Author: Vasik Rajlich
Date: 04:10:46 04/05/04
Go up one level in this thread
On April 04, 2004 at 11:13:18, Aivaras Juzvikas wrote: >On April 04, 2004 at 11:09:05, Vasik Rajlich wrote: > >>On April 04, 2004 at 11:03:07, Vasik Rajlich wrote: >> >>>On April 04, 2004 at 10:55:38, Aivaras Juzvikas wrote: >>> >>>>On April 04, 2004 at 10:48:55, Vasik Rajlich wrote: >>>> >>>>>On April 03, 2004 at 06:29:13, Aivaras Juzvikas wrote: >>>>> >>>>>>On April 03, 2004 at 06:21:51, Vasik Rajlich wrote: >>>>>> >>>>>>>On April 03, 2004 at 04:02:57, Aivaras Juzvikas wrote: >>>>>>> >>>>>>>>i have got this problem here, check extention explodes in the endgame pawns + >>>>>>>>rook or queen, all these open files make it possible to check endlessly. how to >>>>>>>>restrict this? i added smth like this: if ply > depthtoreach + 4 then dont >>>>>>>>extend, where depthtoreach is the depth im searching to. but this is ofcourse >>>>>>>>not very good, because i still get those check extentions where its impossible >>>>>>>>to mate anyway. how did u solve this problem? an idea would be to store info >>>>>>>>into transposition table when not to extend, but how to find that out? >>>>>>> >>>>>>>Are you extending checks, check evasions, or both? >>>>>>> >>>>>>>The third option is quite appealing, but there are some issues. :-) >>>>>>> >>>>>>>Cheers, >>>>>>>Vas >>>>>> >>>>>>when side to move in in check i extend, so i guess it would be a check evasion >>>>> >>>>>So, in certain positions, you are extending every other move. It's true that in >>>>>these positions this will severely hurt your nominal depth, but if you test with >>>>>the extension and without it you should see a clear benefit. I don't remember >>>>>any claims to the contrary. >>>>> >>>>>Vas >>>> >>>>well here is a claim to the contrary: in endgame rook+pawns vs rook+pawns >>>>instead of dealing with passed pawns the engine keeps on extending checks >>>>endlessly thus overall looking very few ply ahead and the result is a loss of a >>>>rook due to pawn promotion. usually (w/o endless check extention stuff) i reach >>>>15-17 ply depth in endgame, with check ext. stuff i get 7-8ply, and again only >>>>when slider like rook or queen is present which can keep on checking no matter >>>>what you do. dont u restrict your extentions somehow? please share with me how >>>>to do this >>> >>>Aha. That's a good point, maybe it's worth testing. >>> >>>My program Rybka always extends every check one full ply. >>> >>>However, maybe it should be a bit less in the endgame, and perhaps even a bit >>>more in the middlegame. (This I did experiment with.) >>> >>>This is worth some tests. >>> >>>It would be interesting if you posted some numbers. (You can always find some >>>good cases, and some bad cases.) >>> >>>Vas >> >>One more thought about this: in the endgame, it may be a question of calibrating >>against pawn push extensions. (Ie rather than decrease the check extension, >>increase the pawn push extension.) >> >>A few months ago I played around with some hyper-aggressive passed pawn >>extensions, up to 3 plies extra for any pawn which reaches the seventh rank in >>an endgame. >> >>It worked nicely on some tactical problems, but there were also some exploding >>searches. >> >>Cheers, >>Vas > >my engine is brand new and instead of many features it has many bugs -> the only >extention i have now is this check (evasion i guess) extention. so meybe instead >of restricting this extention i should add pawn on 7th rank extention and see >how it does. what else is there in endgames to deal with besides mates and >promotions?is there any other must-have extention i am missing? I guess that pawn push extensions are pretty standard. Aside from that, everyone has their own little tricks. For example, in endgames, I reduce moves which capture an advanced passed pawn, and I reduce promotions. I'm not 100% sure that it's a good idea, though. It's very difficult to get good feedback on obscure situations, because any test you devise could be biased in some way. Cheers, 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.