Rhetoric 1.2 Released

[asanjuan]

Hi Alberto,

(...). I managed to add a lot of elo points with only 3 changes: Singular Extensions, Late Move Pruning, and applying a new learning (or tuning) algorithm.(...)

Congrats!

May I ask Which "learning algorithm" you implemented?

I'm a little surprise you attribute the elo increase to (among other things) Singular Extensions. I always thought this added little in terms of raw strength.

Any estimate of the elo gain? I'm interested as I am at the stage of adding these elements to Maverick.

Steve

SE alone gave me around 20 elo.

LMP was a surprise and gave me 65 elo points!! It was a little bit tricky, because you need the null move threat to avoid tactical mistakes.

The big surprise was the implementation of the tunning based on Peter Osterlünd posts. I implemented the sigmoid function, the average error calculation function based on a big epd file, and the rest was a matter of search. You can use whatever technique to search (gradient, G.A, or a simple loop as Peter pointed). The result is a balanced set of parameters for your evaluation. And it gave Rhetoric another 80 or 90 elo more and a very dynamic and positional style. It discovered stunning values: for example:
A pawn in the opening is worth 75 centipawns, but in the endgame the value rises to 110. The mobility values where doubled and almost doubled again for the endgame except for the queen... That values made rhetoric sacrifice pawns for attack or any other positional criteria.

The total amount added: more than 100 elo points in self play. The test against the pool of sparrings that I have been using until now, show even better results. This may put Rhetoric in a good place in the rating lists. I'm not going to give a number but I have good feelings with this version.

S.E. can be tricky also. I implemented it in a very straightforward way. Quite simple, but working. If you want some code, I can send you an email or a PM with the details. I shared this mail with Brian Richardson (author of Tinker), so is not a problem.


Cheers.

[Steve Maughan]

SE alone gave me around 20 elo.

LMP was a surprise and gave me 65 elo points!! It was a little bit tricky, because you need the null move threat to avoid tactical mistakes.

The big surprise was the implementation of the tunning based on Peter Osterlünd posts. I implemented the sigmoid function, the average error calculation function based on a big epd file, and the rest was a matter of search. You can use whatever technique to search (gradient, G.A, or a simple loop as Peter pointed). The result is a balanced set of parameters for your evaluation. And it gave Rhetoric another 80 or 90 elo more and a very dynamic and positional style. It discovered stunning values: for example:
A pawn in the opening is worth 75 centipawns, but in the endgame the value rises to 110. The mobility values where doubled and almost doubled again for the endgame except for the queen... That values made rhetoric sacrifice pawns for attack or any other positional criteria.

The total amount added: more than 100 elo points in self play. The test against the pool of sparrings that I have been using until now, show even better results. This may put Rhetoric in a good place in the rating lists. I'm not going to give a number but I have good feelings with this version.

S.E. can be tricky also. I implemented it in a very straightforward way. Quite simple, but working. If you want some code, I can send you an email or a PM with the details. I shared this mail with Brian Richardson (author of Tinker), so is not a problem.

Great summary.

I wasn't looking for absolute ELO of new version but the estimated deltas for the elements you added.

Since asking the question I found Peter Osterlünd post. I'll take a look. It looks interesting.

I'm getting bogged down in trying to tune Maverick's evaluation function. The domain space of an evaluation function is *massive*. Auto-tuning must be the way to go.

Thanks,

Steve

[Jamal Bubker]

Thank you Alberto for this new release !

[pedrox]

[Graham Banks]

Thanks Pedro. Very useful.