious analogy fails. A classical example is the
where the subjects have to continue a sequence of geometric fig-
conceptualization of a paintbrush as a tool to transfer paint to
ures [4] or numbers and strings [12]. Proportional analogies can
a canvas by pumping paint in contrast to smearing it [14]. In
also be cross domain. In this case, A and B establish a relation
general, verifying hypotheses about allegedly known (real world)
in the source domain which should be applied to a concept C
situation (source) by experiments in a controlled lab environment
of the target domain to get the result X in the target domain.
(target) uses an analogical relation between both domains. The
1 We do not claim that the following classification of analogies is outcome of such an experiment may force us to reconceptualize
complete, nor that it is the only possible one. Rather such a classifi- the source domain.
cation can be useful to specify different properties of analogies. It is often mentioned that metaphors are strongly connected
 source of valuable inspiration.
In some cases, these three phases are supplemented by ad-
ditional steps, such as evaluation of the transferred knowledge,
or the induction of generalized rules, depending on the model
applied and the context in which analogy making is placed.

3 Cognitive Abilities in Analogies

Analogies are paramount for understanding many cognitive phe-
nomena. Figure 1 illustrates the process of analogy making and
Figure 1: Analogies in cognitive processes. relates cognitive abilities to the different stages. In the following,
we describe in detail each of these abilities.
to analogies [7, 14]. Depending on the context one and the
same metaphor can be proportional or predictive. Assume for 3.1 Memory and Adaptation
example a situation in which a teacher is lecturing students
on elementary atom physics by giving the following metaphor The ability to store information and recall it on given occasions
describing the Rutherford analogy: is crucial to an appropriate use of memory. As every new sit-
uation includes a lot of details making it unique and different
(1) Electrons are the planets of the atom. from previous experiences, a retrieval process making memories
usable for that situation heavily relies on analogy. Irrelevant de-
(1) can be interpreted as a predictive analogy: The students tails must be blinded out, old patterns have to be mapped to
learn a new conceptualization of the atom. On the other hand, new domains. So the recall of information from memory is a
for a scientist (1) can be simply interpreted as a (historically sophisticated process with close resemblance to the phases of
important) proportional analogy. retrieval, mapping and transfer described above.
The process of analogy making can be subdivided into sev- Kokinov [15] proposes a cognitive architecture in which
eral, interrelated tasks. Typically, the following three subtasks memory plays the central role, to address cognitive processes
can be identified: retrieval, mapping, and transfer. like analogical reasoning and the estimation of semantic simi-
At the beginning, when exposed to a new situation (the tar- larity. The system consists of a network, the DUAL memory,
get), a source domain has to be identified to which that situation representing the system’s world knowledge. The nodes of the
can be related. Some retrieval technique has to be applied to network, called micro-agents, have an inner structure and are
search the memory for items which seem like candidates for an in a state affecting their behavior in cognitive processes. Such
analogy. In certain settings, such as intelligence tests or teach- processes are triggered by activation potential and marker pass-
ing situations, the source domain may be given explicitly. Some ing along the connections between the micro-agents. An initial
models for analogy making even view this as the standard case activation pattern is determined by context and then propagated
and do not provide special means of retrieval. according to the characters of the individual agents and their in-
The mapping phase aims to establish the analogical relation terconnections. The activation pattern represents the contents
between these two domains, i.e. the alignment of structures of the short-term memory, while the whole micro-agent network
from both domains. In general, there are many possible map- corresponds to long-term memory.
pings and which one is appropriate depends on the context and Based on this memory architecture, different types of re-
the goal of the analogy. Two problems are typically discussed as- trieval and reasoning can be performed. For example, the frame-
sociated with the mapping step. The relevance problem consists work has been demonstrated to be capable of analogical problem
of identifying the parts of the domains that are relevant within solving: given the task to heat water in a wooden vessel in a
the context of the analogy and therefore shall enter the analog- forrest with only matches and a pocket knife the system pro-
ical relation. The representation problem is concerned with the posed, analogous to an immersion heater, to warm up the knife
difficulties in mapping caused by differently structured represen- and then put it into the water.
