Author: Ulrich Tuerke
Date: 14:40:40 01/20/01
Go up one level in this thread
On January 20, 2001 at 17:26:01, Uri Blass wrote: >On January 20, 2001 at 17:03:05, Ulrich Tuerke wrote: > >>On January 20, 2001 at 15:38:01, Uri Blass wrote: >> >>>On January 20, 2001 at 13:30:15, Ulrich Tuerke wrote: >>> >>>>On January 20, 2001 at 09:56:57, Uri Blass wrote: >>>> >>>>>On January 20, 2001 at 06:24:19, Thorsten Czub wrote: >>>>> >>>>>> >>>>>>[Event "Odyssee2001-Tournament"] >>>>>>[Site "k6-400, 40/120"] >>>>>>[Date "2001.01.20"] >>>>>>[Round "1"] >>>>>>[White "Shredder4 Chessbits-style"] >>>>>>[Black "Comet B27"] >>>>>>[ECO "C99"] >>>>>>[Result "*"] >>>>>> >>>>>>1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Ba4 Nf6 5. O-O Be7 6. Re1 >>>>>>b5 7. Bb3 d6 8. c3 O-O 9. h3 Na5 10. Bc2 c5 11. d4 Qc7 >>>>>>12. Nbd2 cxd4 13. cxd4 Nc6 14. Nb3 a5 15. Be3 a4 16. Nbd2 >>>>>>Bd7 17. Nf1 Rfe8 18. Rc1 Qb7 19. Ng3 a3 20. b3 Rac8 21. Qd2 >>>>>>exd4 22. Nxd4 d5 23. Nxc6 Bxc6 24. e5 Ne4 25. Bxe4 dxe4 >>>>>>26. Nf5 Bf8 27. Nd6 Bxd6 28. exd6 Re6 29. Bf4 Rg6 30. Re3 >>>>>>h6 31. Rec3 * >>>>>> >>>>>>after Rec3 shredder4 got a fail-low and comet first considered >>>>>>about Rxg2 with positive score for black, but now has changed >>>>>>mind and says 0.47 for white when playing Rxg2+. now comet >>>>>>says +0.74 for white. >>>>>> >>>>>>anyway - interesting game too. >>>>> >>>>>I see that CometB27 had a lot of fail low by 0.27 at iteration 11. >>>>> >>>>>I think that the decision about this constant number is not a good decision and >>>>>it is better to increase the number after failing low. >>>>> >>>>>Comet needs a long time to find e3 at depth 11 because of this decision and had >>>>>not enough time at tournament time control when other programs have no problem >>>>>to find e3 at tournament time control. >>>>> >>>>>Maybe Uli can change the 0.27 design decision for the next rounds and to use >>>>>bigger numbers after failing low without solving the fail low problem. >>>>> >>>>>Uri >>>> >>>>If I got this right Uri, you are referring to the aspiration window treatment of >>>>Comet. It uses a very small window, basically of constant width, which will be >>>>just shifted in case of a verification search. >>>>This way, one sometimes needs a lot of verification searches in fail low/high >>>>cases. >>>>I think that this is what you are referring to ? >>> >>>Yes >>>> >>>>However, the advantage is that this method speeds up the search im more quiet >>>>positions. >>>>I think that I will still stick to this - at least for a while. >>> >>>I think that the window should not be of constant width. >>> >>>There is no problem with a small window in the first fail low but I think that >>>after failing low and not finding a better move it is better to increase the >>>width of the window. >>> >>>I think that a rule to double the size of the window after failing low without >>>finding a better move may be better. >>> >>>You can also use the previous changes in the evaluation to decide about the size >>>of the window(if there were big changes then it suggest that the position is >>>tactical and it is better to use a bigger window. >>> >>>In the relevant position >>>Comet has a lot of fail low at depth 11 and not only 1,2 or 3 >>> >>>Here is the relevant position: >>> >>>[D]2r3k1/1q3pp1/2bP2rp/1p6/4pB2/pPR4P/P2Q1PP1/2R3K1 b - - 0 1 >>> >>>Comet needed a lot of iterations to get from +1 for black(at the end of >>>iteration 10) to a better move at depth 11. >>> >>>The scores at depth 11 from white point of view >>>-0.86,-0.59,-0.32,-0.05,+0.22,+0.49,+0.76,+1.03 before finding a better move. >>> >>>I am not sure about the +1.03 but I remember that after the +0.76 it failed low >>>the last time and changed its mind to e3. >>> >>>It is better if it can get the following scores >>> >>>-0.86,-0.59,-0.05,+1.03 so it can changes its mind in 4 iterations instead of 8 >>>iterations. >>> >>>Uri >> >>I 've also had a closer look to this position. IMO, the problem is not the >>search algo here, but Comet's typical underestimate of advanced passed pawns. >>Whereas other programs (e.g. Nimzo) have very soon the advance and promotion of >>Shredder's pawn in their PV, this takes quite long for Comet. >>This is rather the consequence of a bad evaluation or a missing the right >>extensions. >>Uli > >I agree that not seeing the promotion of Shredder's pawn in the pv is one of the >problems of Comet in this position but I think that without a constant window >Comet could be faster in finding e3 and being sligtly faster was enough in the >game because Comet did most of the iterations and played Rxg2+ with a positive >score for white based on thorsten's post. I have tried it also with an older version, using the more traditional approach. It has one fail low which seems to last forever. > >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.