Author: Robert Hyatt
Date: 19:15:59 11/10/99
Go up one level in this thread
On November 09, 1999 at 19:41:02, Dann Corbit wrote: >On November 09, 1999 at 10:51:00, Ed Schröder wrote: >[snip] >>It's indeed more complex than that. Set [Chessknowledge = 500] and NPS >>will go down, set [Selective Search = 001] and NPS will drop even more. >> >>Changing one parameter in Rebel (which isn't available for the user) and >>NPS will go up with a factor 3-4 which means Rebel would go over 1,000,000 >>NPS on a fast PC. >> >>The bottom line is that NPS (like ply-depth) is pretty meaningless. >I think that ply depth is an excellent indicator of understanding of the >position unless the program has bugs. That is to say, if one program finds a >best move at ply 10, so will the other most of the time. Certainly, if we have >a complete search (no special pruning) and two programs complete the same ply, >they will have very similar answers most of the time (given that the eval is >half-decent or better). > >If one program gets two piles deeper all the time, I think it will even beat a >program with a somewhat better eval. This isn't always true. IE a null-move program using R=2 will generally search 2 plies deeper than a program that doesn't use null-move. And it will generally take 2 more plies to see some tactical shots that the R=2 searches 'hide'. And then there are programs (Deep Blue comes to mind) that not only doesn't use null move at all, but they do extensions that the rest of us don't consider because of the expense. IE they reported on "singular extensions". I have now seen Hsu mention a new form of this they used, which basically says "If one move is clearly the best, extend a bunch, if two moves are clearly better than the rest, extend, but not so much..." So their 8 ply search likely will see things that everybody else's 12-14 ply searches won't see. Of course, their 8 ply search will have 1000X the nodes of our 12-14 ply searches too. So the comparison is not 'clean'. My favorite is to pick some tactical positions and simply compare the time-to- solution for each. Not depth of solution, not nps while searching the solution, just plain time...
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.