Author: Gerd Isenberg
Date: 13:13:00 10/26/05
Go up one level in this thread
On October 26, 2005 at 07:01:10, Rick Hagen wrote: > >Hi, > >I am running a 8-pawn tournament and I noticed something that surprised me a >bit. >The set-position is pretty well known I guess: > >[D]4k3/pppppppp/8/8/8/8/PPPPPPPP/4K3 w - - 0 1 > >I did a check with various engines: >First number is from the starting postition, second from 8-pawn position: > > KN/s >Junior 9 1460 1415 >Hiarcs 9 218 430 >Shredder 9 365 620 >CT 15 435 580 >Toga II 1.0 760 1355 >Fritz 7 1000 645 >Fritz 9 930 540 >Crafty 19.15 865 855 >Pro Deo 1.1 735 1100 > >P3 3Ghz >HT 128 MB >TBs off (Not installed atm...) > >So, my question: who are the "odd" ones out, and why? > >Thanks, >Rick A few aspects on your interesting nps-compare: The ratio of the number of expensive nodes versus cheap nodes during a search and the (average) ratio of the implementation depending duration of expensive nodes (full eval, all-nodes) versus cheap nodes (repetition detection, lazy eval cuts at leaves, etc.) might be the main factor to explain the nps behaviour. The ratio of expensive nodes versus cheap nodes is depending on the position and of course on the implementation of the engine. If we use forward pruning to cut futile subtrees, we have relative more expensive nodes. If we, probably depending on the state of the game, don't use futility pruning (but probably lazy eval), we will have relative more cheap nodes. More hashhits, better move-ordering favours expensive nodes and therefor lower nps. Also some extension (and reduction) effects may cause the engine to have more cheap or lazy eval-cuts at the leaves. There are a lot of other things of course, pawn-hashing, eval-hashing, move-generation, incremental verus on the fly attack or move-generation, legal versus pseudolegal movegen and whatever else... Movegen and attack generation is usually faster in endings with no sliding pieces also due to less pieces to generate moves for. More sophisticated search statistics would be interesting. Eg. about the number of fullsearch-, [pre[pre]...]-frontier-, horizon- and qsearch-nodes, differentiated by foreward cuts (hashhit-cuts, nullmove-cuts, (lazy)eval-cuts, etc.) and beta cuts and beta cuts after first move. Gerd
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.