2

I read up on Stockfish NNUE and as far as I understand it, Stockfish NNUE was (at least in early stages) trained by positions which were evaluated by the hand-crafted eval. function by stockfish. The question I have is, why the NNUE is stronger than the hand-crafted eval? Since the ANN was trained using the hand-crafted function, it only represents it using an ANN? Is that correct, or is my understanding wrong?

From what I've read, the search speed with the hand-crafted eval. was around 100 Million Positions. Search speed using the NNUE was around 60. Million. Therefore, the hand-crafted eval should be stronger.

What am I missing?

Improve this question
1
  • 2
    IIRC it wasn't train on plain eval but rather a search result from strong SF. The old SF eval is how good SF plays at depth 1 (or 0 if you will). But if you train a NN to play like old SF at depth 10 that would be way way stronger. (and then of course you still use that search on top of NNue)
    – koedem
    Feb 24, 2022 at 12:19

2 Answers 2

Reset to default
1

NNUE was originally trained on depth 8 to 10 handcrafted eval. This is significantly stronger than a single node evaluation.

Improve this answer
0

as far as I understand it, Stockfish NNUE was (at least in early stages) trained by positions which were evaluated by the hand-crafted eval. function by stockfish

And after the early stages the neural network was built up by continual training.

The question I have is, why the NNUE is stronger than the hand-crafted eval?

Because it is more than the hand-crafted eval which was only used to speed up the initial training process.

Since the ANN was trained using the hand-crafted function, it only represents it using an ANN?

No. The finished product is much more than the had crafted evaluation. It incorporates additional evaluations which it "discovered" during training.

Improve this answer
1
  • Okay, thank you. How did the training then continue? I'm guessing that supervised learning wasn't used, because backpropagation couldn't be applied, right? So the ANN was then trained using reinforcement learning?
    – Phil3453
    Feb 26, 2022 at 20:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.