Sandbox for playing with neural nets for Go
Value nets, implementation of MCTS and a real Go engine!
Attached are neural net weights trained from ELF and LZ105 to LZ142 from Leela Zero data, but using the architecture in this repo with 12 blocks, rather than LZ architecture.
Same as v0.1 except with some minor bugfixes - rank arguments to some scripts are actually optional now and the GTP play.py script properly passes move history to the neural net (it didn't before).
Edit: Also uploaded the raw files used by the webserver at neuralnetgoproblems.com, since at least one person requested them. If you want to use them, you'll probably have to write some code to turn them into a format you can use, they're in a packed python pickle format rather than some standard database format.
Mostly finished experimenting with policy net architectures and with incorporating player ranks into the prediction, so going ahead and tagging this revision as a release, before I move on to other experiments (value nets!).
Attached are the weights from one of the final models. It should be a reasonably strong policy net, and can also make predictions for different player ranks besides pro. Yay.
Edit: Use v0.1.1 instead for some bugfixes.