tations of the domains. While it may seem plausible that two A totally different approach to memory are Vector Symbolic
domains are represented in an isomorphic way when they are Architectures (VSAs) [19]. This approach uses high-dimensional
especially prepared for the analogy, this seems unlikely in gen- vectors to store information by simple mathematical operations
eral. Therefore the mapping may guide a restructuring of one or like addition and multiplication. Storage and retrieval can be
both domains and a good deal of the explanatory and creative carried out via the same set of operators which can also be
power of analogies can actually be ascribed to that process of used to realize some kind of analogical access. From a vector
re-representation. that holds the information Paris is the capital of France and
During the transfer phase the analogical relation is used Stockholm is the capital of Sweden, questions like What is the
to translate information between the two domains. Normally relation of Paris and France? and What is the capital of Sweden?
knowledge is transferred from the source to the target domain can be answered. But it is also possible to retrieve statements
and is used there to introduce new concepts or structures, give not explicitly stored in the memory, such as the solution to the
new explanations to phenomena, or solve given problems. This analogy task Paris is to France like X is to Sweden.
new knowledge is in no way logically justified and should merely
be seen as a hypothesis, but when used carefully, it can be the
3.2 Reasoning
Human cognitive reasoning capacities include the classical infer-
ence mechanisms: Deduction infers knowledge given implicitly
by known facts and rules: if the premises evaluate true, deduc-
tion infers according to the rule that the conclusion is true as
well. Inductive reasoning creates knowledge via generalization
over a large set of cases which share a common fact or rule and
induces a general rule. Abduction is reasoning from an effect to
its cause: from a known general rule and a true conclusion it is
hypothesized that the premise is true as well.
Beside these classical reasoning mechanisms, human rea-
soning capacities include analogical reasoning, too. Moreover,
analogical reasoning can imitate the classical forms of reasoning: Figure 2: The analogy between a water pipe system and an
if an analogy completes a causal relation on the target side by electric circuit in a diagrammatic representation. The Figure
a missing premise it resembles an abductive inference and if it contains more information than necessary for an interpretation
transfers the conclusion it resembles a deductive inference. of the metaphorical description.
While a deductive inference is truth-preserving, inductive,
abductive and analogical inferences are not. But the likeliness A prototypical situation in which analogical learning plays
of inductive inferences increases with the number of valid cases. an important role are teaching situations. When the teacher
Therefore the set of cases must be sufficiently large to draw explains an electric circuit by analogy with a water-flow system
sound conclusions. Analogical reasoning, however, requires only (Figure 2), the student needs to establish an analogical relation
two cases: the likelihood of analogical inferences increases with between the involved domains.2 Only if the student successfully
the number of analogical matches between the source and the transfers the flow of water to the target domain she will reach
target [25]. the conclusion that current is flowing in the electric circuit.
We distinguish two types of analogical inferences: Either it A famous approach for such types of transfers is the struc-
comprises only the transfer of a structural relationship between ture mapping theory (SMT) where two graphs are used to rep-
two elements. Or analogical transfer refers to a general rule. resent models of the source and target domain [5]. An align-
Applying the rule from the source domain, new inferences can ment process searches a maximal match of subgraphs. The al-
be drawn in the target domain. However, in both cases the ana- gorithm returns an analogical relation between the two domains.
logical inference is based on the uncertain analogical mapping Learning in SMT is realized by transferring relevant parts of the
which must not necessarily be true. source graph which have not been aligned [6]. The alignment
and transfer process is guided by principles ensuring consistency
3.3 Learning by Transfer of the hierarchical relational structure.

In artificial intelligence, a large number of different learning the-

ories were proposed. Two major types of learning theories can 3.4 Learning by Abstraction
be distinguished: Lazy learning algorithms store each example Whereas a transfer of information from the source to the tar-
explicitly in a database without identifying abstract features. get yields a new conceptualization of the target domain, learn-
Instance-based learning [1], case-based learning [16], or memory- ing in the classical sense is concerned with finding hypotheses
based learning [24] are representatives of this learning type. Ea- that generalize the training examples to unseen test examples.
