By: Paul Hewson
Our dystopian nightmare has finally arrived: The computers have beaten the humans at no-limit Hold’em. The “Brains vs. AI” challenge at Rivers Casino in Pittsburgh has come to a merciful end, with the Carnegie Mellon University poker bot known as Libratus (Latin for “balanced”) beating four professional players for the equivalent of nearly 15 big blinds per 100 hands, which is insane.
For the folks at Carnegie Mellon, this is a major step forward in artificial intelligence. But their goal isn’t to felt everyone in sight; this technology is for “real-world” applications, like figuring out the economy. Meanwhile, the rest of us are trying to figure out what Libratus has under the hood. What can we learn from our new electronic overlords?
Trust the Processor
The most obvious takeaway is how Libratus was put together in the first place. The creators started by doing the programming – 15 million “core hours” of it, using a supercomputer to do the grunt work. Instead of designing outward from a base poker strategy, they let the computer figure it out from scratch, using a variant of the game-theory concept known as Counterfactual Regret Minimization (CFR).
In a nutshell, CFR means never having to say you’re sorry. It’s an iterative, trial-and-error process, where you keep track of all the stuff that went wrong and try to make fewer mistakes next time. Libratus was able to put in trillions of poker hands, playing versus itself, before battling the humans. Then the computer would look over the results from each day of the competition, and keep improving – another four million core hours of work in total.
None of us will be able to put in that much study, but we can do our best with the time we’re given. CFR looks like a recipe for so-called GTO (Game Theory Optimal) poker; the idea is to arrive at a solution that minimizes our losses, and if our opponents play poorly enough, that’ll be the source of our profits. In theory.
In practice, we can’t match the complexity of the Libratus algorithm. Maybe we can pick up a few tricks, though. One of the stranger things about this bot is its propensity for calling on the river with marginal holdings. Bluff-catching, basically. Most humans don’t bluff the river as much as the math suggests they should, so this might not be such a good idea when you’re playing the regular stakes at Bodog Poker. But it sure seems to work well against the pros.