Date: Sun 24 Apr 1988 22:40-PDT From: AIList Moderator Kenneth Laws Reply-To: AIList@KL.SRI.COM Us-Mail: SRI Int., 333 Ravenswood Ave., Menlo Park, CA 94025 Phone: (415) 859-6467 Subject: AIList V6 #80 - Moderator Needed, Credit, PatRec, AI Goals To: AIList@KL.SRI.COM Status: R AIList Digest Monday, 25 Apr 1988 Volume 6 : Issue 80 Today's Topics: Administrivia - AIList Going, Going, ..., History - Demons, AI Tools - Credit Assignment Problem & Conversation Programs & Holographic Pattern Recognition, Opinion - Expert Systems vs. Operations Research & Need for AI and AI Languages ---------------------------------------------------------------------- Date: Sun 24 Apr 88 22:31:23-PDT From: Ken Laws Subject: AIList Going, Going, ... As I mentioned previously, I will not be able to continue moderating the AIList Digest much longer. I have accepted the position of Program Manager, Robotics and Machine Intelligence, at the National Science Foundation (under Y.T. Chien, Division of Information, Robotics, and Intelligent Systems, Directorate for Computer and Information Science and Engineering). This two-year appointment begins at the end of June, and I have a lot to finish up before then. So far there has been exactly one offer of help -- and that was an offer of relaying services if no one volunteered as moderator. So, if anyone wants to take all or part of the AIList stream, the position is still open. If the situation doesn't change, my recommendation is that AIList cease to exist as a digest and that Usenet comp.ai messages be forwarded to the current AIList readers. Submissions can be sent to the gateway address, which will be announced later. (The gateway maintainer has expressed no objection to making it public.) One problem remains. Nearly every digest I send out results in about ten bounce messages (due to mailer problems and people who have abandoned their mailboxes without telling me). If undigested messages are distributed, each message will produce a similar number of error returns -- for a total of perhaps one hundred messages per day! There are two ways to prevent this: digesting and local redistribution. Digesting works, obviously, but puts quite a burden on the new administrator -- especially if it leads to editing and full moderation. The digesting software is also a problem since I use a version written in SAIL, an obsolete language. (There are lists using other digesters, but obtaining one and modifying it would be a bit of a hassle.) Anyway, I have come to favor undigested streams -- we just have to get Arpanet to solve the distribution problem as Bitnet and Usenet have done. Local redistribution means that we should build a tree of relay sites rather than have most hosts connect directly to the new comp.ai relay. Already most AIList addresses are bboards or alias lists, but we need to go further; hosts need to drop from the direct distribution and reconnect to other hosts. The new AIList administrator will then have to tell anyone wanting to sign up to contact his own postmaster, who can contact a postmaster at a secondary relay site if necessary. All this is a hassle to set up and maintain (with no central map of all the connections), but if done properly it can keep the bounce messages from all propagating back to the central administrator. Well, it's up to you. I'm ready to abdicate as soon as we settle on an heir. I'll be around to help out, of course, but AIList will not continue long in its current form unless someone wants to take over the digesting and administrative duties. Meanwhile, I'd appreciate it if some of the host administrators who get this message would offer to take over distribution and signup/drop duties for their principal cliques. -- Ken ------------------------------ Date: 21 Apr 88 02:41:10 GMT From: glacier!jbn@labrea.stanford.edu (John B. Nagle) Subject: Re: demons: was FRL first ? Conniver, circa 1972 (McDermott, MIT) contained a database with similar daemon mechanisms. The Conniver manual appears as an ancient MIT AI lab memo. John Nagle ------------------------------ Date: 21 Apr 88 10:01 PDT From: hayes.pa@Xerox.COM Subject: Re: AIList V6 #74 - Queries, CLOS, ELIZA, Planner, Face Recognition Subject: demons: was FRL first ? I believe that Planner, first partial implementation in MicroPlanner, was the first language to use if-needed, if-added and if-remove demons, called THCONSE, THANTE and THERASE . This is certainly where the concepts originate. Pat Hayes ------------------------------ Date: Thu, 21 Apr 88 09:47:02 -0400 (EDT) From: David Greene Subject: Re: