Author: Dan Newman
Date: 23:06:29 12/20/01
Go up one level in this thread
On December 20, 2001 at 18:44:48, Ren Wu wrote: >On December 20, 2001 at 18:31:41, Dan Newman wrote: > >>On December 20, 2001 at 16:47:29, Ren Wu wrote: >> >>>On December 20, 2001 at 15:31:45, Dan Newman wrote: >>> >>>>I don't see anything wrong with these numbers. I get 92.4% with nullmoves >>>>and 98.7% without on this position. But my program has hash tables. It >>>>also has a fairly complex (and heavily mistuned) eval. When I run it on >>>>all 300 WAC positions at 1s/posn I get much the same result (92 and 98). >>>> >>>>-Dan. >>> >>>What kind of move you search first in case you don't have the move from >>>hashtable? >>> >>>Ren. >> >>If I'm in the PV, I do an internal iterative deepening search and use >>that move. If not I generate the captures, sort them with a SEE, >>and use the top one of those. >> >>-Dan. >> >>P.S. I tried turning off my hash table and got 90.1% with nullmoves >>and 99% without. The searches took quite a bit longer... > >Thanks. Seems a good SEE is really important factor here. > >See my another post >http://www.icdchess.com/forums/1/message.shtml?202825 > >Can you try enable hashtable, but disable SEE? > >Ren. I turned on the hashtable and turned off the SEE and now get 82% with nullmove, 94.9% without. That last seems a bit high, but I don't see anything wrong with my counting... I guess the hash table (or IID) moves are usually pretty good. When I first tried SEE I compared it to MVV/LVA. The SEE wasn't much better for short searches because it was so much more expensive, but even a small improvement in move ordering gets greatly amplified as the depth goes up. Since it was an improvement I kept it. Then later when I used it to order losing captures after killers and to prune the qsearch, it was even better. These two things you can't really do with MVV/LVA (I tried :)). -Dan.
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.