its downside as well:“A clever representation that efficiently represents
alternative solutions,perhaps by excluding clearly unacceptable solu-
tions,will lead to a more efficient search.However,if a representation
‘cleverly’ excludes the best solution,its efficiency is irrelevant....GA
designers walk a thin line between too large a search space on one
side and inadequately sampled solutions on the other”(1994:132).This
tradeoff between extra initial built-in musical knowledge,designed to
give the population a head start and cut down on evolution time,and
unwanted constraints on the search space of musical behaviors,must be
considered as well in evolutionary simulations addressing specific
research questions.Rule-Based Critics
The fitness bottleneck encountered when humans are critics of evolu-
tionary systems,listening to the musical output of each individual in the
population,can be eliminated by creating automated fitness evaluators.
Traditionally,computational evolutionary systems were designed with
readily computable fitness functions in mind.This meant that genetic
algorithms (see Goldberg 1989) and genetic programming methods (see
Koza 1993) generally employed simple rules or more complex rule-based
algorithms to compute the fitness of each member of the evolving pop-
ulation of problem solutions.It is not surprising,then,that the earliest
applications of computational evolutionary methods to music also used
rule-based fitness functions or critics.In what was probably the first
musical genetic algorithm,Horner and Goldberg (1991a,b) used the GA
to search for thematic bridges,sequences of simple operations that would
transform an initial note set into a final desired note set within a certain
number of steps.Both the nature of evolving individuals—sequences
of operations that can be chopped up and mixed back together—and
the specific goal of the fitness function made this musical application
well suited for evolutionary search.As a consequence,the results were
“...musically pleasing to the authors with the usual qualifications
regarding personal taste”(1991b:5).But given its highly structured
inputs,genetic operations,and goals,this compositional aid system could
show little unexpected novelty in its output.
Spector and Alpern (1994) aimed at a more general goal:automatic
construction of synthetic artists that could operate in any specified aes-
thetic tradition.They strove to accomplish this by segregating all aes-
thetic considerations into a distinct critic,and using a method that could
create artists that met those critical criteria.The method they chose is
genetic programming (Koza 1993),which in this application evolves pro-
grams that produce artistic output that is judged by the critic acting as
a fitness function.Spector and Alpern further supplied much culture-370 Peter Todd