still learning and reproducing largely surface-level features of the
example musical input;whereas they should in principle be able to pick
up and use deeper temporal structure,“experiments...show no cause
for optimism in practice”(Mozer 1994:274).In addition,by merely
manipulating surface-level musical aspects of the training set,networks
can come up with new compositions,but they will not be particularly
novel in an interesting way.We do not currently have many good models
of the generation of musical novelty at an individual level.But in an evo-
lutionary computer simulation,we can at least explore the appearance
of novelty in musical behavior at a population level,as we will see below.
First,we must consider how evolutionary systems can act on the rule-
following or learning behavior of simulated individuals and thereby
create successive generations of new individuals.Evolving Musical Systems
To simulate the evolution of musically behaving organisms (or anything
else),we need only construct a rather simple loop:generate,test,repeat.
Basically,we make a bunch of things,test them according to some crite-
ria,and keep the ones that are best according to those criteria;then we
repeat the process by generating a new bunch of things based on the old
ones.This loop continues for possibly many generations until the things
we are making are good enough according to the criteria being used,or
when we have simulated enough of the evolutionary process to answer
our particular research question.The complication comes when we have
to specify what we mean by “generate”and “test.”In natural evolution,
what are being generated are individual organisms through a process of
genetic modification (usually either sexual recombination or asexual
cloning,both with some possible mutation),and the criteria of success
are the forces of natural and sexual selection (i.e.,ability to survive and
reproduce).(Furthermore,in natural evolution there is no “stopping
point”when some criteria have been met;the test keeps changing as a
consequence of continuing evolution of other species as well.) What and
how should we generate and test when dealing with artificial music-
perceiving and -producing creatures?
On the generation side,our system should create artificial organisms
endowed with either a set of musical rules to follow or a neural network
or other learning mechanism,depending on the choice made according
to considerations in the previous section.The means of testing also
depend on the research question.Basically,we must first decide what we
are on the lookout for:do we want to know when,for example,a certain
kind of music-perceiving mechanism first appears in the neural network366 Peter Todd