Sophisticated poker programs are getting tougher and tougher for humans to beat
During the summer of 2004, Binion's Horseshoe was renamed Binion's Gambling Hall. It played host to the World Series of Poker for the final time. Patent attorney Greg Raymer won the bracelet that year, online poker continued its rocket-fueled ascent and a new generation of players, inspired by Chris Moneymaker's 2003 victory, dominated the tables. There was enough activity that it would have been easy to miss the spectacle taking place one floor below the Main Event—and involving a truly new generation of player.
Just downstairs, Phil Laak, a fresh breed of wacky poker pro whose hooded sweatshirt became a virtual trademark that earned him the nickname Unabomber, had busted out of the Series, eschewed lucrative cash games and indulged in a heated donnybrook of limit Texas hold'em against, of all things, a computer. They called it a bot, short for robot, but this was actually a computer program designed to play Texas hold'em as competitively as IBM's Deep Blue played chess. The once high-flying online poker site Golden Palace put up a $10,000 appearance fee for Laak who looked suitably wild with spiky hair and an anarchy "A" symbol splashed in red across his black tee shirt. Laak's girlfriend, the actress and burgeoning pro in her own right, Jennifer Tilly, told me, "Phil's brain is as big as a planet. He's obviously a great poker player and we don't want to see a machine win today."
Looking a tad nervous, as if he had all of humanity's weight on his shoulders, Laak laid out his plan: "I will play a tight, mathematically sound game at the beginning. I want to get an idea of how the bot plays, check out some of its strategies and just breathe a little. Then I will find my rhythm."
Ultimately, Laak beat the poker bot, but it was close and he did require a little luck at the end. Eleven years later, in 2015, he probably would need to get a lot luckier. Especially if he happened to take on Cepheus, the poker bot created at University of Alberta in the Canadian city of Edmonton. Earlier this year it was announced that the computer program, which had been in development for the last 21 years, managed what once seemed impossible: It solved limit Texas hold'em. This means that the bot, created through the use of self-learning software, which essentially taught itself by trial and error via many billions of poker hands, had learned nearly every correct move. It plays an unbeatable form of the game, employing a defensive strategy known as Game Theory Optimal. This is a defensive style of poker in which the player strives to keep himself (or, in the case of Cepheus, itself) from being exploited—as opposed to exploiting opponents.
Michael Bowling, who heads up the Computer Poker Research Group at the University of Alberta, concedes that Cepheus actually remains 0.986 percent away from perfect. But that's okay. He and others in his community, including editors at Science magazine, are willing to let that small detail go and consider limit hold'em to now be a solved game. It's been checked off the list, right alongside backgammon, checkers and tic-tac-toe. Bowling asserts that no human would be able to tell the difference between fully solved and 99.014 percent solved. "The program suggests a good direction of play," understates Bowling, slender and wearing a tweed sport jacket, his head topped with a Brillo-brush of graying hair. "For example, before the flop, it plays aggressively with low-ranked pairs. It raises half the time with pairs of 2s and all the time with 3s. That's more aggressive than most players."
But, according to the countless hands that Cepheus has logged, it is correct—if you want to prevent your opponent from taking advantage of you.
Bowling expresses this to me in the lobby of an Austin, Texas, Hyatt Hotel. It's the site of an academic conference focused on Artificial Intelligence. Grad students and professors from top math and computer-science departments have flown here from all around North America. They're discussing AI's state of the art and attendant applications.
Many of those in attendance have gotten deeply enough into poker that hold'em discussions and seminars take up a good chunk of the schedule. But, for all of their research, few of the scholarly folks here have any real interest in gambling. Bowling tells me that his goals are completely scientific. Another professor confides that he has played some online poker. "But not for money," he clarifies. "That would be illegal."
While poker players would surely have financial incentives to get under the hoods of various applications, Bowling and the others won't allow it. They're having too much fun competing against one another, with their various poker machines, and prefer that secret sauces don't get spilled—the academics see poker as a benchmark. For them, it is something less than a means to an end. "I want to push artificial intelligence and have machines making decisions that humans should ordinarily need to make," Bowling says. He ticks off auctions, cyber security, mano a mano negotiations and even the fighting of diseases as things that his thinking machine can strive to beat. "Poker mimics real-world algorithms that present themselves when you need to make decisions with incomplete information."
That's all well and good, but I'd rather talk about hold'em. Other poker-related discoveries, says Bowling, include the idea of never putting in a fourth (and final) bet before the flop and a propensity to move forward with hands that humans consider to be garbage. "We had a poker pro in Canada looking at Cepheus," Bowling recalls. "He thought the folding region would be a lot higher than it is. With 10, 6 unsuited you call, and, if it is suited, you raise almost three-quarters of the time."
