How Models of Creativity and Analogy Need to Answer the Tailorability Concern - Computational Creativity Research: Towards Creative Machines - page 97

• Across multiple, more-or-less equivalent structural representations of the same input data,

and a wide variety of domains, the matcher still produces the desired results.

• The desired results are still produced when the input is minimal; meaning any redundant

information or structural constructs which might be identified by critics as being used
only to aid the matcher can be removed.

Of course, if a system did happen to require no more than unstructured, natural-
language descriptions as input, or direct sensory data from visual representations,
it would satisfy both of these conditions. This allows our criteria to encompass the
alternate route to answering the TC mentioned by Gentner and Forbus [ 24 ]—a route
which seeks to answer TC by not necessarily having a large database, but by having
one that at least attempts to directly construct structured representations from low-
level sensory or natural-language data. The generality of these conditions allows us
to claim the converse of TCA 2 , leading us to our next iteration:
TCA 3 A computational system of analogy answers the TC if and only if given no more than
either

• unstructured textural and/or visual data, or

• a large, pre-existing database,

andminimal input, it is able to consistently produce useful analogies and demonstrate stability
through a variety of input forms and domains.

One might be satisfied with this set of criteria, which draws its strength from its
lack of commitment to any particular theory of analogy, and its emphasis on large sets
of non-tailored input. But TCA 3 is undeniably ambiguous, and may not be focused
enough to guide any research program. We encourage the reader to take TCA 3 and
develop it further, but first, to close out this paper we will take some steps of our own
to reduce some of its weaknesses.

5.2.1.3 Strengthening the TC with Psychometric AI

We will make two important moves to sharpen TCA 3 . One, we turn to Psychometric
AI (PAI) [ 9 , 13 ], according to which, in a nutshell, commendable AI systems are
those that demonstrate prowess on tests of various mental abilities from psychomet-
rics. Our second move is to embed analogy-generation systems within broader AI
problem-solving systems that make use of additional forms of fundamental reason-
ing in human-level intelligence; e.g., deduction. In particular, we place TC within
the context of the integration of analogical reasoning with deduction, which we dub
analogico-deductive reasoning (ADR), and which we have explained and demon-
strated elsewhere [ 11 ]. An ADR system does use analogy generation, but analogies
are used to guide solutions that can be rigorously verified by proof. The architecture-
sketch of an ADR system that accords with our pair of moves is shown in Fig. 5.2 .
While we don't have the space to provide details here, this system receives puzzles
that are part linguistic and part visual in nature (e.g., so-called seating puzzles ), and