daemon@bartok.Eng.Sun.COM (02/06/90)
Music-Research Digest Mon, 5 Feb 90 Volume 5 : Issue 11
Today's Topics:
Expert system shell needed (for music application)
Knowledge Acquisition
*** Send contributions to Music-Research@uk.ac.oxford.prg
*** Send administrative requests to Music-Research-Request
*** Overseas users should reverse UK addresses and give gateway if necessary
*** e.g. Music-Research@prg.oxford.ac.uk
*** or Music-Research%prg.oxford.ac.uk@nsfnet-relay.ac.uk
----------------------------------------------------------------------
Date: 4 Feb 90 03:41:30 GMT
From: Mitchell Spector <spector%sumax@edu.washington.cs.beaver>
Subject: Expert system shell needed (for music application)
To: music-research@uk.ac.oxford.prg
I am posting this request for someone who doesn't have access to the net.
He's embarking on a project to create an expert system in music (harmony
theory) using Common LISP. Has anybody done any work in this area? Is
there source code available for any small expert system shell written in LISP?
Thanks for any information you can provide. You can send e-mail to me
at spector%sumax.uucp@beaver.cs.washington.edu or get in touch directly with
him: Prof. Will Self, Mathematics Dept., Eastern Montana College, Billings,
MT 59101 (phone #: 406-657-2940).
------------------------------
Date: Mon, 5 Feb 90 08:32:49 EST
From: laske@edu.bu.cs
Subject: Knowledge Acquisition
To: music-research <music-research%uk.ac.oxford.prg@uk.ac.nsfnet-relay>
nswer to Robert Rowe and Eliot Handelman
regarding KNOWLEDGE ACQUISITION
Otto E. Laske
Although I lack Handelman's inquiry, I would like here to
give some information both Rowe and Handelman on my work
with PRECOMP (formerly Pr1Mac). I start with some historical
preliminaries.
In the context of Laske and Truax's OBSERVER program for
knowledge acquisition in music [Laske, Memory and Thought,
1977, ch. 10, and earlier papers], children of ages 7 to 12
were monitored in their problem solving during an elementary
composition task employing electronic sounds; in [Laske 1979,
INTERFACE, vol. 8:207-235] some of the resulting "action protocols"
were analyzed in order to understand the development of the
children's music-syntactic thinking. In a dissertation following
the OBSERVER precedent more than 15 years later--that's half a
research generation!!--, D. Cantor [Cantor 1989, "A Knowledge
Acquisition System for Segmentation in Music," Ph.D. thesis,
C. Sc., Boston University], used G.M. Koenig's PROJECT ONE program
[Koenig 1969/79] for computer-aided composition to create PRECOMP
[called KAIMU by Cantor], a knowledge acquisition tool for under-
standing how composers make intuitive decisions about segmentation
in interpretive composition. (Interpretive composition is based
on computer-generated alphanumeric texts representing compositional
materials [Laske,"Composition Theory: An Enrichment of Music
Theory," INTERFACE, vol. 18, no. 1/2, 45-59].
PRECOMP, by the way, was discussed in two lectures at MIT, in the
fall of 1987 and 1988, but no interest in continuing or extending
this research has been forthcoming.
Why is knowledge acquisition (KA) important? It is the only way
to get away from the ad hoc stipulations now reigning supreme in
"music research" (a nice neutral term that says nothing).
What are some results of working with PRECOMP?
PRECOMP can only be sketched here. It comprises three modules:
score generation, score analysis, and segmentation advisor. The
first module is required for interpretive composition; it creates
the "task environment" for the research. Any other musical task
could be substituted (e.g., defining orchestras in CSOUND). The
second module makes graphic images of the contents of the "event
list" (composition material) generated in the first step. This
is a help since alphanumeric text is opaque and does not yield
immediate insight into the structure of the material generated.
The third component, viz., the segmentation advisor, is the focus
of PRECOMP from a research point of view--I use the two other
components regularly for my compositional work for instruments
as well as tapes.
