Author: Ed Schröder
Date: 08:59:02 11/22/99
Go up one level in this thread
On November 22, 1999 at 10:58:25, James Robertson wrote: >On November 22, 1999 at 06:43:08, Ed Schröder wrote: > >>>Posted by Peter McKenzie on November 22, 1999 at 03:11:18: >> >>Hi Peter, >> >>>Greetings, >>>Here are a couple of interesting positions to try on your favourite programs. >>> >>>Positon 1 >>>The first position comes from the game DarkThought-LambChop at the 1999 >WCCC in >>>Paderborn. Lambchop had an advantage but went astray by allowing a perpetual >>>check. Here is the position: >>> >>>5k1r/1R2bp1p/2p1pp2/7B/8/8/P1q3PP/6QK w - - >>> >>>DarkThought, which must have seen the draw several ply earlier, played Rxe7 >>>which forces the draw (after Kxe7 Qa7+ etc). I tried this position on Chop, >>>which needed a whole 11ply to see that Rxe7 draws. I mentioned this >>>position in channel 64 on ICC, and Bob said that Crafty only needs 6ply. >> >>Same here: >> >>00:00:01 6.00 -0.17 1.Qe3 Rg8 2.Rb8+ Kg7 3.Qg3+ Kh6 >> 4.Qe3+ Rg5 5.Bg4 Qxa2 6.Qh3+ Kg7 (1) >> >>00:00:02 6.23 -0.17 1.Rxe7 >>00:00:02 6.23 0.00 1.Rxe7 Kxe7 2.Qa7+ Kd8 3.Qb8+ Kd7 >> 4.Qa7+ Kd6 5.Qd4+ Kc7 (2) >> >>00:00:02 7.00 0.00 1.Rxe7 Kxe7 2.Qa7+ Kd8 3.Qb8+ Kd7 >> 4.Qa7+ Kd6 5.Qd4+ Kc7 (2) >> >>>I asked Bob if he was doing any fancy stuff in search to see perpetuals >>>faster, but he said he wasn't. Oh well, maybe I have a bug I thought. But looking at >>>the position, its obvious the drawing line is longer than 6ply. The black >>>king can go to d6,c7,c8,d7,d8: quite a number of squares to walk around. >> >>Solved in Q-search. >> >>>Then it struck me why Crafty saw it so fast. Anytime the black king got to >>>the back rant, Crafty would see Qa8+ followed by Qxh8 winning a rook for white. >>>Then it would hit the qsearch (or nullmove and hit the qsearch) where black >>>has no good capture. The difference being that LambChop (which looks at >>>checks in the qsearch) would hit the qsearch, and see Qb1+ forces checkmate. >> >>Just extend Q-search one ply more in such cases. >> >>>So here is a position where Crafty very quickly gets the right move and >>>evaluation, but for the 'wrong reasons'. Bob informs me that this is >>>called the 'inverse horizon effect' :-) >>> >>>So how do other programs do on this position? >> >> >>>Position 2 >>>This one happened in a blitz game Ferret-LambChop on ICC: >>> >>>2r1r2k/1p1q1ppp/pn1p1b2/nNp2b2/2PP4/PP2BN2/Q3BPPP/2R1R1K1 w - - 0 18 >>> >>>Ferret played the crushing Nxd6 which works because white has not one but two >>>pawn forks coming up. The main line being: >>>Nxd6 Qxd6 dxc5 Rxc5 Bxc5 Qxc5 b4, a 7ply line. But of course, the normal >>>horizon effect kicks in so LambChop will then think Rxe2 Qxe2 helps before >>>'realising' the pawn fork is still there. >> >>Rebel Century: >> >>00:00:00 4.00 0.63 1.dxc5 axb5 2.cxb6 bxc4 3.bxc4 >>00:00:00 4.09 0.94 1.Nxd6 Qxd6 2.dxc5 Rxc5 3.Bxc5 Qxc5 >> 4.b4 Rxe2 5.Rxe2 (0) >> >>00:00:00 5.00 1.32 1.Nxd6 Nxb3 2.Qxb3 Qxd6 3.dxc5 Rxc5 >> 4.Bxc5 Qxc5 (0) >> >>Some hints: extend (almost) every recapture. On the horizon: extend one >>ply if the side to move has an interesting double attack such as a fork. >> >>Ed > >How do you detect such double attacks without killing speed? The info comes from Rebel's evaluation function. Ed >James > >> >> >>>So Lambchop needs 9ply, and around 2min and 1.2M nodes on my P133 to see >this little combination. A bit slow for my liking! How does your favourite >>>program do?? >> >> >>>In days gone by, I might have seen it faster due to the recapture extension >>>but >>>these days I try to limit that extension much more so it doesn't help here. >>>Other programs might do some static analysis of the pawn fork to see this one >>>faster. I guess you could include that sort of thing in your evaluation, or >>>perhaps base some sort of extension on it. I think GNUChess had some hung >>>piece >>>code to deal with this sort of thing. I'd be interested in hearing any >>>thoughts >>>from fellow programmers on this issue. >>> >>>cheers, >>>Peter
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.