robocode-tank-rl

MuZero-family policy for Robocode Tank Royale 1v1 (ndriggs-bot), trained with LightZero via league self-play in a bit-exact Python re-implementation of the Tank Royale v1.0.2 physics.

Current: {"mode": "policy", "obs_mode": "threat", "algo": "muzero", "num_simulations": 24, "frame_skip": 2, "num_actions": 45, "obs_dim": 35, "source_ckpt": "gen6.pth.tar", "note": "gen6 @102k: pure 1v1 (knockout), 4-6 vs joshmoody PR3 (up from 0-10)"}