Author: Graham Laight
Date: 15:21:08 05/01/01
Go up one level in this thread
On May 01, 2001 at 18:02:21, Torstein Hall wrote: >On May 01, 2001 at 09:31:11, Robert Hyatt wrote: > >>On May 01, 2001 at 08:59:58, Aaron Tay wrote: >> >>>On April 30, 2001 at 22:19:24, Robert Hyatt wrote: >>> >>>>On April 30, 2001 at 20:42:39, Torstein Hall wrote: >>>> >>>>>On April 30, 2001 at 20:21:44, Sergey pavov wrote: >>>>> >>>>>>On April 30, 2001 at 18:47:28, Robert Hyatt wrote: >>>>>> >>>>>>>On April 30, 2001 at 18:03:26, John Dahlem wrote: >>>>>>> >>>>>>>>On April 30, 2001 at 15:42:44, Robert Hyatt wrote: >>>>>>>> >>>>>>>>>On April 30, 2001 at 14:22:44, Sergey pavov wrote: >>>>>>>>> >>>>>>>>>>Hi Professor >>>>>>>>>> >>>>>>>>>> I am wondering if your perdection of three years ago that programs at best >>>>>>>>>>play 2450 fide holds true for the programs today running on 1 gig processors? >>>>>>>>> >>>>>>>>>Pretty much. Programs _might_ hit 2500 barely today. But I think that is >>>>>>>>>about the absolute max value I would suggest... >>>>>>>>> >>>>>>>>>They have a _long_ way to go. But they _are_ making progress... >>>>>>>> >>>>>>>> >>>>>>>>Can you make another Fearless Forecast and tell us where they will be in another >>>>>>>>3 years (and what kind of hardware will we have then)? >>>>>>> >>>>>>> >>>>>>>Here is the problem: Some of this is _not_ search depth constrained. IE I >>>>>>>have 3 kinds of machines I can run... a quad xeon 400, a quad xeon 550 and >>>>>>>a quad xeon 700. The 700 is nearly twice as fast as the 550, and close to >>>>>>>double the speed of the 400. I use them interchangably with "scrappy" on ICC >>>>>>>and I really don't see much difference. I have asked a couple of GMs if they >>>>>>>can tell when I run on faster or slower hardware and the answer is generally >>>>>>>"no". They might guess, but they are wrong as often as they are right. >>>>>>> >>>>>>>That means that some problems are not going to be solved by faster search. >>>>>>>Yes, some tactical things need more depth, and more depth will _always_ make >>>>>>>the program play stronger overall. But some things simply can't be done by >>>>>>>depth alone, and until the programs recognize these "things" and handle them >>>>>>>in evaluation, they are going to get stomped on by some. >>>>>>> >>>>>>>I have solved some of the problems myself, but nowhere near all of them, or >>>>>>>even a majority of them. But I do have a good idea of how far we have to go >>>>>>>in the evaluation of chess positions, and we have further to go than we have >>>>>>>already come, IMHO. >>>>>>> >>>>>>>And speed won't help many of those problems one bit... >>>>>>> >>>>>>>So in 3 years we might see 4-5 ghz processors or so... but the engines are >>>>>>>not going to feel stronger unless they become "smarter" as well... >>>>>> >>>>>> >>>>>>Don't you think Gambit Tiger has solved many of these problems? It seems to >>>>>>approach chess from a totally different angle, with positive results. I think >>>>>>Chris is the messiah of computer chess. >>>>> >>>>>I'm not sure about this of course, but is it possible to get very much smarter >>>>>using the approach of todays programs? >>>>> >>>>>When they evaluate a position they give a numerical value to each positional >>>>>"feature" they recognize in a position. say free pawn = +1 Pawn weakness -1 >>>>>Bishop pair 1.1 etc. etc. It must be countless positions where the numbers >>>>>cansel each other out, even if one of the positional "features" should be >>>>>counted far above the others in the specific position. Of course search helps >>>>>quite a bit, but not enough I feel. >>>>> >>>>> >>>>>So my prediction is that the programmers are hitting into a brick wall head >>>>>first if they can not find a new approach to positional evaluation. >>>>> >>>>>Torstein >>>> >>>> >>>>I don't think you will find many programs relying on simple first-order >>>>evaluation terms that behave as you describe. Most use at least a second-order >>>>evaluation where pieces influence the score based not only on their own >>>>positions, but on the positions of other pieces as well. I go even further >>>>in some cases with a third order interaction between pieces, king safety, >>>>and so forth... >>>> >>> >>>Sure. But no Chessplayer thinks like that, . As programmers, how are you to >>>know what terms , weights to change or add? Or is everything trial and error? >> >> >>I disagree. I've been playing over 40 years and _I_ think like that. IE >> >>Your g-pawn is missing and you have castled kingside. No light-squared bishop? >>then you have a severe kingside weakness. Can I get my pieces there before you >>can? The weakness is exponentially worse. Can I open lines to get my bishops >>or rooks into the fray? The weakness is near decisive. >> >>IE I don't personally evaluate a knight sitting on g4 with you castled at g1, >>as being very dangerous unless I can get other pieces over to help, you have >>some sort of weakness that the knight exploits, or something. > >Of course I over simplified earlier on, but I still have the feeling that to >numerate every positional factor in the position something must go wrong pretty >often, or at least from time to time. If both sides have weaknesses in the >position, is it not very easy for the program to "underscore" the winning side? >But I'm probaly thinking more like a rule based machine... :-) > >Torstein This represents a major flaw in modern chess programs with their relatively simple eval functions. They have to accurately evaluate up to a million NPS (single processor). This represents a node every 1000 clock cycles - and this can ONLY be done in a relatively simple way. In many positions, the quick eval is enough - but in many other positions - it just isn't. Sometimes, you'll get lucky and get the right eval for the wrong reasons. Other times, you'll have the reasons roughly correct - and still deliver the wrong eval. As for getting "every positional factor in the position", as you've said above - I think we're talking new technology. Something better than a top human, I think. I believe it would be possible, with today's technology, to set up a system that could, somehow, automatically tease out the major positional factors from every position in the Chessbase database. This would make for a superb chess position evaluator. -g
This page took 0.06 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.