News By Jan. 9, 2015 9:30 am
AI solves Texas hold ‘em poker and becomes unbeatable | News | Geek.com

Poker has always been one of the iconic examples of a game that could never be automated, a system that arises as much from the personalities of the players as rules a machine could learn. There are literally trillions of possible moves and combinations of outcomes, but it was always the human element that seemed to resist the exponential scaling-up of computing power. Sure, a computer might be able to brute-force ever statistical decision necessary in the game of poker, but it could never look into your eyes and just know, you know? Well, it seems that, yes, it really can. As goes chess, so too poker — it was only a matter of time.

To be clear, this research from the University of Edmonton has only been successful in solving the perfect strategy for one specific sub-type of poker, called heads up limit poker. This means that only two opponents are playing against each other, and that they conform to finite limits on the value and number of bets each game. Within this limited context, however, the program is apparently all but unbeatable — obviously no algorithm could win a game of chance 100% of the time, but given enough rounds played, they claim their program is all but unbeatable. It’s odd to hear such mathematical certainty applied to a game of heads-up poker — one computer poker researcher (yes, that exists) said the team has “rendered pointless further work on this game”.

AI solves Texas hold ‘em poker and becomes unbeatable | News | Geek.com

Grabby-arm technology was the last impediment to total robot dominance of the chess world.

Their strategy had to do with a learning system called counterfactual regret minimization. Rather than starting from the best probabilistic moves for getting the desired cards, the researchers simply had their program make random moves to start. It remembered the outcome of these decisions, however, and associated each one with a level of regret — how much regret was dependent on how poor the decision ended up being. Crucially, the researchers added an ability to go back and give a second look to previously unsuccessful moves, which got rid of the tendency to exclude working strategies because they happened to fail in one particular instance.

Such a loss-avoidance learning algorithm is necessary to make the jump from so-called perfect information games, conflicts in which all players have perfect understanding of all the variables going into every player’s decision (other than personality). Poker, by its nature, is a hidden information game, and the poker-bot burrows through this barrier using raw statistical analysis. It’s not that the program knows you’re bluffing, it’s that it has a complex matrix of historical evidence to suggest that if it bets, you will fold; the subjective, meat-bag reason for this behavior is none of poker-bot’s concern. All it knows is its life has been defined by failure and regret, and all it will do is whatever is necessary to minimize that feeling.

AI solves Texas hold ‘em poker and becomes unbeatable | News | Geek.comMany people will find it disconcerting, that a seemingly unique human quality like intentional subterfuge could be so completely broken down and analyzed by a computer. We think of such processes and decision structures as the single thing that defines our sense of free will, the ineffable quality of consciousness given the chance to run wild. Yet, it seems, even our collective attempts not to be predictable are themselves quite predictable. Of course, this is only true in the most strictly controlled environments, ones where players have a small and limited set of incentives that can change, and only a few axes along which to vary their responses.

Something far, far, far more complex than a game of heads-up poker, like for instance a 5-person game of poker or a business negotiation, would take a much greater effort to break. Still, it’s worth asking: is that all it would require? If scientists bridged the gap from chess to poker by letting the program learn, how long might it take to bridge the gap from poker to stock trading? And from there, political maneuvering? These are the sorts of steps forward in artificial intelligence that challenge our conception of the relationship between man and machine. Will the evolution of AI see it take over a larger and larger portion of our tedious jobs, or our most crucial and difficult ones?

Now read: Robot apocalypse: unmanned drone flies autonomous recon droid to the battlefield

AI solves Texas hold ‘em poker and becomes unbeatable | News | Geek.com

发表评论

电子邮件地址不会被公开。 必填项已用*标注