Author: Miguel A. Ballicora
Date: 12:14:51 01/25/01
Go up one level in this thread
On January 25, 2001 at 13:37:31, Edward Screven wrote: >On January 25, 2001 at 13:12:55, Miguel A. Ballicora wrote: > >>>applying david's suggestion to a null move implementation would >>>mean reducing the search depth after a null move failed high >>>instead of simply returning immediately with a fail high. >> >>That is what I do with in my program Gaviota (not a good example since >>it is still a beginning project). Instead of returning beta I do >> >>depth = depth - R; >> >>I does not reduce the tree as much as classic null-move but still reduces >>enough. The advantage is that it is not fooled as easily by zugswang-type >>positions. A couple of plies more and it sees it. Some of the positions >>that have been posted here that are tough for excellent null-move programs >>Gaviota was ok. >>Probably not as good as null-move since zugswangs are rare but I just like it. > >what do you do when the null-move fails high, then the reduced >depth search doesn't? do you then re-search at the original >non-reduced depth? Tough luck :-), I returned the value obtained after the reduced search. If it is truly inaccurate, it will be corrected in the following iterations. It is actually like a negative extension. It is risky, as any selective method will be. After some thought I conclude it is less risky than the classic null-move. In fact, using null-move in the same situation I would have returned beta, which is terrible. It is better to have a value with a reduced depth than pruning a branch incorrectly. At least, that is the assumption. I did not do any statistics at all about how many times these things might happen, I just implemented it because I like the idea of not to fall for zugswangs. In the future I might test it seriously since I think that zugswangs are rare and can be disable in the endgame. In fact, I have to disable it anyway as it is now because the performance decreases. I might be better to have a plain null-move that might save me more nodes to search but I do not know. >i vaguely remember trying something like this as a kind of null >move test verification step. i took it out, so i must not have >liked it. i'm pretty sure i heard about the idea from someone >else. If it has been used before and you remember, let me know. Regards, Miguel > > - edward
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.