What I would like to know is, have neural networks been employed to generate chess *problems*. I have in mind two-movers, three-movers, etc., and endgame studies.

For me the difference between an easy problem and a hard one shows up in the time it takes to solve it. For easy problems I can find the key in a few minutes. For hard ones it takes hours; sometimes I give up before finding it. It seems to me that, mutatis mutandis, a neural network could create chess problems of diabolical complexity -- problems that would take even the best solvers days to solve without resorting to computerized assistance.

I found Iqbal's article disappointing. I hoped to see real chess problems composed by AI. The only composition shown (the miniature 3-mover at the top) is trivial.

<We have found many cases where its preferences are not stable over different training runs. We describe one such example in detail, a very important theoretical battleground in top-level human play.>

https://www.pnas.org/doi/10.1073/pn...

(Thanks to <CCastillo> for calling attention to that link over here:

I believe the answer is basically yes. They used it to attract media attention, and as a proof-in-principle demonstration of AI before moving on.

I guess AlphaZero was retooled to become the more general MuZero:

https://venturebeat.com/ai/deepmind...

I think you'll find a good sampling of Deepmind's latest research endeavors here:

https://www.deepmind.com/research

It seems that protein folding is one of current interests.

I wonder if continued training would have found and fixed that weakness.

As far as having little reason to continue training, has AlphaZero answered the question of whether or not White has a forced win yet?

I have a large suspicion that AI is mainly dependent on the power/efficiency of the hardware. AlphaZero and I'm assuming AlphaGo apparently used brute force computing power to in essence find what worked best in almost any situation.

I'm guessing that it was also brute force computing power that targeted in on finding AlphaGo's weakness. Likely stronger power than the Alpha's had.

Of course, AI developers and hypesters aren't about to reveal that 80% (making up a number) of what they produce is due to the hardware.

The AlphaZero algorithm is published in the literature and anyone can read it and see it's not just the hardware, let alone 80% (by some reasonable measurement we'd agree on).

To be precise, it does take powerful hardware, but it's not just using it to achieve deeper complete lookahead (what we generally call "brute force"). If you could somehow reconfigure AlphaZero hardware into an equivalent amount of hardware (again by some measurement we'd agree on) of the top pre-AlphaZero engine, it doesn't mean that engine then plays like AlphaZero.

<Links, we need links!>

OK, allow me - here's one:

https://the-decoder.com/google-deep...

With this additional info:

<

However, difficult chess puzzles still baffle even the strongest chess AI systems, suggesting room for improvement. Researchers at Google DeepMind are now proposing to combine several different AlphaZero agents into an ensemble system, called AZdb, to further improve its capabilities in chess and beyond. AZdb, combines multiple AlphaZero agents into a "league".

<AlphaZero agents are inspired by human collaboration>

Using "behavioral diversity" and "response diversity" techniques, AZdb's agents are trained to play chess in different ways. According to Google Deepmind, behavioral diversity maximizes the difference in average piece positions between agents, while response diversity exposes agents to games against different opponents. In practice, this also means that AZdb's agent will get to see many more different positions, expanding this range of in-distribution data, which should allow the system to better generalize to unseen positions.

But the decoder article was noted on <reddit> from two years ago:

https://www.reddit.com/r/TheDecoder...