Author: Uri Blass
Date: 15:32:22 07/02/01
Go up one level in this thread
On July 01, 2001 at 22:02:24, leonid wrote: >On July 01, 2001 at 21:14:32, Angrim wrote: > >>On July 01, 2001 at 07:48:43, leonid wrote: >> >>>Hi! >>> >>>I see that your program take well those positions. This last one was slow even >>>for my solver. >>> >>>>proved that move f3xh5 wins, 12 turns >>> >>>I am not sure what is "12 turns". Twelve moves deep? >> >>24 ply, this is the length of the win it found, which is usually not >>the shortest possible. >> >>> >>>>PN2:17137399 evals, 368084 expands, 148.10 seconds >>> >>>Yes, and still what is everage NPS for your program since we have almost >>>identical hardware? I am not sure how to read "evals" and "expands". Mine is >>>Celeron 600Mhz. >>> >> >>evals is the number of calls to the leaf node evaluator. >>expands is the number of leaf nodes which have been expanded to >>become internal nodes. >>an expansion of a node consists of generating and evaluating all >>of the positions which can be reached by a legal move from that node. >> >> >>>I don't know exactly NPS (node/per/second) for this position for shortest move, >>>since mine solved it by selective and went by brute force only as far as to be >>>sure how big is shortest move. For selective, for shortest mate, NPS was 697k >>>and for brute force (just one move below the last) was 93k. My program don't use >>>hash, this must make mine NPS somewhat higher that it should be. >>> >>>If I forget your "expand", your average is 115k. Something that still look like >>>we have close NPS numbers. >>> >>>Leonid. >>> >> >>my nps is quite low for this, because my check handling code is just a >>quick kludge to my usual movegen(check does not exist in suicide chess) >>If I was useing the movegen/makemove from crafty, I would expect >>results about 10x as fast. > >Strange! I could expect NPS for Crafty for this position to be around 400k and >maximum up to 500K. If your numbers can go twice as fast, it could be the second >surprise that I saw with NPS for mates. First was when I found that Rebel have >higher NPS for brute force search that for selective, when it look for mate. I >never, ever could expected this! I think that the only correct comparison is comparison of times to do the same task. Nodes may mean different things for different programs. I try to compare my move generator with Crafty's move generator by comparing the times that both programs need to calculate the perft function. The perft function means the number of legal games with n moves. For example in the initial position perft 1=20 because white has 20 legal moves and perft 2=400 because there are 400 different games of 2 plies. Crafty is clearly faster than my program(I compared it by calculating perft 6 in the initial position and calculating perft 5 in another position) but I have plans to improve the speed of my program(I know that Crafty is not the fastest in this task and chest is faster than Crafty in this task). One way to do it faster at big depthes is by using some hash tables in order to see that 1.e4 d6 2.d4 and 1.d4 d6 2.e4 is the same position and not to calculate both but Crafty does not use it when it calculates perft and I do not plan to use it because it is not testing the speed of my move generator but testing some kind of using hash tables. 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.