A reinforcement learning study found that training AI agents on the card game Big 2 offers a clean testbed for one of the field's harder problems: acting under hidden information.
Researchers built a self-play framework around Big 2, a four-player card game where players can't see each other's hands. They held constant the environment, input representation, training budget, and evaluation protocol, then ran four algorithms against each other: PPO, Monte Carlo Q approximation, SARSA, and Q-learning. PPO won. It outperformed the others against random, greedy, and heuristic opponents. Two additional findings stood out: moderate entropy regularization kept PPO from locking into predictable patterns too early, and training against the current policy version of itself produced better results than training against saved checkpoints or fixed opponents.
Most high-profile AI game research gravitates toward two-player zero-sum games like chess or Go, where the math is cleaner. Big 2 introduces complications those settings don't: four players, hidden cards, sparse rewards, and a variable set of legal moves at each turn. A benchmark that captures all of that in one game has real practical value for researchers who want controlled comparisons without the overhead of poker or Mahjong environments.
The honest caveat is that Big 2 remains an academic sandbox — the agents here beat heuristic bots, not human experts — so the gap between "useful benchmark" and "solved problem" is still wide open.