Author: Pat King
Date: 08:31:57 11/30/01
Go up one level in this thread
On November 29, 2001 at 21:11:32, Robert Hyatt wrote: >On November 29, 2001 at 19:51:35, David Hanley wrote: [snip] >>leaf i can just adjust it at each capture/promotion. That seems to work pretty >>well, and i figured that an average on one op per inside node would be a lot >>cheaper than 64 per leaf. >> >>I realized later that many leaves will be reached with material balances far >>outside alpha & beta. So, my thinking was to write code much like the >>following: >> >>if ( material < ( alpha - (2*pawn))) >> return alpha - 1; >>if ( material > ( beta + (2*pawn))) >> return beta + 1; >> [snip] > > >It is a rudimentary approach to lazy eval, plus quiescence futility pruning... > >It definitely is sound, but watch the "window" as you improve your eval. The >window will have to widen. Since the PV score will include the entire eval, I reasoned that it was not terribly important to fully evaluate non-PV branches, and that therefore the window could be quite narrow. Even if the positional score might nudge a given node into the A-B window, it should not affect the PV (assuming you've got the right PV). I squeezed my window down to a 1/4 pawn without apparent ill effect. Admittedly, my eval isn't terribly sophisticated yet, but it sure varies more than 1/4 pawn. Is there some kind of test that can detect "too small" other than test positions, actual play?
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.