T7G Microscope

This is The 7th Guest's "Microscope" puzzle, with the game's own AI - Stauf - extracted and made to run in the browser.

You can use this to beat Stauf in the game itself. Set yourself to Green and your opponent to AshnasBot - Stauf plays Green in T7G, so here you take his side. Play AshnasBot's moves in the real game, then enter Stauf's replies back here.

It started as a mini-solver for the puzzle. Stauf is deterministic, so it wasn't too hard to out-compute with a decent minimaxer, and I moved on to more exotic engines - working through PPO and AlphaZero before ending up on the current engine, a reasonably modern AlphaZero variant with a few upgrades. The search uses a DAG instead of a tree, due to the massive move symmetries in this game, and the non-net components are almost all in optimised C for performance.

The model has trained on roughly 270,000 self-play games, and appears to be close to saturating the signal we can squeeze from the game.

I'd love to hear from you if you beat it!