Question about nnue

[jmartus]

So is every around the 3,000 elo mark about to gain 300 plus elo enough almost automatically with just implementing nnue? So in a few months every engine going to have nnue it seams or lost to the olden era?

[mar]

if they clone then sure and not just 3k elo engines. in fact, lower rated engine would likely gain more, assuming that top engines should already have very decent evals

the critical part that I'm missing is how NNUE is actually trained - fitting to existing eval might serve as a better starting point than small random weights, but how to continue.

one could do something akin Texel tuning (logistic regression), but then you have to train a huge 20m-weight network on tens of millions of positions, not sure how long this would take.

those who are curious will try to understand and likely come up with their own topology and new ways to mix NNs into existing eval other than completely replacing it with "NNUE", which I think is a good thing. it's also very likely that SF devs will find ways to improve further. I think this is just a beginning of something exciting. seems that AB engines are not dead yet after all

[towforce]

So is every around the 3,000 elo mark about to gain 300 plus elo enough almost automatically with just implementing nnue? So in a few months every engine going to have nnue it seams or lost to the olden era?

Yes. Apparently, there are important aspects of chess which humans writing evaluation functions simply haven't been able to capture.

[Tony P.]

We'll see. I do think Arasan needs a neural upgrade to stay competent; fortunately, its elaborate position scoring routines are a natural fit for the preprocessing of positions before they're put into an NN. If I'm not mistaken, the rest of the active top-20 engines (except Booot that's hopeless because Object Pascal has no NN inference infrastructure) already have or are developing neural elements, and far not all of them will be copypastes from nodchip's SF.

[jmartus]

How will that change the style of engine going forward? Will they all become more similar in style?

[towforce]

How will that change the style of engine going forward? Will they all become more similar in style?

Nobody knows yet. However, it probably will if:

1. the knowledge that nets capture that handwritten evaluation functions miss leads to "good moves"

2. in many positions, there tends to be one move which is better than the others

Then it will lead to "the best move" being picked more often, and this may well lead to them having similar "styles"

[jmartus]

Not sure if I like that but always have open tal 1.0 and other stylish engines to play around with as options