Piece weights with regression analysis (in Russian)

[mvk]

My article on a polular Russian IT portal:
http://habrahabr.ru/post/254753/

It is in Russian, but I believe the text is quite understandable even without translation.

Courtesy of Vladimir, I made a CPW page trying a translation with the help of google translate, improvements and suggestions welcome.

https://chessprogramming.wikispaces.com ... n+Analysis

Cheers,
Gerd

Isn't the essence of this method exactly the same what everybody else has been doing now for a couple of years and which has become known as "Texel's tuning method"? (That is, a fit of the evaluation in the percentage domain against the game results?). The only real difference is that we use much larger data sets. All other differences seem not essential at first glance.

[Gerd Isenberg]

My article on a polular Russian IT portal:
http://habrahabr.ru/post/254753/

It is in Russian, but I believe the text is quite understandable even without translation.

Courtesy of Vladimir, I made a CPW page trying a translation with the help of google translate, improvements and suggestions welcome.

https://chessprogramming.wikispaces.com ... n+Analysis

Cheers,
Gerd

Isn't the essence of this method exactly the same what everybody else has been doing now for a couple of years and which has become known as "Texel's tuning method"? (That is, a fit of the evaluation in the percentage domain against the game results?). The only real difference is that we use much larger data sets. All other differences seem not essential at first glance.

You ask me? I just tried to translate the text and try to understand the math
http://cs229.stanford.edu/materials.html
http://cs229.stanford.edu/notes/cs229-notes1.pdf
http://cs229.stanford.edu/section/cs229-linalg.pdf

So far I found the cost function for the logistic regression much more complicated than Peter's average evaluation error function over all features.

Vladimir's input x for the five pieces are several material imbalances during the course of a game with one outcome y...

[Evert]

My article on a polular Russian IT portal:
http://habrahabr.ru/post/254753/

It is in Russian, but I believe the text is quite understandable even without translation.

Courtesy of Vladimir, I made a CPW page trying a translation with the help of google translate, improvements and suggestions welcome.

https://chessprogramming.wikispaces.com ... n+Analysis

Cheers,
Gerd

Isn't the essence of this method exactly the same what everybody else has been doing now for a couple of years and which has become known as "Texel's tuning method"? (That is, a fit of the evaluation in the percentage domain against the game results?). The only real difference is that we use much larger data sets. All other differences seem not essential at first glance.

As far as I can tell, yes.
Still, the explanation is nice (err... if you get it translated from Russian) and the linked utility is also potentially useful.

[Dann Corbit]

There are some minor differences (such as the S curve for fitting) but the main difference is the excellent and detailed explanation of the math. Hence, the web page.

[mvk]

There are some minor differences (such as the S curve for fitting)

Ok, but that is what i mean with "evaluation in the percentage domain" (as opposed to pawn-value domain). Esp for the piece values that is significant and we are all doing that. I'm sorry for being too terse.

[Ajedrecista]

Hello Vincent:

As I see, it's more the value of the pawn who changed more over time (pieces values between them don't change a lot). In my sense, it would be better to set the value of the knight to 300 to have a more clear view of the evolution over time.
Could you do that, please ?

Thanks,
Vincent

I am sorry but I can't. I had a long weekend but I will very busy since right now for do such a big task. I only see a solution: you (or anyone else) can do it but changing regression.cpp file. You can find 100 * theta / theta[0] for various i; I propose that before compiling from sources, you should replace that by 300 * theta / theta[1] (fix N = 300 cp) or by 500 * theta / theta[3] (fix R = 500 cp). I say that because of roundings: an easy output is an easy output.

I know, re-run 234 PGNs is boring and very time consuming.

As a side note, I see that Anand's bishop value is circa 230 cp in Vladimir's article while I wrote 220 cp, keeping the values of the rest of pieces. I guess that we picked the same PGN and one of us made a typo, surely me. Being perfect for 234*4 = 936 values is almost impossible, specially for me.

Regards from Spain.

Ajedrecista.

[Xann]

Hi Marcel,

Isn't the essence of this method exactly the same what everybody else has been doing now for a couple of years and which has become known as "Texel's tuning method"? (That is, a fit of the evaluation in the percentage domain against the game results?). The only real difference is that we use much larger data sets. All other differences seem not essential at first glance.

Sorry to hijack your post but this topic has been bugging me for a while: isn't "Texel's tuning method" just logistic regression that has been used in games for about twenty years (see Buro's papers for instance)?

I don't see any difference between all three apart from QS vs. eval, which seems a minor issue to me. A single-neuron ANN is also the same.

Fabien.

[Bloodbane]

Hi Marcel,

Isn't the essence of this method exactly the same what everybody else has been doing now for a couple of years and which has become known as "Texel's tuning method"? (That is, a fit of the evaluation in the percentage domain against the game results?). The only real difference is that we use much larger data sets. All other differences seem not essential at first glance.

Sorry to hijack your post but this topic has been bugging me for a while: isn't "Texel's tuning method" just logistic regression that has been used in games for about twenty years (see Buro's papers for instance)?

I don't see any difference between all three apart from QS vs. eval, which seems a minor issue to me. A single-neuron ANN is also the same.

Fabien.

Seems that way to me.

[mvk]

Sorry to hijack your post but this topic has been bugging me for a while: isn't "Texel's tuning method" just logistic regression that has been used in games for about twenty years (see Buro's papers for instance)?

I don't see any difference between all three apart from QS vs. eval, which seems a minor issue to me. A single-neuron ANN is also the same.

I don't know. I remember having read about the idea of assigning elo to othello pattern features. Instead of guessing, do you have links to specific papers? (both Buro, and also ANN?)

[Gerd Isenberg]

Hi Marcel,

Isn't the essence of this method exactly the same what everybody else has been doing now for a couple of years and which has become known as "Texel's tuning method"? (That is, a fit of the evaluation in the percentage domain against the game results?). The only real difference is that we use much larger data sets. All other differences seem not essential at first glance.

Sorry to hijack your post but this topic has been bugging me for a while: isn't "Texel's tuning method" just logistic regression that has been used in games for about twenty years (see Buro's papers for instance)?

I don't see any difference between all three apart from QS vs. eval, which seems a minor issue to me. A single-neuron ANN is also the same.

Fabien.

Wow, so simple is that
Slowly the dust settles. Thanks for the insight, yes, a single-neuron ANN.

I don't exactly get that with the cost function requiering gradient descent to find the optimum, i.e. the error formula with least squares as used in Texel Tuning and Deep Thought Tuning and the "wild formulas" in Buro's paper?
http://www.jair.org/media/179/live-179-1475-jair.pdf