Optimization algorithm for the Texel/Gaviota tuning method

[Bloodbane]

I'm not sure what you mean, column F is the engine evaluation. In this case our evaluation function only has one parameter but there is nothing different between W having one entry and W having 1000 entries, it's still equal to the sum(i = 1, n, Wi * existsWeight(Wi, p)). You can choose to not update all but one parameter if you only want to tune one but if your evaluation function uses some parameter it should be added to W. Oh and K should probably be bigger as well, with K = 0.5 convergence over here is far too slow to be practical. This is why I recommend some interactive strategy. Something like multiply K by 1.1 every successful iteration and multiply by 0.5 every failed iteration (of course if an iteration fails W should be restored to the last succesfull update).

[Ferdy]

I'm not sure what you mean, column F is the engine evaluation.

I have presented the table which item you are not sure of?

In this case our evaluation function only has one parameter

No what I mean here is we will tune one parameter, this parameter has default value of 120 cp. Our evaluation function is of course has many parameters.

but there is nothing different between W having one entry and W having 1000 entries, it's still equal to the sum(i = 1, n, Wi * existsWeight(Wi, p)). You can choose to not update all but one parameter if you only want to tune one but if your evaluation function uses some parameter it should be added to W.

Could you illustrate how you do it if you tune 2 parameters? Say add new parameter PawnValue with default value of 100. Putting it on table just like what I did would be better of course.

Oh and K should probably be bigger as well, with K = 0.5 convergence over here is far too slow to be practical. This is why I recommend some interactive strategy. Something like multiply K by 1.1 every successful iteration and multiply by 0.5 every failed iteration (of course if an iteration fails W should be restored to the last succesfull update).

Right I get that. I am not interested whether this will optimize or not, but I am interested on how GD is done.

[Bloodbane]

Here we have an evaluation function which only has two parameters, passedPawn and pawnValue. We compute gradients for both, and update both. If you only want to update one of them then set the gradient of the one you do not want to tune to 0. Easily extendable to as many parameters as you want.

[Ferdy]

That is excellent thank you.