jaw@eos.UUCP (James A. Woods) (04/01/88)
# "When you go chasing rabbits, and you know you're going to fall..."
-- Grace Slick, after Charles Dodgson
"But I have dreamt a dreary dream beyond the Isle of Skye;
I saw a dead man win a fight--and I think that man was I."
-- Raphael Aloysius Lafferty, citing the Ballad of Chevy
Chase, in "The Devil is Dead".
For some time, there has been serious concern (if not outright
consternation) in the chessplayer community over computer chess strength.
Before this morning, I thought I had marshalled enough facts to feel
comfortable about the direction of the asymptote. Two evenings ago an
unlikely conversation struck up at a San Francisco coffeehouse unravelled
some rather frightening information about what appears to be the beginning
of the end of chess as we know it.
This much I had already absorbed: CHIPTEST-m went into ACM 1987 at
around UCSF 2250, after sparring 11-5 with a buggy HITECH. Even Feng Hsu's
machine had some flaky check extension code in the quiescence search
(theoretically worth 100-200 points, according to claims in the ICCA Journal)
only partially in place. Well, the good Professor Berliner wrote to me that
post-ACM HITECH regained the lead against CHIPTEST, posting 6-2 after a
gentlemanly rematch. The recent National Open drove home the consistent
near-2400-level play, amplifying the point that Berliner's chess knowledge
above and beyond Ebeling's VLSI running enchanced alpha-beta (iterative
search, transposition table, killer heuristic, good time control) is worth
a solid 200 points.
Now, the details to follow are still a bit fuzzy to me, as I am not
a hardcore chess programmer. Anyway, it's been bandied about that since late
last year, Feng Hsu has been reworking his chess chip from some rather
standard university CMOS 3-micron technology into 1.5 micron design rules.
Sporting a doubled clock rate, the new chess chip would emerge from foundry
turnaround (underwritten by a anonymous large computer corporation) with
minimal delay. Combined with his recent experiments involving Don Beal's
intriguing "null move" technique (Adv. Comput. Chess vol. 4) and
the intimation that the (low-profile) Ebeling/Berliner HITECH II effort might
use the enhanced Hsu chip, a multi-chip parallel implementation with PV
(principal variation) splitting, would pack quite a punch. Straightforward
extrapolation (baseline 2nd-generation Hsu machine 2300 + 200 points from
Hans' heuristics + 100 points fully debugged quiescence extension + 100-150
points null-move done PHOENIX-style, courtesy Jonathan Schaeffer), not
even counting "massive parallelism" at 10% efficiency -- yields something
powerful enough to be disconcerting even to a grandmaster calibre player.
So much for half-baked speculation. For me, the bombshell dropped on
Wednesday at the aforementioned chess hangout, Cafe Picaro on 16th Street
in San Francisco's Mission District. In between five-minute games,
master Paul Whitehead mentioned his friend IGM Walter Browne appearing
uncharacteristically subdued. Paul mumbled something about a "new chess
machine just completed at Carnegie-Mellon, way beyond HITECH" that Browne
just couldn't fathom. Well, I called Browne's residence that evening, and
his wife (the psychologist) answered that Walter (an acquaintance from U. C.
Berkeley days passed) was really distraught, having been handily beaten during
heretofore secret experiments involving the new beast. She was pretty bummed
too, having coached Browne through the wild era of endgame play with the
various Thompson databases. Apparently Browne never has been too psyched
by machine play.
Well hot damn, that was an earful! The kicker, and this is sure
to get out to the press soon, was that Sergey Kudrin (2625), also no
stranger to HITECH, was having serious trouble, and hadn't even managed
a draw against this thing. Both IGMs were frazzled (the designers
were running out of competition to benchmark the box), and Browne on
Thursday hastily contacted Larry Evans to dig up Fischer in L. A.
That's where I left off on that front. However, yesterday, principals
at CMU pretty much confessed via E-mail that only four extra ply (plus
new positional stuff in the evaluation function) seemed to "crack the wall".
Indeed, there was a hush-hush Ebeling/Hsu/Berliner collaboration,
and they were getting really lopsided results against "old HITECH",
like 87-0-4 win-loss-draw in testing. Of course, the foundry couldn't roll
out enough circuits to try the full-bore 512-node configuration, but
none other than Stephen Wolfram (see the Fortune Magazine of April 11 for
a profile of this off-scale theoretician) was recruited to assist with
the parallel bottlenecks. Turns out he's still slaving away over this
while a Japanese foundry is pounding on yield. What they *did* do was to
temporarily lash up forty-odd 2nd-generation chips in a MINIX configuration
ala the Schaeffer experiments, and *this* was the prototype being rushed
to Bobby!
Man, some folks I know are getting scared ...
James A. Woods
NASA Ames Research Center
(ames!jaw, or jaw@ames)LINDHOLM@pucc.Princeton.EDU (Eric Lindholm) (04/03/88)
In article <490@eos.UUCP>, jaw@eos.UUCP (James A. Woods) writes: [Rather interesting but obviously first-of-April-inspired story deleted] > ...they were getting really lopsided results against "old HITECH", >like 87-0-4 win-loss-draw in testing. A disappointment. I thought you were going to give us the traditional "Hey gang this is April fool's in case you hadn't caught on yet" clue, by making the score 88-1-4. >James A. Woods Eric C. Lindholm 6106306@PUCC (bitnet) "No place is safe for the naive."
jesup@pawl8.pawl.rpi.edu (Randell E. Jesup) (04/03/88)
In article <490@eos.UUCP> jaw@eos.UUCP (James A. Woods) writes: > That's where I left off on that front. However, yesterday, principals >at CMU pretty much confessed via E-mail that only four extra ply (plus >new positional stuff in the evaluation function) seemed to "crack the wall". >Indeed, there was a hush-hush Ebeling/Hsu/Berliner collaboration, >and they were getting really lopsided results against "old HITECH", >like 87-0-4 win-loss-draw in testing. Of course, the foundry couldn't roll Right. <sarcasm> I note that this message was sent a 5am on April 1. A coincidence? >while a Japanese foundry is pounding on yield. What they *did* do was to >temporarily lash up forty-odd 2nd-generation chips in a MINIX configuration >ala the Schaeffer experiments, and *this* was the prototype being rushed >to Bobby! Note that Minix is a PC unix-like operating system, and has nothing to do with Chess. Also, since when does Fisher have ANYTHING to do with chess nowadays, computer or otherwise? A very well done April Fools joke. I compliment you. // Randell Jesup Lunge Software Development // Dedicated Amiga Programmer 13 Frear Ave, Troy, NY 12180 \\// beowulf!lunge!jesup@steinmetz.UUCP (518) 272-2942 \/ (uunet!steinmetz!beowulf!lunge!jesup) BIX: rjesup (-: The Few, The Proud, The Architects of the RPM40 40MIPS CMOS Micro :-)