A Creative Analogy Machine: Results and Challenges
Diarmuid P. O’Donoghue
Department of Computer Science,
National University of Ireland Maynooth,
Co. Kildare, Ireland.
diarmuid.odonoghue@nuim.ie
Mark T. Keane
Department of Computer Science and Informatics,
University College Dublin,
Ireland.
mark.keane@ucd.ie
Abstract
Are we any closer to creating an autonomous model of
analogical reasoning that can generate new and creative
analogical comparisons? A three-phase model of analogical
reasoning is presented that encompasses the
phases of retrieval, mapping and inference validation.
The model of the retrieval phase maximizes its creativity
by focusing on domain topology, combating the semantic
locality suffered by other models. The mapping
model builds on a standard model of the mapping
phase, again making use of domain topology. A novel
validation model helps ensure the quality of the inferences
that are accepted by the model. We evaluated the
ability of our tri-phase model to re-discover several hcreative
analogies (Boden, 1992) from a background
memory containing many potential source domains.
The model successfully re-discovered all creative comparisons,
even when given problem descriptions that
more accurately reflect the original problem – rather
than the standard (post hoc) representation of the analogy.
Finally, some remaining challenges for a truly
autonomous creative analogy machine are assessed.
Introduction
Analogy has a long and illustrious history within creativity,
particularly within scientific and intellectual contexts
(Brown, 2003). Many episodes of scientific creativity are
driven by analogical comparisons (Dunbar and Blanchette,
2001), often involving image related analogies (Clement,
2008). Much progress has been made in cognitive science
on modeling this analogical reasoning process (see below),
prompting the following questions. Are we any closer to
creating an autonomous model of the analogical reasoning
that can generate new creative analogies? What progress
has been made towards such a creative analogy model?
What are the main challenges that lie ahead?
 In this paper we envisage a creative process that can take
any given target description and using a pre-stored collection
of domain descriptions, identify potentially creative
source domains with which to re-interpret the given problem.
This paper explores and evaluates the potential for a
model of analogy to act as a creativity engine.
 While Boden (1992) argues that analogy is effectively
the lowest form of creativity (improbable), we argue that
analogical creativity should be seen a part of a cohesive
human reasoning system. If the inferences mandated by an
analogy contradicts a fundamental belief, especially one
that has accrued many consequent implications, then resolving
this contradiction might well involve the “shock
and amazement” of transformational creativity. As such, it
appears that analogies may drive creativity at any of
Boden’s levels of creativity. Our creativity model is domain
independent and does not include a pragmatic component
or domain context. So, as our model does not use
domain-specific knowledge, arguably it cannot be easily
cast as improbable, exploratory or transformational creativity
(Boden, 1992).
 The current work was driven by three main aims. Firstly,
we wished to assess the creative potential of a three-phase
model of analogy. Secondly, we wished to assess the impact
of using differing knowledge bases upon the creative
potential of our analogy model. Finally, we wished to assess
the wider implications of analogical models for computational
creativity. Is a three-phase model either necessary
or sufficient to function as an engine of creativity?
Can such a model re-discover analogies considered to be
creative by people? Since people often overlook analogies
(Gick and Holyoak, 1980) even when they are present, will
such a model uncover many creative analogies or are creative
analogies, in some way, different and rare?
 We see the current model as being potentially useful in
three distinct ways, but for now we do not commit to using
it in one particular manner. Firstly, it could be used as a
simple model of creativity, yielding creative interpretations
for a presented problem. Secondly, it could be used as a
tool to assist human creativity; suggesting source domains
to people, to enable them to re-interpret a given target
problem. Finally, it could be used as one possible model of
how people analogize in a creative means.
 The paper is structured as follows: first we describe the
Kilaza1
 model for generating creative analogies, briefly
illustrating its operation on the famous atom:solar-system

1
 Kilaza is not an acronym.
International Conference on Computational Creativity 2012 17
analogy. Then we present results that reflect the model’s
ability to re-generate some well-known h-creative analogies
(Boden, 1992). Finally, the implications of these results
are assessed and some remaining challenges are discussed.

Analogy as an Engine of Creativity
An analogy is a conceptual comparison between two collections
of concepts, a source and target (Gentner, 1983),
such that the source highlights particular aspects of the
target, possibly suggesting some new inferences about it.
