Author: Joshua Lee
Date: 00:56:09 03/13/02
Go up one level in this thread
On March 13, 2002 at 02:31:22, Kurt Utzinger wrote: >Hi Tanja > >There is nothing strange: > >You have still not played enough games. I give below an example of a match >[40'/40] I had played some time ago over 100 games between Gandalf 4.32g and >Program_X [I am a beta tester of X] to show what I mean: > >Gandalf 4.32g vs Program X > >Games 1-10 [win Gandalf] >3.0-7.0 (total 3.0-7.0) > >Games 11-20 [win Gandalf] >6.5-3.5 (total 9.5-10.5) > >Games 21-30 [draw] >5.0-5.0 (total 14.5-15.5) > >Games 31-40 [win program X] >3.5-6.5 (total 18.0-22.0) > >Games 41-50 [win program X] >4.5-5.5 (total 22.5-27.5) > >Games 51-60 [win program X] >3.0-7.0 (total 25.5-34.5) > >Games 61-70 [draw] >5.0-5.0 (total 30.5-39.5) > >Games 71-80 [win Gandalf] >8.0-2.0 (total 38.5-41.5) > >Games 81-90 [win Gandalf] >7.0-3.0 (total 45.5-44.5) > >Games 91-100 [win Gandalf] >5.5-4.5 (total 51.0-49.0) > >Can anybody tell me for sure which of the above two is the stronger program?? >And what about if I had only played a 20 games match and these games would have >been those played in rounds 71-90? Then, the result would have been 15.0-5.0 in >favour of Gandalf 4.32g!! Imagine what some testers would have argued about the >strenght of program X? For all these reasons I think that something concrete >about the strength between two programs can only be said if 100, better 200-300 >games or even more have been played. >Kurt So on the same token, wouldn't this apply to Humans Vs Computers? Not that you would be able to get a Super GM to play 100 or more games againts a single Computer but what would be a decent amount of games to show with a the smallest possible margin of error which opponent was better? Obviously 4 games is not enough look at how "Easily" GM Wely beat Rebel Century 4 in those two games of their match. Nor would 6 games be quite enough, Do you really think Deep Blue for instance is "Stronger" Than Kasparov? Especially since some moves by Kasparov can take 19+ Ply for a computer to fully understand; depending on what your deffinition of understanding is... If we assume that Deep Blue is Still Stronger than current top Programs or if you believe otherwise even, A program has to be able to somehow figure out mathmatically what the spirit or essence of a position is to truely know what it is doing, this would mean knowing despite this or that opening move when to defer from book for example since the book was not made by GM's with a Slide ruler or a calculator, i just don't see how a comp could without seeing completely out 20+ moves or 40ply or some number similar just what to do next. Deep Blue didn't see everything out to a +- score in game 6, just analyze backwards and you'll see that DB would have to be much faster than claimed to do this. I realize this is just the openings but not all combinations can be solved as yet either. As far as Rebel Century goes it was already +- (Lost) within the first 17 moves of it's game against GM Wely. I am not sure if it was out of book but that's where i have seen many computers end up losing the thread .... once out of book. Since not every position can be explained mathmatically and computers do not plan , they Calculate, What can be done about actual Understanding/Planning? Is their a way to have a program play more by Themes or some other non calculable element? All i think one can say is that Even the strongest GM's can be beat by a Computer, but does it make the Program GM Strength? If you can get the computer rated it is fairly obvious that a Program can get the Title. This still does not mean the computer understands like the GM or IM and since the Computer Calculates as well as the GM sometimes better, it still doesn't Calculate Deep enough to Understand why it is being outplayed positionally. 9 games were not enough either as Junior was made to look like a 2200-2300 player in it's two losses. So the question still is how many games would you need to make a more realistic estimate?
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.