Decision‑time search has become an effective technique for game AI in both perfect‑information and imperfect‑information games. Collectible card games are imperfect‑information games with enormous belief‑state spaces; Legends of Code and Magic (LoCM) has $2^{101}$ possible belief states. The LoCM champion, ByteRL, plays without any search, and prior work argued that sound enumeration‑based search is infeasible for this genre. We measured three previously defined properties that predict when theoretically unsound perfect‑information Monte Carlo defects are cheap, and found LoCM lies in the favorable region.
Starting from imitation learning of the runner‑up policy NeteaseOPD, we trained a feed‑forward policy network and a value network. Our agent samples worlds from a prior over the opponent’s deck, built from the runner‑up’s draft data, and searches over these worlds. Using the strictest configuration in the battle phase, the agent achieved a 51.35% win rate (95% CI $[50.37, 52.33]$) against ByteRL over 10,000 pre‑registered games under the official LoCM referee and time limit.
Search proved to be a major factor: without search the agent scores 26.8%, and adding search contributes roughly +24.6 percentage points. Although unsound search in imperfect‑information games could be exploitable, we replicated a published best‑response attack against ByteRL and applied the same protocol to two search configurations of our agent. In every iteration our agent resisted the attack better than ByteRL.
Thus, in LoCM unsound search yields a stronger and more resilient AI.
Review: The work demonstrates that, even with astronomically large belief spaces, combining policy/value networks with unsound search can substantially improve performance and robustness compared to a no‑search baseline.