Paradigm shift

[hgm]

It seems that computer chess is experiencing a paradigm shift, but that not everyone already realizes that. There was a time when it was all important how fast you could generate moves, and Chess programs mainly differed in the way they represented the game state and generated moves from that. In the past decade that has hardly be an issue, though; virtually all top engines do it in exactly the same way, and the strength is mainly determined by how well they manage to reduce the late moves. Evaluations were as light as possible.

Now we are entering an era where search is hardly an issue; everyone again does this in a very similar way, which is very hard, if not impossible to improve on. Suddenly the only real difference between top engines is how well they evaluate. As really good evaluations only seem possible with neural nets, designing and training better nets has become the name of the game.

So I think that people who say 'Fat Fritz 2 is just Stockfish, with one tiny parameter changed' are missing the point. The evaluation has been thoroughly changed, as it uses another net, with another size, and different training. In the future that will likely be as large as differences between top engines will ever get. Search will be just as dull a topic as bitboard move generation is now: a solved problem. You either do it as it is known that it should be done, or your engine will suck.

The future holds large (rating) lists of engines that all use the same search (be it verbatim copies of each other, or re-implementations), but are still as different as night and day, because they use different nets.

[Fulvio]

If this were true, there would be a wonderful network that produces amazing results even when used in a dozen different engines. But since we have the opposite, a wonderful engine that produces amazing results with a dozen different networks, I do not understand on what basis your reasoning is based.

[Daniel Shawul]

I wouldn't say search is obsolete now because it is very hard to write a search as efficient as stockfish's. If it was that easy, other engines like komodo
would have done it before NNs arrived to the scene. That is why I think treating stockfish search as a "net runner" is plain wrong.
The proof is that there is not another NNUE engine within 100 elos of stockfish other than Komodo.

The hype about NNUE is simply because stockfish is stronger now, but there really is no innovation there. It is a tiny fast net accelerated by vectorization for the CPU. The chess community is so bad that if you give them a +10 elo engine over the current best chess engine they loose the plot.

It sounds to me like we could be back to the problem of optimizing an engine that is stuck in a local optima, instead of looking for radical alternatives that could bring real paradigm shift like A0's again. That was the real innovation that reduced any game playing to a button click solution of training a net. It is suddenly second rate now for some -- people saying end of an era for NN engines now. IMO NNs are still much better than NNUEs because the former also replace the search with a generic one that is competitive with classical engine's like Stockfish. I still can't believe it to this day, as I am sure anyone who has worked on his AB engine for years would ...

[Raphexon]

Now we are entering an era where search is hardly an issue; everyone again does this in a very similar way, which is very hard, if not impossible to improve on.

If search is hardly an issue, how come Stockfish is leaps and bounds beyond any other (AB) chess engine? Even when you give those other chess engines the exact same network?
Also the last 5 years SF search has been easier to improve on than its (handcrafted) eval.

NNUE changed this for a short while (4 months) but now we're back to a bunch of search patches for every eval improvement.

In fact despite continuous training there hasn't been a better NN for Stockfish since november last year.
If search is the solved problem, why is it showing the most improvement?

So I think that people who say 'Fat Fritz 2 is just Stockfish, with one tiny parameter changed' are missing the point.

And they probably claim you're missing their point...

[Raphexon]

It sounds to me like we could be back to the problem of optimizing an engine that is stuck in a local optima, instead of looking for radical alternatives that could bring real paradigm shift like A0's again. That was the real innovation that reduced any game playing to a button click solution of training a net.

(Incrementally updated) policy head that's fast enough for AB-search could be the next gamechanger.
If move ordering can be improved as much as eval for a similar slowdown, that would make AB incredible even for games with a huge branching factor.

[Guenther]

Now we are entering an era where search is hardly an issue; everyone again does this in a very similar way, which is very hard, if not impossible to improve on.

If search is hardly an issue, how come Stockfish is leaps and bounds beyond any other (AB) chess engine? Even when you give those other chess engines the exact same network?
Also the last 5 years SF search has been easier to improve on than its (handcrafted) eval.

NNUE changed this for a short while (4 months) but now we're back to a bunch of search patches for every eval improvement.

In fact despite continuous training there hasn't been a better NN for Stockfish since november last year.
If search is the solved problem, why is it showing the most improvement?

So I think that people who say 'Fat Fritz 2 is just Stockfish, with one tiny parameter changed' are missing the point.

And they probably claim you're missing their point...

Totally agreed, people completely underestimate SF's for a long time refined search.
CB and AS knew it though...