smoliar@vaxa.isi.edu (Stephen Smoliar) (05/26/90)
In article <16695@phoenix.Princeton.EDU> eliot@phoenix.Princeton.EDU (Eliot Handelman) writes: >"Music compoistion as hypothesis formation: a blackboard concept of >creativity" > >by > >Otto Laske > >--a summary and an informal chat by E. Handelman > I was very glad to see Eliot post this article. I also happen to have a copy of the paper in question (which, as Eliot mentioned only is passing, was delivered at an AI conference in Prague last year). Because some of the points of discussion concern artificial intelligence as much as they do music, I am taking the liberty of cross-posting this to comp.ai. I shall try to make sure that there is sufficient context for new readers. > >Laske has a new paradigm: blackboards. He explains: "...a mental >activity is carried out in parallel by a multitude of agents >communicating with each other via a central short-term memory called a >blackboard." Laske wants to see if this is a good thing for >composition. I'll briefly summarize. commenting below. > In the past (for example JOURNAL OF MUSIC THEORY, Volume 30, Number 1) I have criticized Laske for being too enamored of the artifacts and terminology of artificial intelligence without necessarily having a clear idea of the intentions being such artifacts and terminology. I find Laske's "discovery" of blackboards slightly amusing because in a 1976 JOURNAL OF MUSIC THEORY article (Volume 20) I tried to interpret his work at the time in the context of work in speech recognition . . . particularly the Hearsay effort at Carnegie. This turned into one of the first major applications of a blackboard architecture. Nevertheless, I think it is a big mistake to try to hold up a blackboard as some magic device which is going to solve all sorts of old problems. When all is said and done, a blackboard is nothing more than a vast global data base which may be manipulated by an equally vast repertoire of productions. The Hearsay effort was not so interesting for HAVING a blackboard as it was for the approach it took to STRUCTURING the blackboard. Indeed, we can begin with the observation that such structuring was a good thing (or, conversely, that a vast unstructured pool of data objects is NOT a particularly good thing). Thus, for all intents and purposes, committing yourself to having a blackboard does not amount to very much unless you also commit yourself to how it will be structured. > >Laske wants to make a model of 1. WHAT musicians know, and 2. HOW this >knowledge is used. The former Laske terms "competence", the latter >"performance." > Like many of us, Laske has read his Chomsky. I first came across this distinction in ASPECTS OF THE THEORY OF SYNTAX. Upon returning to this book, I discovered, to my surprise, that Halle and Stevens were discussing the competence-performance distinction with respect to speech recognition in a 1962 article in the IRE TRANSACTIONS IN INFORMATION THEORY. Thus, it would seem that Laske is trying to take some of the things known about speech recognition and adapt them. If we were dealing with LISTENING to music, then there might be some grounds for such analogical reasoning; and this was the sort of case I was trying to make in my earlier JOURNAL OF MUSIC THEORY article. However, as he informs us in his title, Laske is interested in COMPOSITION, rather than LISTENING. Indeed, in another article (from that wonderful issue of INTERFACE which provided the Balaban article which got much of this discussion going in the first place) Laske went so far as to argue that the teaching of composition is placing far too much emphasis on listening. Thus, it is most unclear just what lies at the foundation of Laske's thinking here. Let's put matters another way: blackboards are not a paradigm. Blackboards are simply an artifact of a particularly way of thinking about AI problems. If Laske wants to turn to AI as a paradigm, then he should be looking for some particular means by which AI deals with a specific problem domain and then argue for an analogy between that domain and musical composition. Unfortunately, I do not think he has EVER done this in any of his publications, which means that, for all intents and purposes, it is rather unclear just what he IS doing. >The competence model consists of a blackboard, a knowledge base, and a >data acquistion module. > >Blackboard: post hypotheses about current state of piece here. >Knowledge Base: must know about 1. planning, 2. scheduling, 3. music. >Data Acquisition Module: not explained. > This is all pretty hollow use of vocabulary. There are times when I wish we could call a moratorium on use of the word "knowledge base." It is turning out to be an excellent way to deceive anyone whose working knowledge of AI is weak. I suppose it is this attitude which has driven me towards writers like Minsky, who tend to