Theoria 0.2 is out

[FireDragon761138]

We changed how iterative depth works in search, now the engine checks past depth 10 whether the eval has stabilized and ends the search if a certain ply-to-ply stability threshold is reached. This results in more parsimonious principle variations. A positive side effect is even greater computational efficiency in analyzing a position (at least 20 percent greater on average). The stability threshold is tunable via uci option, but we did some testing and chose a reasonable default for most users:

https://www.theoriachess.org/download/


We are also working on a web app that integrates Theoria with a MLM (medium language model) and retrieval augmented generation to comment on the moves using natural language. And we're experimenting with different analysis tuning parameters for Lucas-Chess, so far a score limit of 1600 and curve of 42 seems to be an improvement over the defaults that seem tuned for Stockfish- it only takes a few tweaks but the result is clearer analysis with less marking of questionable moves and mistakes on positions that are already won or lost.

[Sylwy]

THANK YOU !

Just a logo version (I have many):

[ovenel]

When you say that you are working on a web app to comment on moves, would that mean that your ultimate goal with this is to create something akin to the "coach" and game review on Chess.com or decodechess.com?

Would this be something that you would be hosting, or would it be more for demonstration purposes with the goal of allowing people to run it using their own computer resources (or are you hoping to monetize this work)? My assumption is that hosting such a platform would not be cheap if it turns out to be useful, especially if people are looking for a free alternative to the two that I listed above. Ultimately, I'm curious about your overall goal here. Are you primarily interested in research and providing insights to the broader community as you learn more, meaning that the software and platform is just a bonus, or are you hoping to develop a platform for the automated analysis of a chess game?

[FireDragon761138]

When you say that you are working on a web app to comment on moves, would that mean that your ultimate goal with this is to create something akin to the "coach" and game review on Chess.com or decodechess.com?

Would this be something that you would be hosting, or would it be more for demonstration purposes with the goal of allowing people to run it using their own computer resources (or are you hoping to monetize this work)? My assumption is that hosting such a platform would not be cheap if it turns out to be useful, especially if people are looking for a free alternative to the two that I listed above. Ultimately, I'm curious about your overall goal here. Are you primarily interested in research and providing insights to the broader community as you learn more, meaning that the software and platform is just a bonus, or are you hoping to develop a platform for the automated analysis of a chess game?

We may try making a desktop app, either by modifying Lucas Chess or a stand-alone program. We are trying to iron out all the kinks in the natural language code at the moment so it produces consistent quality output.

Right now the Theoria engine is an intermediate step, and it's GPL and open source so anybody can use it in any project they want. There are several LLM-based open-source projects out there and I have reason to believe it would work better for analysis than most other popular alternatives, since the engine is only focused on analysis quality and efficiency in its development.

[amathur2k]

Hi Firedragon.
I am trying to understand the codebase.
I understand the codebase is pointing to official stockfish nnue's
Is there is a custom nnue you guys are working on, if yes where can we access it ?

[lucario6607]

Hi Firedragon.
I am trying to understand the codebase.
I understand the codebase is pointing to official stockfish nnue's
Is there is a custom nnue you guys are working on, if yes where can we access it ?

Their claim is using leela data somehow makes it better for teaching humans even though that isn’t the purpose of stockfish. But surprise, surprise, stockfish already uses leela data so the only thing they managed to do is make stockfish weaker like most clones. Also seems most of his replies are written by ai sometimes

[amathur2k]

Well i did try with the binaries and it did seem to suggest lines, I as a beginner(1350 chess.com) can get behind even at higher depths. So yes it may be a weaker version of SF, but there is something interesting they are doing for sure. Thats what i wanted to see with the nnue whether its just early stopping or something else.

[lucario6607]

Well i did try with the binaries and it did seem to suggest lines, I as a beginner(1350 chess.com) can get behind even at higher depths. So yes it may be a weaker version of SF, but there is something interesting they are doing for sure. Thats what i wanted to see with the nnue whether its just early stopping or something else.

Placebo

[lucario6607]

Not surprised, that you lied. But you said it doesn’t use the same data as stockfish. Your website says you used a stockfish binpack. So your main claim is that your net is leela data but your data is the same as stockfish.