Read on Twitter
And now for my highly controversial take on the current state of computer chess (thread)
Well outperforming expectations Stockfish, which has a human hand-made heuristics with automatic testing and adjustment, is still stronger than Leela, which uses deep learning for board evaluation.
But this doesn't mean that deep learning isn't helping Stockfish! A lot of what's been happening is that in Leela versus Stockfish games Leela finds weaknesses in Stockfish's board eval to exploit and these artifacts are then observed by human developers, who \\
are then inspired to suggest tweaks to Stockfish which then go to fishtest. The result is an artificial intelligence to natural intelligence to machine learning pipeline without which Stockfish would be much weaker than it is today.
But it isn't possible to cleanly capture everything in simple heuristics. Here's a truly astounding example of a game which it's very hard for traditional engines to understand
On the other hand it's clear that as you get close enough to the endgame the more efficient board eval of Stockfish lets it essentially brute force the rest of the game and the more expensive board eval of Leela is simply inferior.
Most likely the absolute best engine would have some kind of changeover where it uses a more expensive board eval earlier then switches to the cheap one in simpler positions. Leela already has a big concession to this with using tablebases.
(Tablebases are giant lookup tables for what the true positional value is in all positions. Engines currently use 6 piece ones. 7 piece ones have been made but don't fit on a single machine. Watching computer games without them is a terrible spectator experience.)
I'd previously prognosticated that deep playouts in monte carlo tree search were the real secret to deep learning being better because of its better differentiation between draws and wins. This seems to be untrue, but in a way which has only deepened the mystery.
As self-taught deep learning engines have gotten better they've tended to prefer fewer deep searches. This has made their win rate better but resulted in them being poorer at differentiating draws from wins.
Bullet games can still do a better job than the engines of making that distinction, so there's a known artifact of current engines with a simple way of improving it but trying to apply that in practice seems to make overall strength worse, which is a very frustrating situation.
The fix might be something as lame as making the engines have separate draw probability from their metric of which side has the advantage. It would be relatively easy to add this to deep learning engines (but there is a menu of techniques with no obvious winner) \\
but not so obvious how to do it for Stockfish-style engines. This wouldn't be sacrificing the 'zero' part of engines being completely self-taught, but rather explaining to them fully what the actual rules are. Clearly there's work to be done.
And while I'm giving developers advice they didn't ask for: Some of the alternative deep learning engines are doing well based entirely off training from good human games. That could be made to be a 'zero' approach by having an engine play itself at long time control games then \\
train a new neural network to match the play in those games, then have that improved engine play itself, and rinse and repeat. That approach has the benefit that you could force it to play out lots of different openings and get experience with a variety of positions which \\
both would keep its eval from getting stuck in a self-reinforcing rut and result in lots of fun games to go over as part of normal engine development.
Training to computer games also has the benefit that you not only know the preferred moves but the preferred evals as well. Computer chess is far from dead and there's lots more stuff to try.