ger learning algorithms minimize the storage load by an abstrac- This aspect is covered by learning by abstraction: Instead of
tion process. The identification of important (common) features a plain mapping, the following two-layered procedure is consid-
of the examples is a necessary prerequisite for this type of learn- ered. First, identify the common structure of source and target
ing algorithm. Examples for techniques of learning are decision- by a generalization. Second, calculate the mapping of objects
tree learning [21], inductive learning [17], or connectionist-style of the source to objects of the target via the common struc-
learning [2]. ture. Particularly with respect to certain application domains of
Both classes of learning strategies are based on a common analogies like the string domain this abstraction process seems
idea: As input data a (more or less) large sample of examples is to be crucial. In Hofstadter’s analogy model Copycat [12], string
needed to guarantee a successful learning procedure. Contrary sequences like ababab need to be represented at a more abstract
to these approaches, cognitive learning seems to be based on level as 3 ∗ (ab) to recognize the analogical relation to strings
a significantly less amount of training data and occurs often like ccc.
in the form of implicit learning. We propose to explain these A formal theory for computing analogies via an abstraction
discrepancies by analogical learning. process is proposed in [11]: Heuristic-Driven Theory Projection
In the case of analogical learning a conceptualization of the (HDTP) is an algorithmic approach to calculate a common gen-
source domain is sufficient to generate knowledge about the eralization of a source and a target. The underlying mathemat-
target domain. Moreover, a projection of attributes and relations ical model for such generalizations is anti-unification [20]. Con-
from the source to the target can productively introduce new straints are given by the problem structure or by the functional
concepts on the target domain. As a result cognitive agents role the objects play in a given structure.
can learn a new conceptualization of the target domain without
2 c.f.
perceiving a huge number of examples. http://hyperphysics.phy-astr.gsu.edu/hphys.html
 endeavors were not very successful so far. Perhaps analogical
reasoning can provide a way out of this dead end.3 Exam-
ples of analogy related creativity can be seen in natural language:

(2) He reached the maximal height of his career.

(3) Gills are the lungs of fish.
(4) My brain is a little foggy.
(5) Juliet is the sun.

These metaphors govern the creation of new concepts to a

different extent. Height of career in (2) seems to be a conven-
tional metaphor, i.e. the constituent is lexicalized and therefore
an idiom. In (3) the new concept gills is strongly determined by
Figure 3: A graphical representation of learning stages. the source domain. In (4) the meaning of the adjective foggy
has to be adapted to be applicable to brain states. Shakespeare’s
In HDTP’s logical framework, source, target, and gen- metaphor in (5) is at a very abstract level and it remains rela-
eralization are given by theories T hS , T hT , and T hG re- tively vague which properties are transferred.
spectively. Comparing attracts(sun,planet) from T hS to at- Another domain for which creativity in analogies is an im-
tracts(nucleus,electron) to T hT in a logical representation of portant aspect are examples from the physical domain like the
the Rutherford analogy (section 2), this procedure would re- former mentioned water-current analogy.4 Figure 2 depicts the
sult in the generalization attracts(central body,orbiter) in T hG , situation between a water-flow circuit and an electric circuit.
where central body and orbiter are variables. (1st level Although there is no possibility to observe current flowing in
of Figure 3). By calculating the abstraction, the mapping an electric circuit directly, humans are able to establish such an
{central body 7→sun/nucleus, orbiter 7→planet/electron} is ob- abstract concept in analogy to water flowing in a water pipe
tained which represents two substitutions: Θ1 from T hG to system.
T hS and Θ2 from T hG to T hT . The generalized theory T hG How can analogical reasoning explain these creative abili-
comprises the common structures of source and target: in the ties? An answer can be provided by combining the transfer of
Rutherford analogy the generalization of the solar system and facts and rules from the source to the target (section 3.2) with
the atom can be conceptualized as a central force system. abstraction and generalization (section 3.4). In (6), current is
In HDTP, learning occurs as a side-effect of the model- explicitly associated with water, but the properties of current are
ing and can be considered as being implicit: the generalization not specified. In fact, even if a student has no idea which prop-
process yields new conceptualizations of the target domain at erties should be ascribed to current (for example, in a learning
an abstract level for no additional costs. In a uniform way, it scenario), (6) could be the cue to get the flowing conceptual-
provides a structural description of the underlying domains in ization of current by transferring properties of water from the
terms of a generalized theory. source to the target. This transfer enables the learner to create
Analogical learning in HDTP does not end with a successful properties of current like flowing-in-circuit or the necessity of
establishment of an analogical relation, but learning continues in a pressure source, provided the modeling and the background
stages. After the most specific generalization is established (1st knowledge is fine-grained enough.
