Author: Omid David Tabibi
Date: 13:05:34 08/21/03
Go up one level in this thread
On August 21, 2003 at 15:20:40, Robert Hyatt wrote: >On August 21, 2003 at 10:29:15, Omid David Tabibi wrote: > >>On August 21, 2003 at 10:10:54, Mihaly Szalai wrote: >> >>>[D]8/8/8/4p1p1/6N1/5p1p/5K2/7k w - - 0 1 >>> >>>All the programs I've tried see the mate >>>except Junior 7 and 8. They go to depth=63 >>>and score=0.00 in a few seconds then stop. >>>Why? One knight is not enough to mate? >> >>Interesting case. I guess when one side is left with a knight or a bishop, >>Junior sets beta = 0, i.e., that he cannot achieve more than a draw. It is a >>kind of tradeoff: in many positions this knowledge saves a lot of search (by >>pruning the tree), but in very rare examples like the above, it produces >>erroneous results. >> >>I believe that the risk is negligible in comparison to the considerable gain. >>However, as a matter of principle I'm against such "assumptions"... > > >There are ways to make it "safe". > >IE do it in the eval, so that if you have only a bishop, you say "that side >can't win" and you limit the score to no more (or less depending on which side >has the lone piece) than zero. I assume that Junior's main gain by using this assumption is its pruning functionality. Applying it in evals only will correct the problem, but will also take away most of the profit. > >That way the search will find the mate _before_ the evaluation gets a chance >to say "draw". If you do it via "interior node recognizers" then you have to >handle it directly or you lop the search tree off at that point and say draw >when it is not.
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.