The segmentation advisor (SA) is a set of rules whose slots are
set by the user/"subject"/expert; each rule concerns one aspect
(parameter) of the event list material; each rule has several "slots"
regarding the interestingness of the parameter for the task of
segmenting--finding structure in--the compositional material, as
well as "filters" and "multipliers" permitting to the expert to
state how a particular parameter should be treated in finding
suitable breakpoints in the event list. All rules are processed
for all events to find the most highly weigthted parameters as
well as event, the latter being the first of the 10 most highly
ranking potential breakpoints. Since the rules and their slots
can be reset as often as desired and in whatever order, a real
time protocol is taken, to reveal the expert's thinking during
the session. The output of the SA consists of:
1. a log of the session showing the steps taken (procedural
information)
2. a computed list of interestingness indeces showing what
parameters were foremost on the expert's mind when he decided
on how to segment the event list
3. a list of weighted events, showing a ranking of the most
highly weighted breakpoints [both (2) and (3) are declarative
material].
As in OBSERVER, the session log --protocol--needs to be inter-
preted by a cognitive musicologist; automatic protocol analysis
would be preferable, and a very interesting research task. (There
are many Ph.D. theses in this research, believe me.)
What's the point???
Here is a slender program--written in Pascal and Lisp, using
Hypercard, running on a Mac--that, for once, takes advantage of
the plethora of now existing computer-aided tool for making music,
to track what the expert/users actually DO with the tool, thus
generating research material for cognitive musicology. Dealing
with this material would permit to develop empirical theories
regarding the knowledge used in musical tasks, and theories of
expert knowledge used in these tasks.
In fact, such monitoring and knowledge acquisition modules
should be part of every tool built today, since only with such
material--rather than the informal musings of users--tools
can be systematically improved, and criteria for their evaluation
can be developed.
What are the results from work with PRECOMP. Yeah, everybody
wants results, before even having understood the purpose. There
are results but the repertory of materials so far is too small
to generalize them. There is a lack of people working on this
kind of project, and there is only so much I personally can do.
So, I invite collaboration.
What interests me in this research is not only the abstract
theoretical insights into expert's knowledge, but equally, if not
more, the insights that lead to better composition tools. The SA
in PRECOMP, for instance, could be used for any score-generating
program, to aid the composer in designing a definitive score. It
is thus an aid in compositional design.
Similar tools are possible for other musical tasks, there is
no limit.
We know what happened in expert systems, when knowledge acquisi-
tion was restricted to "rapid prototyping" (i.e., =nil). Maybe
we should try to avoid the errors made, now, 18 years after the
first knowledge acquisition tool, OBSERVER, was build.
Any comment?
------------------------------
Date: Mon, 5 Feb 90 09:44:10 EST
From: laske@edu.bu.cs
To: music-research <music-research%uk.ac.oxford.prg@uk.ac.nsfnet-relay>
nswer to Eliot Handelman
Otto Laske
Dear Eliot,
I got your--pretty condescending--letter to which I respond to
ask you to adopt a more kindergarten-unlike talk. I am not used
to such loose talk.
In part, your problems may stem from the fact that the articles
I have written about composition theory have not appeared in
American journals, or have not yet appeared at all. For this
reason, here are some pointers:
1991 "Capturing Musical Intuitions by Way of Rules." READINGS IN
ARTIFICIAL INTELLIGENCE AND MUSIC. A joint publication of
papers from U.S. and European workshops on AI&M 1987-1989, Mira
Balaban,Kemal Ebcioglu, Otto Laske eds.
1990 (a) "The Computer as the Artist's Alter Ego," LEONARDO, New York,
N.Y.: Pergamon Journals, vol. 23, no. 1 (Spring 1990).
(b) "Three Prototypical Approaches in Artistic Composition." (Otto
Laske, ed.) Interface, vol. 19, 1990. Amsterdam, Holland: Swets & Zeitlinger.
(c) "On the Acquisition of Compositional Knowledge." Proceeedings,
A.I. '90. Prague, CSSR.
1989 (a) "Composition Theory in Koenig's Project One and Project
Two." THE MUSIC MACHINE (C. Roads ed.), 119-130. Cambridge,
Massachusetts: The M.I.T. Press. Reprinted from Computer Music
Journal, vol. 5.4:54-65. Cambridge, MA., U.S.A.: The M.I.T. Press.
1981.
(b) "Composition Theory: An Enrichment of Music Theory." Interface,
vol. 18, no. 1/2, 45-59. Lisse, Holland: Swets & Zeitlinger. Reprint from Proceedings, The Arts and Technology II: A Symposium. New
London, CT.: Connecticut College, 101-119.
(c) "Composition as Hypothesis Formation: A Blackboard Model of
Compositional Creativity." In (V. Marik, ed.) Proceedings,
Applications of Artificial Intelligence - A.I. '89. Prague, CSSR,
reprint in Knowledge-Based Systems, Guildford, Surrey, U.K.:
Butterworth Scientific.
