marek@iuvax.UUCP (02/10/86)
> [Stan Shebs:] In article <3600036@iuvax.UUCP> marek@iuvax.UUCP writes: > ...one of the most misguided AI efforts to date is > taxonomizing a la Michalski et al: setting up categories along arbitrary > lines dictated by somebody or other's intuition. If AI does not have > the mechanism-cum-explanation to describe a phenomenon, what right does it > have to a) taxonomize it and b) demand that its taxonomizing be recognized > as an achievement? > > I assume you have something wonderful that we haven't heard about? I assume that you are intentionally jesting, equating that which I criticize with all that AI has to offer. Taxonomizing is a debatable art of empirical science, usually justified when a scientist finds itself overwhelmed with gobs and gobs of identifiable specimens, e.g. entymology. But AI is not overwhelmed by gobs and gobs of tangible singulars; it is a constructive endeavor that puts up putatative mechanisms to be replaced by others. The kinds of learning Michalski so effortlessly plucks out of the thin air are not as incontrovertibly real and graspable as instances of dead bugs. One could argue, I suppose, that taxonomizing in absence of multitudes of real specimens is a harmless way of pursuing tenure, but I argue in Indiana U. Computer Science Technical Report No. 176, "Why Artificial Intelligence is Necessarily Ad Hoc: Your Thinking/Approach/Model/Solution Rides on Your Metaphors", that it causes grave harm to the field. E-mail nlg@iuvax.uucp for a copy, or write to Nancy Garrett at Computer Science Department, Lindley Hall 101, Indiana University, Bloomington, Indiana 47406. > Or do you believe that because there are unsolved problems in physics, > chemists and biologists have no right to study objects whose behavior is > ultimately described in terms of physics? > > stan shebs > (shebs@utah-orion) TR #176 also happens to touch on the issue of how ill-formed Stan Shebs's rhetorical question is and how this sort of analogizing has gotten AI into its current (sad) shape. Please consider whether taxonomizing kinds of learning from the AI perspective in 1981 is at all analogous to chemists' and biologists' "right to study the objects whose behavior is ultimately described in terms of physics." If so, when is the last time you saw a biology/chemistry text titled "Cellular Resonance" in which 3 authors offered an exhaustive table of carcinogenic vibrations, offered as a collection of current papers in oncology?... More constructively, I am in the process of developing an abstract machine. I think that developing abstract machines is more in the line of my work as an AI worker than postulating arbitrary taxonomies where there's neither need for them nor raw material. -- Marek Lugowski
shebs@utah-cs.UUCP (Stanley Shebs) (02/12/86)
In article <3600038@iuvax.UUCP> marek@iuvax.UUCP writes: >... Taxonomizing is a debatable art of empirical >science, usually justified when a scientist finds itself overwhelmed with >gobs and gobs of identifiable specimens, e.g. entymology. But AI is not >overwhelmed by gobs and gobs of tangible singulars; it is a constructive >endeavor that puts up putatative mechanisms to be replaced by others. The >kinds of learning Michalski so effortlessly plucks out of the thin air are not >as incontrovertibly real and graspable as instances of dead bugs. Now I'm confused! Were you criticizing Michalski et al's taxonomy of learning techniques in pp. 7-13 of "Machine Learning", or the "conceptual clustering" work that he has done? I think both are valid - the first is basically a reader's guide to help sort out the strengths and limitations of dozens of different lines of research. I certainly doubt (and hope) no one takes that sort of thing as gospel. For those folks not familiar with conceptual clustering, I can characterize it as an outgrowth of statistical clustering methods, but which uses a sort of Occam's razor heuristic to decide what the valid clusters are. That is, conceptual "simplicity" dictates where the clusters lie. As an example, consider a collection of data points which lie on several intersecting lines. If the data points you have come in bunches at certain places along the lines, statistical analysis will fail dramatically; it will see the bunches and miss the lines. Conceptual clustering will find the lines, because they are a better explanation conceptually than are random bunches. (In reality, clustering happens on logical terms in a form of truth table; I don't think they've tried to supplant statisticians yet!) >Please consider whether taxonomizing kinds of learning from the AI perspective >in 1981 is at all analogous to chemists' and biologists' "right to study the >objects whose behavior is ultimately described in terms of physics." If so, >when is the last time you saw a biology/chemistry text titled "Cellular >Resonance" in which 3 authors offered an exhaustive table of carcinogenic >vibrations, offered as a collection of current papers in oncology?... Hmmm, this does sound like a veiled reference to "Machine Learning"! Personally, I prefer a collection of different viewpoints over someone's densely written tome on the ultimate answer to all the problems of AI... >More constructively, I am in the process of developing an abstract machine. >I think that developing abstract machines is more in the line of my work as >an AI worker than postulating arbitrary taxonomies where there's neither need >for them nor raw material. > > -- Marek Lugowski I detect a hint of a suggestion that "abstract machines" are Very Important Work in AI. I am perhaps defensive about taxonomies because part of my own work involves taxonomies of programming languages and implementations, not as an end in itself, but as a route to understanding. And of course it's also Very Important Work... :-) stan shebs
tgd@orstcs.UUCP (tgd) (02/19/86)
Taxonomic reasoning is a weak, but important form of plausible reasoning. It makes no difference whether it is applied to man-made or naturally occurring phenomena. The debate on the status of artificial intelligence programs (and methods) as objects for empirical study has been going on since the field began. I assume you are familiar with the arguments put forth by Simon in his book Sciences of the Artificial. Consider the case of the steam engine and the rise of thermodynamics. After many failed attempts to improve the efficiency of the steam engine, people began to look for an explanation, and the result is one of the deepest theories of modern science. I hope that a similar process is occurring in artificial intelligence. By analyzing our failures and successes, we can attempt to find a deeper theory that explains them. The efforts by Michalski and others (including myself) to develop a taxonomy of machine learning programs is viewed by me, at least, not as an end in itself, but as a first step toward understanding the machine learning problem at a deeper level. Tom Dietterich Department of Computer Science Oregon State University Corvallis, OR 97331 dietterich@oregon-state.csnet