Author: Robert Hyatt
Date: 19:31:08 09/11/99
Go up one level in this thread
On September 11, 1999 at 19:57:09, Dave Gomboc wrote: >On September 11, 1999 at 18:01:30, Alessandro Damiani wrote: > >>On September 11, 1999 at 17:58:59, James Robertson wrote: >> >>>On September 11, 1999 at 17:54:55, Alessandro Damiani wrote: >>> >>>>On September 11, 1999 at 15:56:10, Ed Schröder wrote: >>>> >>>>>On September 11, 1999 at 15:42:03, Robert Hyatt wrote: >>>>> >>>>>>On September 11, 1999 at 11:42:29, Ed Schröder wrote: >>>>>> >>>>>>>>Posted by Robert Hyatt on September 11, 1999 at 10:19:19: >>>>>>>> >>>>>>>>In Reply to: Re: Interesting mate test for hashing posted by Ed Schröder on >>>>>>>>September 11, 1999 at 01:43:12: >>>>>>>> >>>>>>>>On September 11, 1999 at 01:43:12, Ed Schröder wrote: >>>>>>>> >>>>>>>>> >>>>>>>>>Do not underestimate the idea that in case there is no bestmove from the >>>>>>>>>hash table you do a full static evaluation of all nodes first and based >>>>>>>>>on that you pick the bestmove as being the first move you are going to >>>>>>>>>search for this (new) depth. The very early Rebel's (1981) worked that >>>>>>>>>way and I remember (although the system is very time consuming) it was >>>>>>>>>superior to all other systems I tried. >>>>>>>>> >>>>>>>> >>>>>>>>I'm not underestimating it. I was simply saying that this approach can >>>>>>>>be applied when the position is encountered and there is no 'best move' >>>>>>>>in the hash table. Rather than doing it when the hash entry is stored, >>>>>>>>and we are not even sure that this hash entry will ever be used again or >>>>>>>>that it won't be overwritten before it is needed. >>>>>>> >>>>>>>Right. >>>>>>> >>>>>>>>>I later removed the system because hash tables + bestmove was more powerful >>>>>>>>>at least for Rebel. But I wouldn't exclude the possibility such a system >>>>>>>>>can have a positive effect on the speed of the search. >>>>>>>>> >>>>>>>>>Actually I didn't remove the system but I replaced it with a faster one >>>>>>>>>that is: >>>>>>>>> >>>>>>>>>- generate all legal moves; >>>>>>>>>- for all moves do a (very) quick evaluation; >>>>>>>>>- sort all moves based on the quick evaluation. >>>>>>>>> >>>>>>>>>This (move ordering) system (for Rebel) is still superior. >>>>>>>>> >>>>>>>>>Ed >>>>>>>> >>>>>>>> >>>>>>>>Do you use killers, history, etc? >>>>>>> >>>>>>>Just the normal stuff. >>>>>>> >>>>>>>Order... >>>>>>> >>>>>>>- hash table move >>>>>>>- winning captures (ordered by expected material gain) >>>>>>>- promotion >>>>>>>- equal captures (QxQ etc) >>>>>>>- killers (4 of them) >>>>>>>- remaining moves ordered by the intelligent move generator >>>>>>> >>>>>>>The Killer History from Jonathan Schaefer gave no improvement for me. >>>>>>> >>>>>>>Ed >>>>>> >>>>>> >>>>>>It probably won't help if you keep 4 killers. I didn't find any improvement >>>>>>in Cray Blitz either... but I did even more with killers. I tried the current >>>>>>ply killers then the killers from _other_ plies if they were legal... adding >>>>>>history did nothing for me... I probably ought to re-check Crafty again as it >>>>>>might be extra overhead for nothing now... >>>>> >>>>>I use 2 killers from the current ply (that's the normal way) and the 2 from >>>>>2 plies back. The latter gave me 5%. >>>>> >>>>>Ed >>>> >>>>I am surprised: history doesn't help? I think that the static ordering is good >>>>enough then. Perhaps history is good for those with a bad static ordering, like >>>>me? >>>> >>>>Alessandro >>> >>>History didn't help me at all either. I just ended up with a lot lower NPS.... >>> >>>James >> >>Am I the last one with history heuristic here? :-) >> >>Alessandro > >I doubt it, but it has to be said that KH (killer heuristic) and HH (history >heuristic) are competing with each other. Having both may easily be slower than >just one of them. > >Dave Killer is basically a proper subset of the history heuristic. When I added it the tree size didn't change.. but the reason I use them is I can try them _before_ generating any moves. And if they cause a cutoff, the avoid the move generation at that ply completely. As a result, they save me 5-10% of total search time, although the size of the tree doesn't go up or down when they are removed/added...
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.