Hacker News Reader

Comment by cellis | Hacker News Reader

Translation: Alpha Zero, we're ready for a rematch!

Alph zero/mu at similar computation resources was weaker for a while iirc.

Right, my recollection is that stockfish + nnue has been stronger than leela since stockfish 13, and leela is stronger than A0.

Isn't the difference though, that Leela is a pure RL implementation whereas NNUE/SF still uses some tablebase stuff?

No, NNUE doesn't necessarily need tablebases and Leela uses tablebases too. The difference is that their network architecture are different and their search algos are different.

Leela uses mcts, sf does not. SF also trains on human positions, while I think Leela is self play only.

SF14's net was actually trained on lc0 data. Human data is way too weak for SF to learn from.

Yes and I think it's really generous of the LC0 to "donate" that data. On the other side LC0 uses SF during the training process since a while.

It makes sense. They're both open source programs working towards better understanding of chess. There are actually a few people who develop for both. They are obviously very different types of engines, but they are much closer to friendly rivals than enemies.

AlphaZero and co only trained via self-play because that was their research goal, and they looked unwinnable because they were the first to get this kind of neural net working, but it doesn't seem like it's the best option.

AZ notably got completely tilted once it started losing, doesn't necessarily recognize strange positions you can't normally get into, and doesn't care about its win margin at all.

but.. they're using a pretty big neural net themselves (NNEU) as far as I can tell? With datasets of hundreds of Gigs. Doesn't this remove the significance of the matchup, that was supposed to be about Deep Learning vs more traditional chess bot methods?

Claiming that something that uses a neural net trained on hundreds of gigs of data isn't deep learning .. I mean it's possible, I don't know the details.

What is it about now, open vs closed source? Different methods of deep learning and big data fighting? (Both of these are also interesting ofc)

NNUE is much smaller than Leela's net, and has a much different architecture that's optimized more for CPU. Additionally, Leela uses Monte Carlo Tree Search and Stockfish uses Alpha/Beta pruning.

They say that the dataset is hundreds of Gigs worths of games, so the net must still be really pretty big.

Though definitely not directly comparable, dataset of GPT2-xl is 8 million web-pages. What I mean to say is that this is clearly deep learning.

Size of training set is not enough to make it deep learning, right? Doesn't deep learning imply at least one hidden layer?

NNs without at least one hidden layer are rarely used.

They're used all the time, we just call it logistic regression.

Are you saying you read that it didn't have a hidden layer?

My point is that having such a huge dataset would not be extremely useful without using a deep neural net (of at least one hidden layer)

Its actually really small, mostly because bigger networks take longer to evaluate which slows down the search making it shallower and ending in a less clever algorithm.

Are you involved in the project ? Can I ask what your source is? Great if it's the case.

I am involved in lc0 development and fairly aware of SF dev. NNUE is a very small (3 layer dense) cpu only net.

NNUE is a 4 layer (1 input + 3 dense) integer only neural network.

It's just over 82,000 parameters.[1] That's a very shallow, small NN - by comparison something like EfficientNet-B1[2] is 7.8M parameters, and that's considered a small network.

[1] https://www.chessprogramming.org/Stockfish_NNUE#NNUE_Structu...

[2] https://proceedings.mlr.press/v97/tan19a/tan19a.pdf

You can have a relatively small model and still benefit from using a gigantic training set to train the model.

> They say that the dataset is hundreds of Gigs worths of games, so the net must still be really pretty big.

This isn't true. The size of the training data doesn't imply anything about the size of the neural network.

In the case of Stockfish, the NN is quite shallow, and implemented using a custom framework designed to to run fast on CPUs.

See https://news.ycombinator.com/item?id=26746160 for previous commentary on this.

> Though definitely not directly comparable, dataset of GPT2-xl is 8 million web-pages.

This is irrelevant. You can train GPT3 on a smaller dataset, or a smaller model on the same dataset as GPT3.

> What I mean to say is that this is clearly deep learning.

It's been clear that neural network models are superior since Alpha Go. There's not "Deep Learning vs <something else>" anymore because the <something else> isn't competitive and no one is really working on it.

You can find out more about Stockfish's NNUE architecture and internal working here:

https://github.com/glinscott/nnue-pytorch/blob/master/docs/n...

It's pretty interesting read.

Reply on news.ycombinator.com