The University of Alberta’s Poker Research Group announced they have solved heads-up limit hold’em by creating an effectively unbeatable strategy.
What does this mean for poker players? Not much, at least in the short term. Heads-up limit hold’em had already largely died out. Heads-up (one-on-one) poker is generally not offered at all in casinos, and increasingly less popular online. Likewise, limit poker has largely given way to no limit or pot limit games. So heads-up limit hold’em is at this point a fairly obscure poker variant for human players.
Nonetheless, the achievement is a landmark for computer poker and has some interesting implications. I had been thinking recently about the question, why are computers better at chess than at poker? Arguably this new announcement calls into question whether that’s even the case, but I think it’s still safe to say that computers are indeed better at chess. Computers have been much better than the best human chess players for years now, while it is unclear if they are better than the best human players in many popular forms of poker.
I imagine many people’s first answer will be that poker rewards skills such as bluffing and reading opponents that computers are not good at. But in fact, computers bluff just fine, and while it is indeed difficult to teach a computer to read cues such as facial expressions, there is still no advantage there for a human adversary: the human can’t “read” the computer at all.
Maybe one reason computers have been bad at poker is that humans are so bad we don’t really know what to teach them. As with chess, computers are getting good at poker not by emulating humans, but by excelling in their own computer-y way. The Alberta team started with a program that bet completely at random, then re-evaluated its play and adjusted. Countless iterations of this process eventually produced an unbeatable strategy. In other words, they started with no preconceptions about how the machine “should” play. Poker is so strange and confusing that the best way to learn may turn out to be to teach computers to teach themselves so that they can teach us.
I’ve found that it is possible, more or less, to turn your brain into a chess machine. If you play and study enough you get to a point where the best moves often seem natural and obvious. This seems to be less true with poker. While experience will help you beat a human opponent (when you sense he’s uncomfortable, you raise), it may not help you much if you’re trying to play “optimal” poker, the kind you’ll need to beat a computer. In many cases, the best strategy involves mixing together different options. For example, maybe you bet 60% of the time and check 40%. This idea – that neither betting nor checking are right or wrong, but must be mixed in the right frequency to form an optimal strategy – is very difficult for humans to understand and implement.
The biggest reason that computers are better at chess is probably a boring one: people have been working on it for longer. Building a chess-playing machine was an early grail for computer programming, and many of computer science’s brightest lights worked on it at one time or another: Alan Turing, Claude Shannon, and many others. It is only much more recently that the same kind of attention has been paid to poker.
It’s natural to wonder if computers’ mastery of poker might eventually kill the game. This doesn’t seem to be happening in chess: if anything, computers have made chess more interesting for humans by improving analysis. But of course, in poker money is at stake, so there is more of a concern that players will use computers for a competitive advantage. My guess is that this will be a big problem in online poker, maybe eventually making it unviable, but won’t have much impact on live poker. Most people, it seems, play however they like, without much concern for theory.