Author: Uri Blass
Date: 14:28:15 11/25/00
Go up one level in this thread
On November 25, 2000 at 14:54:01, Peter Fendrich wrote: >On November 24, 2000 at 11:05:04, Uri Blass wrote: > >>On November 24, 2000 at 10:31:15, Peter Fendrich wrote: >> >>>On November 24, 2000 at 10:02:04, Uri Blass wrote: >>> >>>>On November 24, 2000 at 09:05:15, Peter Fendrich wrote: >>>> >>>>>On November 24, 2000 at 02:47:09, Uri Blass wrote: >>>>> >>>>>>On November 23, 2000 at 19:06:29, Peter Fendrich wrote: >>>>>> >>>>>>>On November 23, 2000 at 16:10:54, Uri Blass wrote: >>>>>>> >>>>>>>>On November 23, 2000 at 12:33:12, Peter Fendrich wrote: >>>>>>>> >>>>>>>>>On November 22, 2000 at 12:10:18, Uri Blass wrote: >>>>>>>>> >>>>>>> >>>>> >>>>>-- snip -- >>>>>>The reason is simply for analysis of correspondence games. >>>>>> >>>>>>I generate a small tree of moves and I want to use the evaluations after search >>>>>>in the leaves of the tree to decide about my move. >>>>>> >>>>>>If I use a preprocessor I cannot compare between the evaluation in different >>>>>>leaves because they are based on different tables. >>>>>> >>>>>>If I do not use a preprocessor I can do it. >>>>>> >>>>>>Uri >>>>> >>>>>In that case I think you definition is perfectly allright. >>>>>Even small gradually changes between rootnodes will give >>>>>different evaluations, not comparable beteween rootnodes. >>>>>Terra is preprocessing! >>>>>Another thought: These numbers can't be of much value in corr. games. I mean the >>>>>positional/strategic knowledge in chess programs isn't much to rely on. >>>> >>>> >>>>I think that the positional knowledge of myself is also not much to rely on. >>>>If I think that the evaluation of programs is wrong then I try to generate the >>>>tree that is going to convince the program that it is wrong. >>>> >>>>If I fail to generate the relevant tree then probably my evaluation is the >>>>wrong one. >>>> >>>>>When material is the difference however, then the preprocessing has no impact >>>>>anyway. >>>> >>>>I think that there are cases when the evaluation is changed by a pawn because of >>>>preprocessing so preprocessing may influence also in this case. >>> >>>Absolutely not in my program... >>>The evaluation might change even more, but the contribution from preprocessing >>>solely, will never be by that much. Maybe after 10 more moves ahead in the game. >>> >>>I can't imgine that it is a good strategy to let the preprocessed information >>>affect the evaluation that much between two adjecent game moves. The program >>>will aim for one thing at move x and something completely different at move x+1, >>>even if the opponent answered with the expected move. >> >>I agree that it seems not to be a good strategy but I saw it happens with >>Gambittiger and Fritz5.32 >> >>Here is an example for the change in the evaluation in Gambittiger >> >>The scores before 44.bxc6: >> >>3.54 depth 11 >>2.96 depth 12 >>2.40 depth 13 >>2.30 depth 14 >>1.80 depth 15 >>1.68 depth 16 >> >>Gambit's main line is always 44.bxc6 Rc7 >> >>The scores after bxc6 Rc7: >>1.66 depth 9 >>1.72 depth 10 >>1.54 depth 11 >>1.12 depth 12 >>1.32 depth 13 >>1.30 depth 14 >>1.18 depth 15 >> >> >> >>I cannot be sure about the exact difference because I do not know to translate >>depthes because it is possible that there are different extensions but it seems >>that the difference in the static evaluation of part of the positions is more >>than one pawn after 2 plies. >> >>Uri > >Does it really has to be preprocessing? >There might be another explanation here. Before the move 44. bxc6 at depth 15 >GT says 1.80 and in the next iteration 1.68. This about the same level of >evaluations that comes after Rc7. The small differences after that could be >different use of extensions or nullmoves (or something related). >//Peter Depth 15 and depth 16 before bxc6 Rc7 is not eqvivalent to depth 9 and 10 after bxc6 Rc7. The difference between the position is only 2 plies and I expect it to be even less than it because of capture extensions. If I assume that it is between 1 and 3 plies then The 2.96 at depth 12 before 44.Rxc6 is eqvivalent to one of the numbers 1.66,1.72,1.54 and you get more than 1 pawn difference. The difference is smaller if you use the last iteration and in this case the 1.68 is eqvivalent to one of the numbers 1.32,1.3,1.18 but you also get almost .5 pawn difference. 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.