Engines that have a hybrid evaluation

[Madeleine Birchfield]

What engines use both neural networks and handcrafted features in their evaluation function? Examples I could think of are Stockfish, Winter, Fizbo, and Wasp.

[AndrewGrant]

Ethereal's HCE contains a neural network dedicated to Pawns+Kings.

[Madeleine Birchfield]

Ethereal's HCE contains a neural network dedicated to Pawns+Kings.

But the dev version of Ethereal currently only uses neural networks, right?

[AndrewGrant]

Ethereal's HCE contains a neural network dedicated to Pawns+Kings.

But the dev version of Ethereal currently only uses neural networks, right?

No; Dev version either uses the NNUE, or the HCE+PawnKing-NN.

[jdart]

Arasan switches between NNUE and classical eval depending on the position, as Stockfish does (not with the same logic, however).

[Branko Radovanovic]

What engines use both neural networks and handcrafted features in their evaluation function? Examples I could think of are Stockfish, Winter, Fizbo, and Wasp.

There is a nice overview here: https://wiki.chessdom.org/Engines_and_authors (it's perhaps not 100% up to date, e.g. Wasp last competed as a classical engine, but has gone NNUE since)

[jstanback]

FYI, Wasp 5.0 uses a NN, but it is not the NNUE architecture.

[jtwright]

Andrew, if you have a moment, I was curious about this. IIRC from the Ethereal github the condition for switching to HCE is:

- being in check, or
- > 400 cp of imbalance in the PSQT (using EG values)

How often does the second condition seem to come up and how much does it make an impact? From glancing at the PSQTs, it seems like there would have to be *significantly* more pieces on one side than the other, which I would often expect to be a crushing endgame anyway.

I'm curious if anyone else can weigh in on this as well.

[AndrewGrant]

Andrew, if you have a moment, I was curious about this. IIRC from the Ethereal github the condition for switching to HCE is:

- being in check, or
- > 400 cp of imbalance in the PSQT (using EG values)

How often does the second condition seem to come up and how much does it make an impact? From glancing at the PSQTs, it seems like there would have to be *significantly* more pieces on one side than the other, which I would often expect to be a crushing endgame anyway.

I'm curious if anyone else can weigh in on this as well.

Perfect example for you:

Recently I took my training set which generated the current best net, and made a copy that excluded positions such that 400cp material imbalance occurred. No positions in the dataset were incheck. Results: 1,846,808,245 => 1,466,369,570 positions.

400cp could be a single Knight. Realistic during the game? No. During the search tree? Yes. Although its worth noting that I once measured deleting the entire evaluation, thousands of lines, only lost like 3 points of elo. It is not clear to me the value of the HCE, aside from some sanity when dealing with extreme positions.

[jtwright]

Perfect example for you:

Recently I took my training set which generated the current best net, and made a copy that excluded positions such that 400cp material imbalance occurred. No positions in the dataset were incheck. Results: 1,846,808,245 => 1,466,369,570 positions.

400cp could be a single Knight. Realistic during the game? No. During the search tree? Yes. Although its worth noting that I once measured deleting the entire evaluation, thousands of lines, only lost like 3 points of elo. It is not clear to me the value of the HCE, aside from some sanity when dealing with extreme positions.

Thanks for the explanation. I should have paid a bit more attention and realized it was `psqtmat`, not just `psqt`. Being more than 400 cp in material + PSQT is significantly more common. Also, always good to keep in mind that many positions unlikely to come up in an actual game will still be seen very often in the search tree.

Interesting that the difference in elo is so small between hybrid and NNUE only in Ethereal. I've been debating the value of a hybrid approach myself for the future so this knowledge is very valuable. Thanks.