In creative analogies, an productive source domain conjures
up a new and revolutionary interpretations of the target
domain, triggering novel inferences that help explain
some previously incongruous phenomena or that help integrate
some seemingly unrelated phenomena (Boden, 1992;
Eysenck and Keane, 1995). Creative analogies differ from
“ordinary” analogies primarily in the conceptual “distance”
between the source and target domains (i.e., these two domains
may never have been linked before) and the usefulness
of the resulting comparison. Both creative and mundane
analogies appear to use the same analogical reasoning
process, as described in the following section, but different
in their inputs and outputs.
 Kekulé’s is famous for his analogy between the carbonchain
and a snake biting its own tail. But this analogy
could have been triggered by many alternative and more
mundane source domains – from tying his own shoe-lace
to buckling his belt. While many source domains could
have generated the creative carbon-ring structure, Gick and
Holyoak (1980) have shown most people (including Kekulé)
frequently fail to notice many potential analogies.
This highlights one potential advantage of a computational
model, in that a model can tirelessly explore all potential
analogies, returning only the most promising comparisons
to a user for more detailed consideration. Thus, computational
models could potentially act a tools helping people
overcome one barrier; namely, their failure to perceive
analogies when they are present.
Kilaza Analogical Creativity Engine
Keane (1994) presented a five-phase model of the analogical
reasoning process, which recognises the distinct phases
of representation, retrieval, mapping, validation and induction.
While other authors describe slightly different
subdivisions of this process, there is broad agreement on
these phases. Our computational model encompasses the
three central phases of analogy (see Figure 1). We highlight
that Walls & Hadamard subdivide creativity into the
phases of preparation, incubation, illumination and verification
(Boden, 1992), which is reminiscent of several
multi-phase models of analogy.
 The heart of our creativity model is the central mapping
phase and this borrows heavily from Keane and Brayshaw’s
(1988) IAM model (see also Keane, Ledgeway &
Duff, 1994). Our model of the retrieval phase attempts to
overcome the semantic bias suffered by many previous
models, improving the diversity of the source domains that
are returned. It was intended that this diversity might address
the quality of novelty (Ritchie, 2001) associated with
creativity, retrieving more “unexpected” and potentially
creative sources. Finally, our model of the validation phase
attempts to filter out invalid inferences, addressing the
quality (Ritchie, 2001) factor associated with computational
creativity.
 Ritchie (2001) identifies the essential properties of creativity
as being directed, novel and useful. We argue that
our model is directed in that it focuses on re-interpreting
some given target domain. Our model addresses the novelty
property by its ability to retrieve potentially useful but
semantically distant, even disconnected, source domains.
Finally, the useful property is addressed through a validation
process that imposes a quality measure on the inferences
that are accepted by the model.
Figure 1: Kilaza is a three-phase model of Analogy
Analogical Retrieval Phase-Model
Existing models for analogical retrieval suffer from the
limitations in the range of possible retrievals because their
they either (i) focus exclusively on domain semantics (like
MAC/FAC; Forbus, Gentner and Law, 1995) or (ii) focus
primarily on domain semantics (like HRR; Plate, 1998).
Other models -- such as ARCS (Thagard et al, 1990) and
Rebuilder (Gomes et al, 2006) - supplement domain representations
by elaboration from external sources (like
WordNet) to widen the net to include more semantically
non-identical sources. However, all of these approaches
arguably over-constraint retrieval for the the proposes of
creativity. We argue that a creative retrieval process must
allow semantically distant and even semantically disconnected
sources to be retrieved, ideally without overwhelming
the subsequent phase-models with irrelevant domains.
Figure 2: Topology is a key characteristic in retrieving creative
source domains
Gentner (1983) mentions two specific qualities are required
of analogical comparisons: semantic similarity and
structural similarity. The model presented in this paper
performs retrieval based exclusively on structural similarity,
performing retrieval based exclusively on the graph
structure (or topology) of each domain description. This
design decision was taken to overcome the semantic narrowness
that constrains existing models, with the hope that
retrieval mapping validation
loves loves
tom jo
liz
loves
(b) Unrequited Love
loves
loves
loves tom jo
liz
(a) Love Triangle
International Conference on Computational Creativity 2012 18
this would increase the possibility of retrieving surprising
and creative source domains. As the example in Figure 2
illustrates, semantics and domain topology are often intertwined.

