Browser demo of our Clash Royale RL environment: a 5.6k-parameter REINFORCE policy learns defensive placement against a brute-force optimum [P]

A browser-based demo showcases a 5.6k-parameter REINFORCE policy learning defensive placements in a Clash Royale RL environment. The developers observed that a strong local optimum exists for 'Giant vs Cannon,' with 5 out of 6 runs getting stuck there using a constant entropy coefficient of 0.01. A linear anneal from 0.1 to 0.005 over 10k tries improved this to 1 out of 6 runs. However, the 'Battle Ram vs Valkyrie' pairing proved challenging, with no setting achieving more than 55% of the optimum.