Author: Pham Hong Nguyen
Date: 06:37:32 07/23/02
Go up one level in this thread
On July 23, 2002 at 05:04:56, Bo Persson wrote: >On July 23, 2002 at 02:15:56, Pham Hong Nguyen wrote: > >>>It might also win you a few games. I have seen "short sighted" underpromotions >> >>I guess that in the real matches probability when underpromotions really help to >>win is around 1/million games. The results may be converted to an extra of 0.001 >>elo ;) >> > >If we are to speculate, I can see that promotions could be ocurring more >frequently than we expect in the search. I once had a bug in my unmake code, so >that capture left&right with promotions actually restored *two* pawns (blush). >With silly moves from the opponent, you only need 10 plies from the initial >position to reach such a case. > >My point is that even though promotions (and under-) does not happen that >frequently in actual game play, they are rather frequent in the search, >especially if we favor captures/promotions in the move ordering. The purpose of >the search is not just to find good moves, but also to quickly reject the silly >ones. Ignoring some valid moves, however unlikely in actual play, may delay this >rejection and cause the game tree to grow. Not good! > >For move ordering, a promotion to a queen that is not immediately captured can >be tried first, even before the winning captures. Underpromotions might be put >among the apparently losing captures, or even at the very end of the move list, >as a sort of last chance. Just skipping them does not seem to be such a good >idea to me. > > >Bo Persson >bop2@telia.com I have other point of view. It is very simple: when searching, if underpromotions help only one over a large number of useless cases, say 1/millions (you can have the real proportion for your search function by adding few lines), it should not be implemented. The probability of errors will be so small to ignore (I think it is smaller than errors from qsearch, nullmove, razoring and so on). One of good rules for coding evaluation function is that: try to add general cases and avoid so specific cases. Logically, in search function, we should apply the same rule: avoid so rare cases such as underpromotions.
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.