Author: Sandro Necchi
Date: 15:44:42 02/22/05
Go up one level in this thread
On February 22, 2005 at 16:58:55, Uri Blass wrote: >On February 22, 2005 at 15:53:49, Kurt Utzinger wrote: > >>On February 22, 2005 at 14:42:20, Sandro Necchi wrote: >> >>>On February 22, 2005 at 14:19:06, Kurt Utzinger wrote: >>> >>>>On February 22, 2005 at 13:45:48, Sandro Necchi wrote: >>>> >>>>>On February 22, 2005 at 13:31:28, Kurt Utzinger wrote: >>>>> >>>>>>On February 22, 2005 at 13:13:01, George Sobala wrote: >>>>>>[...] >>>>>>>So this result is due to bad opening performance in particular circumstances by >>>>>>>Junior 9, and does not in any way reflect on Shredder 9 v Shredder 7.04 relative >>>>>>>strengths. >>>>>> >>>>>> I fully agree. Such games should be deleted and not count or >>>>>> even better: the book learning function should be disabled if >>>>>> we want to know something about playing strength instead of >>>>>> to good implementation of book learning. If one of these games >>>>>> should have occurred at WCCC, it is for sure that Amir Ban >>>>>> would at once change the book line and the game would never >>>>>> be played again. This example shows for me the nonsense of >>>>>> book learning and repeating the same book line again and again. >>>>>> Such computer matches can leave a complete wrong impression -:) >>>>>> Kurt >>>>> >>>>>Hi Kurt, >>>>> >>>>>I do not agree with you because what the people want is a test of the best >>>>>setting one can get from an available program to find out how strong it is. >>>>>The learning feature is very important as if you play against a program and this >>>>>one loses and play it again and again you would get disappointed as people did >>>>>when this extremely important (to me) feature is. >>>>>We spent a lot of time, me and Stefan to discuss on the learning feature and to >>>>>make it better especially to please the customers, but also to get better >>>>>results. >>>>>If one program does not have this function or not as good it should show up in >>>>>the tests and not otherwise. >>>>>I do not think that only because some programs do not use the time good we >>>>>should use ponder off or off learning and so on...why not to switch everything >>>>>off then... >>>>> >>>>>Also to people trying to know everything from a single match I tell them that in >>>>>order to find out a realistic rating one needs to test the program against >>>>>several different program and for many games...exactly what SSDF do as they have >>>>>been doing this from several years and they know what they are doing.. >>>>> >>>>>Sandro >>>> >>>> >>>> Hi Sandro >>>> Nothing against your statement. All is correct and your efforts >>>> in implementing an excellent learning function is appreciated. >>>> Personally I am looking at this question from a more practical >>>> point of view: if my opponent has played me out in the opening, >>>> I will no longer use the same line thus giving him never in my >>>> life the possibility to beat me with the bad opening line. And >>>> therefore I see no reason why we should allow computer programs >>>> to repeat the same games. This makes a match rather worthless >>>> in my opinion and obviously leads to "wrong" Elo's. >>>> Kurt >>>Hi Kurt, >>> >>>we are saying the same thing in a different way. >>> >>>I mean that the learing feature is essential to avoid the program to repeat a >>>game like a human do, so if the program is not doing it than it should be >>>penalized rather than the opposite. >>> >>>THIS WEAKNESS IS PART OF THE ELO of a program and to hide it is not correct to >>>evaluate a program strength like you would do removing those games. >>>This weakness would come out in a tournament too if one finds the way to beat >>>the program, so I understand you may not be interested in those games, but >>>looking things from a Elo strength it MUST be included. >>> >>>The same is true for ponder on and openings books. >>> >>>Can you ask Kasparov not to think while the opponent is evaluating the move or >>>not to use the openings books or home work while playing? >>>The answer is no, so the same is true for a designed chess player program. >>> >>>This is why the SSDF do not remove those games and why me and very many people; >>>most of them, believe the SSDF list is the most accurate one to know a program >>>strength. >>> >>>Sandro >> >> >> Hi Sandro >> Again nothing much to add. But in the given match vs Junior9 >> the learning feature of Shr704 did not avoid bad moves but >> repeated won games and Junior9 was unable to avoid many unnecessary >> losses. This is what I mean when saying such matches do not >> give evidence of playing strength of Junior9. Had the learning >> function of Junior9 worked properly the outcome would not have >> been the same ... and therefore it's wrong to rate such games >> in my opinion. >> Kurt > >I think that it is correctly right to rate the games for Junior9 unless it is >proved that the learning problems are because of some bug in the autoplayer. No, it is not a 232 problem. This would not happen on Fritz 8 or Deep Fritz 8. > >If chessbase release a new version with a learning function that does not work >then they should suffer the consequence. The truth is that the strenght of a program is given by book+engine+learning function+ponder+table bases and of course harware used. Here we can see the evidence that the not well working learning function do have an effect on the Elo. If we want programs to improve we have to force the programmers to do it by effecting the performance of these things and not remove parts which do not work or switch off options...is it clear Kurt? In fact if only one book would be good and you give it to everybody you would damage this program indirectly by giving an advantage to the competitors... > >Uri Sandro
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.