Terrible Go player and right now I don’t play enough to get better. I did make the Spectral bot you can find here on OGS though. The goal was to play human-like games at any kyu rank on minimal hardware.

Spectral still falls short by messing up ladders, not playing at lower kyu ranks, and not quite fitting in RAM on a single Raspberry Pi or free Google Cloud instance (although any individual rank is fine). I hope to fix these problems eventually.

If you’re interested in the technical details, the bot is essentially a 10 block, 128 filter version of the supervised benchmark in the AlphaGo Zero paper. This means it has a stack of convolutional layers with residual connections and a policy and value head at the end. A single network was trained on progressively higher-level games from OGS (500k in total) and finally on professional games from GoKifu (50k). The networks which play at different ranks are from different points in this training progression. Hardware used was a Tesla V100 GPU, and the entire training sequence took around 15 days. I hope to make the code and maybe the data for this project available eventually.