What are the strategies of Alphazero? [closed]

Question

Alphazero is an amazing chess engine, one of the best in the world. It also has a more "human" playing style than other engines. Furthermore, understanding this engine might be vital to improving the way we play chess. Of course, one might argue that engines play the best move using brute force calculations and combinations. Human just can't do that. However, I argue that there are some clear strategies of alphazero. These are just some of them.

1. fianchettoing bishops
2. controlling space and giving opponents no moves(pawns)
3. forcing the opponent to create creating weaknesses (checkmate threats) and taking over the holes in the position. (when pawn g6 is played to stop mate, the dark squares are a weakness.

There are much more than these strategies. I have not yet seen how Alphazero defends, for example. Also, I would request u to go to this link (https://www.chess.com/forum/view/general/alpha-zero-vs-alpha-zero-10-lesser-known-alpha-zero-lines). This link shows the openings when alphazero plays both white and black. It would be nice if somebody could add understanding of it.

I know this post seems outdated, but Alphazero is so ahead of its time. It is still the best engine! I think we should understand it and help the chess community.

Answer 1

I don't think there are any "strategies" in the human definition of "strategy".

When you use the "common" computer engine (basically, min/max tree search), you can always check how and why the engine arrived to some particular position evaluation, but still, this is usually limited to 10-12 half-moves, beyond that it's still hard for people to grasp what was the threat, or what is the forced mate or why this particular line evaluated higher than another line: "Mate in 20?" -- "Yeah, I saw that one a few moves back" =)

When you talk about AZ or any similar NN-based software, it becomes practically impossible to reason about the moves. Once the training is over, all you get is the bunch of multi-dimensional matrices, which, once multiplied/added, give you some answer that leads to some policy chosen in this particular position.

And this behaviour is not specific to chess only. Same things happen in image recognition, pattern analysis, language processing or any area where NN are in favour last few years. You get the results, but not the explanation.

So, right now, I would not recommend to blindly follow any of the AZ patterns, because no one has a clue, er... clear understanding, why and when those patterns should be followed. It might be as minor as "if black pawn is on h6, fianchetto bishop on b2 (or g2, or both)" -- we'll never know. Not in the nearest future, at least.