Author: Uri Blass
Date: 15:02:51 10/14/01
Go up one level in this thread
On October 14, 2001 at 14:36:36, Dieter Buerssner wrote: >On October 14, 2001 at 14:15:43, Rafael Andrist wrote: > >>If an underpromotion is clearly worse than the promotion to a queen, this line >>will be pruned anyway (Nullmove+AlphaBeta). Problems are IMO only these cases, >>where the score of the promotions is nearly equal (mostly because the promoted >>piece can be captured, as in example given by Eduard). > >I agree, that this is the most likely case. > >>So it is sufficient to >>include a rule like: if all promotions have (nearly) equal scores, prefer the >>promotion to a queen (or to the piece with the highest value from the ones with >>equal score). > >I think, this won't be sufficient. It is not easy to explain for me. Let me >start again with my hash table argument. I think, we all have seen cases, where >an engine can find the correct solution to a problem at different depthes, when >different hash table sizes are used. Both cases can occure: the misleading hash >info was overwritten, or, luckily, the hash info that leads to the correct >solution was still there. This can make the scores very different at the same >search depth (I don't mean few centipawns). And sooner or later, due to this, we >will see wrong underpromotions again. I do not understand the exact way that the hash works but Here is a possible idea that may help. First thing calculate pronmotion to queen and get a main line for it. If the main line include capturing the queen in the first ply you can prune underpromotions at the root because the opponent can get at least the same by capturing the promoted piece. A possible problem is a case when the queen is captured not in the first ply of the main line(for example a line like 1...h1Q 2.g4+ Kf4 3.Nxh1). In this example you can in the analysis the possibility 1...h1B 2.g4+ Kf4(white can reply Nxh1 and get the same position) and it is possible to define a general rule to generalize this example. I need to think about it more because it still seem that there is a problem here and the main line may change to 1...h1B 2.Nxh1 after pruning. 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.