Of these, 1989b is available, and 1990a will shortly be available. Also,
I am available for lectures.
****************************************************************************
Now to the questions as far as I can make them out and answer to them
without too much repetition.
Question 1: "...how you sat down with a living composer and studied his compositional processes." I see two ways to do this: (a) you choose a
computer-aided environment, (b) you do not. When (a), this is done, e.g.,
by Ron Roozendaal at the Center for the Arts, Utrecht, Holland, you get
a lot of data --verbal, musical, video--that is hard to assess because
you haven't bounded the task beforehand. When (b), which is what I have
done, the composer and the task are (so to speak) predetermined by the
computer-aided tool you use. The advantage in (b) is that you have a
rather "well-defined" task, but, by choosing an aspect only, such as
a segmentation problem in PRECOMP (1991, 1990a,c), you limit your scope
severely. This, I think, is an advantage if you know what you are leaving
out. Because computer-aided tools prestructure the task environment of
the expert, and findings from that task environment, however one-sided,
can serve as points of departure for understanding unstructured, "ill-
defined" tasks such as "composing a tune", "composing a fugue."
Question 2: "what kinds of questions would you ask?" This is the problem
of knowledge acquisition in (applied) A.I. and it's a whole discipline
that works on that question today. For one, I would not ask any question,
since verbal reports are not the best way to capture action knowledge,
such as musical knowledge. This is well documented in cognitive psychology.
Also, what experts say and what they DO are two separate worlds. Read
the knowledge acquisition literature. I don't need to ask questions
since my knowledge acquisition system, here PRECOMP, "asks" them for me
and monitors the "answers" (which are mouse klicks). Read my response
to Robert Rowe to understand better.
Question 3: "what is it that you think you do have to offer?" This is
a pretty condescending question to ask somebody who has written about
musical epistemology and cognitive musicology for 20 years, but I
forgive you. The question says more about you than about me. You could
also have asked this question to other people in the field, such as
Bernard Bel, Marseille, Ron Roozendaal, Utrecht, or Stephen Smoliar.
What do we have to offer? Simply that, with more or less experience,
we are investigating composition by way of the scientific tools now
available, due to A.I. (Maybe you guys should grab a proceedings on
knowledge acquisition to at least appear less ignorant?). In my view,
it takes a special kind of expertise to do this research: compositional
expertise of one's own, as well as scholarly expertise in A.I. and
cognitive psychology. (I am not sure Princeton is such a good place
to get good at knowledge acquisition, the A.I. climate there being,
as in most places, pretty formalistic.)
As to the tool builders, the group that build M is actually not
the worst I know, since they shared a lot of experiences before they
set out to build the tool, and a composer, Joel Chadabe, was
instrumental in tying their experiences together. But again, they
left out the monitoring device, the KA module that would have traced
what people do with M, since that wouldn't sell too well, and so,
we'll never know haw M is used, although that would be very inter-
esting to know for the sake of understanding improvisational composi-
tion.
Question 4: "why should the process of one composer be relevant, or
even necessarily related to, the process of another composer?" Good
question, this. Indeed, a compositional process by X in music may
have more of a relation to a design process by Y in architecture,
that is quite possible. Music composition is a discipline of design,
and design occurs in many arts, and not only in the arts. But then,
design in music--called composition--requires a certain kind of
competence, and therefore, the "knowledge base" of composers inter-
sects to some degree. Also, the tools in one society tend to share
common traits, the music paper is a common task environment, or else
the computer-based orchestration language. So, even if the composers
think wildly differently--which they do, just look at how different
composers trade off between what they invest in a CSOUND orchestra
and what they invest in a CSOUND score--they are grouped sociologically
according to their "task environment", i.e., the tools they use. They
are also related by their education, and thus, the kind of processes
they come up with. Hard to imagine, maybe, for a young composer who
thinks he is soooo unique, but nevertheless true. In short, what is
a "relation" of one "process" of composition to another "process"
is a scholarly question, not a rhetorical one (as in Handelman).
