Using Freeware AI and Dynamically Generated Endgame Tablebases

[zug234zwang]

A neural net can make mistakes with code that is entirely bug-free ...

So if the programmer+grandmaster can't explain the losing move it might not be a bug in the code. However, it could still be a coding bug, couldn't it? Further, if the programmer+grandmaster feel that the Neural Net has so many games (i.e. so much data) in its history that the losing move is very surprising, would that not suggest that the problem is a coding bug rather than insufficient learning?

A sample bug might be if the original code indicated that the two bishops are a disadvantage rather than an advantage.

Perhaps I am mis-interpreting your response and mis-interpreting what the Chess-AI is doing. If a subtle bug was made, would the Chess AI's learning be so self-correcting that eventually the coding bug would be neutralized? Does this mean that in such a situation, the programmer would assume that further AI-learning would neutralize the effect of a coding bug?

[brianr]

Long answer:
I don't think you appreciate how the nets learn, particularly the so-called "zero" approach.
Zero here means almost no chess knowledge at all.

The nets are not told what a bishop is, much less that two are better.

Nobody, including the programmer, knows how many games/samples are enough (although there is something known as over-fitting, but that would not apply here).

In more general terms, the net training code does not even know what game it is playing.
The same framework is and can be used for almost any game.

At the most basic level, it is very roughly like training a net to do image recognition.
It looks a many pictures and learns to predict that a particular image is "a cat" with some probability.
Think of the chessboard like an image and it learns to predict the "best" move given that image and that move is sort of like the label cat for the more general images.

There is a quite humorous example of this from the Silicon Valley show where they develop a phone app that can predict if it is shown a hot dog or not. There is code out there you can look at and try. Chess is pretty complicated, but the principals are the same.

Short answer:
It's magic

[zug234zwang]

I don't think you appreciate how the nets learn, particularly the so-called "zero" approach.

The smoke clears; thanks for explaining this. Very interesting subject.