there must be a better way

[ddyer]

Regarding using neural nets for games. Doesn't it seem odd that we take a domain
of perfectly discreet objects - black stones, white stones, squares on the board, and
re-represent it as floating point numbers with thousands of parameters?

It seems like there ought to be an approach that uses similar ideas for training and
evaluating neural nets, but uses discreet functions to define activation levels.

[Rémi Coulom]

You can use Binarized Neural Networks:
https://arxiv.org/abs/1602.02830