Author: Ed Schröder
Date: 10:47:25 07/22/00
Go up one level in this thread
On July 22, 2000 at 11:23:50, leonid wrote: >On July 22, 2000 at 09:35:12, Robert Hyatt wrote: > >>On July 22, 2000 at 08:46:25, leonid wrote: >> >>>Hello! >>> >>>Is the nodes per second number tend to grow when program search position more >>>advanced in the game? When there are less pieces on the board. >>> >>>I just speeded my program in around 30% and tried to find where my nps is now. >>>Went to the Fritz 6 counter (Fritz engin inside of Hiarcs package) and was >>>disappointed. Fritz still keep me at the safe distance at the end of the game. >>>This is due to the strange (for me) fact that Fritz numbers goes up at the end >>>of the game. This time in mine nps goes down. In mine program less are the >>>pieces on the board, less efficency in the search of the legal moves. Wanted to >>>see if there are some general tendency in all programs. Before asking here my >>>question, went to see Rebel 10. Its numbers looks like to be very stable. No big >>>difference between heavy initial and final positions. Once again difference with >>>my program. >>> >>>Thanks for response, >>>Leonid. >> >> >>This is way too hard to answer directly, because it varies by program and >>by approach. My NPS doesn't change much, because of the way my move generator >>works. I have noticed that Ferret is just the opposite and speeds up >>significantly in the endgame. >> >>The question is, as the game progresses, how much work do you save when a piece >>comes off the board? IE it might be almost as expensive to ask "is there a >>white bishop to generate moves for?" as it is to actually generate those >>moves. In such a case, you won't speed up very much in endgames... > >Probably your response is right. I save practically nothing at this stage with >exception of best moves found for each ply (in my program each ply have specific >procedure). Probably hash tables (have nothing like this right now), or >something of this nature make all this strange difference. But once again, why >your program goes like Rebel? > >Thanks! >Leonid. NPS in Rebel in general is quite equal between mid-game and end-game. If the end-game becomes domain-specific NPS will raise sometimes over a factor 2.5. My guess is I (maybe) have a more overhead during search than others (info needed for the selective part) plus the fact the amount of knowledge for the end-game is more then for the mid-game. I think this behavior differs a lot per program. I also think that NPS is a bad indicator for the speed of the program. When I set Rebel to brute-force NPS goes up with 25% average. When I lower the [Chess Knowledge] parameter NPS goes up, when I set this parameter to its maximum NPS may drop with a factor of 2. It all depends how you count and that in combination with all the tricks you use to speed the thing. Meaning to say, don't bother too much about it. Ed
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.