production or perception;here I describe their use in a compositional
role,which has been more common in recent evolutionary music simu-
lations,but the reader should keep in mind that they can typically be
inverted into an analysis or perceptual role as well.The choice of which
mechanism to use in a particular simulation is determined by the
research question the simulation is designed to answer.
Probably the most intuitive means of creating an artificial system for
music composition or analysis,particularly one to be implemented on
a computer,is to come up with a set of compositional rules for the com-
puter to follow either in producing or assessing music.The rules in a
musical algorithm can be very simple,as for example,in musical dice
games developed by Mozart and others in which precomposed phrases
were merely combined in new random orders (Loy 1991).They can
also embody complex knowledge about specific musical styles,as in
Ebcioglu’s (1984) collection of rules for chorale harmonization.Because
the computer is constrained to follow the rules it is given,its composi-
tions will generally be well formed according to those rules,and thus will
attain at least some minimal level of musicality.(Indeed,Mozart’s dice
compositions cannot help but sound reasonable.) On the other hand,we
pay the price for this rule-following lawfulness:compositions from rule-
based systems are unlikely to be surprising.One could hardly be shocked
by the combinations produced by Mozart’s dice music.
Perhaps more discouraging,coming up with rules to put into the algo-
rithm in the first place is no simple matter.For centuries,scholars have
tried to specify fully the rules involved in particular musical styles,such
as counterpoint;but whenever a set of rules is nailed down,exceptions
and extensions are always discovered that necessitate more rules (Loy
1991).This is the other price of a highly structured composition system—
the cost of creating the rightstructure.This cost must be paid in any
system that is evolving rule-based artificial composers as well.Most prob-
lematical,many artists question whether creativity can be captured by a
set of rules at all.If not,we may well want to explore musical evolution
using other computational models.
A second approach to construction of artificial musicians is to train a
learning system to create new pieces of music.Rather than requiring
the development of a set of musical rules,a learning composition system
can simply be trained on a set of musical examples.These examples are
chosen to represent the kind of music that the user would like the com-
position system to create new instances of (or at least mimic old instances
of):for a waltz-composing system,train it on a corpus of waltzes;for a
Bach-Hendrix amalgamator,train it on melodies from both composers.
Thus,the big advantage of a learning composition system over a rule-
based one is that,as the saying goes,the system builder does not have to364 Peter Todd