Author: Vasik Rajlich
Date: 09:56:07 08/29/04
Go up one level in this thread
On August 29, 2004 at 12:16:10, Uri Blass wrote: >On August 29, 2004 at 12:09:32, Anthony Cozzie wrote: > >>On August 29, 2004 at 11:13:49, Vasik Rajlich wrote: >> >>>On August 29, 2004 at 09:52:25, Uri Blass wrote: >>> >>>>On August 29, 2004 at 09:37:13, Uri Blass wrote: >>>> >>>>>On August 29, 2004 at 08:58:46, Vasik Rajlich wrote: >>>>> >>>>>>White to play and avoid Rg5. According to my calculations, you'd need a 120-ply >>>>>>search to understand that Rg5 draws. >>>>>> >>>>>>Vas >>>>>> >>>>>>[D] 8/6k1/4p1r1/4Pp1R/5P2/5K2/7P/8 w - - 0 1 >>>>> >>>>>You do not need to see that Rg5 draws in order to avoid Rg5 >>>>>It is enough to find a move that is evaluated as better. >>>>> >>>>> >>>>>Stupid Movei does not know much about endgame and does not use hash for pruning >>>>>but it avoid Rg5 from the first second. >>>>> >>>>>I guess that after I add some knowledge it may have a problem to avoid Rg5 >>>>> >>>>>depth=18 +1.75 h2h4 g7f7 h5g5 g6h6 h4h5 h6h8 f3g3 h8h6 g3h4 h6h8 g5g2 h8h7 g2g3 >>>>>f7e7 g3g6 e7d7 h4g5 >>>>>Nodes: 286711354 NPS: 769922 >>>>>Time: 00:06:12.39 >>>>>depth=18 +1.75 h2h4 g7f7 h5g5 g6h6 h4h5 h6h8 f3g3 h8h6 g3h4 h6h8 g5g2 h8h7 g2g3 >>>>>f7e7 g3g6 e7d7 h4g5 >>>>>Nodes: 315137376 NPS: 780448 >>>>>Time: 00:06:43.79 >>>>>depth=19 +1.77 h2h4 g6h6 h5g5 g7f7 h4h5 h6h8 f3g3 h8d8 h5h6 d8h8 g5g7 f7f8 g7g6 >>>>>f8f7 g6f6 f7e7 g3f2 h8h7 f2e3 e7d7 f6g6 d7e7 >>>>>Nodes: 393358321 NPS: 781884 >>>>>Time: 00:08:23.09 >>>>> >>>>>Uri >>>> >>>>There was a reduction at depth 20 so I was interested to see if movei may find >>>>Rg5 but it was only a temporary reduction >>>> >>>>depth=21 +1.93 h2h4 g6h6 h5g5 g7f7 h4h5 h6h8 f3g3 h8d8 h5h6 d8h8 g5g7 f7f8 g7g6 >>>>f8f7 g6f6 f7e7 g3f3 h8g8 h6h7 g8a8 f6h6 a8h8 h6h2 e7f7 >>>>Nodes: 1287416720 NPS: 795708 >>>>Time: 00:26:57.95 >>>> >>>>Uri >>> >>>True - really the problem should be for the position after Rg5: >>> >>>Black to play and report a drawing score. >>> >>>Vas >>> >>>[D] 8/6k1/4p1r1/4PpR1/5P2/5K2/7P/8 b - - 0 1 >> >> >>Such positions are incredibly hard - if the whole position is moved back one >>rank, white wins: >> >>1. f5xg6 Kg8-g7 2. Kf4-f5 Kg7-h6 3. h3-h4 Kh6-g7 4. h4-h5 Kg7-h6 5. Kf5-f4 >>Kh6-g7 6. Kf4-e4 Kg7-h6 7. Ke4-d5 Kh6-g7 8. Kd5-c6 f6-f5 9. Kc6-d5 Kg7-f6 10. >>Kd5-d4 f5-f4 11. Kd4-e4 Kf6-g7 12. Ke4xf4 Kg7-h6 13. Kf4-e5 Kh6-g7 14. Ke5-d5 >>Kg7-h6 15. Kd5-c6 Kh6-g7 16. Kc6-d7 Kg7-f6 17. h5-h6 Kf6-f5 18. Kd7xe7 >> = MAT28 Depth: 26/37 00:00:07.67 8126kN (1059 KN/s, 0 splits, 0 aborts) >> >>(With the help of a little backwards analysis) >> >>If you want a _really_ embarrassing position, try: >> >>[D]8/1kP5/1P6/6p1/6Pp/5K2/8/8 w - - 0 1 >> >>1. Kf3-f2 Kb7-c8 2. Kf2-g2 Kc8-d7 3. Kg2-f3 Kd7-c8 4. Kf3-f2 Kc8-b7 5. Kf2-f3 >>Kb7-c8 >> = (5.67) Depth: 23/38 00:01:41.43 36269kN (358 KN/s, 0 splits, 0 aborts) >> >>Which zappa scores as almost +6, and is of course drawn. >> >>anthony > >This is not so embarassing and movei only has scores near +2 for white. > >If you want really embarassing position then choose some drawn tablebases drawn >position of KQ vs KP or KBP vs K or KNN vs KP. > >Uri IMO the way to tackle this is to have some concept of progress in the endgame, ie. bring the eval score toward 0 if some conditions are met: 1) Scores are similar at depth X, X+1, X+2, X+3 2) Eval terms are similar 3) Pawns haven't moved 4) Very simple ending (KK, KBKB, maybe KRKR) In some cases, it might be enough to focus the search rather than adjust the eval. Here is an example - h4-h5 will be seen with enough search, so some sort of a "blocked-position" reduction might help. (Rybka needs 23 or 24 ply & 2 billion nodes IIRC) [D] 8/2k4p/1p4p1/1Pp2p2/P1P2P1P/2K3P1/8/8 w - - 0 1 Vas
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.