I rebuilt AlphaGo's architecture for a game of hide-and-seek (Graph Neural Network)

Executive Summary

5 detectives chase an invisible fugitive on a 199-station graph, seeing only which ticket he plays. I trained both sides. The detectives are an R-GCN that eats an HMM-style belief state plus distance maps. BC warm start, then PPO: 25% → 79% win rate against the same MCTS Mr. X. Then Mr. X got a GNN ...