Author: Uri Blass
Date: 11:10:55 10/14/01
Go up one level in this thread
On October 14, 2001 at 11:27:24, Dieter Buerssner wrote: >On October 14, 2001 at 09:01:32, Uri Blass wrote: > >>>Black played 53...h1=B :-)))) >> >>This is not the first time that I see this kind of mistake. >> >>I suspect that it may be better for practical games if programs stop using brute >>force and start to use selective search(not negative extensions but pruning at >>the root). >> >>Bishop or Rook promotions are the first candidates to prune at the root. > >This is most probably true for practical games. But it will make the program >look stupid in some studies. One could make this optional, but it would be error >prone (from a user point of view). > >And, as you mention correctly, negative extensions for this will probably not be >good, and may very likely even be counterproductive (at the root). > >Accidently, I have yesterday observed a similar game, with promotion to rook, >after which Yace could reach a draw, in probably totally lost position. > >Other then the total pruning of such moves, there I can see no easy cure for >this behaviour. (For example: The program will different hash info available >when processing 2 different promotion pieces, and therefor can come to a >different score, that can wrongly prefer an underpromotion). > >Regards, >Dieter I think that you can use the following rule in order to prevent bad underpromotions: If the program wants to play an underpromotion that when you search promotion to queen imagine that the real score of the underpromotion is sligtly lower. It means that the program may change it's mind from +2.7 for promotion to bishop to +2.6 for promotion to queen because the program imagines that the real score of promotion to bishop is 2.5 and not 2.7 when it searches promotion to queen. Usually when underpromotion is needed there is a tactical difference and not positional difference between promotion and underpromotion. 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.