Author: Gerd Isenberg
Date: 11:12:47 04/14/04
Go up one level in this thread
On April 14, 2004 at 12:38:25, Volker Böhm wrote: >Hi, > >we (Ralf Schäfer and me) are currently developing a new engine from the scratch >and we have currently implemented the search. >We desided to use 0x88 and attack-tables. The attack-tables are build up >incrementally, also in q-search. Here is why: > >1. We are nearly three times faster in implementing attack-tables incrementally >than fully, even with a undo feature. >2. Lots of positions are positions with king in check because of check >extensions. With attack tables you can implement a move generation with allmost >no illegal moves with much less costs than a full move generation. >3. The legality of a position can be calculated at no cost. >4. We use attack tables for king security. As king security will produce high >positional values it would´nt be a candidate for lazy eval. >5. We use attack tables for mate in 1 detection at every ply >6. We use attack tables for a simple SEE (not yet implemented but an easy table >lookup). > >On the other hand, attack table generation is really expensive, even if it is >done incrementally. Without any eval (just material) we are just as fast as >ruffian. > >Greetings Volker Hi Volker, yes, incremental updating attack tables is of course interesting stuff. I did that too in my former DOS-Program. With such stuff one should IMHO do legal move generation, a huge eval, sophisticated move ordering - and probably to avoid lazy eval cuts in qsearch by more sophisticated futility or SEE pruning. That makes the update cost relative smaller, but your nps drops a bit. 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.