level), the inductive refinement (2nd level) adjusts the parame- In particular, analogies in qualitative reasoning applied to
ters in order to find a new (and finer) conceptualization of source the physics domain are good examples for creative concept for-
an target. The 3rd level aims to identify general principles that mations: a famous and extensively discussed example is the cre-
can be applied in a variety of domains. ation of the concept heat in a water-flow vs. heat-flow scenario
In plain structure mapping as well as in mapping via abstrac- [22, 9]. In this scenario, the abstract concept heat-flow needs to
tion, the greatest common sub-structure of source and target be creatively generated by the analogy, although heat is neither
must be identified. But in mapping via abstraction, the com- observable nor measurable. Projecting the situation in a water-
mon structure is not forgotten but represented explicitly. Instead flow system to the situation in which the temperature of a cup
of mapping one object directly onto another, it is mapped via of coffee is decreasing and the temperature of a metallic cube
the role it is playing in the common structure. is increasing (both are connected by a metal bar), the success-
ful establishment of an analogy introduces a concept heat-flow,
explaining the changes in the setting.
3.5 Creativity
Creativity as defined in [3, p. 205] is “the production of an idea,
action, or object that is new and valued”. Analogies can pro- 4 Conclusions and Future Directions
vide a means for creativity, as they can introduce new concepts
In this article, we argued that analogical reasoning is a crucial
into a domain via analogical transfer. The principles guiding the
issue for cognition. Classical cognitive abilities of human agents
transfer ensure that the concepts are relevant in order to under-
such as memory capacities, learning by transfer, learning by ab-
stand the target domain. Furthermore analogies can induce new
straction, forms of reasoning, and creativity can be considered
insights about a domain by reconceptualizations.
From the very beginnings, artificial intelligence struggles 3 Examples of creative aspects of analogies can be found in [14].
4A
hard to develop models for creativity. Nevertheless these precise modeling of this analogy can be found in [10].
as necessarily interconnected with and in an important respect [14] B. Indurkhya. Metaphor and Cognition. Kluver, Dordrecht, The
governed by the capacity to establish analogical relations. In Netherlands, 1992.
a certain sense these capacities (typically considered to be dis- [15] B. Kokinov. The context-sensitive cognitive architecture DUAL.
tinct) can be explained by analogies in a rather unified manner. In Proceedings of the Sixteenth Annual Conference of the Cog-
Clearly the proposed modeling of a variety of cognitive nitive Science Society. Erlbaum, 1994.
phenomena can be interpreted as a reductionist position [16] J. Kolodner. Case-based reasoning. Morgan Kaufmann, San
of cognition where several seemingly incoherent aspects of Mateo, CA, 1993.
cognition can be reduced to one general principle underlying the [17] S. Muggleton and C. Feng. Efficient induction of logic programs.
whole system. Although we clearly do not claim that cognition In First Conference on Algorithmic Learning Theory, pages 368–
as a whole is nothing else but analogy making, the possibility 381, 1990.
of a reduction to a main principle allows to speculate about the [18] Oppenheimer. Analogy in science. American Psychologist,
chances to build integrated large-scale cognitive systems: the 11:127–135, 1956.
vast amount of information, the different computing paradigms [19] T. A. Plate. Distributed representations and nested composi-
in AI for modeling particular aspects of cognition, and the tional structure. PhD thesis, Department of Computer Science,
problem of modeling dynamic changes of a cognitive system University of Toronto, Toronto, Canada, 1994.
make it difficult to come up with a large system. Analogical rea- [20] G. Plotkin. A note on inductive generalization. Machine Intelli-
soning makes a step towards the very possibility of this endeavor. gence, 5:153–163, 1970.
[21] J. Quinlan. Induction of decision trees. Machine Learning, 1:81–
Acknowledgement 106, 1986.
