Final results:bob wrote:Ok, The jury is back, and the verdict is "don't do this." It is only worth about 2-3 Elo to not do back-to-back nulls, and the reason seems to be efficiency. After the cluster test showed the idea to be slightly worse, I ran some test positions with some additional statistics gathered. A null after a null does not fail high very often. And when it doesn't, it is a total waste of time. If you disallow a null right after a null, you gain just a bit of zugzwang protection, but you lose a little more in efficiency due to the extra null-move searches that fail low and don't refute the previous null-move.
I'm sticking with the current approach. I will test both of these versions by adding verification, for completeness, but it seems like a very slightly worse way of doing null-move searches, although it is not very significant overall. But as I have often said, give away a couple of Elo here, a couple there, and soon you are working on a patzer.R08 is the version with double-nulls allowed, and nulls can be done even in pawn-only endings. It always looks a bit better until the 15-20,000 game point where it has dropped back to about the same.Code: Select all
Crafty-23.2R08 2653 6 6 6974 61% 2569 23% Crafty-23.2R01-1 2651 3 3 30000 61% 2570 23% Crafty-23.2R07-1 2648 3 3 30000 60% 2570 23%
Note that R01 does win one more game out of every 100 than R07 (the double-null version).. I'll post the complete results in a couple of hours when I get out of class... Test should have finished by then.
Code: Select all
Crafty-23.2R01-1 2645 3 3 30000 61% 2564 23%
Crafty-23.2R08-1 2644 3 3 30000 60% 2564 23%
Crafty-23.2R07-1 2643 3 3 30000 60% 2564 23%