How does AI learn chess?

Hey,

It probably depends on which AI we talk about, but I assume that they just play each "turn" up to x depth to determine the best move + some special cases.

I was wondering...Since humans can learn in chunks?(openings, mid game, end game)
Can't AI also learn ONLY openings then mid game and lastly end game?

I saw a post somewhere stating how Alpha Zero(or some other AI) played 1 billion games and someone mentioned that it still is far from reaching Magnus Carlsen in terms of games played(meaning that Carlsen learned more with less games played).

But the problem with this statement is that humans don't just play chess, they study from books, think in their head and so on.

Doesn't that mean that the AI would have to do similar in order to be compared to the human in terms of games played?

So this is where I thought about having AI focus on specific parts of chess game, such as openings/end game.

And maybe even most common openings/positions.
Instead of doing it all(I assume that AI training program is much more complex and they probably do something like this)

What do you think?

The

a chess forum is probably not a good place to learn about AI and stuff. Most of us don't know how to program and the ones who do know probably don't know how machine learning works.

That isn't entirely true. Most historical engines work that way, but newer neural net based engines absolutely learn how to play the game. Alpha Zero is the most famous example; it was just programmed with the piece movement rules and learned how to play by playing itself.

Mu Zero learned with no rules.

https://deepmind.com/blog/article/muzero-mastering-go-chess-shogi-and-atari-without-rules

Thank you for sharing that interesting and informative article.   I did keep abreast when Alpha GO and Zero were mastering GO.  There were some excellent informative articles about them.  It was fascinating to see how the AI learned.  In GO literally from playing stones on the first line and seeing them die to more complex situations. As I recall it took a few days too master GO.  

It is ironic that just the year before that, the best computer program could only play at best at a strong amateur level.  

https://mlinsightscentral.com/ with code. Learn Machine Learning. #Suscribe

Engines, such as Alpha Zero, learned by playing a ton of games. While it can't verbalize meanings as it wasn't designed for that, it was able to learn concepts and apply them in games.

It's not unreasonable that program could be taught to explain those concepts as well.

I think Stockfish uses a neural network for the static position analysis too. It would be way more effective than tweaking it by hand.

IT uses knowledge from online apps and video and games and it uses that information to move a piece

It would be possible teach an AI concepts from master players and/or texts and take lines/evaluations to help explain how those concepts apply to those lines.

I won't claim it's easy to do, just that it's theoretically possible to do via AI processes in combination with other systems.

Martin's correct. This forum isn't the best place to discuss it but if you want to learn more then you should start here: https://en.m.wikipedia.org/wiki/Explainable_artificial_intelligence