 Each domain description is mapped onto a location in an
n-dimensional structure space (Figure 3), where each dimension
represents a particular topological quality of that
domain. Structure space is somewhat akin to feature vectors
(Yanner and Goel, 2006; Davies, Goel and Yanner,
2008). Image related analogies are often involved in creative
comparisons (Clement, 2008) and a variety of imagebased
analogy models has been developed, focusing on
specific topics such as; geometric proportional (IQ type)
analogies (Evans, 1967; Bohan and O’Donoghue, 2000),
geo-spatial comparisons (O’Donoghue et al, 2006), spatial
representations of conceptual analogies (Davies et al, 2008;
Yanner et al, 2008) and reasoning about sketch diagrams
(Forbus et al, 2011). Our model performs a single retrieval
process for each presented target, in contrast to the iterative
retrieval and spreading activation phases employed by
KDSA to retrieve semantically distant sources (Wolverton
and Hayes-Roth, 1994).
 Specific topological features used by our retrieval model
include quantifying the number of objects and predicates
(first order and higher order) and number of root predicates
etc. Thus, the representation in Figure 4 might be mapped
onto the location (4 0 2 2 0 0 1) in structure space – 4 object
references, 0 high-order predicates, 2 unique firstorder
relations, 2 first-order relations and 2 root predicates
etc. The distinction between unique and non-unique relations,
for example, distinguishes between domains repeatedly
using a small number of relations and domains that
typically have one instance of each relation in its description.
One advantage of this scheme is that the distance between
domains is not impacted by the number of domains
contained in memory so the retrieval system should scale
reasonably well. For the retrieval results presented later in
this paper a maximum retrieval distance of 10 is imposed –
and only candidate source inside this threshold are considered.

Figure 3: Displacing the Locus of Retrieval within a 3D representation
of n-dimensional Structure Space. Only source domains
within the displaced boundary are retrieved and passed to the
remaining phases of analogy.
 Topologically similar (i.e., homomorphic as well as
isomorphic) domains are mapped onto similar locations
within this topology-based structure space (O’Donoghue
and Crean, 2002). To account for the inferences that were
sought from any inspiring source domain, the locus of retrieval
was slightly offset to account for this additional
source domain material. Included in this offset is the desire
for sources containing additional first-order relations and
high-order relations. However, this offset has relatively
little impact on the final results.
(heavier nucleus electron)
(attracts nucleus electron)
Figure 4: Simplified Model of Rutherford’s Problem
Analogical Mapping Phase-Model
The model for the mapping phase is based on the Incremental
Analogy Machine (IAM) model (Keane & Bradshaw,
1988; Keane et al, 1994). It consists of the three subprocesses
of root-selection, root-elaboration and inference
generation. Mapping proceeds as a sequence of rootselection
and root-elaboration activities, gradually building
up a single inter-domain mapping. Typically a domain description
will consist of a small number of root predicates,
each controlling a large number of (partly overlapping)
lower-order predicates.
Root selection Root selection identifies “root predicates”
within a representation, which are typically the controlling
causal relations in that domain. Each root predicate lies at
the root of a tree of predicates and each root is be seen as
“controlling” the relations lower down the tree. In our implementation
of IAM, the root-selection process examines
the “order” of each predicate. Objects are defined as order
zero and first-order relations that connect two objects are
defined as order one. The order of a causal relation is defined
as one plus the maximum order of its arguments.
Mapping begins with the highest order relations and maps
any unmapped low-order root-predicates last.
Root elaboration Root elaboration extends each rootmapping,
placing the corresponding arguments of these
relations in alignment. If these arguments are themselves
relations, then their arguments are mapped in turn and so
on until object arguments are mapped. Items are only
added to the inter-domain mapping when they conform to
the 1-to-1 mapping constraint (Gentner, 1983).
Inference Generation Each analogical comparison is
passed to the inference generation sub-process. Analogical
inferences are generated using the standard algorithm for
pattern completion CWSG – Copy With Substitution and
# root predicates
# object references
nucleus electron
heavier attracts
potential
source
domains
International Conference on Computational Creativity 2012 19
Generation (Holyoak et al, 1994). In effect, additional information
contained in the source domain is carried over to
the target, creating a more cohesive understanding of that
target problem.
Analogical Validation Phase-Model
The third part of our tri-phase model is focused on analogical
validation. Validation attempts to ensure that the analogical
inferences that are produced are correct and useful.
 O’Donoghue (2007) discusses the accuracy of this validation
process, using human raters to assess the goodness
of inferences that were rated as either valid or invalid.
However, this paper did not assess the models ability to
discover creative analogies.
