Author: Vincent Diepeveen
Date: 18:56:59 12/31/02
Go up one level in this thread
On December 31, 2002 at 17:49:52, Uri Blass wrote: >From Ed programmer stuff >Killer-One [current ply] 110 >Killer-One [current ply-2] 108 >Killer-Two [current ply] 106 >Killer-Two [current ply-2] 104 > >I until today used only > >Killer-One [current ply] >Killer-Two [current ply] From 1995 (1 year after i started diep) the best what worked for me was indexing based upon the actual depth. so not depthleft, but real number of moves done. then the best works: killer[0][realply+2] killer[0][realply] killer[1][realply+2] killer[1][realply] killers more near to the root i used initially too, but nowadays i give a very little bonus for them. they hardly work for me. also there is a big difference between first and second killer. the 'long term' killer is really doing very bad compared to a very active killer. In fact i do not lose quality in move ordering if i leave out long term killers. The static move ordering is much better than the statistical chance that a long term killer is going to work here. >I am interested to know if using 4 killers is a new idea or maybe this idea is >known to be used by other programs. >I did try part of the idea that Ed suggested without clear results. >I changed the order of moves in movei to >Killer-One [current ply] >Killer-One [current ply-2] >Killer-Two [current ply] > >instead of > >Killer-One [current ply] >Killer-Two [current ply] > >I found that it is better only in part of the cases and have not clear results >if it is better or worse than previous order but I had a bug in the >implementation and I checked killer[ply-2] even in cases when ply-2<0. > >It is surprising that the program did not crush and even performed better in >part of the cases. >I still do not use check bound software. >I asked in a previous discussion about checking bounds but I solved the >problem that caused me to ask about it and I also read a claim that if a >varaible is out of bound the program should crush. > >I also looked for a software that will help me under visual C++ but after I >asked to get it for free evaluation and I only got an email that suggest me to >contact them by fax or telephone I did not respond(I responded by email but my >email was blocked for some reason and I decided that the subject is probably not >very important). > >I think now that it may be important because a chess program may even play well >inspite of the fact that it calls killer[-1] so it is possible that I have more >mistakes like that. > >I also see that the number of nodes in small depth also may be the same. > >Uri
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.