Author: Richard Pijl
Date: 06:23:41 10/18/04
Go up one level in this thread
On October 18, 2004 at 09:11:54, GeoffW wrote: >Hi Richard > >Thanks for the response > >>Depends on the type of quiescense search you have: Lean and mean (like Crafty), >>or more extensive (like many others). The more precise the score returned from >>your quiescense search, the more errors in it will be bothering you (as you will >>not have the additional plies of regular search). > >I think lean, mean and dense would be an accurate description of my qsearch :-) > > >> >>This looks wrong to me, even with a small qsearch. I assume you have the eval >>score anyway, so better to me looks something like >>evalscore+seescore+margin<alpha >>Margin is needed to compensate for the changes of the evaluation by performing >>the move. >>I'm not doing this in the Baron though. > >Yes looks like it needs a margin adding. The question then is what value is >correct? As I mentioned in my other reply, I seem to need a far too large a >value from initial experiments. Depends on how big a difference one move can make in your evaluation. If that's just a piece-square table evaluation perhaps 0.5 pawns will do, if quite complex with advanced passer evaluation, you may need 5 pawns. The 1.5 pawns as stated by Martin in another response sounds ok though. >Incidentally, The Baron was one of the 8 test programs I was using for my SEE >experiments. Baron was scoring about 70% against my program and beat all the >other 7 programs in my Arena gauntlet test tourney. The Baron uses SEE for ordering all capture/promotion moves in the whole tree. I'm very carefull with pruning based on SEE though. For instance, I only skip moves based on SEE when the see-score is < -1.2 pawns, allowing pawn-sac's in Qsearch. I found raising the minimum score to 0 would reduce playing strength. And no alpha based pruning. So perhaps that makes the Baron a good test engine for tuning your SEE based pruning ... Richard.
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.