Solving Google's Tic-Tac-Go puzzle turned out to be a lot harder than it looks. I went through several approaches — breadth-first search, reinforcement learning with deep Q-networks — before landing on the insight that solving complex search problems rarely comes down to a single perfect algorithm.
The final solution combined heuristic search, beam search, and a supervised CNN, reaching a 93% solve rate on real puzzle boards. Full writeup on Dev.to below.