Author: Robert Hyatt
Date: 21:34:19 12/07/00
Go up one level in this thread
On December 07, 2000 at 08:14:25, Severi Salminen wrote: >>>You certainly need to clear the history table sometime. What happens when one >>>entry overflows? If you don't ever clear your history table your move ordering >>>is going to get really funky. >>> >>>History is a move ordering heuristic... because a move made 20 turns ago was >>>good has nothing whatsoever to do with that move being good now. Keep the >>>history table between iterative calls to search, sure... but don't keep it >>>between moves, in my opinion. >> >>It is probably a bit better to decrease not clear the history table after a >>move. History from 20 moves ago is certainly not worth much, but from 2 moves >>ago ... yes. >>Most increase counter by depth*depth btw. For me history doesn't help at all > >Thanks for both of you. Now it works smoothly. I must though experiment with >many possibilities (how much to weight, when to weight (only if failhigh or even >if score>alpha), how much to decrease after search...) > >>>A killer move is a great move at this ply. Like promoting a pawn or checking >>>the king. THe idea is that you want to try this move again at the same depth in >>>another branch of the tree earlier (and therefore cutoff earlier in the other >>>branch). I do not make capture moves killers because, as Bob Hyatt said, >>>captures are often not repeatable at different branches of the tree. I make any >>>non-capture that results in a beta cutoff a killer move for its ply. I keep 2 >>>killer moves and order these first before all other non-capture moves. I clear >>>killer moves between top-level calls to Think (iterative search). >>> >> >>Just one more note: if you use 2 killers and get a new one make the 2nd 1st and >>replace the second one. > >Yep, so far I only use 1 killer. I think one could use counters. And replace the >lower one maybe. But then again, is the one with most counts better that the >recent one? Testing to do... > >I also think that one should not clear all killers after search, but to move >them to previous ply level. What do you think? It sounds logical: if a move was >great at ply 5 it should still be great at ply 4 after making a move? > >>>Your numbers for an 8 ply search of that position look sane. > >I had to check the difference between insane and sane in dictionary... :) > >Severi One killer probably is not enough. You will find positions where nearly every move has the same killer, but one or two have a different one. Letting this displace the more popular killer will wreck those sub-trees. 2 works well (this was what slate/atkin used way back in the 1970's... I have been using 2 ever since.
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.