Author: Robert Hyatt
Date: 20:02:48 12/03/02
Go up one level in this thread
On December 03, 2002 at 19:16:39, Pat King wrote: >On December 03, 2002 at 17:49:01, Robert Hyatt wrote: > >>On December 03, 2002 at 13:09:42, Pat King wrote: >> >>>On December 03, 2002 at 11:29:56, Robert Hyatt wrote: >>> >[snip] >>>Perhaps there's a middle ground here. You can sort the scores of the fail-lows >>>at the root, cost O(B^2) at worst, and then probe the hash table for the >>>corresponding variations. You'd know the PV was worth (say) exactly 314, >>>variation 1 < 271, variation 2 < 141, etc. As a practical matter, how useful are >>>those less-than scores? >>>[snip] >>> >>>Pat >> >> >>How can you do that? You don't know _exact_ scores for the moves that failed >>low. You >>only know an upper bound on each, as the real scores for each of those moves >>could be even >>worse had the entire sub-tree been searched. >> >I understand that. I can see a user, however, being interested in knowing that >what he views as the best move is at least (say) half a pawn worse, or that >several moves may have scores nearly equal to the PV's. What could be the killer >is thinking that sorting on those upper bounds means anything. All you know is this: score(best) = X Score of any other root move = X - rnd(constant). Why rnd()? Because all we do is prove that each move is worst than the best, and the "stopping point" for each of those searches is simply the _first_ value that is <= X. Displaying those scores would be so terribly misleading that it would be a gross idea to purport that they are anything but random offsets downward from the best move score. > >>Fail-low scores don't contain any information that can be compared between them, > >So your position would be that sorting them, and presenting (say) the "top >five", would be invalid. "misleading" is the word that comes to mind, because what do you get if you sort random numbers??? > >>all >>they are useful for is to compare against the best move... and we already know >>and have >>proven they are worse, just not by how much. > >We've proven at LEAST how much. That could be useful while doing interactive >analysis. What seems dodgy to me, perhaps, is assigning any merit to the >resulting order. That I would agree with....
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.