A word on random opening

[xr_a_y]

In the aim of generating training data with/for Minic, I am using random move for n-ply, then starts to play.
If n is odd (meaning that white plays one more random move than black) then black score is better than white. This looks natural.

But if n is even (i've try 6, 8 and 10), white is also loosing it advantage, while white is still the "first" to play a not random move.

Does someone already notice that ?

[Ras]

Does someone already notice that ?

I don't use random moves for openings, but it seems logical: White's advantage is having the initiative, but White needs to actively maintain and foster that, and random moves give it away.

[Laskos]

In the aim of generating training data with/for Minic, I am using random move for n-ply, then starts to play.
If n is odd (meaning that white plays one more random move than black) then black score is better than white. This looks natural.

But if n is even (i've try 6, 8 and 10), white is also loosing it advantage, while white is still the "first" to play a not random move.

Does someone already notice that ?

4 plies random openings still have some 0.20 in classical SF eval median white advantage, which is in line with the expected white advantage. The spread is fairly large, that is what you mean, that often one can find 0.40 or -0.20 openings?

Here is the distribution of 4 plies random openings in SF eval:

[Madeleine Birchfield]

4 plies random openings still have some 0.20 in classical SF eval median white advantage, which is in line with the expected white advantage. The spread is fairly large, that is what you mean, that often one can find 0.40 or -0.20 openings?

Here is the distribution of 4 plies random openings in SF eval:

I would like to see the same test done with the current dev version of Stockfish.

[Laskos]

4 plies random openings still have some 0.20 in classical SF eval median white advantage, which is in line with the expected white advantage. The spread is fairly large, that is what you mean, that often one can find 0.40 or -0.20 openings?

Here is the distribution of 4 plies random openings in SF eval:

I would like to see the same test done with the current dev version of Stockfish.

It won't look very different, maybe a bit compressed. This was 2-3 years ago with the old SF dev at 100ms / positions. The individual values vary quite a bit with time per position used, but the shape of the distribution overall is fairly constant. Median white advantage of some 0.20 is always there and it would probably be just a bit smaller with the current SF dev.

[xr_a_y]

I'm talking in game results not evaluation just after the random moves (but those games were played at fixed depth search).

[Laskos]

I'm talking in game results not evaluation just after the random moves (but those games were played at fixed depth search).

I think SF tests do have a white advantage in results.