🚀 go-pugleaf

OK, lots of recent bashing of "knowledge engineering" but I'd like to
clarify a few things.... while "ke" is overused and misused, and I'm not
sure what the hell it means anymore one shouldn't necessarily retreat to
behavioristic approaches (that is, an approach that models intelligent
behavior as has been recently suggested here). 

What it does mean, is that once cannot do ke for music because there aren't
any decent KRs yet (knowledge representations). I think developing a good kr
for musical "rules", both in terms of basic music theory (e.g. what scales
seem to sound good, or why modulating the tonic to the fourth and fifth is a
good progression..) and for certain stylistic information is necessary
before any real work can proceed anywhere. In some sense this is an argument
against neural net approaches, etc. because while one may be able to come up
with a system which happens to behave nicely, one cannot in any sense
describe what the system actually "know"s. KE, for music, then, consists of
taking this kr, from which one can theoretically come up with music charts,
and actually encode the special (stylistic/performance) and general
(theoretic) knowledge.

The goal would not be, of course, to simply have a prolog-like program that
could generate a valid musical piece, but rather to have a prolog-like
program that could act as an intelligent arranger/compositional enhancement
tool. (not that the UI would in any sense be prolog-like of course, I would
simply hope to encode the knowledge in this fashon). Something of an expert
system in the general sense, although the techniques would actually come
more from the planning and natural language domains (since a peice of music
is very much like a plan to be executed: one wants to acheive certain goals,
they must be done in some temporal order, and things like duration are
crucial to acheiving particular effects.

My work is much too sketchy at this point to present anything too cogent
here (e.g. the rule structures), but the point is, as someone in the AI
research field, I don't think one should discard non-behavioristic
approaches too quickly. They have the theoretical advantages of being able
to build a system that can be well understood, both in terms of
construction, programming, and interpreting of results. Connectionist
approaches may well make a better music composer in the long term, but until
problems like "appropriate motivation" for these systems are solved, I'm not
convinced we will personally appreciate the output. (i.e. why do you think a
system that doesn't enjoy music would write good symphonies?)

Provocatively yours,