LAWS@SRI-AI.ARPA (12/09/84)
From: AIList Moderator Kenneth Laws <AIList-REQUEST@SRI-AI.ARPA> AIList Digest Sunday, 9 Dec 1984 Volume 2 : Issue 174 Today's Topics: AI Tools - UNSW Prolog, Books - Pitman AI Series, Cognition - Childhood Memories, Expert Systems - Optical Disk Memories, Machine Translation - Folklore, Knowledge Representation - Nonverbal Meaning, Seminar - Reinforcement Learning (CMU) ---------------------------------------------------------------------- Date: Thu, 6 Dec 84 13:08:58 PST From: Adolfo Di-Mare <dimare@UCLA-LOCUS.ARPA> Subject: UNSW Prolog Date: Mon 26 Nov 84 23:21:34-PST From: Michael A. Haberler <HABERLER@SU-SIERRA.ARPA> Subject: UNSW Prolog interpreter To: info-ibmpc@USC-ISIB.ARPA I have ported the University of New South Wales Prolog interpreter to an IBM PC running MS-DOS 2.0. It implements all built-in predicates of the Unix version and can call your favorite editor or the command line inter- preter. UNSW Prolog is closely patterned after Prolog-10, but has no compiler. I got the permission to redistribute the interpreter from the author of the Unix version, Claude Sammut of UNSW. If you want to obtain a copy, sign the license which can be FTP'ed from [SIERRA]<HABERLER>PROLOG.LICENSE, and send the license with 2 DSDD diskettes to the address below. Neither Claude nor I charge anything for it. Michael Haberler Computer Systems Laboratory ERL 403 Stanford University, Stanford CA 94305 (415) 497-9503 Adolfo /// ------------------------------ Date: Fri 7 Dec 84 17:32:57-EST From: SRIDHARAN@BBNG.ARPA Subject: Pitman AI series now a concrete reality! Many of you know that Derek Sleeman and myself are the two Main Editors for the Pitman AI research notes series. The series was conceived and developed over the past 18 months. Some of you also saw the Pitman booth at the AAAI-84 trade show. Finally, the series appears to have taken concrete reality. I have received the first "book" in the series. Another six books will be out within the month. THe first title is Perry Miller, A Critiquing Approach to Expert Computer Advice: ATTENDING. The other titles are listed below. Paul Cohen, Heuristic Reasoning about uncertainty. A. Palay, Searching with probabilities. Y. Ohta, Knowledge based interpretation of outdoor natural color scenes. R. Korf, Learning to solve problems by searching for macro-operators. P. Poliltakis, Empirical Analysis for Expert Systems. J. Kender, Shape from texture. The series covers the whole spectrum of AI and publishes research materials suitable for use in graduate courses, seminars, and reference material for individuals working in this field. The aims of the series are (a) rapid publication in softback form; (b) worldwide exposure for significant research results; and (c) low cost - usually under $20. Authors are encouraged to get in touch with one of the main editors either Sridharan@BBNG or Sleeman@SUMEX. [...] ------------------------------ Date: Fri, 7 Dec 84 21:30:31 est From: utcsrgv!dciem!mmt@uw-beaver.arpa Subject: Childhood memories I have many memories dating back to as early as my 2nd birthday, and can clearly remember large parts of the floor plan of the school I attended from 3-5. But ALL these memories are pictorial, not sequential. I cannot remember happenings until about 5, when I remember my first introduction to French verb conjugation. Perhaps the truth is that children are not capable of sequential logical operations until around 5, and therefore cannot remember events of that kind, whereas pictures are more readily preserved if you happen to grow up to be imagery-oriented. Martin Taylor ------------------------------ Date: Thu 6 Dec 84 19:19:58-EST From: Wayne McGuire <MDC.WAYNE%MIT-OZ@MIT-MC.ARPA> Subject: Personal Assistants & Optical Disks Re: Dietz's speculation about optical disks: Optical disks will clearly impact information technology in general (microforms, magnetic tape, commercial databases, book publishing, etc.) and microcomputers in particular in many revolutionary ways. One potential use would be to integrate the optical disk with AI-based integrated software in a microcomputer product which would be a powerful general purpose idea processor and personal assistant. We already see a trend towards general purpose idea processors in such micro products as Framework, Symphony, Thinktank, Clout, Dayflo, Factfinder, and The Desk Organizer. This trend is likely to continue and to accelerate as new generations of microprocessors rapidly come online and make available ever greater random access memory for personal computer users. Framework and Symphony are the crude precursors of general purpose personal assistant programs of 1MB, 5MB, and more of memory. A sign of the times: Mitch Kapor, the founder of Lotus, recently commented in an MIS Week interview that the next key step for his company would be to explore current AI research in depth, and to develop new more powerful products that were capable of sophisticated qualitative, not just quantitative, information processing. Optical disks would nicely interface with the next generation of general purpose idea processors. With them one could easily store, retrieve, and manipulate all the vital information and minute details in one's life: financial transactions, notes for miscellaneous projects, diary entries, address books, medical records, rough drafts, datebooks, electronic mail, shopping lists, statistics, papers, bibliographies, administrivia, programs in progress, graphs, abstracts and full-text documents downloaded from commercial databases, etc. Every individual record or key chunk of information in one's personal digital archive could be uniquely identified by a date and time stamp, and every personal database, structured and/or free-form, could be integrated into a single richly