Author: Uri Blass
Date: 15:30:52 09/06/02
Go up one level in this thread
On September 06, 2002 at 17:55:20, Dave Gomboc wrote: >On September 06, 2002 at 16:27:08, Uri Blass wrote: > >>On September 06, 2002 at 15:50:25, Robert Hyatt wrote: >> >>>On September 06, 2002 at 15:15:29, Uri Blass wrote: >>> >>>>On September 06, 2002 at 12:14:00, Robert Hyatt wrote: >>>> >>>>>On September 06, 2002 at 10:56:54, Uri Blass wrote: >>>>> >>>>>>On September 06, 2002 at 07:23:15, Yuriy Lyapko wrote: >>>>>> >>>>>>>Which chess engine will see that Rb1 is winning in this position? >>>>>>> >>>>>>>[D] b4bk1/p2p4/p1p1p2p/P1P1Pp1P/3P3p/2N4N/1r3PP1/4R1K1 w - - >>>>>>> >>>>>>>Best regards, >>>>>>>George >>>>>> >>>>>>Movei0076 can see a small advantage for white with Rb1(p850,32 Mbytes hash) >>>>>>in less than 10 minutes. >>>>>> >>>>>>I imagine that stronger engines can find it faster. >>>>>> >>>>>>scores at the last iterations(from white point of view): >>>>> >>>>> >>>>>I don't make any claims about being "stronger" or "weaker" than much of >>>>>anything, but Rb1 is the _only_ move Crafty considers. >>>>> >>>>>It isn't going to like white's position, of course, until it sees white >>>>>winning material, which suggests that the bishop scoring needs more control >>>>>on the lack of mobility in this position. >>>>> >>>>>> >>>>> >>>>>The score starts out at -1.0 (good for black) and drops to about -.5 quickly. >>>>> >>>>>Which is totally wrong. >>>> >>>>Movei does not consider pair of bishops in it's evaluation. >>>> >>>>If we look at pawns and pawn structure then the score is about 0.4 pawns for >>>>black based on Movei's evaluation(black has one pawn more but movei reduce 0.3 >>>>pawns for every case that there are 2 pawns in the same file). >>>> >>>>Movei evaluates usually bishop as slightly more than knight but >>>>in this position it is not the case because the knights have better squares >>>> >>>>Movei gives white a small positional advantage because of better mobility except >>>>the root position when the black rook is more active than the white rook. >>>> >>>>I guess that this is the reason for the scores that are close to 0.00 score. >>>> >>>>At big depth movei can see advntage for white that is increasing >>>>Here is movei's output for depth 18-21 >>>> >>>>Depth 18 0.36 for white >>>>Depth 19 0.45 for white >>>>depth 20 0.47 for white >>>>depth 21 0.54 for white >>>> >>>>debug file(the negative number of nodes is not a problem that >>>>is important for me today) >>> >>>Hey!... >>> >>>weren't you one of the ones urging me to go to the 64 bit node counter >>>in crafty, for longer analysis? >>> >>>Get that node counter fixed... :) >> >>I remember that I said that I prefered Deep Fritz for deep search because of >>having correct number of nodes. >>I also know that the history table may be wrong if I search many nodes but >>I read that history table is not very important when I use hash tables and >>killers and some even say that it is counter productive. >> >>I guess that I will fix the number of nodes in the future but today it is not in >>the top of my priorities. >> >>Uri > >But since it should take you about three seconds to fix, why don't you go ahead >and fix it? > >Dave I do not think that it is going to take 3 seconds. I agree that it can be relatively easy because I already printed 64 bit numbers in the function that calculated perft but The problem is that I do not have the exact version of 0076 and only versions few days before it and few days after it(I did not follow christophe's advice closely and saved only 0075 and 00761). I do not consider it as a big problem because the difference in strength between 0076 and 0075 is not big. 00761 seems to be weaker than 0076 inspite of the fact that I fixed a bug (0076 may display the wrong pv) It is also possible that 0076 is weaker than 0075(the difference is very small and 0076 is only a bug fix of 0075 that may return wrong mate search from the qsearch that may cause it to lose on time in very rare cases but it is possible that the bug in 0075 cause it to be stronger and at least in the GCP test suite 0075 is the best so far). I expect when I am going to continue to work on movei that I am going to do something clearly better than 0075 or 0076 and today I decided not to work on movei before learning more about C. 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.