connor_mcmonigle wrote: ↑Sat May 15, 2021 4:43 am
I, personally, didn't really find anything conclusive when I read over the evaluation function code outside of the general layout being fairly similar to Stockfish and I'd agree that many top engines share such evaluation terms.
Within the Stockfish family Fire 8.1's highest depth-1 similarity is 64.21% with Stockfish 8 so I compared its sources with this version and I must agree with you. There are things which are common in many top open source engines like the previously mentioned ones and I would say that a couple of them are much more influenced by Stockfish than Fire is but that's it. If I didn't know Houdini is a Stockfish derivative and Fire is based on Houdini, I wouldn't think Fire's evaluation function could be considered derived from Stockfish at all. If anyone is interested I can make a post with the notes I took during the comparison.
Yes. Houdini's evaluation function was undoubtedly the most original part of the engine (though there is some speculation that it was RE'd from Komodo's evaluation function). Obtaining the Houdini source code is rather difficult now, but if you're able to compare against that, the similarities are striking.
Indeed ...
I'm not lying. Perhaps you missed this?
I agree completely...Houdini's eval was quite unique
I've said that many time, yet many prefer to continue the falsehood often re-iterated by yourself and Andrew that Houdini is a simple SF clone
"..., the similarities are quite striking"
"... Indeed"
Thomas Miller later asked to edit the post you're now quoting in that very thread as well...
Perhaps you're confused, he's referring to the similarities between Fire eval and Houdini's (not Houdini and SF)
"I wouldn't think Fire's evaluation function could be considered derived from Stockfish at all. "
"Yet, Fire improved 174 elo when nnue was added...
much more than any Segio V. RL net would provide"
So what's true now? Did it gain 174 elo or did it not? You realise not both can be true?
The stronger the "base engine" is the less nnue will generally gain. That's why SF gained less from it - which seems rather obvious.
"the falsehood that Houdini is a simple SF clone" - both the source code AND simex disagree with you on this.
You're fighting a battle which was lost before it even started.
kranium wrote: ↑Thu Jun 17, 2021 6:35 pm
Perhaps you're confused, he's referring to the similarities between Fire eval and Houdini's (not Houdini and SF)
Sure, what's the difference? If you cloned Stockfish, Komodo or literally any other engine's evaluation function your engine would still have a copied evaluation function and not be eligible for participation in tournaments (directly counter to MikeB's initial assessment). I fail to see how which engine you copied is relevant.
ChickenLogic wrote: ↑Thu Jun 17, 2021 6:36 pm
"Yet, Fire improved 174 elo when nnue was added...
much more than any Segio V. RL net would provide"
So what's true now? Did it gain 174 elo or did it not? You realise not both can be true?
The stronger the "base engine" is the less nnue will generally gain. That's why SF gained less from it - which seems rather obvious.
"the falsehood that Houdini is a simple SF clone" - both the source code AND simex disagree with you on this.
You're fighting a battle which was lost before it even started.
yes it gained ~174 elo
There is no version of SF eval that compares closely with Houdini eval...I've looked closely
also
as far as I know, sim does not show Houdini as highly similar to SF...
Last edited by kranium on Thu Jun 17, 2021 6:45 pm, edited 1 time in total.
kranium wrote: ↑Thu Jun 17, 2021 6:35 pm
Perhaps you're confused, he's referring to the similarities between Fire eval and Houdini's (not Houdini and SF)
Sure, what's the difference? If you cloned Stockfish, Komodo or literally any other engine's evaluation function your engine would still have a copied evaluation function and not be eligible for participation in tournaments (directly counter to MikeB's initial assessment). I fail to see how which engine you copied is relevant.
we're talking about similarity and uniqueness, not rules for tournament entry
besides, I'm fairly certain that Houdini was disqualified from TCEC primarily by word of mouth (you and Andrew & cucumber maybe?) not sure
but certainly not by an impartial expert
ChickenLogic wrote: ↑Thu Jun 17, 2021 6:36 pm
"Yet, Fire improved 174 elo when nnue was added...
much more than any Segio V. RL net would provide"
So what's true now? Did it gain 174 elo or did it not? You realise not both can be true?
The stronger the "base engine" is the less nnue will generally gain. That's why SF gained less from it - which seems rather obvious.
"the falsehood that Houdini is a simple SF clone" - both the source code AND simex disagree with you on this.
You're fighting a battle which was lost before it even started.
yes it gained ~174 elo
There is no version of SF eval that compares closely with Houdini eval...I've looked closely
also
as far as I know, sim does not show Houdini as highly similar to SF...
Then you're just misinformed or willfully ignorant...
Raphexon wrote: ↑Thu Jun 17, 2021 6:44 pm
Anyway, could you tell me who this "Sergio Viera" is?
a scientist/researcher from the National University of Singapore who pretty much perfected nnue reinforcement learning
here's his download page https://www.comp.nus.edu.sg/~sergio-v/nnue/
Raphexon wrote: ↑Thu Jun 17, 2021 6:44 pm
Anyway, could you tell me who this "Sergio Viera" is?
a scientist/researcher from the National University of Singapore who pretty much perfected nnue reinforcement learning
here's his download page https://www.comp.nus.edu.sg/~sergio-v/nnue/