Author: Robert Hyatt
Date: 20:44:29 01/06/00
Go up one level in this thread
On January 06, 2000 at 20:29:49, Michael Fuhrmann wrote: >On January 06, 2000 at 17:28:21, Robert Hyatt wrote: > >>On January 06, 2000 at 11:09:31, Graham Laight wrote: >> >>>On January 06, 2000 at 10:26:58, Robert Hyatt wrote: >>> >>>>I disagree there. If you let a GM play a computer over and over, and _then_ >>>>play the same program in some serious games, the computer is going to catch >>>>hell. Humans evolve. Computer programs do not. Once he finds a hole, he >>>>will exploit it over and over, while a human would 'learn'. >>>> >>>>That will be a weakness for another 25 years or more. >>> >>>Is the number of such "holes" infinite? >> >>Depends on your definition of 'infinite'. If that is just hyperbole for "large" >>then yes, the number is 'large'. The problem is the exceptions. And as you >>code for more exceptions, things interact in bad ways. As the code gets bigger, >>it gets harder to keep these interactions under control. It is all about time, >>which is the precious commodity here. >> >> >> >> >>> >>>And can the GMs learn an infinite number of "holes" that they must avoid in >>>order to get the opportunity to put the computer in its hole? >> >>The GM doesn't have to avoid an infinite number of holes. He only has to >>stumble into _one_ and then the program has big trouble. I can name several >>programs that have _no_ clue about pawn majorities. And I have watched GM >>players exploit this over and over, although luckily these programs are all >>non-automatic, so a GM will thrash them twice and the operator won't play him >>again. Put 'em on automatic, and the holes become glaring, because the IM and >>GM players _will_ exploit them. I've been exploited. I know. :) >> >> >> >> >> >>> >>>The upper limit to the number of "holes" a GM can know about in his head at any >>>given time is probably about 50,000 - based on studies showing that GMs have >>>knowledge of about this number of "piece patterns" (or "positional themes"). >> >>true. The upper limit on the various things a chess program knows (not counting >>special-case hardware like deep blue) is probably closer to 500-1000. That is >>a _huge_ gap. >> >> >> >> >>> >>>However, this doesn't stop Bob from being right about this. There may be a huge >>>number of such "holes" - and it may be that computers v humans becomes a cat and >>>mouse game of covering holes (computers) vs digging new holes (the humans). >>> >> >>And there is the problem that as the human finds a hole in "his" style, he >>will fix it. The computer won't. That is a real problem. I suspect that if >>I just let Crafty play on ICC for 3 months, its rating at the end would be 2-300 >>points lower than at the start. This assuming I made _no_ changes of any kind. >>At present, crafty "learns" by my fixing problems as they show up on a daily >>basis. Sometimes even between games. >> >> >> > >So is it fair to say that many (most?) of the "improvements" you make to crafty >don't actually make the program better (i.e. stronger vs humans, which would >show up in higher ratings), but just prevent it from getting worse (i.e. easier >for humans to beat)? That is like trying to answer "So, have you stopped beating your wife yet?" :) You could define "better" in multiple ways. One is "not getting worse" perhaps. There is little doubt the holes I plug are problems that cause it to lose games it should be able to win or draw. And plugging those holes clearly makes it play better, but only in those situations where the "plug" fits. One example I gave a couple of weeks ago about two split (isolated) passed pawns normally being worse than two connected passed pawns, until all the pieces are removed, then the two split passers win usually. That doesn't happen very often, until someone finds that you mis-evaluate it. Then it becomes pretty easy for a GM to "let" the program reach an endgame with those superior connected passed pawns, only to lose again and again. > >> >> >>>But I also think that it's possible that computers simply become so good that no >>>human will be able to beat them. If this happens, then on the current trend it >>>will take less than 25 years. >>> >>>-g >> >> >>I think this will happen. I am just not sure about the time-frame. However, >>I was around when the programs were barely 1400. I think they are now >>approaching 2500. That is about 1000-1100 points in 30 years. In another >>25 years at that same pace, they will most definitely be unbeatable. if that >>pace can be maintained...
This page took 0 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.