 Phineas (Falkenhainer, 1990) is a multi-phase model of
analogy that incorporates a post-mapping verification
process. To achieve this Phineas incorporates a model of
the target domain – qualitative physics simulation - illustrating
the power of embedding an analogy model within a
specific problem domain. However, this qualitativesimulation
process effectively limits Phineas to reasoning
only about physical and physics-related analogies.
 The validation model presented in this paper is relatively
simple, aimed at rejecting those predicates that are deemed
invalid – rather than guaranteeing the validity of those inferences
that are accepted. This approach helped maximise
the creative potential of this model, by resisting the rejection
of potentially plausible inferences. Of course, a more
complex validation process could make use of problemspecific
domain knowledge (where available). In the absence
of such domain-specific knowledge verification and
validation of the analogy could be carried using user feedback,
employing Kilaza in a tool-like way.
 The validation phase-model is composed to two main
parts. The first performs validation by comparing the
newly generated inference to predicates already stored
somewhere in memory. The second mode of validation is
more general and driven in part by the functionally relevant
attributes that play a role in analogical inference
(Keane, 1985).
Validation by Predicate Comparison The validation
process compares newly inferred predicates (produced by
CWSG) to the previous contents of memory. Inferences are
firstly compared to predicates in memory, with both the
agent and patient roles potentially being validated independently.
This validation mechanism thus has access to
the entire contents of memory, accessing predicates from
any of the domains stored in that memory. This model of
validation captures the advantages of simplicity and generality,
but it does of course mean that dependencies between
arguments are not captured. This limitation was deemed
acceptable within the context of our desire for a creativity
engine. While many simple inferences were validated by
this mechanism, many creative inferences were not. This
may be partly attributed to the relatively small number of
predicates contained in memory and to the novelty associated
with creative inferences. To address this shortcoming
validation using functional attributes was introduced.
Validation with Functional Attributes Functional attributes
specify necessary attribute requirements for each role
of a predicate – being inspired by the functionally relevant
attributes of Keane (1985). Functional attributes are intrapredicate
constraints that ensure each predicate appears to
be a plausible combination of a relation coupled with each
of its arguments.
 It should be pointed out that functional attributes have
only been used with first-order predicates – those whose
arguments are objects. Although validating higher-order
(causal) relations might make use of the spatio-temporal
contiguity associated with causality, but this cannot be
relied upon (Pazzani, 1991) and is not enforced by our
model. Thus our model treats all causal inferences as implicitly
valid.
 Functional attribute definitions connect each role of a
predicate directly into an attribute hierarchy, whereby arguments
filling those roles must conform to these attribute
constraints. Kilaza stores functional attributes for both the
agent and patient arguments of each relation independently.
More general relations (part-of, next-to) typically
have few functional attributes, whereas more specific relations
(hit, eat) possess a greater number of attribute restrictions.
For example the agent role of hit might require the
hitter to be a physical object, whereas the agent of an eat
relation might have to be a living organism or an animal.
Relations that are more specific are seen to be more amenable
to the validation process, while their more general
counterparts are more difficult to validate accurately.
 In addition, functional attributes have also been used to
support a form of inference adaptation. This allows an inferred
relation to be adapted to a semantically similar relation
that better suits the arguments that pre-existed within
in the target domain. Adaptation uses the functional attributes
to conduct a local search of the taxonomy, to identify
a more semantically suitable relation that better fits the
given arguments.
Data Sets
Three datasets were used to conduct experiments using the
described model. These are referred to as the Professions
dataset, the Assorted dataset and an Alphanumeric dataset.
The dataset contained a total of 158 domains and our creativity
engine attempted to find creative source analogues
for a given number of target problems. It was hoped that
the differing natures of these collections would provide a
reasonable grounds on which to evaluate the computational
model – and to assess its potential to act as a creativity
engine.
Professions Dataset consists of descriptions of fourteen
professions, including accountant, butcher, priest and scientist.
These are rather large domain descriptions created
by Veale (1995) and range in size from 10 to 105 predicates
(M=55.4, SD=29.3). One important feature of the
Professions dataset is its reliance on many different instances
of a small number of relational predicates, including
control, affect, depend, and part. The domains range
from using just 6 distinct relational predicates (ignoring
International Conference on Computational Creativity 2012 20
duplicates) to the most diverse domain that uses 15
(M=8.9, SD=2.2). Another important feature is that this
dataset does not appear to use a set of clearly identifiable
high-order relations (such as a cause, result-in or inhibit)
between first-order predicates.
