specific knowledge that the artificial artist could draw on in a case base
of prior works in the particular genre of interest.In the application they
describe,the case base is a library of bebop jazz melodies.The evolved
musician programs could take examples of melodies from the case base
and alter them with a set of predetermined transformation functions,
such as INVERT,AUGMENT,and COMPARE-TRANSPOSE,which
are also largely culture specific.
The fitness function in this bebop case consisted of five critical crite-
ria gleaned from jazz improvisation techniques,rules that looked for a
balance of novel tonal material and material taken from the case base,
or for rhythmic novelty balance,and so on.Musician programs that gen-
erated new bebop melodies meeting these criteria would have more off-
spring in the next generation (created by reproduction and crossover
alone,to scramble existing combinations of transformation functions).
Spector and Alpern ran their system with a case base of five Charlie
Parker song fragments of four bars each.After 21 generations of popu-
lations with 250 evolving composers,individuals emerged that could
produce four-bar “improvisations”that were found highly satisfying by
the five-rule critic.The system’s creators,however,were less impressed:
“Although the response...pleases the critic,it does not please us[the
authors] particularly well”(Spector and Alpern 1994:7).They do not see
this as a failing of the evolutionary artist construction method in general.
Instead,they believe that with proper choice of critical rules,the
approach can be made to succeed,and “nobody said it would be easy to
raise an artist”(p.8).
But a deeper problem remains with rule-based critic approaches in
general,as people found earlier with rule-based composers.Artificial
critics who go strictly by their given rules,as opposed to more forgiving
(or sloppier) human critics,are generally very brittle.They may rave
about the technically correct but rather trite melody,while panning the
inspired but slightly off passage created by just flipping two notes.In fact,
for good composers it is critical to know when to break the rules.As a
consequence,for critics it is imperative to know when to letthe com-
posers break the rules.Rule-based systems,by definition,lack exactly this
higher-level knowledge.Critics based on learning methods such as neural
network models,on the other hand,can generalize judgments sufficiently
to leave (artificial) composers much-needed rule-breaking “wiggle
room,”although this too can end in cacophony,as we will see.Learning-Based Critics
To remove (or at least transform) the necessity of human interaction in
the algorithmic composition process further,critics used in evolving arti-
ficial composers can be trained using easy-to-collect musical examples,371 Simulating the Evolution of Musical Behavior