Author: Robert Hyatt
Date: 09:30:57 04/02/04
Go up one level in this thread
On April 01, 2004 at 16:18:55, Artem Pyatakov wrote: >Hi Robert, > >Thanks for replying as usual. > >>That is how everyone does killers I suspect. It is exactly how Slate described >>it in Chess Skill in Man and Machine... If you only use 2, the counters can >>probably be eliminated (I did this a long while back). > >I don't know if that's what everyone does - I bet a lot of people have adopted >the approach used by Crafty. Also, according to my test and common sense, the >counters do make a difference even for 2 moves, because as opposed to blindly >replacing one of the moves, the program will actually replace the less >frequently used one. Of course the difference between approaches becomes much >more significant when you go to a larger number of killers (and a lot of people >do use 4, so it becomes significant). Note I dont "blindly replace". I keep the list ordered... The most recently used killer is #1, next most recently used is #2. #2 gets replaced... > >>What is the overhead for keeping the extra killers, updating the counts, finding >>the best killer, etc? IE it isn't free. Hopefully it doesn't cost 10% which >>would make this "break-even"... > >Obviously the answer would vary from program-to-program, but I think the answer >is clearly no - there is no way that this operation takes 10% of the total >program time! Even with the most naive of implementations, we are talking about >an extra 10-iteration/10-comparison loop, and a tiny memory hit (10 moves versus >2, plus a couple bytes for the score = int per move -> 8 new moves+scores*4bytes >each=32 bytes/ply!) There is also an equivalent for every cutoff that occurs, >and that also has at most 10 comparisons. I think this is a tiny price to pay >for 10% less nodes. You are overlooking some important details. I use killers before generating moves. I therefore have to do a legality check on each killer. It isn't free. etc... If you generate moves first, you miss an important savings of avoiding the move generator overhead when a killer works... > >Artem
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.