Credit Assignment Problem Obviously, it will depend on the technique and domain you are involved with, but Holland and Smith both offer some interesting insights into some of the issues... especially with regard to genetic algorithms and classifier systems. Holland,J.H. "Escaping Brittleness: the Possibilities of General Purpose Learning Algorithms Applied to Parallel Rule-Based Systems" in Machine Learning: An Artificial Intelligence Approach, volume II, R. Michalski, J. Carbonell, and T. Mitchell (Eds.), Morgan Kaufmann, 1986. Smith, S.F. "Adaptive Learning Systems" in Expert Systems, Principles and Case Studies, R. Forsyth (Ed.), Chapman and Hall, Ltd., 1984, chpt. 11. Hope these are useful. -David dg1v@andrew.cmu.edu ------------------------------ Date: Fri, 22 Apr 88 15:42 EST From: PGOETZ%LOYVAX.BITNET@CUNYVM.CUNY.EDU Subject: Conversation programs Someone asked for sources on programs like ELIZA & SHRDLU: R.C. Parkinson, K.M. Colby, W.S. Faught. "Conversational Language Comprehension Using Integrated Pattern-Matching & Parsing." Artificial Intelligence 9, 1977, p. 111-134. Also found in a recent (1986?) collection from Morgan Kaufman, Understanding Natural Language (or the same words in some other order). Parry: a simulation of a paranoid patient. Program outline. Michael Dyer. Understanding Natural Language. 1983. Boris: A system to summarize & answer questions about narratives. About 400 pages. Talks about emotional scripts (ACEs or AFFECTs, I forget), memory organization, extensive use of demons. Joseph Weizenbaum. "ELIZA - A Computer Program for the Study of Natural Language Communication Between Man and Machine." Communications of the ACM, Jan 1966 V9 #1 p. 36-45. Weizenbaum. "Contextual Understanding by Computers." CACM, Aug 67 V10 #8, p. 474-480. Note that Weizenbaum's extensions to ELIZA let it do much more than the sample ELIZAs you see popping up in magazines every now & then, including learning & answering queries. Terry Winograd. Understanding Natural Language. 1971. SHRDLU. Whoever asked about Racter - it was originally written on, surprise, the Apple IIe. ------------------------------ Date: Thu 21 Apr 88 08:57:36-PST From: Ken Laws Subject: Re: holographic pattern recognition Thanks for the comments you tacked onto my comp.ai.digest query about holographic pattern recognition. > ... Field target-recognition systems are likely to use holograms or > acoustic-wave devices because they are faster than digital techniques > and more robust than complex lens systems ... Holographic systems > storing dozens of different views of tanks and aircraft have been > demonstrated. Can you point me to any reference on this stuff, or is it all classified? Raymond Lister Most of it isn't classified, but it's so widely distributed that I hardly know where to begin looking. We're talking about entire fields of 2-D matched filtering, optical target queueing, correlation matching, character recognition, etc. I remember seeing conference papers on these tank/aircraft recognizers, but would need about a day to track them down. The SPIE and CVPR conferences would be good places to start. You might like the March '87 Scientific American article by Abu-Mostafa and Psaltis on Optical Neural Computers, although they emphasize associative memory rather than recognition. (Recognition simply taps a different plane in the optical system.) David Casasent at CMU is active in this field. I have one of his papers that's relevant: Optical Word Recognition: Case Study in Coherent Optical Pattern Recognition, SPIE Optical Engineering, Vol. 19, No. 5, Sep/Oct 1980, pp. 716-721. Another paper that comes to hand, although not a great illustration, is Mendelsohn, Wohlers, and Leib, Digital Analysis of the Effects of Terrain Clutter on the Performance of Matched Filters for Target Identification and Location, SPIE Vol. 186, Digital Processing of Aerial Images, 1979, pp. 190-196. Some early papers on correlation matching and Fourier signatures can be found in Computer Methods in Image Analysis, a book of reprints edited by Aggarwal, Duda, and Rosenfeld. Two examples are Horwitz and Shelton, Pattern Recognition using Autocorrelation, and Lendaris and Stanley, Diffraction-Pattern Sampling for Automatic Pattern Recognition. For somewhat more recent work see Agrawala's book of reprints, Machine Recognition of Patterns. Examples are Preston's A Comparison of Analog and Digital Techniques for Pattern Recognition, and Holt's Comparative Religion in Character Recognition Machines. I think I should emphasis that these correlation-based