Nullmove vs classic selective search
Posted: Sat Aug 04, 2012 10:56 pm
The last months I have been working on my old "static selective search" system, found improvements (tuning basically) and now it outperforms (full) nullmove so convincingly (61%) I would like to bring it under your attention. I think that for those who like a REAL challenge there must be some good elo improvement as a reward for perseverance.
In short, the search is split into 2 parts
1. Nullmove part
2. Static selective search part.
Typically a 10 ply search (for the middle game) is split into 3 plies nullmove and 7 plies selective search, a 12 ply search 4 plies nullmove and 8 plies selective search, etc. The system is now fully parameter driven for easy tuning. Setting all the values to its maximum will result in a normal modern nullmove searcher.
The nullmove part is used to as a safety net to catch the most important failures of the selective search. Static selective search is risky business but it can be done, I think that especially Richard Lang has proven that by 10 (or so) consecutive world titles.
The static selective search decides either to search the current node at full depth, make a reduction or even prune merciless. The system works independently of futility pruning and LMR, they happily coexist.
I already described the inner workings in 2004 and although much of the stuff has been changed the basic principles still stand as a rock and are explained, I just don't know if I have explained it well enough.
Take a deep breath first.
http://www.top-5000.nl/authors/rebel/ch ... #SELECTIVE SEARCH
In short, the search is split into 2 parts
1. Nullmove part
2. Static selective search part.
Typically a 10 ply search (for the middle game) is split into 3 plies nullmove and 7 plies selective search, a 12 ply search 4 plies nullmove and 8 plies selective search, etc. The system is now fully parameter driven for easy tuning. Setting all the values to its maximum will result in a normal modern nullmove searcher.
The nullmove part is used to as a safety net to catch the most important failures of the selective search. Static selective search is risky business but it can be done, I think that especially Richard Lang has proven that by 10 (or so) consecutive world titles.
The static selective search decides either to search the current node at full depth, make a reduction or even prune merciless. The system works independently of futility pruning and LMR, they happily coexist.
I already described the inner workings in 2004 and although much of the stuff has been changed the basic principles still stand as a rock and are explained, I just don't know if I have explained it well enough.
Take a deep breath first.
http://www.top-5000.nl/authors/rebel/ch ... #SELECTIVE SEARCH