know much about music,only what he or she likes.This means that the
process of evolution does not have to shape a system made up of a set
of musical rules,but rather must sculpt a music learner,and possibly a
corpus of musical examples for each learner to be exposed to.(This
corpus can also arise through interactions of each generation of artificial
composers,for example,or could be evolved through a cultural process
of learning and handing down musical examples across generations.)
Early instances of the learning approach to algorithmic composition
analyzed a collected set of musical examples in terms of their overall
pitch-ransition probabilities (Jones 1981;Loy 1991).Based on how often
particular pitches followed each other in the examples,new compositions
could be constructed with similar statistical structure.Such Markov-
process music sounds good over the short term,reflecting the note-
by-note structure in the original input.Novelty is also introduced
through the probabilistic nature of the composition process.This matrix
representation is also easy to represent and evolve computationally.But
a difficulty arises when we listen to music generated in this way over
the long term:it has no structure beyond the moment,and the novelty
of randomness often accumulates and leads compositions to wander
aimlessly.
The development of new neural network learning algorithms (Rumel-
hart and McClelland 1986) led to the possibility of connectionist music-
composition systems (Todd 1988,1989;Todd and Loy 1991;Griffith and
Todd 1998).Feedforward and recurrent neural networks can be trained
to produce successive notes or measures of melodies in a training set,
given earlier notes or measures as input.Once they have learned to
reproduce the training melodies,they can be induced to compose new
melodies based on the patterns they have picked up.Neural networks
can be made to learn more abstract and long-term patterns than typical
Markov-process systems,allowing them to incorporate a greater amount
of musical structure from the example set.In addition,they can have
additional structure built into them,including psychologically motivated
constraints on pitch and time representation (e.g.,Mozer 1991,1994)
that help their output to be more musically appropriate.Furthermore,a
reasonable amount of research has been conducted into the ways that
neural networks can be represented and manipulated in evolutionary
simulations (e.g.,Miller,Todd,and Hegde 1989;see chapter 2 in Mitchell
1996,for a review).
Yet,despite the increasingly sophisticated neural network machinery
being thrown at the problem of composition,results to date have been
rather disappointing.As Mozer commented about his own CONCERT
system,outputs are often “compositions that only their mother could
love”(Mozer 1994:274).Much of the problem is that these networks are365 Simulating the Evolution of Musical Behavior