Fruit reloaded seems to benefit from stockfish null move

[Dann Corbit]

Code: Select all

   Program                          Elo    +   -   Games   Score   Av.Op.  Draws

 1 Fruit-reloaded-sf-null-pruning : 3175   31  30   788    83.0 %   2900    5.6 %
 2 Fruit-reloaded                 : 2917   25  25   788    34.5 %   3029    5.7 %
 3 Fruit-reloaded-mingw           : 2907   25  25   788    32.6 %   3034    6.0 %

Current standings.
I suspect that I have some configuration problem if others cannot reproduce the result.

[Laskos]

Code: Select all

   Program                          Elo    +   -   Games   Score   Av.Op.  Draws

 1 Fruit-reloaded-sf-null-pruning : 3175   31  30   788    83.0 %   2900    5.6 %
 2 Fruit-reloaded                 : 2917   25  25   788    34.5 %   3029    5.7 %
 3 Fruit-reloaded-mingw           : 2907   25  25   788    32.6 %   3034    6.0 %

Current standings.
I suspect that I have some configuration problem if others cannot reproduce the result.

Did you tune for Fruit the expression of R, or just left it untuned from Stockfish? I guess this highly tuned for Stockfish formula, left untouched, will only weaken Fruit. Even a tuned one for Fruit will not give anything like 500 or 250 ELO points.

[Dann Corbit]

Code: Select all

   Program                          Elo    +   -   Games   Score   Av.Op.  Draws

 1 Fruit-reloaded-sf-null-pruning : 3175   31  30   788    83.0 %   2900    5.6 %
 2 Fruit-reloaded                 : 2917   25  25   788    34.5 %   3029    5.7 %
 3 Fruit-reloaded-mingw           : 2907   25  25   788    32.6 %   3034    6.0 %

Current standings.
I suspect that I have some configuration problem if others cannot reproduce the result.

Did you tune for Fruit the expression of R, or just left it untuned from Stockfish? I guess this highly tuned for Stockfish formula, left untouched, will only weaken Fruit. Even a tuned one for Fruit will not give anything like 500 or 250 ELO points.

Simple cut and paste.
I was expecting 10-25 Elo.

[supersharp77]

Dann Corbit

Current standings.
I suspect that I have some configuration problem if others cannot reproduce the result.

--------------------------------------------------------------------------------
Great work Dann C. (A little bit of "something" is better than a whole lot of "nothing")...............Thx AR

[Steve Maughan]

I implemented a similar version in Maverick and it looks as if at lightning chess (game in 10 sec + 0.02 sec) new vs. old is 54% vs. 46%. This is about 30 ELO - nice!

I suspect it may give better results at longer time controls where the depth is higher (so reduction is bigger).

I'll try to tune it for better results.

Steve

[bob]

Code: Select all

   Program                          Elo    +   -   Games   Score   Av.Op.  Draws

 1 Fruit-reloaded-sf-null-pruning : 3175   31  30   788    83.0 %   2900    5.6 %
 2 Fruit-reloaded                 : 2917   25  25   788    34.5 %   3029    5.7 %
 3 Fruit-reloaded-mingw           : 2907   25  25   788    32.6 %   3034    6.0 %

Current standings.
I suspect that I have some configuration problem if others cannot reproduce the result.

Has to be something broken. I restored the old adaptive null-move code from an old version of Crafty, one which is significantly less aggressive than Stockfish however, and it MIGHT have added 10-15 Elo over the straight R=3 version.

Using the same exact formula I used back in the 90's when this first came along, I get the following R values in a game using 90 seconds for this move:

null-move (R): 3/75.0M 4/6.7M 5/210.7K 6/7.4K 7/66

In this search, R was 3 in 75M NM searches, 4 in 6.7M and so forth. For this data, the nominal search depth was 28 plies, and without fractional plies, the numbers could be anywhere between 3 and 9, but
apparently > 7 was never hit for whatever reason, perhaps LMR as much as anything.

The formula is very simple and I have not yet tried anything more complex since I have completely rewritten the parallel search and have been tuning that for a while... Here's my simple one:

int R = null_depth + depth / null_divisor;

Currently both parameters are tunable, with null_depth being set to 3 as always (back when it was 2 of course) and null_divisor is currently 4, where it was 6 back in the 90's as well. However, except for unusual searches back then, the usual values were 2 to 3, but with fractional plies (back then, not now) it could vary in that range fractionally as well.