matching methods are rather fragile. Casasent has done a lot of work on recognizing patterns that may be rotated or scaled, but most of these techniques require exact matches of standard, isolated characters against uniform backgrounds. They will not recognize handwritten characters, for instance. -- Ken ------------------------------ Date: 21 Apr 88 22:54:47 GMT From: mcvax!ukc!its63b!epistemi!edai!ceb@uunet.uu.net (Colin Bridgewater) Subject: Re: Expert Systems in the Railroad Industry. In article <8816@agate.BERKELEY.EDU> lagache@violet.berkeley.edu (Edouard Lagache) writes: ....for those interested in > computers and trains: what sort of expert systems have developed for > the railroad industry? It seems to me that there are a number of > promising areas: > > 1.) Scheduling. > > 2.) Optimal switching moves and train assembly. > > 3.) Cargo routing and loading. > > 4.) Equipment Maintenance. > > Does anyone know of what work (if any) has been done by railroads > or A.I. outfits in this area? Interestingly enough, Dreyfus would > probably claim that the first 3 areas would be very promising domains > for expert systems. Just to get my two penn'orth in, whatever happened to dynamic programming for scheduling, cargo-space optimisation and inventory control etc ? This well-worn technique is quite adequate for the majority of purposes envisaged by EL. I mention this to raise a wider issue which was possibly not in the mind of the original sender, namely that of the desire to throw ever more complex solution procedures at the simplest of problems.... Why should we want to implement an expert system, when adequate techniques exist already ? That is, is the application of expert system technology appropriate to the magnitude and complexity of the problem ? Should we be advocating the application of such 'high-tech' solutions to all and sundry ? I have no doubt that such systems could be made to work, don't get me wrong on that, I just question whether the level of technology required in order to do so is justified. Surely it is better to apply the simplest solutions when- ever possible. Having said that, I too, would be interested to hear of any research, actual implementations etc that are around. As an engineer involved in AI, I look for simple solutions, in the (vain ?) hope of being able to debug them when things go wrong.......... Colin Bridgewater Univ of Edinburgh P.S. there is an expert system around that diagnoses faults and discusses repair strategies on diesel-electric locomotives. Unfortunately, I don't have any references to hand, but I hope that this jogs someones memory. The Happy Hacker loves to go a-wandering, it's legal in the UK (official). ------------------------------ Date: Thu, 21 Apr 88 19:55 EST From: INS_ATGE%JHUVMS.BITNET@CUNYVM.CUNY.EDU Subject: AI -- Reasons Carole Hafner pointed out that one reason why we pursue AI is curiosity about what computers can do. Another equally valid reason is the possibility of finding out what -we- as intelligent systems can do, and possibly -how- we do it. Not all of AI is directly relevant to psychological and neurological study, but some parts of it is. It definately provides a way to determine the relative complexity of problems using certain AI algorithms, and thus when we find that the computer is has trouble doing what we easily and quickly do, we know that the brain isn't thinking in that manner. (That is, AI provides both positive and negative evidence to psychological theories). Computational neuroscience has already had an effect on modern physiological psychology. In the future, with neural networks and other "natural-like" AI systems, we might learn even more. -Thomas Edwards from the positivist school for good technology ------------------------------ Date: 21 Apr 88 19:58:56 GMT From: moss!ihlpa!tracy@att.arpa (Tracy) Subject: Re: Prof. McCarthy's retort In article <8804180635.AA09224@ucbvax.Berkeley.EDU>, prem@RESEARCH.ATT.COM writes: > > This is a very cute, and compact retort, but not very convinving; it admits > of very many similar cute and compact retorts... The essence of JMC's retort was not to be convincing, but rather to show that they missed the point of why AI (or LISP, for that matter) is useful. Clearly, you could not convince someone that the problem could not be solved in assembly language, because in theory it could be done. It just is not easy. --Kim Tracy AT&T Bell Laboratories, Naperville, IL, ..ihnp4!ihlpa!tracy But of course, it's only my opinion! ------------------------------ End of AIList Digest ********************