Question to Orion 0.8 NN ...

[Frank Quisinsky]

Hi there,

at the moment my FCP Tourney-2022 is still running.

I got from two persons the information that Orion 0.8 NN used a Stockfish net.
Is this right?

Best
Frank

[Sylwy]

Hi there,

at the moment my FCP Tourney-2022 is still running.

I got from two persons the information that Orion 0.8 NN used a Stockfish net.
Is this right?

Best
Frank

Nope ! Original.



Links are here:
========================================
https://www.chessprogramming.org/Orion

https://github.com/david-carteau/cerebrum

========================================

[Frank Quisinsky]

After all I remember the first Beta version from Orion used an SF net.
Honestly I forgotten ...

Thank you!!

[David Carteau]

Hi there,

at the moment my FCP Tourney-2022 is still running.

I got from two persons the information that Orion 0.8 NN used a Stockfish net.
Is this right?

Best
Frank

Sorry for the late answer, I haven't visited the forum for months now, due to intermittent connection problems from France (403 forbidden errors) and... mainly due to a lot of work

I think that it is important to understand what has been done for the neural network embedded in Orion 0.8 : as I tried to explain on my personal blog (see signature) :
- Orion uses its own trained network, with its own architecture, and 100% of its code is original ;
- I even prepare and release a Python script and a C library to help people who are interested to train and embed their own Orion-compatible networks ;
- but then comes the question of the data used to train the network...!

To train my own network, I used a set of 360 million positions which were evaluated as follows : take a position, use the Stockfish 12 neural network to statically evaluate it (no search). The Stockfish engine has not been used during this process as I had developed my own code to read the network file (for the Orion 0.7 NNUE experiment). The Orion search code has not been used either (again, each position was statically evaluated).

Orion's network is not a copy/paste from the SF12's network : it has been trained to "mimic" SF12's network.

I know that some people can then consider that, in a certain way, Orion 0.8 is not 100% original (it is from a code and network point of view, but isn't if you consider that Orion has learned how to play chess with a good level from a famous teacher).

I'm trying for several months now to apply another approach, inspired by the Seer engine : take a set of endgame positions, evaluate them thanks to Syzygy tablebases, and try to retro-propagate evaluations to starting positions. I poorly failed for the moment

In parallel, I worked on the next version of Orion (v0.9), which is actually ready since... November 2021, but not released. The new version embeds a new and tiny network (24x smaller !) and has SMP support (a big evolution !). The network has been trained in the same way than for v0.8 (from "teacher" Stockfish 12 ) and has more or less the same strength in my own tests.

Except if someone is interested, I'm not planning to release it before I achieve to train a network not using "external knowledge" from another engine. Be patient !

Kind regards,

David

[matejst]

David,

I liked Orion 0.8 very much and I would be very interested by a new Orion.