Author: Rajen Gupta
Date: 09:24:47 05/24/01
Go up one level in this thread
On May 24, 2001 at 12:14:38, Uri Blass wrote: >On May 24, 2001 at 08:24:47, Rajen Gupta wrote: > >>On May 24, 2001 at 07:55:37, Mogens Larsen wrote: >> >>>On May 24, 2001 at 07:36:12, Rajen Gupta wrote: >>> >>>>something to do with the learning function, not just in openings but also i >>>>suspect during the game itself when it plays an alternate move, which may turn >>>>out to be better. >>> >>>When did Frans Morsch implement position learning? >>> >>>Mogens. >> >>i dont know but i have noticed it in a few games. dont ask me for examples, i >>dont have them >> >>rajen > >1)I do not know about what you are talking. >I know that if I ask Fritz to analyze position and after it load another engine >fritz does not learn from previous analysis. >There are other programs who do learn from previous analysis(Shredder) so it >seems that Fritz does not have position learning. >I do not have the autoplayer and maybe when Fritz is using the autoplayer it >does something different. > >2)I do not think that position learning is important in the ssdf matches >because the openings are not the same. >My guess is that Fritz wins long matches because of the fact that it has bigger >book than the opponents. > >If one program has bigger book than it is going to find it more easy to repeat >good opening for itself even if it does not repeat games. > >For example suppose one program can score good results with 1.h4 when the >opponent can score good results with 1.a4 > >Suppose Deep Fritz has 1.h4 in it's book and after enough games it discovers >that 1.h4 is good for it and can repeat it(it cannot repeat exactly the same >game but it is not important because the games are similiar). > >The opponent has not 1.a4 in it's book for white because it knows that 1.a4 is a >bad move so it has no chance to discover that Fritz does not know how to play >against 1.a4. > >This example is an extreme example and books usually do not include these bad >moves but basically a big book that includes also a lot of bad moves can help a >program to win if the match is long enough because it is going to learn to avoid >the practically bad lines after enough games when the opponent can do nothing >against the good lines and learning by position is not very important if you get >a position that your evaluation is wrong. > >It is better to give programs bigger books if the matches are longer. >The matches of the ssdf are not infinite so having too big books is also not a >good idea. > >Uri hi uri: you are probably right, fritz does not have positional learning. i assumed it had, because i had seen a few ssdf games by fritz 6 where i thought it played a different move someway through a game but i guess it must have still been in book and merely played a different line.it would however be worth having a second look at close ssdf matches involving fritz6 and above to find out whether it does better in the second half or not rajen
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.