Why? "Cepheus doesn't get into it the way you or I might. Its answer would be that it makes more money when it plays that way."
Considering the hard work and years of research devoted to all of this, I tell Bowling that I'm surprised he hasn't turned to poker as a way to make some extra cash. "I've given up on [extra cash] a long time ago," he says. Then, with a bit of an air, he adds, "I think there would be other ways for me to make money than grinding it out at poker."
For some, grinding it out at poker is a perfect occupation, especially when you have a computer doing the work for you. Gregg Giuffria, the former keyboardist for the 1970s metal band Angel, has reinvented himself as a producer of casino gaming machines. He used technology developed by a Norwegian engineer, Fredrik Dahl, to market a limit hold'em machine to casinos. It resides on gaming floors, right alongside the traditional video-poker machines. Unlike those, though, in which you play against a pay-table, winning money for making certain hands, in this game, you actually take on the machine. It's no different from playing heads-up poker against an individual. The machine, though, is much tougher than most individuals. Particularly amazing is that Giuffria's game, simply called Texas hold'em Heads Up Poker, manages to beat most players without ever adjusting its style of play.
So, it takes on Phil Ivey in the same way it would take on any amateur player. The machine strategizes in a completely defensive manner, playing a level of Game Theory Optimal poker that its Scandinavian creator describes as "almost paranoid poker." Despite the fact that it is in the casino and proving its mettle by going up against all comers for money, Bowling remains unimpressed. "I think Gregg Giuffria is a lot of marketing," says Bowling. "He can't even prove that he has solved poker. What has he added to game theory?"
Giuffria does not want to add anything to game theory. His goal is different from that of Bowling. He's not trying to prove a point. He just wants to create a poker brain that can make money by beating opponents—without beating them too badly or too consistently. He's told me that an early version of the machine needed to be dumbed down in order to work in casinos. In its original form, the game won so frequently that playing it was no fun. Still, it's done a pretty good job of fleecing players.
The point here is that a machine does not have to be perfect or to have definitively solved a poker format in order to consistently win. This was the point, about 10 years ago, when the Texas-based billionaire banker Andy Beal was playing a group of professionals, dubbed The Corporation, at ultra high-stakes heads-up limit hold'em with eight-figuresums routinely changing hands. A former blackjack player (a man we'll call Steve) had coded a program that could play online as a bot—masquerading itself as a human but making the kinds of calculations that no human brain could pull off. Or, for a diligent learner like Beal, it could be used offline as a training tool. Steve contacted Beal and tried negotiating a deal for Beal to use the software in order to tune up for play against The Corporation. Like the bots of Giuffria and Bowling, Steve's system did not adjust to opponents because it was doing enough things right—via its Game Theory Optimal style—and making enough correct plays that no human could reasonably beat it.
Steve says that Beal showed interest and a date was set for them to meet in Texas. "Then," he remembers, "the night before we were going to fly down, we got a call from Beal's mathematician. He asked how the strategy changes based on each opponent's style. We explained that it doesn't change and that it doesn't need to. He said it was impossible to create a winning poker program without adapting." But adaptation was something that Giuffria and Bowling and others seemed to have innately accepted as not being a problem. "Beal sounded disappointed and told us not to come," Steve says. "He said we didn't know what we were talking about."
At the moment, Bowling is basking in his achievement and not sweating over anyone closing the loop by perfecting that last 0.986 percent. What really might show him up, though, is the solving of no-limit poker, an infinitely more complex variant. "In no-limit, there are too many decision points," he insists. "There are 10 to the 160th power of decisions." To put that number into perspective, he adds, there are only 10 to the 70th or 80th power of atoms in the universe, so perfect no-limit isn't happening.
Bowling's pessimism has done nothing to stop people from trying. Tuomas Sandholm, a professor of computer science at Carnegie Mellon University, appears undaunted. "Ten years ago, it seemed impossible that limit poker could be solved, and now it is almost there," reasons Sandholm, seeming to savor the word almost. "They say it is essentially solved. But I think that counts. My question, though, is this: Was it essentially solved three years ago?" His point being that maybe you do not need to get that close for a game like poker to be essentially solved. Maybe there is more room for the unknown than people think.
Whatever the case, no-limit seems to be where the big challenges reside right now. The appeal lies in the seeming impossibility of nearing perfection. It also is likelier to be applied to real-life poker scenarios (limit heads-up just does not get played very commonly). In fact, one of Sandholm's PhD students, Sam Ganzfried, happens to be a keen heads-up no-limit player. While he says that what he knows about poker has informed the program, he admits that the program has informed him as well. "One surprising thing it does is limp pre-flop, which few humans would actually do," says Ganzfried. "It limps with aces a small percentageof the time, always folds the bottom hands, and uses several raise sizes before the flop." Humans typically raise two- or three-times the pot in an effort to keep from giving away information. "The bot thinks it's optimal to mix bet sizes."
