tuning for the uninformed

[jdart]

If your calculus is rusty, you can use this online tool:

https://www.derivative-calculator.net/

I handle midgame/endgame weighting as follows: each parameter has an entry in an array that is a structure and includes how it should be scaled.

Each time a gradient is updated, I call a scale method that applies the scaling factor for the associated parameter.

See tune.cpp and specifically the scale method here:

https://github.com/jdart1/arasan-chess

--Jon

[fierz]

Is there anyone who has a large epd file for download somewhere with the results of the games?

cheers
Martin

[AlvaroBegue]

Is there anyone who has a large epd file for download somewhere with the results of the games?

Give this a try: https://bitbucket.org/alonamaloh/ruy_tu ... th_results

[jdart]

I have done something similar.

It has occurred to me, though, that the current tuning method would work just the same if each position had a winning probability, not just a 0, 0.5, 1 result. So if you really want to burn a lot of CPU cycles, you could run a match of say, 10 games with each position. The result would be a training set with more accurate labels.

--Jon

[AlvaroBegue]

I have done something similar.

It has occurred to me, though, that the current tuning method would work just the same if each position had a winning probability, not just a 0, 0.5, 1 result. So if you really want to burn a lot of CPU cycles, you could run a match of say, 10 games with each position. The result would be a training set with more accurate labels.

--Jon

I don't have a strong argument for why I think this, but I think if you had 10 times the CPU budget to spare, you would be better off having 10 times as many positions labelled by a single game.

[flok]

[I downloaded games from CCRL, took positions from those games and analyzed them with my program RuyDos. I saved positions on which the evaluation function was being called after searching 1000 nodes. I then labelled each position by running one very quick SF8-vs-SF8 game.

https://bitbucket.org/alonamaloh/ruy_tu ... th_results

EDIT: In that file each position has been replaced by the position from which quiescence search got its score.

In that file, we can see for example:

6k1/8/p4P1B/3b4/1pp3B1/7P/1R4P1/6K1 b - - 1-0

Now is that 1-0 for win for white or for black? I'm asking as it is now black that can move.

[AlvaroBegue]

1-0 for white. I use the same format of the "Result" tag in PGN.

[flok]

Something I forgot to ask:

for
totalError += pow(value_from_fen - calculateSigmoid(eval_score), 2);

value_from_fen = 1.0 for 1-0, 0.0 for 0-1 and so on

but eval_score, should it be from the point of view of the fen-string? or from white? or...?

[Sven]

Something I forgot to ask:

for
totalError += pow(value_from_fen - calculateSigmoid(eval_score), 2);

value_from_fen = 1.0 for 1-0, 0.0 for 0-1 and so on

but eval_score, should it be from the point of view of the fen-string? or from white? or...?

Obviously the sigmoid function (and therefore also eval_score) must represent a value from the same viewpoint as value_from_fen. Otherwise you would get a high "error" e.g. if value_from_fen = 1, it is black's turn, and your eval function returns a high score in favor of white so that the sigmoid function returns a value close to 0 (from black viewpoint).

[flok]

Something I forgot to ask:

for
totalError += pow(value_from_fen - calculateSigmoid(eval_score), 2);

value_from_fen = 1.0 for 1-0, 0.0 for 0-1 and so on

but eval_score, should it be from the point of view of the fen-string? or from white? or...?

Obviously the sigmoid function (and therefore also eval_score) must represent a value from the same viewpoint as value_from_fen. Otherwise you would get a high "error" e.g. if value_from_fen = 1, it is black's turn, and your eval function returns a high score in favor of white so that the sigmoid function returns a value close to 0 (from black viewpoint).

Ok, is what I thought to be honest but as the results are currently garbage, I wondered if I maybe did it wrong.

Thanks!