Author: Dann Corbit
Date: 13:50:16 10/13/99
Go up one level in this thread
On October 13, 1999 at 14:30:30, Ratko V Tomic wrote: >Well, say a factor 100 is gained in speed. That would be at most gain of 3-4 >plies in depth. Given that DB already computes the near term combinations better >than GK, and GK has better planning and can see long term consequences better, >the 3-4 plies will, of course increase the percentage of tactical shots which >exists in a position, which are seen by DB and missed by GK, but the question is >whether that figure isn't already nearly maxed out with the existent hardware >(i.e. if the3 current DB already saw 99.9% of near term existent tactial shots, >which missed by GK, and the 100xDB sees 99.999% it isn't much of a difference). >That is, the strenghtening DB in that capability in which it was already much >stronger than GK (or any human), it won't necessarily improve perceptibly its >results over GK. By a more careful choice of openings and positions he enters, >GK (or other human champions) could easily offset such minor adjustment in the >percentage due to hardware speedups on the top of essentially brute force >searcher. > >In checkers, for example, Tinsley vs program Chinook, won the match by seeing >plans 35-40 plies long, which were far beyond Chinook (which could see around 20 >plies ahead). Speeding up Chinook 100 or 1000 times wouldn't have made much >difference, since Chinook still couldn't see (not even remotely) what Tinsley >could. > >One should also consider that if the computer-human chess competition gains in >importance (among the top chess players), along with the sufficient incentives >for the humans to win, the anticomputer strategy will evolve as well. Currently >that area of chess theory is quite undeveloped (a handful of outdated guidlines, >proposed by isolated individuals who found themselves on the spot on few >occasions), but given the right incentives to the professional chess community, >it could develop fast and set the computers back another few decades or longer. >The top human player isn't necessarily the best person to advance such theory >(I'd guess that Karpov or Fischer would do better against programs than Kasparov >or Anand). In few years we may have a human champion (or two) and an >anti-computer champion (just as postal chess is its own separate world). By the time 30 seconds have passed, DB starts doing special extensions. With 100x the power, I think it would see strategic manouvers. I believe (though I may be mistaken) that DB would do real, live long-range plans. Even if that were not the case [ignoring extensions], consider this: At 200M NPS, it probably sees about 15 plies ahead (just a guestimate based on C.A.P. data) during a tournament game. 100x faster would give probably 2-3 more plies. That's about nine full moves on a crowded board (without any screw-ups.) When the board is fairly sparse (say ten pieces+pawns) I suspect super DB would get 20-30 plies most of the time. Look at programs that can [be set to] calculate only 2 plies. Your 10 year old kids can beat it. If we go to 5 plies, it is quite a good opponent. At 10 piles, they play brilliantly. At 18+ plies they would be "the gifted of the gifted of the gifted" as far as tactics are concerned, and I think such a machine would actually have strategic power from time to time (depending on the board condition). Look how hard they would make an expert work. 17% of the time, a new ply gives you a better move. With 3 plies deeper, that is nearly 50% of the time you get a better move [(1-.83)^3] on average. If you are not prepared for it, you will have to think really hard about why a particular choice was made. I don't think there is any drop-off in the value of another ply. Just pit an old M68000 against an Alpha and see how often the old hardware wins.
This page took 0.03 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.