Hyperparameter tuning

[andyljones]

All the hyperparameters right now were arrived at through 'grad student descent', which is to say they're garbage. Top of the list to tune:

• Main: buffer length, batch size, env count
• Opt: learning rate, warmup duration... beta1 & beta2? Do I need momentum at large batch sizes?
• MCTS: sim count, c_puct, Dirichlet noise intensity and concentration
• League: field size, stable size, prime/challenger split, prior intensity
• Network: width, depth.

This should all wait until I've got a general algorithm I'm happy with though.

[andyljones]

I did a lot of tuning ad-hoc, and the current settings have been ossified by the fact I've done a bunch of experiments with them. Might want to vary them and do re-reuns in a future, better-resourced version of this project. Might also need to do some light variations to check we're approximately at optimal - the fact we hit quickly perfect play on 9x9 suggests we are, but reviewers will likely disagree.