Author: Gerd Isenberg
Date: 13:06:09 10/31/02
Go up one level in this thread
On October 31, 2002 at 16:00:16, Gerd Isenberg wrote: >On October 30, 2002 at 16:13:36, Gerd Isenberg wrote: > >>Position from Shark - IsiChessMMX, Leiden Round 10 >> >>[D] r4rk1/2n1b1pN/pqb1p1Qp/1p6/P1p3n1/N1P1B3/1PB2PPP/R2R2K1 b - - ; bm Rxf2 >> >>Even if the position is already won by black, finding the spectacular 24...Rxf2 >>seems quite difficult for some engines. Ruffian is clearly the fastest i tried >>so far on my 2.1XP+ : >> >>Ruffian 1.0.0 2 seconds >>Shredder5 14 seconds >>IsiChessMMX 47 seconds >> >>May be a bad testposition due to Nxe3 wins too. >> >>Cheers, >>Gerd > >Hi all, > >Thanks for all your replies. > >IsiChess is even better now - 6 seconds at depth 7. > >A matter of lazy eval threshold - and a design error. > >I hold a recursive variable MaxPositionalScore, member of CNode, which got >initialized at the root and copied from parent to child nodes during the search. > >If MaterialBalance + MaxPositionalScore is less alpha or MaterialBalance - MaxPositionalScore is greater beta > (together with some other >conditions and some depth dependencies), i use lazy eval. > >If the absolute positional score, determined by some interior node eval, >exceeds this value, i used this maximum so far for shallower nodes. > >But in positions like this, if both kings have some trouble, two relative huge >kingsafety penalties compensate each other and gave a relative small positional >sum. So unfortunately at the critical leaves no increase of MaxPositionalScore >occured in the path and lazy eval got triggered with a minimum threshold of only >2.5 Pawn units, which is obviously to less if you sacrificed a queen. > >Maybe the whole idea of dynamic lazy eval threshold sucks due to alpha-beta >dependency, even if i look whether max(whitePos, blackPos) exceeds >MaxPositionalScore. > >Actually i threw it out and use a rather huge constant instead. > >Cheers, >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.