interconnected knowledgebase. The set of storage optical disks for a program of this kind would constitute for anyone, in compact and efficient form, an extremely thorough journal of his or her life. Write-once optical disks would actually be preferable for this archival purpose than disks which could be erased and written over. Subsets from the master archival disk(s), of any desired information or complex combination of records, could be transfered at will to working floppy or hard disks. The technology for the greatest revolution in the history of personal information management is already solidly in place. It is not likely that the total information processed by a personal assistant for an average person over a lifetime would occupy more than one or two disks. Even for someone whose personal information needs were much greater than average--say, a Harvard economics professor who is a dedicated teacher, a prolific scholar and author, holds a cabinet-level post (not concurrently with his teaching responsibilities, of course), and has an active globe-trotting social life--under 100 disks would probably neatly archive a lifetime of rich intellectual, professional, and social activity. Our professor would be able to pinpoint in a few minutes those two sentences in which x remarked about y in a private communication twenty years ago, or that small note of last year which captured a flash of insight about how to improve a formula in an econometric model of the Venezuelan oil industry. (Literary scholars analyzing the biodisks of future Walt Whitmans or Virginia Woolfs would be able to reconstruct in microscopic detail the evolution of their subjects' works and themes, and the interaction of quotidian life events with their imaginative creations.) AI-based personal assistants and optical disks seem to be made for one another. I wouldn't be surprised to see prototype products on the market within the next two years. By the '90s we may well wonder how we ever got by without them. -- Wayne McGuire (mdc.wayne@mit-oz) ------------------------------ Date: 7 Dec 84 16:22:46 EST From: Allen <Lutins@RU-BLUE.ARPA> Subject: more on translation... I understand that a similar attempt with a Chinese/English translator yielded the following results: English input: "Out of sight, out of mind" Translated response: "Blind and Stupid" I did have the occasion to "speak" with a Japanese student using a Sharp(?) hand-held translator. Surprisingly, general ideas were conveyed quite well. However, I think we're still a long way off from getting a computer to translate a language any better than an eight year old bilingual person can. -Allen ------------------------------ Date: Fri, 7 Dec 84 09:58:10 pst From: Douglas young <young%uofm-uts.cdn%ubc.csnet@csnet-relay.arpa> Subject: Nonverbal meaning Following my enquiry in AIList 62, a few people have asked what I mean by "nonverbal meaning". It seems appropriate to reply to them, and to explain to any others who may not understand the significance of the term, through the medium of AIList. While until quite recently Wittgenstein, Frege, Quine, and Chomsky might have seemed nearer than any other philosophers of language ( or anyone else, for that matter) to providing a firm foundation from which to represent meaning, none has been willing to go systematically "deeper" than using words, ultimately, as that foundation. They have written only of very unspecific and vague concepts and structures. But Jerry Fodor's recent and exciting book, " The Modularity of Mind", made, in my view, a major leap ahead, in at least recognising that meaning is founded upon nonverbal, cognitive modules, although he did not suggest either the exact form that such modules might take, nor just how they could be applied to providing nonverbal meaning. We have been working here for several years on the theory and foundations of a system by which word and sentence meaning could be represented nonverbally in a natural language understanding system. The principles of this system arose from clues derived from some neuro- -physiological experiments I conducted during 1976-78 (in which recordings were taken from the pulvinar complex, a part of the brain that in man is involved in language but that also exists as far back the phylogenetic tree as in the rabbit).During the following six years, further neurologic and psychological research provided the detailed foundations of a system by which we could represent the meaning(s) of any word or sentence, in English (but that is essentially transportable to any other major natural language), wholly nonverbally. Some of the neurological and psychological grounds, for both the semantic and the syntactic base of the system were described in two papers published in Medical Hypotheses in 1982,83, but, as I mentioned in my previous communication, the original systems of modalities and coding described in these papers were so long ago superseded that they (but not the grounds) are of little significance now. We are currently in the early stages of designing the software for a prototype of the modal system, and some of the reults of this work should be published during the latter part of 1985. In order to explain as concisely as possible the principles and some of the techniques employed, it may be helpful to take people back to basics: Try to explain by words alone the meaning of any one of a range of different words (eg, MUG, DIFFERENTIATE, WALK, OR, INTERNATIONAL, QUICKLY). You will succeed in providing several sets and trees of dictionary-type definitions; but, in the end ,if you continue to ask yourself the meaning of each new word in each succeeding set of definitions, you will either get into an endless cycle of using the same words with which you began your definitions, or you will reach