The work was supported by the German Research Foundation [22] U. Schmid, H. Gust, K.-U. Kühnberger, and J. Burghardt. An
(DFG) through the project ”Modeling of predictive analogies algebraic framework for solving proportional and predictive analo-
by Heuristic-Driven Theory Projection” (grant KU 1949/2-1). gies. In F. Schmalhofer, R. Young, and G. Katz, editors, Pro-
ceedings of the European Conference on Cognitive Science, pages
295–300. Lawrence Earlbaum, 2003.
References [23] U. Schmid, U. Sinha, and F. Wysotzki. Program reuse and ab-
straction by anti-unification. In Professionelles Wissensmanage-
[1] D. Aha, D. Kibler, and M. Albert. Instance-based learning algo-
ment - Erfahrungen und Visionen, pages 183–185. Shaker, 2001.
rithms. Machine Learning, 7:37–66, 1991.
[24] C. Stanfill and D. Waltz. Toward memory-based learning. Com-
[2] C. Bishop. Neural Networks for Pattern Recognition. Oxford
munications of the ACM, 29(12):1213–1228, 1986.
University Press, 1995.
[3] M. Csikszentmihalyi. Creativity. In R. A. Wilson and F. C. Keil, [25] A. I. Ujomov. Die Hauptformen und -regeln der Analogieschlüsse.
editors, The MIT Encyclopaedia of the Cognitive Sciences. MIT In G. I. Ruzavin, editor, Studien zur Logik der wissenschaftlichen
Press, 1999. Erkenntnis, pages 307–360. Akademie-Verlag, Berlin, 1967.
[4] M. Dastani. Languages of Perception. PhD thesis, University of
Amsterdam, 1998. ILLC Disseration Series 1998-05. Contact
[5] D. Gentner. Structure-mapping: A theoretical framework for
analogy. Cognitive Science, 7:155–170, 1983. PD Helmar Gust, Ulf Krumnack, MA,
[6] D. Gentner. The mechanism of analogical learning. In S. Vosni- Prof. Kai-Uwe Kühnberger, Dr. Angela Schwering
adou and A. Ortony, editors, Similarity and analogical reasoning, Email: [hgust | krumnack | kkuehnbe | aschweri ]@uos.de
pages 199–241. Cambridge University Press, New York, 1989. Institut für Kognitionswissenschaften
[7] D. Gentner, B. Bowdle, P. Wolff, and C. Boronat. Metaphor is Universität Osnabrück
like analogy. In D. Gentner, K. Holyoak, and B. Kokinov, editors, Albrechtstr. 28
The Analogical Mind: Perspectives from cognitive science, pages 49076 Osnabrück
199–253. Cambridge University Press, 2001.
[8] M. L. Gick and K. J. Holyoak. Analogical problem solving. Cog-
nitive Psychology, 12:306–355, 1980.
[9] H. Gust, K.-U. Kühnberger, and U. Schmid. Ontologies as a cue
for the metaphorical meaning of technical concepts. In A. Schal-
ley and D. Khlentzos, editors, to appear in: Mental States: Evo-
lution, Function, Nature. John Benjamins Publishing Company,
to appear.
[10] H. Gust and K.-U. Kühnberger. Explaining effective learning
by analogical reasoning. In R. Sun and N. Miyake, editors, Pro-
ceedings of the 28th Annual Conference of the Cognitive Science
Society, pages 1417–1422. Lawrence Erlbaum, 2006.
[11] H. Gust, K.-U. Kühnberger, and U. Schmid. Metaphors and
heuristic-driven theory projection (HDTP). Theoretical Com-
puter Science, 354:98–117, 2006. The analogy group at the Institute for Cognitive Sciences in
[12] D. Hofstadter and T. F. A. R. Group. Fluid concepts and creative
Osnabrück (from left to right: Kai-Uwe Kühnberger, Ulf Krum-
analogies. Basic Books, New York, 1995. nack, Angela Schwering, and Helmar Gust). Analogy has been
[13] J. Hummel and K. Holyoak. Distributed representation of struc- a topic of interest for the last years and has lead to the devel-
ture: A theory of analogical access and mapping. Psychological opment of HDTP, a formal model for analogy making.
Review, 104(3):427–466, 1997.