It’s been a while since I looked into this, but when I did, the big problem was behavioral modeling.
All the math in poker is local—the state of the game in one hand influences the state of the game in the next only insofar as it affects the amount of chips players have on hand. That makes it easy to build a poker AI that’ll wipe the floor with innumerate players. But in also narrows the scope enough that it’s possible for more sophisticated human players to mentally solve for their probability of winning a given hand, or at least approximate it pretty well, and many do. An AI can’t do much better than that with pure statistics, the only way to squeeze more comparative advantage out is for it to become better at gauging playstyle and hiding information about its own patterns of play than human players are. And that’s a much harder problem than building a tree of possible moves in chess.
Some great resources on poker AI: University of Alberta Computer Poker Research Group. Papp 1998 in particular goes into detail about what makes it difficult, briefly: multiple opponents, imperfect knowledge, risk management, agent modeling, deception, and dealing with unreliable information. To these I would add the distinction between optimal and maximal play:
In chess AI, it never really matters what you expect your opponent to do like it does in poker. In chess, you just always try to move the board into the most favorable possible state. A win is a win in chess, but in poker the optimal strategy makes less money than an exploitable, but maximal strategy. But if you’re playing an exploitable strategy, then your opponent can turn around and play a strategy to beat you...much of poker strategy is figuring out how to be one “level” above your opponents for as many hands as possible. And getting an AI to do that is difficult.
Trying to figure out why poker AI hasn’t been solved (in the sense that chess AI has been solved).
It’s been a while since I looked into this, but when I did, the big problem was behavioral modeling.
All the math in poker is local—the state of the game in one hand influences the state of the game in the next only insofar as it affects the amount of chips players have on hand. That makes it easy to build a poker AI that’ll wipe the floor with innumerate players. But in also narrows the scope enough that it’s possible for more sophisticated human players to mentally solve for their probability of winning a given hand, or at least approximate it pretty well, and many do. An AI can’t do much better than that with pure statistics, the only way to squeeze more comparative advantage out is for it to become better at gauging playstyle and hiding information about its own patterns of play than human players are. And that’s a much harder problem than building a tree of possible moves in chess.
Some great resources on poker AI: University of Alberta Computer Poker Research Group. Papp 1998 in particular goes into detail about what makes it difficult, briefly: multiple opponents, imperfect knowledge, risk management, agent modeling, deception, and dealing with unreliable information. To these I would add the distinction between optimal and maximal play:
In chess AI, it never really matters what you expect your opponent to do like it does in poker. In chess, you just always try to move the board into the most favorable possible state. A win is a win in chess, but in poker the optimal strategy makes less money than an exploitable, but maximal strategy. But if you’re playing an exploitable strategy, then your opponent can turn around and play a strategy to beat you...much of poker strategy is figuring out how to be one “level” above your opponents for as many hands as possible. And getting an AI to do that is difficult.