Question 5: "We are the guys who can tell you something about com-
position..." Well, I don't want to be told, I want to be SHOWN. You
can tell me anything you like, I have listened to many experts, not
only in music, and, having been a "knowledge engineer" for 5 years
now, I have found that there is a decisive difference between
"espoused theory" and "theory-in-use" (See Chris Argyris, Action
Science, Jossey-Bass Publishers, San Francisco, 1985) for some
enlightenment. The point is just not to have to rely on talk alone,
since talk is an action, and not the most conducive one to understand
musical knowledge. For this reason, already in OBSERVER (1972-77),
we adopted "action protocols", and "rules" in PRECOMP. The difficult
point is just to get beyond what is said, to the "theories-in-use"
(i.e., actually used).
Also,there is the difference between novices (such as Eliot) and
experts (such as Paul Lansky, say); it would be important to find
out what their differences are when they do the same task.
Conclusion (for now): I value Eliot's response but, as I say, I think
we need a different tone (on his part). We also need more lectures
and experiments regarding composition theory, because the only guys
who can enlighten us about composition (in music) are the "composers"
(weirdos or not;they are not as alone as they think to themselves,
there are lots of designers out there, in fields other than music.)
Otto Laske
------------------------------
End of Music-Research Digestrowe@mit-amt.MEDIA.MIT.EDU (Robert Rowe) (02/08/90)
In reply to my query about the knowledge acquisition process in PRECOMP, Otto Laske makes some observations and asks for comment. I would like to explore further some of his remarks, such as >Why is knowledge acquisition (KA) important? It is the only way >to get away from the ad hoc stipulations now reigning supreme in >"music research" (a nice neutral term that says nothing). Why is KA the only way to improve music programs? This is what I fundamentally do not understand about your position, and the above statement does not help me. >In fact, such monitoring and knowledge acquisition modules >should be part of every tool built today, since only with such >material--rather than the informal musings of users--tools >can be systematically improved, and criteria for their evaluation >can be developed. You have built a knowledge acquisition section into PRECOMP, which records how an expert user chooses to segment musical material generated by the other parts. I'm sure you realize that it is a trivial step to add a protocol recorder to a composition program - the reason no one does it is that it is so unclear why it is worth saving all those megabytes of data. Recording a protocol in itself, in any case, is hardly knowledge acquisition. You save the interesting part of the process for the cognitive musicologist who will examine and evaluate the data later on, and for the designer who sets up the task in the first place. >As in OBSERVER, the session log --protocol--needs to be inter- >preted by a cognitive musicologist The kind of decision-making you will find recorded by your segmentation advisor is, I wonder if you would agree, highly directed by the structure of the task. Anyway, you have this data which says which parameters an expert uses to segment musical material. So - what are they? If you tell me what parameters your users favored to segment material, that is information I could use as I continue development of my system. If you tell me I have to record protocols of people using my system, save the data, and hand it over to a cognitive musicologist so he can interpret that "knowledge" for me, that is considerably less informative. >What are the results from work with PRECOMP. Yeah, everybody >wants results, before even having understood the purpose. Guilty on both counts. I know you think the purpose is glaringly, painfully obvious, but I am genuinely having trouble discerning it. Let me try to outline what I take you to be saying: Composition tools are inadequate because they do not reflect the cognitive processes composers use. Knowledge acquisition is the only way to uncover those processes, and, thereby, improve the tools. I am writing a computer program for composition and performance. I have been doing such things for many years. As the program is used for more compositions, and by more musicians, it improves. Score of other composers are in the same state of affairs. Still, I do not believe that the operations I encode will be useful for all composers. I do expect that the knowledge representations and manipulations I take myself to be developing will be of interest to other composers and musicians involved in similar pursuits; I do expect, and in fact, have seen, other musicians use the system to make music. But I do not expect that the knowledge acquisition process as you describe it would have any particular impact on the development of my, or many other composers', tools. According to you, knowledge acquisition has to provide us with a substrate of tools capturing the common compositional cognitive heritage for anyone to make any progress. >What interests me in this research is not only the abstract >theoretical insights into expert's knowledge, but equally, if not >more, the insights that lead to better composition tools. The SA >in PRECOMP, for instance, could be used for any score-generating >program, to aid the composer in designing a definitive score. It >is thus an aid in compositional design. >Similar tools are possible for other musical tasks, there is >no limit. This is where the purpose becomes unclear to me. Do you believe that the segmentation advisor captures a universal cognitive skill, applicable to the tasks of all composers? If you do believe it, what are other such universal, or at least, very widespread compositional strategies? If you do not believe it, why is any one composer's structuring of a musical task supposed to provide useful tools for all the others? - Robert Rowe - Music & Cognition Group - MIT Media Laboratory