Date: Sat 8 Oct 1988 14:59-EDT From: AIList Moderator Nick Papadakis Reply-To: AIList@AI.AI.MIT.EDU Us-Mail: MIT LCS, 545 Tech Square, Rm# NE43-504, Cambridge MA 02139 Phone: (617) 253-6524 Subject: AIList Digest V8 #96 To: AIList@AI.AI.MIT.EDU Status: RO AIList Digest Sunday, 9 Oct 1988 Volume 8 : Issue 96 Spang Robinson Report Philosophy: Re: common sense "reasoning" Followup on JMC/Fishwick Diffeq Re: Newell's response to KL questions ---------------------------------------------------------------------- Date: Sun, 18 Sep 88 08:07:46 CDT From: smu!leff@uunet.UU.NET (Laurence Leff) Subject: bm965 The Spang Robinson Report on Artificial Intelligence, August, 1988, Volume 4, No. 8 Lead Article is on the "New AI Industry" Revenue List Expert System Development Tools 1987 139 million 1989 278 million Natural Language 1987 49 million 1989 95 million Symbolic Processing Languages 1987 51 million 1989 145 million AI Services 1987 150 million 1989 336 million Symbolic Processors 1987 170 million 1989 161 million General Workstations 1987 81 million 1989 277 million Number of companies selling AI technology or applications 1986 80 1988 ~160 Discussions aof Carnegie Group, Inference, IntelliCorp, Teknowledge and Lucid (Lisp) Lucid revenues in 1988 were 1.4 million. ________________________________________ Neural Netowrks: Discussion of various neural network products. They have a 10,000 u9it installed base. It took 30 months to achieve 10000 units in Expert systems contrasted to 13 months for neural networks. MIT sent out 7000 copeis of the software in Explorations in Parallel Distributed Processing. NeuralWorks published 1000 copies of its NeuralWare tool. They range from $195 to $2995.00 Neuronics hs sold 500 units of MacBrain, TRW sold 40 units but is third in dollar volume. ________________________________________ Hypertext and AI. CogentTEXT is a hypertext system embedded in Prolog. Each hypertext button causes execution of an apporpriate segment of Prolog code. This system is a "shareware" product. It can be obtained from Cogent Software for $35.00 (508 875 6553) ________________________________________ Third Millenium is a venture capital fund still interested in AI start ups (as well as Neural networks). )))))))))))))))))))))))))))))))))))))))) Shorts: IntelliCorp reports profit of $416,000 for its fourth quarter. Lucid has a product called Distill! which will remove the develpment environment for the runtime execute. SUN renewed its on-goiing OEM agreement. Lucid has sold a total of 3000 products with 2000 went to SUN. CSK will be selling LUCID in Japan. Neuron Data has integrated Neuron OBJECT with ORACLE, SYBASE and Ingres. The interfaces cost $1000 each. KDS has released a version of an expert system shell with Blackboard. Logicware ported MPROLOG and TWAICE (expert system shell) to IRIS systems. Flavors TRechnology has introduced a system to real-time inference 10,000 rules in 10 milliseconds. A Japanese company ordered the product. Inference has ported ART to IBM mainframes and PC (under MS-DOS). The Spang Robinson Report has a two pag list of AI companies broken down into each of the following fields: Expert System Tools, Expert System Applications, Languages (e. g. PROLOG), natural language systems and hardware ------------------------------ Date: 26 Sep 88 14:59:56 GMT From: jbn@glacier.stanford.edu (John B. Nagle) Reply-to: glacier!jbn@labrea.stanford.edu (John B. Nagle) Subject: Re: common sense "reasoning" Use of the term "common-sense reasoning" presupposes that common sense has something to do with reasoning. This may not be the case. Many animals exhibit what appears from the outside to be "common sense". Even insects seem to have rudiments of common sense. Yet at this level reasoning seems unlikely. The models of behavior expressed by Rod Brooks and his artificial insects (there's a writeup on this in the current issue of Omni), and by Hans Moravec in his new book "Mind Children", offer an alternative. I won't attempt to summarize that work here, but it bears looking at. I would encourage workers in the field to consider models of common sense that don't depend heavily on logic. There are alternative ways to look at this class of problem. Both Brooks and Moravec use approaches that are spatial in nature, rather than propositional. This seems to be a good beginning for dealing with the real world. The energetic methods Witkin and Kass use in vision processing are another kind of model which offers a spatial orientation, an internal drive toward consistency, and the ability to deal with noisy data. These are promising beginnings for common-sense processing. John Nagle ------------------------------ Date: 28 Sep 88 2:21 +0100 From: ceb%ethz.uucp@RELAY.CS.NET Subject: Followup on JMC/Fishwick Diffeq >From ceb Wed Sep 28 02:21:09 MET 1988 remote from ethz >for Robots Interchange Apropos using diffeqs or other mathematical models to imbue a robot with the ability to reason about observation of continuous phenomena: in John McCarthy's message , JMC states that (essentially) diffeqs are not enough and must be imbedded in "something" larger, which he calls "common sense knowledge". He also state that diffeqs are inappropriate because "noone could acquire the initial [boundary?] conditions and integrate them fast enough". I would like to pursue this briefly, by asking the question: Just how much of this something-larger (JMC's framework of common sense knowledge) could be characterized as descriptions of domains in which such equations are in force, and in describing the interactions between neighboring domains? I ask because I observe in my colleagues (and sometimes in myself) that an undying fascination with the diffeq "as an art form" can lead one think about them `in vitro', i. e. isolated on paper, with all those partial-signs standing so proud. You have to admit, the idea as such gets great mileage: you have a symbolic representation of something continuous, and we really don't have another good way of doing this. Notwithstanding, in order to use them, you've got to describe a domain, the bc's, etc. This bias towards setting diffeqs up on a stage may also stem from practical grounds as well: in numerical-analysis work, even having described the domain and bc's you're not home free yet - the equations have to be discretized, which leads to huge, impossible-to-solve matrices, etc. There are many who spend the bulk of their working lives trying to find discretizations which behave well for certain ill-behaved but industrially important equations. Such research is done by trial-and-error, with verification through computer simulation. In such simulations, to try out new discretizations, the same simple sample domains are used over and over again, in order to try to get results which *numerically* agree with some previously known answer or somebody elses method. In short, you spend a lot of time tinkering with the equation, and the domain gets pushed to the back of your mind. In the case of the robot, two things are different: 1. No one really cares about the numerical accuracy of the results: something qualitative should be suffficient. 2. The modelled domains are *not* simple, and do not stay the same. There can also be quite a lot of them. I would wager that, if the relative importance of modelling the domain and modelling the intrinsic behavior that takes place within it were turned around, and given that you could do a good enough job of modelling the such domains, then: a. only a very small subset of not scientifically accurate but very easy to integrate diffeqs would be needed to give good performance, b. in this case, integration in real time would be a possibility, and, c. something like this will be necessary. I believe this supports the position taken by Fishwick, as near as I understood it. One might wonder idly if the Navier-Stokes equation (even in laminar form) would be among the small set of subwager a. Somehow I doubt it, but this is not really so important, and certainly need not be decided in advance. It may even be that you can get around using anything at all close to differential equations. What does seem important, though, is the need to be able to geometrically describe domains at least qualitatively accurately, and this `on the fly'. I am not claiming this would cover all "common sense knowledge", just a big part of it. ceb P. S. I would also be interested to know of anyone working on such modelling --- this latter preferably by mail. ------------------------------ Date: 30 Sep 88 04:06:59 GMT From: goel-a@tut.cis.ohio-state.edu (Ashok Goel) Subject: Re: Newell's response to KL questions I appreciate Professor Allen Newell's explanation of his scheme of knowledge, symbolic, and device levels for describing the architecture of intelligence. More recently, Prof. Newell has proposed a scheme consisting of bands, specifically, the neural, cognitive, rational, and social bands, for describing the architecture of the mind-brain. Each band in this scheme can have several levels; for instance, the cognitive band contains (among others) the deliberation and the operation levels. What is not clear (at least not to me) is the relationship between the two schemes. One possible relationship is colinearity in that the device level corresponds to the neural band, the symbolic level to the cognitive band, and the knowledge level to the rational band. Another possibility is containment in the sense that each of band consists of (the equivalents of) knowledge, symbolic, and device levels. A yet another possibility is orthogonality of one kind or another. Which relationship (if any) between the two schemes does Prof. Newell imply? A commonality between Newell's two schemes is their emphasis on structure. A different scheme, David Marr's, focuses on the processing and functional aspects of cognition. Again, what (if any) is the relationship between Newell's levels/bands and Marr's levels? Colinearity, containment, or some kind of orthogonality? --ashok-- ------------------------------ End of AIList Digest ********************