Komodo 14.1 Release and Dragon Announcement

[AndrewGrant]

I think Minic maybe swapped in it’s own qsearch, otherwise I’m not aware of any net author other than myself that has replaced the qsearch. I can’t speak to dragon, but I replaced the qsearch for training night nurse with a very simple one. You’re not looking for the best qsearch for playing, just quiet positions for training.

So, it is the same training code then, except you've cut out the SF bits?

For night nurse, yes.

Are you not willing to say that Komodo is trained with separate code? I would think you would say so, because that would set Komodo apart.

[Madeleine Birchfield]

Well, Halogen's original networks are trained in Pytorch. Halogen's current network is trained by a private NN/NNUE trainer I wrote; but Halogen is NOT using the NNUE King-Piece, Shogi Style, whatever you want to call it, framework. All Halogen networks are trained on game results I believe, not evaluations, so there is no SF involved at any step.

NNUE is a much bigger framework than just king-piece or half king-piece; king-piece/half king-piece is only one NNUE architecture amongst many that are possible. There exist NNUE architectures (currently unused as far as I know) that do not use king-piece whatsoever for input features, such as mobility or piece count NNUE architectures.

Stockfish developers themselves have recently been talking about moving away from half king-piece to a slightly different NNUE architecture and adding more input features like passed pawns to their network.

[IanKennedy]

Great news but as Frank already mentioned Arena bundles Dragon by Bruno Lucas even if its not actively developed now, so can you call it something else please.

[Madeleine Birchfield]

Great news but as Frank already mentioned Arena bundles Dragon by Bruno Lucas even if its not actively developed now, so can you call it something else please.

Komodo Dragon would be a good name.

[dkappe]

Are you not willing to say that Komodo is trained with separate code? I would think you would say so, because that would set Komodo apart.

Not for me to say.

[AndrewGrant]

Are you not willing to say that Komodo is trained with separate code? I would think you would say so, because that would set Komodo apart.

Not for me to say.

Odd. I'm surprised by such a statement. Will Mark or Larry say?
Otherwise people will assume this is just SF's trainer, like most/all other NNUEs in play at the moment.

[Madeleine Birchfield]

I am not familiar with what Seer is doing, although I have talked to the author a fair amount.

Regarding Seer, this is what Connor had to say about Seer's architecture vs Stockfish:

The tensor product is a generalization of the notion of an outer product (read: in this context they mean the same thing). Consequently, the question is a bit meaningless. I've not seen halfkp features described by a tensor product before, but one could consider it a tensor product of piece and king features if they wanted to.

If we let vec(x) denote flattening x, a X b represent the outerproduct of a and b, B represent the board features in {0, 1}^12x8x8 (a boolean "tensor"), and let B[0, :, :] and B[1, :, :] (using slice notation) denote white and black king planes respectively, then Seer's input is (vec(vec(B[0, :, :]) X vec(B)), vec(vec(B[1, :, :]) X vec(B))). This differs from the input features currently used in SF networks which would be (vec(vec(B[0, :, :]) X vec(B[:2, :, :])), vec(vec(B[1, :, :]) X vec(B[:2, :, :]))).

Current Stockfish input features are referred to by the name "halfkp" (king X piece). AFAIK, some individuals experimenting with switching to use Seer style input for Stockfish networks have coined the name "halfka" (king X all). Also notable is that, in Stockfish, the (black king) X (piece) features are rotated (though the rotating is an accidental, possibly elo gaining, bug as mirroring was intended). Seer doesn't use mirroring/rotating as it uses separate affine transforms ("feature transformers") for the white and black halves of the halfka input.

[dkappe]

Are you not willing to say that Komodo is trained with separate code? I would think you would say so, because that would set Komodo apart.

Not for me to say.

Odd. I'm surprised by such a statement. Will Mark or Larry say?
Otherwise people will assume this is just SF's trainer, like most/all other NNUEs in play at the moment.

You like to push things to a crisis, don’t you? Let me just say here, as I have many times in the SF discord and elsewhere, that the sooner SF can move away from a Hand crafted c++ CPU-based trainer to a tensorflow or pytorch, gpu based trainer, the better. I see they’ve started taking some baby steps in that direction.

[AndrewGrant]

The sooner SF can move away from a Hand crafted c++ CPU-based trainer to a tensorflow or pytorch, gpu based trainer, the better. I see they’ve started taking some baby steps in that direction.

I agree with you on that. Pytorch is problematic, but I had assumed you trained your networks with Tensorflow.
That was a wrong assumption for Bad Gyal.

[dkappe]

The sooner SF can move away from a Hand crafted c++ CPU-based trainer to a tensorflow or pytorch, gpu based trainer, the better. I see they’ve started taking some baby steps in that direction.

I agree with you on that. Pytorch is problematic, but I had assumed you trained your networks with Tensorflow.
That was a wrong assumption for Bad Gyal.

Bad Gyal is a leela style resnet initially trained on tensorflow, now trained on a pytorch based system. It has nothing to do with NNUE.