Copy Link
Add to Bookmark
Report
AIList Digest Volume 4 Issue 070
AIList Digest Wednesday, 9 Apr 1986 Volume 4 : Issue 70
Today's Topics:
Queries - BKG Backgammon & LISP Machines & Games,
Applications - Machine Translation,
Correction - Research Credits for Aviation Week Survey,
AI Tools - Common Lisp Systems & Borland Prolog,
Book - Machine Learning: A Guide to Current Research,
Databases - Nonmilitary AI Jobs & Reference Database on Logic,
Techniques - Rete Algorithm Survey
----------------------------------------------------------------------
Date: 27 Mar 86 16:37:00 GMT
From: pur-ee!uiucdcs!convex!ti-csl!dnichols@ucbvax.berkeley.edu
Subject: BKG request
I am interested in obtaining a copy of Hans Berliner's
famous BKG program. Does anyone know of an implementation
in LISP or for UNIX?
I would also love to have a copy of the source for studying.
Can anyone help or can anyone tell me if Mr. Berliner is
on the net and how to reach him?
Please respond to me rather than flooding this list.
*USNail* *electronic*
Dan Nichols USENET: {ctvax,im4u,texsun,rice}!ti-csl!dnichols
POB 226015 M/S 238 ARPA: Dnichols%TI-CSL@CSNet-Relay
Texas Instruments Inc. CSNET: Dnichols@Ti-CSL
Dallas, Texas VOICE: (214) 995-6090
75266 COMPUSERVE: 72067,1465
He o shite shiri-tsubome!
------------------------------
Date: Wed, 2 Apr 86 15:46:09 EST
From: reiter@harvard.HARVARD.EDU (Ehud Reiter)
Subject: LISP machines
Has anyone done a price/performance comparison of LISP machines with
conventional workstations running LISP? If so, could they please send
me the results of their investigations? I will summarize to the net if
there is a lot of interest.
My interest is academic (price/performance of different computer architectures)
not practical. My initial hypothesis, based on looking over Richard Gabriel's
book PERFORMANCE AND EVALUATION OF LISP SYSTEMS and on talking to people is
that special LISP processors offer a 2-3 fold speed advantage over a SUN 3 or
MicroVAX II class workstation, but at 2-3 fold greater cost. Microcoded
architectures like Xerox's D-machines seem to offer little performance
improvement.
Please note that I am NOT interested in software issues like how good an
environment a machine provides. This is strictly a hardware comparison.
Thanks.
Ehud Reiter
reiter@harvard.ARPA
reiter@harvunxh.BITNET
harvard!reiter.UUCP
------------------------------
Date: Tue, 8 Apr 86 08:12 ???
From: Black holes are where God is dividing by zero
<SHERZER%ti-eg.csnet@CSNET-RELAY.ARPA>
Subject: Wanted: info on game playing systems
Can anyone give me any information on game playing AI programs? I
am especially interested in systems that play games where there is
a great deal of uncertainty.
Poker (or any card game) would be a good example. This is because
a Poker player does not have complete information about the other
players hand. The player is therefore forced to deduce the other
players hand by observing his play.
Chess would be a bad example because there is no missing
information. All possible moves for both players are known with
total certainty.
I would also be interested in any programs that build models of
a users behavior (especially a hostile one) with the goal of
guessing future behavior.
Thanks in advance
Allen Sherzer
SHERZER@TI-EG.CSNET
------------------------------
Date: 8 Apr 86 09:49 EST
From: Gocek.henr@Xerox.COM
Subject: Re: Machine Translation of Documents
I read a similar report that said machines are translating 100,000 pages
of text per year for various applications, and in some cases reach 95
percent accuracy. The article I read, which was printed in the
Rochester Democrat & Chronicle on Sunday, 4/6/86, appeared to be
prompted by Xerox's use of machine translation. (Xerox is big in
Rochester.) The 95 percent accuracy was reached only in very
specialized applications, though. Highly technical applications where
the technical jargon is unambiguous is a good application for machine
translation. The European Common Market is trying to use a machine
translation system and is not obtaining 90 percent accuracy.
Gary
Gocek.Henr@Xerox.Com
------------------------------
Date: Tue 8 Apr 86 13:19:44-PST
From: GARVEY@SRI-AI.ARPA
Subject: Re: Aviation Week Technical Survey
I think you should have given credit where credit is due: for example,
the DARPA Pilot's Associate program is also jointly supported by
Lockheed-Georgia and McDonnell Aircraft Company, since they together
are providing approximately half of the total $20 million. Likewise,
the Autonomous Land Vehicle is jointly supported by DARPA and
Martin-Marietta and the first Navy Battle-Management Program (FRESH)
is partially supported by TI.
Cheers,
Tom
------------------------------
Date: 25 Mar 86 1257 PST
From: Les Earnest <LES@SU-AI.ARPA>
Subject: Common Lisp systems
We have been reviewing Common Lisp implementations that run on Sun workstations.