Assorted Dataset consists of a large number of smaller
and more varied domain descriptions, including many of
the frequently referenced domains in the analogy literature;
such as the solar-system, atom, heat-flow and water-flow
domains. It also includes an assortment of other domains
describing golf, soccer and story-telling. The 81 domains
of the Assorted dataset use 108 distinct (ie non-repeated)
relations. Each of these domains contains between 1 and 15
predicates (M=4.16, SD = 2.9). The average number of
distinct relational predicates in each domain is M=3.48,
indicating that most relational predicates are used just once
in each of the Assorted domains.
Alphanumeric Dataset One final dataset contained 62
semantically constrained domains. However, these domains
contained a great deal of topological diversity. It
was hoped that this mixture of topologies might support
some novel comparisons and inferences and provided a
counterpoint to the semantic richness of the other domains.
Example: p-Creative Re-Discovery of Rutherford’s
Analogy
Before presenting detailed results, we will first see how
Kilaza can re-discover Rutherford’s famous solarsystem:atom
analogy. We highlight that this is a test for the
p-creativity (Boden 1992) of our model – though not necessarily
a model how Ernest Rutherford actually conducted
his own reasoning.
 The traditional representation of this analogy (Figure 5)
is heavily based on a post hoc description of the domains
involved. These descriptions and are heavily influenced by
the analogy itself. We shall first look at the traditional representation
of this domain, before examining how our
model can also deal with more realistic version of how
Rutherford might have thought of the target problem before
arriving at his famous comparison.
Figure 5: Traditional representation of Rutherford’s Solution
First, the semantically impoverished target problem (Figure
4) is mapped onto its location in structure space. We
highlight that the “locus of retrieval” is slightly displaced
from the targets original location to account for the additional
information that one expects to be found in a useful
source domain. In this instance the desired source was retrieved
at a distance of just over 6 “units” in structure
space. The desired source (the solar-system domain) and
all other candidate sources near the locus of retrieval were
passed in turn to the mapping and validation phases of the
model.
 In total 10 other candidate source domains that were
retrieved also generated inferences, most yielding only one
inference each. Three domains generated more than one
candidate inference – but all three were different versions
of the solar-system domain. We point out that our semantic
“free” retrieval process can also trigger identification of the
same source, even if it was represented in a number of alternate
ways (O’Donoghue, 2007). Our mapping model
successfully generated the correct inter-domain mapping
and CWSG generated the desired inferences without adaptation.

Representation Issues in de novo Discovery of
p-creative analogies
We argue that the traditional presentation of Rutherford’s
analogy is a simplified pedagogical device (Figure 5). This
description of the target problem effectively removes much
of the complexity of the real discovery task as encountered
by Rutherford. The description of the target problem uses
terminology specifically designed to accentuate the semantic
(and structural) similarity that is the result of Rutherford’s
comparison – and should not be treated as an input
when re-creating this creative episode.
 This distinction between the problem domain as it would
have existed before the creative analogy and its subsequent
representation after discovering that analogy is a serious
problem - one that is easily overlooked. Any model that
attempts to re-discover known creative analogies must address
the original problem, not just the representation that
accentuates the desired similarity. Differences in domain
terminology and topology are central to the distinction
between elaborating a given analogy, and the much more
difficult task of generating a novel h-creative (or pcreative)
analogy (Boden, 1992).
 We argue that generating Rutherford’s analogy using the
representation in Figure 6 is a far better test of a models
creative ability, than the normal post hoc representation in
Figure 5. Terminological differences are particularly
prevalent in distant between-domains analogies as the firstorder
relationships describing the problem domains originate
in different disciplines. When modeling analogical
creativity, we must expect to encounter these differences in
terminology, and our models of retrieval, mapping and
validation must be able to overcome these problems.
 Ernest Rutherford would most likely have thought of the
target relation between the nucleus and electron as electromagnetic-attraction,
and not the more generic
attracts relation. The corresponding relationship between
source’s sun and planet is gravitation. It is only
after he found the analogy (which involved mapping
revolves
nucleus electron
attracts heavier
and
cause
International Conference on Computational Creativity 2012 21
electromagnetic-attraction with gravitation)
that these relationships can be generalized to a
common super-class like attracts (Gentner, 1983).
 We point out that our model can operate successfully on
either the simplified or more realistic domain descriptions.
This is primarily the result of our retrieval and mapping
models using domain topology, rather than using
identicality (or similarity) between the predicates in both
domains.
