Author: Uri Blass
Date: 03:58:29 11/08/01
Go up one level in this thread
On November 08, 2001 at 05:50:10, Thomas Mayer wrote: >Hi José, > >> Kein Problem, nur eine Frage. Ist es nicht langweilig, die gleiche Partie >> wieder zu gewinnen? > >you missed the point - compare those two games - one with 30 minutes, another >one with 60 minutes - the moves were not all the same, but the idea behind the >game was the same. He uses several ideas to crush those beasts and they seem to >work on almost any level. The science is to get the computer humbling in those >traps. You forget that the match of Kramnik is not blitz and on a better machine. It is possible that Fritz7 is not going to do the same kind of mistakes in 120/40 on a fast machine. I am sure that kramnik can win if he has the program but I expect longer games and I am not sure if the ideas of nemeth are going to work against it and it is possible that kramnik will have to prepare different ideas for longer games. It seems it is always possible - now think about that Kramnik prepares >some of such ideas - and he can do that for sure much better then Eduard. > >By the way, Eduard has played some games on the F7-Server against my program >Quark - he lost because he do not know the program. But it was extremely >interesting to watch those games, Eduard moves around doing nearly nothing and >trying to get the engine making some stupid moves he wants for his attack... >that is very cool - it's like knocking on all the stones of the engines wall to >find a hole. I believe when giving my program to Eduard and let him try a day or >two Quark will not win any game against Eduard except I use maybe a completely >different opening book with some strange lines... 1)I believe that these conditions are unfair conditions. playing against a program when the program does not know about the game and cannot change the evaluation after the game is unfair. Humans learn to change their evaluation after games and if you do not give the same opportunity for the program to learn then it is unfair. 2)I believe that it is possible to teach programs to learn from losses and not only learn by position. Programs should learn to generalize from their loss about the reason that they lost and to change the evaluation in some conditions that were common to their losses or learn to extend lines that were common to their losses. I do not think that it is impossible. I think that it may be more easy for the extension part. programs can find lines that prove that they lose and extend similiar lines in the next game Example: if they learn that extending the lines hxg4 hxg4 Nd2 Qh4 f3 g3 ,hxg4 hxg4 Ne1 Qh4 f3 g3 could help them to avoid losing they should extend these lines in the next game and the relevant lines can be found by giving the program to play against itself and taking back moves until they see quickly that the first move was the losing move. I know that Crafty solve the problem by a different way but there may be other cases and I think that extension learning from games may be more productive because of it's generalized nature. Pograms cannot do it immediatly after a game and need time for this learning. Uri
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.