Null move Pruning

[DaOnlyOwner]

Hello,
I have developed an engine and I recently implemented null move pruning.

However I don't understand why we pass - beta + 1 as an argument in the function call (like this):

Code: Select all

-alphabeta(depth-1-R,-beta,-beta+1)

Why not how it's always done with -alphabeta(depth-1-R,-beta,-alpha)?

[Ras]

However I don't understand why we pass - beta + 1 as an argument in the function call (like this):

Code: Select all

-alphabeta(depth-1-R,-beta,-beta+1)

Why not how it's always done with -alphabeta(depth-1-R,-beta,-alpha)?

Because with null move, your alpha is (beta-1), hence -alpha = -(beta-1) = -beta+1. You're trying to see whether you get a fail-high, i.e. whether you get a score of at least beta returned. Anything below beta would not be a fail-high, so you make the window as small as possible, that's why you choose alpha as beta-1.

[hgm]

If the null move doesn't score above beta, you won't use its score, and consequently you don't care wheter is scores above or below alpha. Requesting the search to determine the exact score, just to discard it, is a waste of time.

[Henk]

null window searches may pollute killer administration? Or better bad killer moves than no killer moves

[Desperado]

Hello,
I have developed an engine and I recently implemented null move pruning.

However I don't understand why we pass - beta + 1 as an argument in the function call (like this):

Code: Select all

-alphabeta(depth-1-R,-beta,-beta+1)

Why not how it's always done with -alphabeta(depth-1-R,-beta,-alpha)?

Another point to the already mentioned ones is, that most people do not use nullmove at a pv node.
In PVS that means you only do nullmove when the window is already closed, which then, is equivalent to "-beta,-alpha".

Finally you want to test against a bound and normally you test against beta.
But you do not need to do this. If you decide to test against beta +- margin you can do that too.

You should think of what the idea behind nullmove is. The idea is to check for a serious threat or
more general you want to know if you have a move that will refute the predecessor (e.g. by beating beta).
The assumtion then is, that if the nullmove move will do it, a real move will do it too. (if we exclude Zugzwang situations).

"Testing" for a bound in a search means that you will always use a nullwindow (bound-1,bound) because you are only interested if the value is >= bound.

[Henk]

I tried to use null moves to test if there were threats on both sides. If no threats than stop searching. But too slow I guess. So I now believe null moves are too slow for using extra threat detection.

[DaOnlyOwner]

Thanks guys, I think I finally got it.