of a simulated creature? If so,we must devise an automated fitness-
evaluation routine that can check for the presence of this mechanism
and allot more offspring to those creatures that have something closer
to this mechanism than others.Or do we want to see how a particular
music-generating ability appears and spreads through the population? In
this (more commonly explored) case,we could have one of several dif-
ferent types of critics evaluating individuals in the population and assign-
ing them fitness values according to their musical ability.We can use
humans directly as such critics,listening for certain behaviors,or we can
again use automated critics that are themselves rule based,or learning,
and possibly coevolving with the music-creating individuals themselves.
Examples of each of these types of fitness-determining critics are given
below.
Whatever type of generating process and fitness-testing function we
use,the evolutionary process as a whole will embody several common
features.These evolutionary systems are based on the general framework
provided by Holland’s original genetic algorithm (GA;Holland 1975;see
also Goldberg 1989,and Mitchell 1996,for general introductions),either
directly,or indirectly by way of the genetic programming paradigm of
Koza (1993) in which chunks of computer program code are evolved.
In nearly every case,new populations of potential solutions to some
problem (here,one related to musical behavior) are created,generation
after generation,through three main processes.First,to make sure that
better solutions to the problem will increase over time,more copies of
good solutions than of bad solutions from one generation are put into
the next generation (this is fitness-proportionate reproduction,because
fitter solutions have proportionally more offspring).Second,to introduce
new solutions into the population,a low level of mutation operates on
all acts of reproduction,so that some offspring will have randomly
changed characteristics.Third,to combine good components between
solutions,sexual crossover is often employed,in which the “genes”of two
parents are mixed to form offspring with aspects of both.
Evolutionary simulations,like evolving populations in nature,are good
at exploring the space of possible solutions to the posed problem,
because they can consider several such solutions in parallel and combine
aspects of the best.However,evolution is not often described as being
fast,although it can be in some cases (see e.g.,Weiner 1994),and patience
is commonly called for in artificial evolutionary systems.It can take many
generations of artificial creatures,each of which must be evaluated in a
time-consuming fashion (by whatever type of critic we are using,espe-
cially those that must “listen”to lengthy musical output,leading to what
Biles [1994] and others identify as the “fitness bottleneck”),before any
interesting behavior comes along.The main reason for this sometimes367 Simulating the Evolution of Musical Behavior