AlphaZero Connect Four, trained from scratch
Residual policy-value network (374,810 parameters) after 34 self-play iterations. Load with build_net() from train.py. Scores against a random mover, a one-ply win/block heuristic and the untrained network are in history.json.
Final-iteration scores (win = 1, draw = 0.5): {"random": 1.0, "one_ply": 1.0, "uct1000": 0.9453, "iteration_0": 0.9844, "policy_vs_one_ply": 0.5391}