Regarding Deep Learning/NIM

[Harvey Friedman]

How well can the new programs which do so well with Chess and Go and
visual recognition, and so forth, play NIM?

I think the answer is probably not perfectly, which means that they
pale in comparison to humans.

But then might it be interesting to start with small parameter NIM
initial positions and gradually raise them to much larger NIM initial
positions. And see when the program does discover the perfect
strategy.

There are two obvious parameters. The maximum number of stones in each
pile. The number of piles.

Probably the number of piles should be investigated first.

If the number of piles is 2 then I would assume that the program would
very quickly learn the winning strategy.

What if the number of piles is 3? Do these programs play this
perfectly? And 4,5,6,... ?

Say it doesn't play n piles perfectly. Now use the other parameter,
the maximum number of stones in piles.

QUESTION. How can we design a program that can play perfect NIM with
only being given the rules of play?

QUESTION. How good is this question?

Harvey Friedman