scrupulously AVOID words like "knowledge" when writing about artificial intelligence. If Laske believes he can have a knowledge base which "knows about" planning, scheduling, and music, then he is probably sucker enough to buy the Brooklyn Bridge (probably from a sales representative of your favorite commercial AI house). Any AI student would be able to tell him that we still have a long way to go in teasing out just what we mean by knowledge of either planning or scheduling. We may be able to do impressive things in appropriately restricted problem domains, but we are still some distance from being able to generalize any successes of the past into any theories concerned with knowledge of either planning or scheduling. Being sloppy about terminology is sort of like eating potato chips. Once you start, it's hard to stop. Having sailed over planning and scheduling as if we already knew all about them, Laske sees no trouble in adding music to his knowledge base. Does he have the slightest idea what music knowledge is? Do any of us? Some of us are still trying to get a handle on what constitutes musical BEHAVIOR. We might grudgingly admit that musical knowledge amounts to whatever competence enables such behavioral performance; but, if we are at all honest, we also have to admit that our knowledge of that behavior is SO limited that we are a far cry from approaching any questions of competence. In fairness to Laske, however, it SHOULD be noted that he seems to want to give "equal time" to matters of performance: > >The performance model consists of 1. an interpreter and 2. a >scheduler. > >1. interpreter -- knows how to carry out compositional plans. >2. scheduler -- post, retract and alter hypotheses. > Unfortunately, these are more shallow words. Furthermore, they seem to indicate that Laske feels that talking about competence is prerequisite to talking about performance. I would argue that this is absolutely the wrong way to approach the issue. The fact is that performance is about the only thing for which we have any concrete evidence at our disposal. Laske should be worrying about how to talk about evidence of musical behavior. (To some extent, he has done this in the past. He has even done protocol studies in which he has attempted to collect such evidence systematically.) Until we converge on some language for talking about performance, it makes little sense to try to talk about competence. The danger is like trying to build a bridge from both banks of the river without a clear idea of where the two pieces are supposed to join. Another way of putting it is that a language for talking about competence should REFLECT the language for talking about performance rather than impose PROPTER HOC constraints. Many of these comments actually overlap some of Eliot's concluding assessments. However, I felt it was necessary to review some of these observations in a slightly different light. Still, Eliot has made one very powerful point which music be emphasized: > >o Model stresses composition as "problem solving," though nature of problem > isn't described. Method of solution is described. First identify > problem, then describe solution. > There is a standard joke about AI results: each is a solution in search of the right problem. Laske has fallen into the standard trap. He would like to follow where technology leads in hope of finding something. In general, it is healthier to start out with some goals . . . even if those goals must be subsequently abandoned. > >o Model is a masturbation fantasy concerning the future of AI. > Well, it is nice to see Eliot close with the imprimatur of his style of argument! Actually, one could also read this as a commentary on the sad PRESENT of AI. It is very easy to misunderstand what AI is all about. Unfortunately, such misunderstandings are often cultivated by the very people who practice AI. While this is often done in the name of attracting customers or funding agencies, it is often overlooked that there may be innocent bystanders who are caught up in all this confusion of accomplishment and wishful thinking. Laske seems to have been such an innocent bystander, and I worry about how many more are out there who have been similarly deluded. ========================================================================= USPS: Stephen Smoliar USC Information Sciences Institute 4676 Admiralty Way Suite 1001 Marina del Rey, California 90292-6695 Internet: smoliar@vaxa.isi.edu "By long custom, social discourse in Cambridge is intended to impart and only rarely to obtain information. People talk; it is not expected that anyone will listen. A respectful show of attention is all that is required until the listener takes over in his or her turn. No one has ever been known to repeat what he or she has heard at a party or other social gathering." John Kenneth Galbraith A TENURED PROFESSOR