Author: Eugene Nalimov
Date: 11:40:44 06/03/99
Go up one level in this thread
The good strategy is to maintain 2 killers, and when the move causes the cutoff, check: if it is killer#1, do nothing; if it is killer#2, swap killers; if it is new one, move killer#1 to killer#2, and that new move to killer#1. Also, if the move is capture of the opponent's piece that just was moved, do not store it as killer, as it's likely that move will not be good in the next position (that means that you have to try those captures first, before considering killers - very reasonable, as in the half of nodes we have to try every move, and after majority of moves you can easily win the material by just capturing moved piece). Eugene On June 03, 1999 at 14:29:23, Andrew Williams wrote: >On June 03, 1999 at 12:59:02, William Bryant wrote: > >>Every time I try to improve my program, I seem to break it in innumerable ways. >> > >Join the club :-) > >>The killer move heuristic should be rather easy to implement. >> >>At every ply, any move that generates a cutoff (I interpret as a beta cutoff, a >>fail high), this move is added to the current killer table. >> > >This is what I do. > >>Is it an alpha cutoff (score > alpha) more appropriate? >> >>Also, does this move automatically displace previous killer moves, or do you use >>some qualifier such as the search score, keeping the moves with the greates >>score? > >What I use is a count for how many times the killer move has been successful >(caused a cutoff). The least successful gets replaced. > >> >>Finally, in using the killer moves for move ordering the ranking I am using is: >>1. Hash Table Move >>2. PV Move (in not the same as #1) >>3. Captures with a net positive score (captured piece > capturing piece) >> Promotions >>4. Killer moves and Captures with an equal score (exchange) >>5. Loosing Captures >>6. All other moves--sorted by the history heuristic. >> >>Any help will be appreciated, I am missing something simple because this is >>slowing down, not speeding up my search. >> > >How many killers are you using? (I find that changing the number of killers >makes quite a difference). Are you checking killers before generating your >moves? This (I think) was the original justification for killer moves; you can >try them before even generating moves at the current ply. If you do this, you >must be careful to ensure that the killer is valid in the current context, as it >may have arisen from a quite different position. > >I hope others will also answer this question, as I have never found killers >to be a particularly big win either. Normally I just use the history heuristic. >At the moment, I'm keeping a couple of killer moves and checking for them >when I generate moves. This isn't in the "spirit" of the killer heuristic, but >it seems marginally better. > > >Andrew Williams
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.