How Models of Creativity and Analogy Need to Answer the Tailorability Concern - Computational Creativity Research: Towards Creative Machines - page 92

Fig. 5.1 One version of the
popular “heat flow” example
from Falkenhainer et al. [ 17 ]
used to test many analogical
systems
For example, the well-known water-flow/heat-flow analogy (Fig. 5.1 ) has been
used as a demonstration of many models of analogy [ 17 , 27 , 31 , 43 ]. But little to
nothing is written about how the structural representations used in examples such
as these are acquired in the first place. One approach is to model the acquisition
of structured representations through sensory data (e.g., see [ 16 ]), and another is to
presume the existence of a large database of already-structured data (such as that to
which a neurobiologically normal adult might be expected to have access), and some
sort of filtering process that ensures that from this database, proper representations
are selected and any unnecessary data that would produce incorrect matching results
are excluded. 1 Yet even when such filtering processes are proposed, they are not
put to the test and proven to perform well with a large database containing enough
knowledge to match that of a child's, much less an adult's. The TC rightfully attempts
to refocus efforts on these filtering processes, by requiring that they demonstrate the
ability to produce clean source and target analogs as required by the analogical
mappers.
The field of case-based reasoning (CBR), which overlaps quite heavilywith that of
analogical reasoning (AR), also deals with some of the issues raised by the TC. There
are differing opinions on what features distinguish the CBR and AR approaches (see
[ 1 , 14 , 34 ]), but two common themes are that CBR tends to deal with source and
target cases that come from the same domain, and cases are selected and adapted with
some clear pragmatic goal in mind. AR approaches, on the other hand, try to be more
generally applicable across different domains, and tend to focus more on the mapping
process that actually determines analogical similarity. CBR approaches, then, deal
with the TC by trading generality for effectiveness, so that a program designed to
work well in one domain (medical diagnosis, for example, is a popular field) may
not work so well in another without a significant amount of human assistance.
Unfortunately, the CBR approach of restricting generality does not sufficiently
answer the TC. Analogy research can be seen as centering around fundamental
1 It would be less appropriate to urge a careful treatment of the TC and tie it so closely to large
semantic databases if they weren't available. But over the past few years, natural-language process-
ing and semantic-web technologies have been progressing to the point where we now have access to
large collections of semantic databases containing wide-ranging general knowledge. These include
Cyc [ 35 ], Freebase [ 7 ], and DBPedia [ 3 ]. Many of these have easy-to-use interfaces.