Knights or Bishops? What AlphaZero Thinks

No need for the thought experiment since we were already on the same page here. My argument mostly on this was that we cannot claim the values  objectively. The values are for aiding in calculation. However, it would be interesting. I'd love to see my conclusion of the piece values. I'd be curious on my own innovation of the values that I may put on each pieces.

He'd probably become as terrible as Hikaru or even as terrible as GothamChess... unless someone from his team helps him to re-learn the values.

Nice. I am also self-taught although I did learn basics like piece values from the internet, but I never read a book and the only opening I learned in my whole life is the maxlange opening that stems from the italian opening. I also favor knights over bishop. It's just a preference thing. 

Does your dad got a FIDE rating?

I was not aware that conventional engines "completely destroy AlphaZero every time"; can you provide proof (matches between A0 and conventional engines)? And if I am not mistaken, top engines like Stockfish, Dragon or LC0 do not use an evaluation function "adjusted" by humans, but a self trained NN. Of course, the developers have to create the NN in the first place, to program the board representation, the search function (usually alpha-beta), etc. As of Knight vs Bishop, I too agree that their value is about the same but it depends on the position (closed or open, Knight on outpost or not, good or bad bishop), and that modern masters in general only favour the bishop pair, which is usually evaluated about 50 centipawns more than two light pieces.

I see. I am sorry for your loss. 

Thanks for the compliment, but I did play over 10 thousand chess games in my whole life and that's excluding the blitz games or bullets. So I guess it’s mostly the result of my experience. 

You are very much mistaken. They use human guided evaluations and checking the trees of variations directly, that's the entire way they work. Of course it's mostly the checking variations (brute force) but there is still a lot of human intervention in figuring out what to prefer or what branches to cut off at what point. I have no idea know why you would think they use a "self-trained NN" and nothing else. That's what Alpha Zero does and it does not compete in any computer competitions. Alpha Zero is an experiment and a bit of a gimmick. No serious competitive computer engine does this or works that way. I hope you understand now and stop talking nonsense.  

Please, don't be rude. Of course there is a lot of human intervetion, for example to make the engine decide weather to go deeper on a branch and how to manage time; you can tune up the engine to be more aggressive, or more selective, etc. But we were talking specifically about the evaluation function that every engine uses on leaf nodes. Traditional engines use a crafted-by-humans function, while others use an NN or an NNUE. Indeed, the NNUE used by Stockfish cannot be considered a "zero" network, because it is trained on positions already evaluated with the traditional evaluation function, but LC0 uses a NN like A0 does:

" Traditional chess engines have a very-filnely-crafted-by-humans value and policy generation system; unlike traditional engines, Leela uses its neural network trained without human knowledge for both value and policy generation. Then, Leela expands the tree to get a better understanding of the root node, the current position."

https://lczero.org/dev/wiki/technical-explanation-of-leela-chess-zero/

As of A0, the very fact that it does not compete cannot make you state that traditional chess engines "destroy it every time". The only matches of which I am aware of are the two against SF, both won by A0, especially the second one. Of course, SF has improved since then, but we cannot say that the evaluation functions of LC0 and A0 are crap. By the way, returning to the topic, many strong traditional engines do not value the bishop much better than the knight; AFAIK often it's a matter of few centipawns.

Regards

I feel like you are still framing it wrong. Alphazero would not be a match for traditional top chess engines like Stockfish because otherwise why would they refuse to ever compete? LC0 incorporates both traditional chess evaluation techniques and neural network techniques, it's a hybrid. 

Recently, as in July of 2020, Stockfish started to incorporate neural networks in it as well to help its evaluation function. However 1) it was already at the top in 2020 before it started to do this and 2) it is still predominantly a traditional chess engine that has been created by human intervention.   

https://en.wikipedia.org/wiki/Top_Chess_Engine_Championship

https://en.wikipedia.org/wiki/Stockfish_(chess)

So if OP wants to talk about something, he should be talking about these engines. AlphaZero sucks compared to these, that is my point.  

I see your point, but I don't know why Deepmind "refuses" to compete; maybe they are just not interested in spending more resources in the developement a chess engine. As of LC0, the developers clearly state that the NN does not use human knowledge to evaluate positions and to prioritize moves, hence the "Zero" in the name. You can also check the Wikipedia page:

https://en.wikipedia.org/wiki/Leela_Chess_Zero

Anyway, as for the OP, to me the fact that A0 considers the Knight roughly the same value as the Bishop, just confirms what we already knew by experience and in fact often there is no big difference between N and B in traditional evaluation functions, too.

I agree about how it's fascinating that knights and bishops are relatively equal in value, even though their movements are so different. It's an interesting point to note.

Most programmers agree that the latest SF is superior to AZ, based on AZ's performance against those earlier iterations of SF.

But I also find it likely that, if AZ were left to continue self-training (instead of being stopped at a million self-training steps, or wherever it was stopped), it would likely continue to find marginal, incremental improvements in strength, just as SF has done with each new version.

The most recent version of AZ is actually "MuZero", which mastered chess without even being taught the rules. It was just shown the board and pieces, and was given no other information (except for being informed when an attempted move was illegal).

Even given no information at all, MZ still taught itself chess and reached AZ's level of mastery in a short amount of time ...

MZ also found improvements over AZ, in the various other games it was applied to.

So it's possible (likely?) that MZ also found chess improvements, too. Apparently, having to discover the rules of the game somehow yields greater ideas, in the eventual gameplay ...

 i think it all depends on how they are placed and coordinated with other pieces, but overall bishops are stronger

Since i will never play anyone or anything rated 3800, it makes no difference to me.