Figure 6: More realistic representation of Rutherford’s Analogy
Results of Individual Phase Models
We shall first briefly examine the performance of the retrieval
and validation models in isolation, before looking at
their combined performance in the next section. We shall
briefly examine the results of the mapping model, but our
focus will remain on the inferences that it produced. Results
were produced from a memory containing the three
previously described datasets.
Retrieval Results Retrieval was performed in structure
space. The distance between domains in structure space
varied from 2.645 to 230 (M= 80, SD=57.3), with a large
number of domains being given a unique structural index
in this space. A small number of locations contained multiple
domains – these mostly involved small domains of just
a few predicates from the Assorted dataset.
Retrieval and Mapping A broad tendency was identified
between structure-based retrieval and the size of the resulting
inter-domain mapping, although the correlation was
low. A range effect was identified between structure space
and the size of the resulting mapping, indicating that larger
distances between domains in structure space tend to produce
smaller inter-domain mappings. This indicates a weak
connection between structure-based retrieval and the size
of any resulting mappings.
Validation Results Although the validation model was
very simplistic, it proved surprisingly effective. For example
with the inferences generated on the Professions dataset,
the average (human) rating awarded to predicates that
Kilaza categorized as valid was M=2.62 (SD=2.09), while
the average rating awarded to the invalid predicates was
M=1.57 (SD=1.23). As ratings were given between 1 and 7
with 7 representing clearly valid inferences, this indicates
that many of the generated inferences were of rather poor
quality.
Adaptation Results In addition, 24 inferences were passed
to the adaptation process and 20 of these were adapted.
While we cannot realistically assess if these adapted inferences
matched what was “intended” by our analogy model,
we did assess the validity of these inferences using two
human raters.
 When we look at human ratings for the 20 adapted predicates
before and after adaptation, we see that the average
ratings were increased by the adaptation process - from
1.57 (SD=1.23) to 2.57 (SD=1.70). The average ratings of
the adapted predicates was broadly in line with the predicates
from Kilaza’s valid category above (M=2.62,
SD=2.09). Before adaptation, 18 of the 20 (90%) predicates
were given rated as invalid and after adaptation just
12 (60%) were rated as invalid. Thus, adaptation has a distinct
influence on improving the ratings of the rejected
inferences.
It may well be argued that this adaptation process is itself
somewhat creative – identifying new relations that
better fit the available target arguments. In contrast to the
top-down nature of the creative analogy approach, predicate
adaption is a very much a bottom-up process that is
motivated by the detection of a potential analogical comparison.

Creativity Test Results
To assess the creative potential of our model, we assess its
performance at the p-creative task of re-creating some
well-known h-creative analogies (Boden, 1992). These
include some of the famous examples of creative analogical
comparisons including the Rutherford’s solarsystem:atom
analogy, the heat-flow:water-flow and the
tumour:fortress analogies. Our descriptions are based on
the standard representation of these domains as found in
the analogy literature.
Creative Retrieval We now examine the performance of
our model on the creative retrieval task. We presented our
model with the target domain of each of 10 creative analogies,
together with a memory of 158 source domains. From
this memory of 158 potential sources, the retrieval model
selected a number of these domains as candidate sources.
Only the selected candidate sources were passed to the
mapping and validation phase-models. Evaluating only the
selected source domains was necessary in order to avoid an
exhaustive search through all possible analogical comparisons.
While computationally feasibly in this instance, an
exhaustive search would be impractical on a larger collection
of domains.
 Before looking at the results, we point out that many
comparisons did not generate a viable inter-domain mapping.
Furthermore, most analogies did not generate any
valid inferences. The following results ignore these unproductive
comparisons and we focus only on the productive
analogies.
 All of the desired creative sources were among the candidate
sources that were retrieved by the model. This gives
nucleus electron
electromagnetic
attraction
heavierthan

sun planets
gravitation
more-massive
than
International Conference on Computational Creativity 2012 22
our retrieval model a recall value of 100% on this creative
retrieval task. While a large number of other candidate
sources were also retrieved, this was still a pleasantly surprising
result. The distance within structure space between
the target and the creative sources ranged from 3.1 to 7.9,
suggesting that structure based retrieval was reasonably
accurate in locating candidate sources.
 The precision of the retrieval processing is summarised
in Figure 7. As can be seen, precision was above 0.2 for
two problems showing that few other sources were located
near the structural index of those targets. However, precision
was much lower for most problems, indicating that the
desired source was merely one of a larger number of candidate
sources that had to be explored.