Former Google engineer Eric Jackson has been building a no-limit poker bot of his own. He lacks the super-computer and staff of grad students that Sandholm has at his disposal, but he still seems to be doing pretty well with his card-playing machine—to the point that Sandholm expresses admiration for the program. Like the guys at Carnegie Mellon, Jackson has noticed out-of-whack raises that can go to seven-times the big blind. "When I play against my bot, it is very frustrating," he says. "I don't know how to respond to those large bets. But it can't be right to fold everything except the top four hands."
The bot created by Sandholm, Ganzfried and the Carnegie Mellon group currently ranks as the top no-limit program out there (Jackson's machine, which he nicknamed Slumbot, finished fourth in a competition of no-limit hold'em poker machines). Considering that the Carnegie Mellon program has outplayed other computer programs, one question does linger: How would it fare against a top no-limit pro?
Olivier Egger and Johannes Levermann, partners in a Sweden-based company called Snowie Games, have wondered the same thing. The pair of programmers disrupted backgammon hustling in the 1990s when they introduced Snowie, which basically showed all the statistically correct moves for that particular game. Most losing players either got good or realized that they were hopeless and gave up. With fewer people to beat, a lot of the strongest backgammon players—guys like Erik Seidel and Abe Mosseri—gave up the game and moved to poker. Now, the recently introduced PokerSnowie, which promises AI coaching in much the same way that the original Snowie did, could conceivably dent poker.
In the hope of seeing just how well PokerSnowie could compete, Levermann and Egger recruited heads-up specialist Daniel "jungleman12" Cates to play a session against the program. Cates is famous for having accrued some $10 million on the old Full Tilt Poker site, and is fearsome at highest-stakes heads-up no limit hold'em. "I didn't think it would be very good and was somewhat surprised at how well it played," says Cates, acknowledging that their 2,500-hand showdown is not enough to iron all the luck out of the game. "I beat it, but 500 hands from the end, I was not beating it."
Similarly, a quartet of top pros recently took on Sandholm's bot and won. Playing for theoretical chips, they would have snagged a total of $732,713 from the computer. Though Sandholm insists it was a "statistical tie," the humans still took down a $100,000 prize put up by Pittsburgh-based Rivers Casino and Microsoft.
Cates noticed that, like the Carnegie Mellon bot, Snowie has a tendency to overbet, which he acknowledges was a tricky way of playing. It's tricky enough that he's not crazy about confronting Snowie if it improves very much beyond its current level—underscoring the fact that a program does not need flawless execution in order to crush human opponents. "I'd rather play Tom Dwan than the computer," he says in response to a question of how difficult it is to beat Snowie as compared to the skilled pros he's gone up against. "But I'd rather play the computer than Isaac Haxton."
It's worth noting that perfect play is not what Levermann and Egger are pursuing. While Bowling and the other hobbyists may be trying to prove that a machine can make more-or-less mistake-free decisions, Egger and Levermann are going for something else altogether. "We're not trying to solve the game," emphasizes Levermann, pointing out that he's not so interested in beating other machines. "We're trying to make a system that should play as good or better than any human." Up against Snowie's level of Game Theory Optimal, he says, "the best humans cannot beat the program even though the program is not playing perfectly."
For those in the trenches striving to play winning poker against tough opponents, just good enough is, well, good enough. Nikolai Yakovenko, a former Google colleague of Jackson's and creator of the app ABC Open- Face Chinese Poker, would like to see a fast-and-dirty version of Snowie for lots of poker variants. As a medium high-stakes veteran of Bobby's Room, the elite poker enclave at Bellagio, he's used to playing what is known as mixed-games—eight different forms of poker, rotating over the course of a given time period. Pointing out one of the more obscure games, badugi (you receive four cards and the object is to get the lowest possible hand), he says, "I'd like to have eight badugi robots, each playing a little differently, just like the guys in Bobby's Room. Even if they weren't all that good, they would be extremely useful. If a program has a good strategy, plays well and nobody can beat it, that's just fine."
If this is what Yakovenko, a skilled computer programmer in his own right, needs and it doesn't have to be perfect, why doesn't he code one himself? The Russian poker player hesitates for a beat and softly replies, "Who says I'm not?"
Michael Kaplan is a Cigar Aficionado contributing editor.
Log in if you're already registered.
Search our database of more than 17,000 cigar tasting notes by score, brand, country, size, price range, year, wrapper and more, plus add your favorites to your Personal Humidor.