June to November 2026
An AI opponent for SETI: Search for Extraterrestrial Intelligence, a strategy board game I love playing.
- Rewrote an existing web implementation of the game as a headless Rust engine, so that training runs fast. Along the way I caught and fixed 45 bugs in the original implementation.
- Combined game knowledge with ML techniques (expert iteration, weight tuning and an MLP leaf evaluator) to guide MCTS and beam-search agents that run on limited compute.
- The agents reproduce the scores of experienced players across games and positions, and compete evenly against a from-scratch self-play agent that was trained on a 64-core cloud server for a week.
The code is private for now.