How Models of Creativity and Analogy Need to Answer the Tailorability Concern - Computational Creativity Research: Towards Creative Machines - page 94

5.2.1.1 What Is a 'Good' Analogy?
Though the widely influential Structure-Mapping Theory [ 22 ] offers the systematic-
ity principle and the one-to-one constraint as indicative features of a good analogy,
it does not provide a clear quantitative measure of analogy quality. SME evaluates
match scores by combining the scores of the evidence provided by its match rules;
this allows for a comparison between different matches of the same problem. But
the resulting match score is not normalized, and as a result match quality between
different problems cannot be compared [ 21 ]. Other analogical models do not help
much in this regard. Holyoak and Thagard's [ 27 ] Multiconstraint Theory, for
example, introduces additional criteria to evaluate what makes an analogical match
a good one, making cross-domain analogical match quality more difficult to assess.
This is especially problematic when considering TC. Given a set of predefined
problems and desired answers for all of them, the good analogy is simply the one
that performs as intended. But when applied to problem sets where the best analogy,
or even the existence of a good analogy, is not clear even to the persons running the
experiment, do the guidelines currently available still apply?
Paul Bartha [ 4 ] offers a comprehensive theory of analogy that can be useful
here. His goal is to produce a normative account of analogy, and in the process he
sheds light on the proper role of analogy within a larger context of reasoning. His
ArticulationModel is based on the idea that there are two features common to all good
analogical arguments: a prior association (a relationship in the source analog that is
transferred to the target), and a potential for generalization. But perhaps most relevant
to our present discussion is his claim that by virtue of the analogical argument, the
resulting hypothesis inferred through an analogical argument contains no more than
prima facie plausibility, which can be understood as something like a suggestion
that the inferred hypothesis is worth exploring further. If there is an independent
reason to reject the hypothesis, such as a deductive argument showing it leads to a
contradiction, or contrary empirical evidence, then it can be abandoned.
The idea that the proper role of an analogical argument is to do no more than
provide prima facie plausible hypotheses [ 37 , 38 ] suggests that the relationship
between an analogy's match quality and its tendency to produce hypotheses which
can be independently verified may not be as simple as it might seem. In the end,
a model of analogy is a good one only if it produces good analogies, and an analogy
is a good one only if it produces plausible hypotheses. 2
This complicates things further. In toy examples, the information available to
the analogical matcher is very limited; a hypothesis in these cases is plausible if
the information available to the analogical matcher does not provide a reason to
reject that hypothesis. But if we are suggesting that the information available to
the analogical system (or more specifically, the filtering process) is actually a large
2 We might leave room here to exclude models of analogy that have psychological or neurological
plausibility as their primary end goals. In these cases, it might be the goal of the model to replicate
poor analogical reasoning as well, if it matches human performance. But it is our assumption (at
least in the present inquiry) that the ultimate goal of AGI research is not to model poor human
reasoning.