eliot update


Subject: eliot update
From: Eliot Handelman (eliot@generation.net)
Date: Fri Aug 19 2005 - 01:02:01 EDT


If I may be so immodest as to announce where I am in work & research, I
announced a while back that
I'd begun an AI composing program. What really happened was this. I kept
feeling my conceptualiztion of
rhythm wasn't adequate and so I worked for some months trying to build
computational models of whatever I
came up with as basic, testing those ideas with a huge folksong database
(through which I became enamored of old
german folksongs), evolving diffrent ideas, etc, which was threatening
to turn into a full-scaled detailed research jaunt of several years.
I began to feel I was getting nowhere. One of the things I persistently
studied were tunes generated by
David Temperley's tune generator "melisma" (on net) since these have a
very fake, localized feel with no overall motivic logic,
shapeless and irritating. My progam was about trying to represent and
evaluate the difference between
this and a random german folktune, saying which is human and which
machine, so as it were
running a turing test on another computer program. I got fed up with
Temperley (which is based on "arch-like contour" and "leaps have to be
filled
because someone theorizes we expect this to happen.") and decided to
write some really dumb melody generators. The fifth of
these clicked and turned into the big AI program. It's one voice at a
time for now, which chapter 2 nearly done (I call the levels "chapters"
each based on the previous chapter) and it's just making short pieces.
Everything is in c major so far (the output is supposed to be
more or less tonal). Some of the tunes are suprisingly nice and
characterful -- about 7% -- with easily 70%-80% acceptable. The
variability of the music is
good -- with 500 test runs (1000's more generated for testing) easily
25% have an individual character. It's just about to
learn how to vary some basic rhythmic shapes.

AT present the program has an extensive repertory of shapes that it can
combine and reanalyze, varying features or transforming
in some pattern of shape. Shape is recursive and is applied at many
levels. A piece consists of many patterns controlling many
different aspects of the music -- of the shapes. Thus the output
occupies something like a "pattern space".

One problem with getting the dimensionality of pattern space high is
that it's
essentially an np constraint problem, so the compromise is a sort of
random walk through pattern space inhabited by a relational vocabulary
of shape. Ok, this is getting too verbose, so I'll break here, otherwise
I'll start going on about my favorite subject, parallels.

This has been about 20 years to get to square two, I'm hoping the next
few chapters get done in a year or two. Time's
catching up -- I've just gone five-o says freeze describing my age and
bank account.

-- eliot



This archive was generated by hypermail 2b27 : Sat Dec 22 2007 - 01:46:10 EST