an impasse. If, however, you then ask yourself, and consider carefully, the subjectively experienced nonverbal significances of those same words, several ideas will come to mind. For example, in respect of MUG, you will likely notice the fact that it has both aspects of "appearance" (such as its visual shape, or the interorientation of its parts to one another) and of "function" (such as the motor and kinaesthetic sequences of events that enable you to drink from a mug). The same kind of thoughts may also occur to you when you consider a word like QUICKLY or UP, for example. Abstract words, like the verb MATCH, and "long" words, like INTERNATIONAL, will require some or many levels of verbal "unfolding" of their meanings in order for you to be able to reach any of their nonverbal foundations; but these words also can be nonverbally represented, by means of the "mental modalities". In fact, all of these nonverbal aspects of meaning can be represented by means of a whole range of modalities. The system incorporates 32 different modalities, of which 27 are neurologically based (such as visual detection of movement (VDM), verbal expression (VXP), kinaesthetic (KIN), central autonomic proprioception and control (CAP)), and 5 are the mental modalities, for which there are no neurological, only cognitive, grounds (such as cognitive mental acts (CMA), metaconceptual (MET), emotive mental states (EMS)). Codes within these modalities, grouped together as a frame of generic parts of function and/or appearance, and closely interrelated, can provide a nonverbal meaning representation for any word. The meaning of a sentence is provided through an interactive syntactic process that, both anteroactively and retroactively, interrelates appropriate segments of those modal code frames, so as to disambiguate both the individual word meanings and their "use-categories" (i.e., "object", "activity", "characteristic", or "relation"). By this method, it is possible to represent the meaning(s) of any sentence nonverbally, and at the same time provide access, up to any depth required of a particular system application, to direct and associated knowledge re that sentence. The modal system seems both versatile and quite powerful, and it has the advantage over some other systems of NLU that it reduces memory and storage requirements by taking advantage of the many cognitively equivalent modal aspects in descriptions of similar objects, activities or characteristics. One rather satisfying aspect of the mental modalities is that the cognitive mental act modality not only provides for the nonverbal meanings of such words as ASSOCIATE, NEGATE, SYMBOLIZE, MATCH or CONJUNCT, but also provides the means of executing the relevant logical activity. Incidentally, another feature of the system is that it can provide for both metaphor and idiom; but work on this will almost certainly be delayed until 1986 due to the need to complete the basic system software for the prototype. It would be inappropriate in the AIList to do more than try to provide with sufficient background an idea of the general characteristics of the system. I hope, however, that what I have written will be sufficient to explain at least what sort of thing I am referring to by "nonverbal meaning" As mentioned in AIList 62, I would be most interested to hear about, and/or to receive copies of any papers from other projects in this or any allied area of natural language understanding. Douglas A,Young Dept of Computer Science University of Manitoba Winnipeg Manitoba, R3T 2N2 CANADA ------------------------------ Date: 7 Dec 84 11:48:09 EST From: Steven.Shafer@CMU-CS-IUS Subject: Seminar - Reinforcement Learning (CMU) [Forwarded from the CMU bboard by Laws@SRI-AI.] Richard Sutton, from U. Mass., will be speaking at next Tuesday's AI Seminar. WeH 5409 at 3:30 pm. If you'd like to speak with him during his visit, please contact Geoff Hinton. REINFORCEMENT LEARNING: LEARNING METHODS FOR COMPLEX SYSTEMS Reinforcement learning is the process of learning to make decisions based on the observed results of previous decisions. It is distinguished from other forms of machine learning in that it does not require instruction as to what the learning system should do, only evaluation of what it does do. In this sense reinforcement learning requires less help from its environment and is more powerful and robust than other forms of learning. In complex learning systems it is particularly difficult to specify in detail what the learning system should do, and reinforcement learning is particularly relevant. Nevertheless, reinforcement learning has been studied very little. This talk will present computational experiments comparing the performance of many previously-studied algorithms and several new ones. In many cases the previously proposed algorithms were found to perform very poorly, much worse than the new algorithms. Since in many cases the new algorithms are only slightly different from the old, these results suggest that the space of possible reinforcement learning algorithms is mostly unexplored. Among the previously-studied algorithms compared are those due to Minsky, Rosenblatt, Farley and Clark, Widrow, Samuel, and Michie and Chambers. The most sophisticated of the new algorithms appears to be a refinement and generalization of the algorithm used in Samuel's celebrated checker-player to modify and improve its static evaluation function. This talk will emphasize (1) the difference between reinforcement learning and other basic forms of learning which have already been thoroughly studied, (2) the demonstration of improvement over previously-studied methods, and (3) areas of possible application of reinforcement learning methods. ------------------------------ End of AIList Digest ********************