Author: Uri Blass
Date: 14:28:23 09/18/00
Go up one level in this thread
On September 18, 2000 at 16:17:07, Gian-Carlo Pascutto wrote: >On September 18, 2000 at 15:37:47, Uri Blass wrote: > >>It is possible that the program will prune lines like 1.xx yy 2.zz tt 3.ww aa >>4.rr uu 5.ff because of the fact that 5.ff is not a threat of 3 plies when 5.ff >>is an important move because it is a threat of 5 plies and if you search 16 >>plies without pruning you will find that 5.ff is good. > >This is just the same as what happens with standard nullmove. Dynamic >nullmove is even better here, because it is more carefull when searching >smaller depths. > >>>Probably true. However Dan is right about the fact that extensions are >>>gambling. Sometimes they see it, sometimes they don't. Using a large >>>standard search always sees it, save the nullmove/razoring failures, which >>>are really small. >> >>Not extending is also gambling because you can miss a long tactical idea. >>I prefer the gambling of extensions and programmers know that extensions are >>productive. >> >>Chess is a tactical game and I am almost sure that 14 plies with extensions are >>better than 16 plies without them. > >I am not so sure. There is a certain depth above which tactical threats >get less and less important and positional play takes over. I believe 16 >ply to be over that depth. Also, I didn't totally disable extensions, just >reduced them. Normal Crafty also reduces extensions if the search goes >above a certain depth compared to the nominal depth i.e. it will start >getting more carefull extending if it is on ply 32 when the search was >only 16 ply to start with. I think this 32 ply may be a bit too much, >because of the exponential nature of the search. I looked at part of the positions that Dan corbit posted and they are tactical positions. I believe that most programs cannot avoid Nxe6 in a reasonable time in one of the positions because they cannot see deep enough. I know from my correspondence games that tactics is important. Here are some examples that I posted as hard test positions for chess programs: [D]r4rk1/2p3pp/p7/1p1pq3/8/2P2N2/PPQ2KPP/R1B5 w - - 0 1 My opponent blundered and played Kg1 that is losing against Rxf3. Genius3 does only at most 12 plies of extensions so it cannot see that Kg1 is losing and after 48 hours on p100 the score was 0.00 inspite of the fact that the main line after Kg1 was correct. other programs do more extensions so they can find that Kg1 is losing faster. [D]3r1b1k/pp6/2n2q1p/2Bp1B1Q/5P2/1P4p1/P7/4R1K1 w - - 0 1 igor findior surprised my friend with Re6 in a correspondence game and got a draw when other moves are losing. [D]7r/1k2P2P/4p3/1pp5/8/2pp4/1n3PP1/R4K1R w - - This position could happen in my correspondence game. programs like the move e8Q and need a long time to find that g4 is winning(again programs have problem with tactics) [D]r1bqr1k1/pp1p1ppp/5n2/1N2n3/2P1PB2/6P1/PP1Q1K1P/4RB1R b Again this position could happen in my game. The right move is d6. I am interested how much time does crafty need to find the right moves with the normal extensions and with your limited extensions. > >Almost all programs extensions are tuned to a certain 'expected' ply >depth. If hardware continues to improve, I'd be not surprised if people >started making programs with less, or smarter extensions. I believe this >is necessarry because extensions don't scale well by nature. I agree that there is a need for smarter extensions but it means also more extensions in some cases. The programs that I know cannot find the draw in game 2 of the match Deeper blue-Kasparov because they do not use these smarter extensions. A team of human+computer have no problem to find the draw in few hours because they use smarter extensions. 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.