Author: Robert Hyatt
Date: 10:10:25 11/03/98
Go up one level in this thread
On November 03, 1998 at 07:48:50, Ernst A. Heinz wrote: >On November 02, 1998 at 16:27:39, Robert Hyatt wrote: >> >>On November 02, 1998 at 09:49:34, Ernst A. Heinz wrote: >> >>>On November 02, 1998 at 08:24:08, Robert Hyatt wrote: >>>> >>>>On November 02, 1998 at 02:49:14, Koen van Dijken wrote: >>>> >>>>>mr Hyatt, >>>>> >>>>>I was comparing the speed in nps of my own program against that of Crafty. I >>>>>have the sources for 14.2 and 15.20. I noticed that in Search there is a >>>>>difference in node counting between 14.2 and 15.20. In 14.2 you do not count >>>>>illegal moves, in 15.20 you do. Is there a reson for this? >>>>> >>>>>15.20 >>>>> } else tree->nodes_searched++; >>>>> UnMakeMove(tree,ply,tree->current_move[ply],wtm); >>>>> >>>>>Is this maybe the way nodecounting is done in other programs as well, for >better >>>>>comparising? >>>>> >>>>>Koen van Dijken. >>>> >>>> >>>>That was the main idea. However it turned out to have no effect... because now >>>>I increment nodes_searched at the top of search, and I *never* get there unless >>>>the position is "legal"... so for search() it doesn't matter. In quiesce, I >>>>do count illegal positions... but the point is counting "work" really... and if >>>>I do a MakeMove() and so forth, I "searched" that node whether it was legal or >>>>not. >>> >>>We do *not* count illegal positions for our NPS numbers. >>> >>>IMO counting illegal positions is not a good idea because they require so much >>>less work than legal nodes. Actually, there is not much to do beside >>>make/unmake move at illegal nodes. >>> >>>=Ernst= >> >>What about in the quiescence search? Do you cull illegal captures before >>making them? > >Some yes, some not. > >>Or not count them when you notice they are illegal. > >Exactly -- we do not count them although we called "expand_node" and made the >move. This way your NPS number goes up if you generate less illegal moves. > >>I think >>counting or not-counting makes little difference overall since in my case, >>the number is already "mixed" since normal search weeds illegal moves out >>while the q-search does not... > >If you do not count illegal positions you can actually measure the improvement >made by generating less of them (see above). > >BTW, do you roughly know how many of your reported nodes are illegal ones? > >=Ernst= at one time I did count illegal positions when I hit them, and that is what led to the "GenerateCheckEvasions()" procedure I use to generate moves when I start off in check. It only generates legal moves, and *greatly* cut down that illegal position count to the point I stopped counting them. It is small, but I'll try to go back and add a counter to see what they look like. But I'd suspect very few overall...
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.