Stakeholder Groups in Computational Creativity Research and Practice - Computational Creativity Research: Towards Creative Machines - page 26

serve as a driving force behind future creation, allowing someone to work towards
goals that they have set themselves and strengthening claims of intentionality.
Despite this, little work has been done to build systems which can generate aes-
thetic preferences of their own and apply them intelligently. One reason for this may
be the uncomfortable clash between the subjective and the objective that so often
affects research in Computational Creativity. The notion of 'optimality' in many
creative domains, particularly those associated with the arts, is a contentious one
and leads to much criticism of systems which attempt to quantify the quality of an
artefact. The idea of having a system quantify the quality of an opinion on creative
artefacts is equally controversial, if not more so. Similarly, in the past, the question
of how to quantify the degree to which a system is creative was also a subjective
and controversial task. In this case, researchers such as Ritchie found it useful to use
metrics which dealt with abstract notions of creativity without directly laying out
objective measures of quality for any particular artefact or medium. Ritchie's criteria
are described in [ 55 ], and have been used in many evaluations of creative systems in
a variety of different fields and media.
We propose here a similar set of criteria which apply to aesthetics or preferences
rather than creative systems. By using abstract metrics, we can avoid talking about
aesthetic measures in objective ways, while retaining a meaningful vocabulary with
which to describe different kinds of aesthetics. These metrics can be used to evaluate
aesthetic comparator functions , namely binary functions which take two examples
of a type of object, and then return 1, 0 or 1 depending on whether the first object

is preferred less, the same as, or more than the second object. Assuming we have
an aesthetic function f , and a set of objects the function expresses a preference
over, O , we define the following criteria which can be used to differentiate aesthetic
functions from one another. Note that these metrics do not necessarily represent a
linear gradient of quality—different types of aesthetic function may be desirable in
different scenarios.
The first metric is specificity . Specificity captures the degree to which the aesthetic
represents a total order over the set of objects O . If an aesthetic can offer a definite
preference (that is, a nonzero result) for many of the objects, it will have a high
specificity, and vice versa. High-specificity aesthetics might suggest the aesthetic is
experienced or well-developed in some way, if it is able to make clear distinctions
between many different artefacts.
The second metric is transitive consistency . This captures how self-contradictory
the aesthetic function is. Suppose we have three artefacts: A , B and C , and our
function f . We can write A

−

< B to indicate that B is preferred to A . We might

expect that if A C . Transitive consistency measures what

proportion of O this holds for. In some scenarios, wemight want a high transitive con-
sistency, as this indicates a lack of contradictions in the preferences being expressed.
However, in some scenarios, preferences can be complex and multi-objective, and it
might be the case that transitivity does not hold for highly subjective opinions about
artefacts produced by creative acts.
The third metric is agreement . Instead of being expressed in terms of a single
aesthetic function, agreement is expressed about two different aesthetics, which we

< B and B < C then A <