Figure 7 –Precision of retrieval for 10 Creative Analogies
Creative Inferences Next we summarise the inferences
that were generated by each of these comparisons (Table
1). These results implicitly encompass a productive interdomain
mapping between the target and each candidate
source in turn. Kilaza generated and validated the correct
inferences for 9 (70%) of the creative analogies. The cycling:driving
analogy correctly generated no inferences.
Target
Correct
Inferences
Validated
Inferences
Atom: Solar-System y 4
Atom-Falkenhainer: Solar-SystemFalkenhainer
y 3
General: Surgeon y 4
Heat-flow: Water-Flow y 4
Leadbelly : Caravaggio y 4
Love-triangle: Triangle-Directed y 0
Requited-love: Love Triangle y 3
Fish : Bird y 4
Vampire : Banker y 3
Cycling: Driving n 0
Table 1 – Number of Inferences generated by different analogies
 One of these analogies also required one inference to be
adapted. The bird:fish analogy generated the inference
(flies-through fish water), which was correctly
adapted to (swim fish water).
Conclusion
We presented a three-phase model of analogy, adapting it
to function as a tool for discovering creative analogies.
This model encompasses the three central phases of analogy,
namely retrieval, mapping and validation. We argue
that a model encompassing these three core phases of analogy
is the minimum required to be considered a model of
analogical creativity.
 Our retrieval model overcomes the semantic bias of previous
retrieval models, helping retrieve new and surprising
source domains. This helps to improve the novelty of the
source domains identified by our creativity engine. Our
model of the post-mapping validation phase attempts to
filter out any clearly invalid inferences, thereby improving
the quality of the analogies identified as being creative. We
note that novelty and quality are two attributes strongly
associated with creativity (Ritchie, 2001).
 Our three-phase model of analogy successfully rediscovered
10 examples of creative analogies, including
the heat-flow:water-flow and solar-system-atom analogies.
In doing so, the model retrieved the correct source from a
large memory of potential sources. It then developed the
correct mapping and successfully validated (and adapted)
the resulting inferences. We point out that these analogical
comparisons, if produced by a human analogizer, would be
considered creative.
 Our focus on creative analogies rather that the more
normal (or pedagogical) analogies had a far-reaching impact
on the model. Terminological differences are particularly
prevalent in creative between-domains analogies, as
the first-order relations describing each domain originate in
different disciplines. When modeling analogical creativity,
we must expect to encounter these differences and cannot
rely heavily on the presence of identical relations. Our
model successfully created Rutherford’s famous solarsystem:atom
analogy, even when the target was represented
in a more realistic and challenging from. Our model
shows that very significant progress has been made towards
an autonomous creativity machine, re-discovering
many creative analogies.
 We briefly outline three remaining challenges to analogical
creativity, beginning with the issue of knowledge
representation. Our results illustrate a trade-off between the
specifity and the generality of domain descriptions. Overly
specific representations make comparisons more difficult
to discover, but overly general representations appear too
profligate and can overwhelm the validation (and subsequent)
processes. Perhaps multiple representations of each
domain might offer a useful avenue for progress. Multiple
representations might also help explain why exerts are
more fluent in their use of analogy within their own domains
(Dunbar and Blanchette, 2001). Our model does not
currently include an explicit re-representation process
International Conference on Computational Creativity 2012 23
highlighting “tiered identicality” (Gentner and Kurtz,
2006).
 It seems that the greatest challenge to computational
analogizing might lie with the post-mapping phases. Challenges
include assessing analogical inferences for validity,
evaluating the significance of an analogy and considering
the implications of creative comparisons. Surprisingly little
attention has been given to this phase – partly because of
its ultimate dependency on the target problem domain.
Phineas (Falkenhainer, 1990) and also Rebuilder (Gomes
et al, 2006) showed that integration of the analogy and
case-based reasoning within the target domain can have
very positive effects. While tight integration of all target
domains into an analogy model seems most unlikely, Kilaza
has show that a generic validation model can play a
part improving the quality of the inferences that are accepted.

 Overall, the results presented in this paper highlight that
a three-phase model of analogical reasoning can operate
successfully as a model of analogical creativity. Our results
highlight the improbability of finding a suitable source
domain to re-interpret a given target in a creative manner.
Extending this model will necessitate a tighter integration
of the analogy process with other facets of intelligence.
<references_biblio/>
References
Boden, M. 1992. The Creative Mind. London: Abacus.
Bohan, A. O’Donoghue, D.P. 2000. A Model for Geometric
Analogies using Attribute Matching, Proc AICS, Galway,
Ireland, pp 110-119.
