Search speed vs evaluation complexity

[osvitashev]

My understanding is that fast search combined with fast and simple evaluation function has won out. Engines which rely on things like piece-square tables seem to be dominating because they can search some insane number of plies.

I am curious if there are any examples from the other branch of this evolutionary fork.
Are there any strong modern chess engines which examine (relatively) few plies and instead invest the computational time into complex evaluation of the few positions they do examine?

[Steve Maughan]

I believe you're incorrect. The latest engines focus on two factors: a highly selective reduction-based search coupled with a neural network to evaluate the positions. Neutral network evaluations are far from simple and are relatively expensive to calculate. The NN evaluation is better than any hand crafted evaluation.

— Steve

[Ciekce]

to expand on "better than any hand-crafted evaluation" - for the majority of HCE engines (so, anything that's not stockfish classical) implementing a decently strong NNUE is *multiple hundreds of elo*.

[Jouni]

Current SF speed is about 40% of HCE version. But +300 Elo better.

[syzygy]

My understanding is that fast search combined with fast and simple evaluation function has won out. Engines which rely on things like piece-square tables seem to be dominating because they can search some insane number of plies.

Wrong. NNUE engines clearly outperform other alpha-beta engines due to their superior but heavy evaluation. Not to mention GPU-based engines such as LC0.

[Frank Quisinsky]

Hi Jouni,

Stockfish 200731 dev (last HCE) is around 160-170 Elo stronger as Stockfish 16 NN.
With longer time controls, not more!

And I believe again 20-30 lesser if I double the time control I am using so far for engine-matches (at the moment I am testing with 90 and 150 minutes games).

Best
Frank

[smatovic]

Stockfish NNUE make some million NPS on CPUs, Lc0 CNN makes some ten kilo NPS on GPUs. Initial SF NNUE had about 10 million weights, initial A0 had about 50 million weights. So here we have a knowledge:search trade off with neural networks.

--
Srdja