The principal alternatives appear to be those made by Lucid and marketed
by Sun (415 965-780), Franz Inc. (415 769-5656) and Kyoto University, which
is marketed by Ibuki (415 949-1126). We expect to be getting some of each
of these implementations for various purposes. Ibuki's product
description is attached.
Les Earnest
********************************************************************************
KCL PRODUCT DESCRIPTION
Kyoto Common Lisp (KCL) is a full implementation of Common Lisp. It
contains all the Common Lisp functions, macros and special forms defined
in the Common Lisp Reference Manual. It has both a compiler and an
interpreter. Full sources are available for modification.
KCL was developed at the Research Institute for Mathematical Sciences,
Kyoto University, Kyoto, Japan by Masami Hagiya and Taiichi Yuasa.
It is used throughout Japan for building expert systems and conducting
research in Artificial Intelligence.
THE FEATURES OF KCL
-- KCL is complete: It supports all Common Lisp functions, macros and
special forms defined in the Common Lisp Reference Manual; COMMON LISP:
THE LANGUAGE, by Guy L. Steele et al, Digital Press, 1984.
-- A complete KCL is small: It is only 1.4 MB with interpreter and
compiler loaded. For customers with source code, this core image may be
made even smaller by separating the compiler, intepreter and run-times
and making everything inessential autoloadable.
-- KCL is efficient: Its compilation time (including the two passes)
and run time (both of compiled and interpreted code) have times
comparable with the other Commmon Lisps present on the market
(benchmarks appear in the KCL report).
-- The kernal of KCL is written in C and the rest in Common Lisp itself.
Thus KCL is totally embedded in the C language and provides clean
access to the underlying operating system.
-- KCL uses C and the standard C libraries as the interface to the
operating system. Using the standard I/O facilities greatly enhances
the portability of KCL.
-- The KCL compiler is a two pass compiler with a first pass from LISP
to C and a second from C to compiled code. This allows the use of
any optimizing C running on the machine to create efficient code which
is totally compatible with preexisting compiled C code.
-- Having a kernel written in C and compiling to C, KCL is highly
portable and independent of the machine and operating system. It
currently runs on the machines of six manufacturers and more are being
added soon.
-- All KCL versions are made from the same sources. This means that
all versions behave the same and any Common Lisp code can be cross-
compiled (by the KCL compiler) and the C code generated can be used
on any of the systems running KCL.
-- The runtime efficiency of interpreted code has been as important a
design criterion as the efficiency of compiled code. This, together
with its small size makes KCL appropriate for teaching. Educational
discounts are available.
IBUKI is dedicated to providing high quality software that is fairly priced
and allows the people using it maximal flexibility to get their problems
solved. We believe in symbolic computing and want to make it available
on a wide scale. For this reason we provide source code and simple,
inexpensive licensing arrangements.
Versions for VAXes and SUNs running UNIX 4.2 bsd are currently available
in the US and are being distributed by IBUKI. For commercial use,
distribution fees are $700 per CPU for the object code and an additional
$700 for the sources. For educational institutions the distribution fees
are $450 object and sources respectively. Quantity discounts are available.
For further information about ordering, contact
IBUKI
399 Main Street
Los Altos, CA 94022
Phone: 415 949-1126
Telex: 348369
Netmail: KCL@SU-Carmel.ARPA
------------------------------
Date: 31 Mar 86 21:33:29 GMT
From: dual!islenet!jayf@ucbvax.berkeley.edu (Jay Fields)
Subject: Borland Prolog
I just read in today's Infoworld that Borland has announced
a new Prolog for the IBM priced at 99.95. They didn't say,
"Sorry, one per customer," either.
Aloha,
J Fields
PRC, Honolulu
...ihnp4/islenet/jayf
/* The usual disclaimers go here */
------------------------------
Date: 3 April 1986 1616-EST
From: Jaime Carbonell@A.CS.CMU.EDU
Subject: Yet another ML book...
[Forwarded from the CMU bboard by Laws@SRI-AI.]
Not to be confused with "Machine Learning Vol I" and "... Vol II",
Kluwer Academic Publishers is coming out with a book titled:
"Machine Learning: A Guide to Current Research", which contains
a zillion (i.e. 77) very short papers -- rather than a lot fewer, but
much more detailed papers of the two ML volumes. Thus, the
Kluwer book is very useful as a survey and guide to the symbolic
machine learning field, but not as useful for in-depth analysis
of techniques, ideas or applications. Most of the short papers
are revised versions of those presented at the 1985 Machine
Learning III workshop.
[...] There's a 30% discount on the 39.95 price and no shipping cost
(hence: 27.95) for prepaid orders received "soon" (ignore the April 1
date on the form).
------------------------------
Date: Tue, 1 Apr 1986 14:23 EST
From: HENRY%OZ.AI.MIT.EDU@XX.LCS.MIT.EDU
Subject: AI Jobs
A while back on this list, I mentioned a job bulletin board
sponsored by High Technology Professionals for Peace. It is
now available. The number is (617) 969-2273, and hours of
operation are after 5 PM Eastern time weekdays and all day weekends.