Brown, T.L. 2003. Making Truth: Metaphor in Science.
University of Illinois Press, New York, USA.
Clement, J. 2008. Creative Model Construction in Scientists
and Students: The Role of Imagery, Analogy, and
Mental Simulation. Dordrecht: Springer.
Davies, J. Goel, A.K. Yaner P.W. 2008. Proteus:Visuospatial
analogy in problem-solving, KnowledgeBased
Systems 21(7): 636-654.
Dunbar, K. and Blanchette, I. 2001. The in vivo/in vitro
approach to cognition: The case of analogy, Trends in
Cognitive Sciences, 5(8): 334–339.
Falkenhainer, B. 1990 A Unified Approach to Explanation
and Theory Formation, in Shrager, J. & Langley, P. (eds.)
Computational Models of Scientific Discovery and Theory
Formation, Morgan Kaufman: CA. pp 157-196.
Forbus, K. Gentner, D. Law K. 1995. MAC/FAC: A Model
of Similarity-based Retrieval, Cognitive Science, 19(2):
141-205.
Forbus, K. Usher, J. Lovett, A Lockwood, K. Wetzel J.
2011. CogSketch: Sketch Understanding for Cognitive
Science Research and for Education, Topics in Cognitive
Science, 3(4): 648–666.
Gentner, D. Kurtz, J. 2006. Relations, Objects, and the
Composition of Analogies, Cognitive Science, 30(4): 609–
642.
Gentner, D. 1983. Structure-mapping: A theoretical
framework for analogy, Cognitive Science, 7(2): 155–170.
Gomes P, Seco N, Pereira FC, Paiva P, Carreiro P, Ferreira
JL, Bento C. 2006. The importance of retrieval in creative
design analogies, Knowledge-Based Systems, 19(7): 480–
488.
Holyoak K. J. Novick L. Melz E. 1994. Component Processes
in Analogical Transfer: Mapping, Pattern Completion
and Adaptation, in Analogy, Metaphor and Reminding,
Eds. Barnden and Holyoak, Ablex, Norwood, NJ.
Keane, M.T. 1985. On Drawing Analogies When Solving
Problem, British Journal of Psychology, 76: 449-458.
Keane, M.T., Bradshaw, M. 1988. The Incremental Analogical
machine: In D. Sleeman (Ed.) 3rd European Working
Session on Machine Learning, Kaufmann CA: 53–62.
Keane, M.T., Ledgeway, T. Duff, S. (1994). Constraints on
analogical mapping, Cognitive Science, 18: 387-438.
O'Donoghue, D.P. 2007. Statistical Evaluation of ProcessCentric
Computational Creativity: 4th International Joint
Workshop on Computational Creativity (IJWCC), Goldsmiths,
University of London, 17-19 June.
O'Donoghue, D.P, Bohan, A. Keane M.T, 2006. Seeing
Things: Inventive Reasoning with Geometric Analogies
and Topographic Maps, New Generation Computing,
Ohmsha Ltd. and Springer 24(3): 267-288.
O'Donoghue, D.P. Crean, B. 2002. Searching for Serendipitous
Analogies, ECAI - Workshop on Creative Systems,
Lyon, France, 21-26 July.
Pazzani, M. 1991. A Computational Theory of Learning
Causal Relationships, Cognitive Science, 15(3): 401-424.
Plate T. 1998. Structured Operations with Distributed Vector
Representations, in Advances in Analogy Research,
New Bulgarian University, Sofia, Bulgaria.
Ritchie G. 2007. Some Empirical Criteria for Attributing
Creativity to a Computer Program, Minds & Machines, 17:
67–99.
Ritchie G. 2001. Assessing Creativity, Proc. AISB Symposium
on AI and Creativity, York, March.
Thagard, P. Holyoak K. J. Nelson, G. Gochfeld, D. 1990.
Analogue Retrieval by Constraint Satisfaction, Artificial
Intelligence, 46: 259-10.
Veale, T. 1995. Metaphor, Memory and Meaning, PhD
Thesis, Trinity College, Dublin, Ireland.
Wolverton M. Hayes-Roth, B. 1994. Retrieving Semantically
Distant Analogies with Knowledge-Directed Spreading
Activation, Proceedings AAAI, pp 56-61.
Yaner, P.W. Goel A.K. 2006 Visual analogy: Viewing
analogical retrieval and mapping as constraint satisfaction
problems, Applied Intelligence, 25(1): 91-105.
International Conference on Computational Creativity 2012 24