rather than constructed using difficult-to-determine musical rules.
Baluja,Pomerleau,and Jochem (1994),for instance,working in the visual
domain,trained a neural network to replace the human critic in an inter-
active image-evolution system similar to that created by Sims (1991).The
network “watches”the choices that a human user makes when selecting
two-dimensional images from one generation to reproduce in the next
generation,and over time learns to make the same kind of aesthetic eval-
uations as those made by the human.When the trained network is put
in place of the human critic in the evolutionary loop,interesting images
can be evolved automatically.With learning critics of this sort,whether
applied to images or to music,even less structure ends up in the evolved
artificial creators,because it must get there indirectly by way of the
trained fitness-evaluating critic that learned its structural preferences
from a user-selected training set.We can thus expect a great degree of
novelty in compositions created by this approach,but how will they
sound?
Spector and Alpern (1995) extended their earlier rule-based system
to find out.They expected that a neural network trained to make aesthe-
tic evaluations of a case base of melodies would be able to evaluate
the musical output of evolving composers at a deeper structural level
than their rule-based critics could.This time their composers were to
create single-measure responses to single-measure calls in a collection
of Charlie Parker melodies.The composers were again evolved in the
genetic programming paradigm,but using more abstract (less musically
specific) functions than before.The critic neural networks were trained
to return a positive evaluation of one measure of original Charlie Parker
followed by the correct next measure.They were also trained to return
negative evaluations of one Charlie Parker measure followed by differ-
ent kinds of bad continuations:silence,random melody,or chopped-up
Charlie.To evaluate a given composer program,the program was given
an original Charlie Parker measure as input,and both that input and the
composer program’s one-measure output were passed to the neural
network critic.The critic then returned a fitness value indicating how well
it thought the composed measure followed the original measure.
One advantage of such a system is that new critical constraints can be
added simply by training the neural network critic on additional musical
examples,rather than by constructing new rules.The problem,though,is
that one can never be sure the network is learning the musical criteria
one would like it to,as Spector and Alpern discovered.As in their earlier
work,a composer program with very high fitness value was found
quickly,in fact,after only a single generation of evolution.But as before,
its performance did not meet the standards of its human overseers:in
response to a simple measure of eight eighth-notes,it returned a mon-372 Peter Todd