It lists employers recruiting for non-military jobs. Later
versions of the system will provide keyword retrieval.
------------------------------
Date: 3 Apr 86 22:22:39 GMT
From: allegra!mit-eddie!think!harvard!seismo!mcvax!ukc!dcl-cs!nott-cs!
abc@ucbvax.berkeley.edu (Andy Cheese)
Subject: Reference Database on Logic
I currently post out a reference database on functional and logic
languages, denotational semantics and formal methods to various people.
It is never up to date but i add more when i have the time. If anybody
is interesting in recieving a copy, i post it at the beginning of every
month, please reply and i will add you to my distribution list.
Andy Cheese
Department of Computer Science
University of Nottingham
University Park
Nottingham
NG7 2RD
England
ARPA : abc@uk.ac.nott.cs
UUCP : ukc!nott.cs!abc
Andy Cheese
------------------------------
Date: Wed 2 Apr 86 10:46:43-PST
From: Matt Heffron <BEC.HEFFRON@USC-ECL.ARPA>
Subject: Rete query summary
Thanks to all who replied to my query about Rete algorithm info.
Here is a summary of the replies:
From: Dan Scales <SCALES@SUMEX-AIM.ARPA>
I'm doing a master's thesis on modifying the Rete network
implementation in OPS5 to be more efficient for an AI architecture
called SOAR built on top of OPS5. The main references for the Rete
network itself are:
Forgy, C. L., On the Efficient Implementation of Production Systems.
PhD thesis, Dept. of Computer Science, CMU, February, 1979.
Forgy, C. L. Rete: A Fast Algorithm for the Many Pattern/Many Object
Pattern Match Problem, Artifical Intelligence 19(1), September 1982,
17-37.
Also, you should try to get the OPS5 (or other OPS) source code. I
assume it is freely distributed, since we have it here at Stanford.
Unfortunately, it is not commented at all.
________________
From: Duke Briscoe <duke@mitre.ARPA>
... The person in the office next to mine has implemented the Rete
algorithm. It doesn't sound like he had too much trouble doing it.
I guess the tricky part is keeping track of variable bindings for
different invocations of a rule.
________________
From: Robert Farrell <farrell@YALE.ARPA>
Lee Brownston, Elaine Kant, Nancy Martin and I have a book called
"Programming Expert Systems in OPS5" available that describes the
algorithm in some detail. Also look at Forgy's AAAI article about
how to implement them in assembler and his thesis from CMU.
Or you can contact Forgy directectly at Forgy@CMU-CS-A.
Also Liz Allen (used to be at MD) has hacked up one in the YAPS system,
so she would be of help. Please don't contact me - I'm too busy.
________________
FROM: E1AR0002%SMUVM1.BITNET@WISCVM.WISC.EDU
In response to your query regarding Rete algorithms, here
is a reference to a conference that will be published in April.
It may prove useful to you:
D BOOK22 Applications of Artificial Intelligence\
%I Society of Photo-Optical Instrumentation Engineers\
%D 1-3 April 1986\
%N 635\
%C Orlando
%A L. Lafferty
%A D. Bridgeland
%T Scavenger: an Experimental Rete Compiler
%B BOOK22
%K AI01
________________
From: Dan Miranker <DAN@CS.COLUMBIA.EDU>
A cornerstone of my thesis, which I am just completing, is
the development of a new production system algorithm, TREAT,
and its comparison to RETE.
The preliminary results are just coming in. Even though
TREAT was motivated by the algorithmic requirements of parallel
processing it is doing better even in a sequential environment.
I have an OPS5 implementation just coming to life. It appears that
TREAT reduces the number of comparisons to do variable binding by
about 30%. (TREAT does more work on an add to wm, but eliminates all
the work RETE has to unwind when doing a delete). TREAT also doen't
use any of the "beta memories", which can be combinatorially explosive
in size. So it does better in space as well. The absolute speed of
the two OPS5 implementations,(mine and Forgy's) is currently roughly
the same, but we haven't yet made any attempt to cleanup and speed up
our code.
The TREAT algorithm is also much easier to implement. Our run
time interpreter is 4 pages of LISP compared to Forgy's 12.
The TREAT algorithm was described in the 1984 International
conference on fifth generation computing, held in Tokyo.
There is a slight error in the algorithm as published. If you
think you will be implementing TREAT let me know and I'll finally
insert the correction into the tech report version and send
that to you.
________________
From: Jim Wogulis <wogulis@ICSE.UCI.EDU>
We have a production system here that was developed my a number
of people over a long period of time. Currently, Pat Langley
has taken over maintaining/improving the system. It is written
in Franzlisp, and I have ported it to Interlisp-D.
Prism uses a rete net to store all the partial matches from
the rules and facts. We will send it to anyone who is willing
to pay for the taping charges (I think $100 for tape or floppy
and $30 for the manual). This might help since there would
be code to look at.
________________
Matt Heffron
------------------------------
End of AIList Digest
********************
