The Gathering Breaks AI

Limits of Poker AI and State Space 00:00

  • Research on AI in games like poker has been advanced due to manageable state spaces with limited hidden information.
  • In heads-up no-limit Texas Holdem, there are 1,326 possible states due to two hidden cards per player.
  • When playing six-handed poker, the number of possible hidden states increases linearly with more players, but remains computationally tractable.
  • AI for poker works by enumerating all possible states, assigning probabilities, and using neural nets for action selection.
  • As the number of hidden cards increases (as in Omaha poker), or in games with exponentially more possible states (like Stratego with 40 factorial states), these poker-based approaches become unworkable.

Challenges for AI in Complex Games Like Magic the Gathering 01:36

  • Existing poker AI techniques do not work "out of the box" for games with vast state spaces such as Magic the Gathering or Stratego.
  • There is ongoing research on coping with high-complexity hidden-information games, as enumerative approaches break down at large scales.
  • Model-free reinforcement learning (RL) does not suffer from the same state space explosion and may be adaptable for superhuman play in Magic the Gathering.
  • The more promising avenue is developing general reasoning AI approaches that are not restricted by the limitations of poker-style search techniques.
  • With progress in general reasoning and RL, an AI capable of superhuman Magic the Gathering gameplay is expected to emerge eventually.