This article originally appeared on VICE Sports UK.
The year 2010 seems a long time ago now, but it was then that Claude Sammut, a computer science and engineering professor in Australia, made the bold claim that 2050 would be the year that robot footballers would supercede the abilities of mere humans. It’s safe to say that, since then, the fears over job automation that keep supermarket checkout workers lying awake at night have not done the same for Lionel Messi. Or, for that matter, Steve Sidwell. Aside from the slightly wacky nature of the claim, there’s another problem – who would pay to watch robot footballers? It would be quite easy to develop a robot that could beat the world’s best at snooker or darts, but who would be watching?
Videos by VICE
There is one sport, however, where that doesn’t matter, because it would be difficult to detect whether the player was a human or a robot at all. It is a mind sport where the level of money involved is vast, and only peripherally linked to the amount of people watching. It is, of course, poker, and the start of the year saw one of the most significant developments in the game and artificial intelligence.
Everybody’s heard of Deep Blue beating Garry Kasparov, but chess is a game of complete knowledge, where everything is known and there are clear mathematical lines to follow. Poker is the complete opposite – a game of limited knowledge, bluffs, pattern recognition and psychology, where the mathematics and logic involved are there to be wilfully manipulated. It should have been considered a far more earth-shattering moment, then, when a poker AI, Libratus, soundly and comprehensively defeated a selection of the world’s best players in a lengthy series of heads-up, one-on-one games.
READ MORE: Sinking The Blues – How Snooker Helped Me Cope With Severe Stress
The dominance shown by Libratus certainly seems to be beyond doubt. Jason Les, one of the participants, was pretty clear about the abilities of the AI. “I don’t think there’s any doubt that Libratus is superior to human players right now, and it’s doubtful that will change. It’s hard for me to imagine humans improving enough to ever overtake it.”
Perhaps the most terrifying thing about Libratus is how it was built. The old maxim of a computer only being as intelligent as the man who programmed it now appears to be dead. While it may be true for the bots of old, the way this AI worked was very different, as Lee Jones – a poker author, PokerStars spokesman and computer science bachelor – explained to VICE Sports.
“The bots we have encountered so far are simply algorithmic. That is, they’re taught to play poker by their developers, and they do just that. They play poker no better and no worse than they were programmed to play. An AI learns as it goes. Libratus was taught only the rules for poker, not the strategy. It ‘learned’ to play by playing against a copy of itself, and over trillions of hands, became an expert player.”
PA Images
For Les, this produced something both unpredictable and impossible to exploit. “Libratus played like a human player should strive to play. It had a balanced game, it bluffed, value bet, and defended against bets appropriately. However, it achieved this balance by playing hands differently than a lot of humans would. It would check and bet hands that humans typically wouldn’t. It used very large bet sizes much more frequently than humans. When playing Libratus you faced a lot of larger bets in tougher spots than you ever do versus a human.”
The fact that Libratus lacks human cunning was no advantage, either. “At the start of the challenge we were attempting to find holes in its strategy that we could exploit. Unfortunately we ultimately were not able to find anything we could sufficiently exploit for profit. We also attempted to use off-tree bet sizes, bet sizes that the AI did not have in its strategy. However, what we didn’t know was that the AI was learning these sizes overnight so while they might have helped us for a little, ultimately the Libratus was able to learn and understand those bet sizes.”
That’s perhaps the scariest part: any minor advantage gained by the players was quickly erased, since Libratus was able to quickly adapt and cut out any predictable patterns in its own play. As Jones says: “The point is that as it continues to play, it will become better and better by finding its own weaknesses and removing them. Furthermore, [a different AI] could be built which would find and exploit weaknesses in its opponents.”
Now, clearly AIs are not going to be appearing on late-night poker TV shows anytime soon, no matter how dark the sunglasses. But online poker is of fundamental importance to professional players in that it allows for multiple games to be played at once, meaning far more hands are played per hour. If poker AIs can reach the levels of Libratus, is that now under threat?
READ MORE: Can Anyone Beat Michael Van Gerwen, The World’s Best Darts Player?
Both Les and Jones are highly doubtful. There are three major barriers, the first of which is the sheer complexity of Libratus. “You need something like a $10 million, 40,000 core computer to run Libratus,” Jones points out. Even aside from the initial investment, it still wouldn’t make economic sense. “It’s extremely expensive to run Libratus because of supercomputer power and time,” adds Les. “It would not be profitable to run it on the internet.”
That might be the case now, but assuming computers continue to make drastic improvements in processor power, will that last? Another potential saving grace for professionals, however, is the of game chosen. Heads-up poker consists of just two players, face to face. But hold on – isn’t this even more scary? Table games, particularly online where players come and go frequently, often allow for a more methodical, predictable of play. When there are a lot of players holding cards, and the bet sizes start rising drastically, you know there are only a limited number of hands opponents will be holding. In contrast, pretty much any hand is playable in heads-up – it’s a game of wild-eyed bluffing, psyching out opponents, and one where almost anything is possible. Isn’t that a much more complicated game? And, perhaps more worryingly, one where you’d expect humans to have a much bigger edge?
Apparently not, says Jones. Libratus simply doesn’t ‘think’ in the same way a human player would, while what is complicated for us may be simple for an AI, and vice versa. “In no way do the developers teach it to ‘be aggressive’ or ‘play tight’ or anything like that. They simply say: ‘Start playing and doing things randomly. When you find things that cause your score to go up, do more of them. Things that cause your score to go down, do less of that.’”
“The AI’s whole understanding of the world is based on two things,” says Lee. “Number one, what happened in the past, and number two, the possible paths it could go down in the future. Well, if you’re heads-up, then the tree only involves a single opponent. Think of it being on the river, and you check. Your opponent can check, or bet from one BB to its whole stack. If she checks, the hand is over. If she bets, then you can call, fold, or raise. And so on. Now add another player. If A checks, then B can check or bet. If A bets, then B can call, fold, or raise. You can see how the tree just explodes as players are added.”
PA Images
In other words, most online players don’t need to be worried yet. “Specifically heads-up players will have to be concerned about bots as the computing power gets more feasible in the future,” Les points out. “However, since the majority of internet poker traffic is six max, full ring cash games and tournaments there is not really a concern until the AI makes significant advancements in those games.”
Oh, and the final hurdle is the fact that online casinos have long been wise to the tricks of bots. Far cruder systems than Libratus have existed for a long time, but as well as scepticism over their efficacy, casinos have been able to detect and ban many of them. “We consider bots to be one of the most serious threats to the integrity of online poker games,” says Jones. “For that reason, PokerStars has invested a huge amount of time, money and people into bot detection technology and techniques – and into game integrity on the whole. We are working with some of the best experts in the business to improve our proprietary detection of both traditional algorithmic bots and more advanced AIs. We don’t expect the bot developers to give up, but we intend to stay one step ahead of them.”
So, Libratus might end up meaning more for AI as a whole than for poker itself. A machine that can learn such a complex game on its own and quickly rise to dominate people who’ve spent thousands of hours playing it is clearly a scary prospect, but the limitations of applying it to the real world are also apparent. “We expect there to be ample opportunity for serious profitable poker players for years to come,” insists Jones. Whether that will continue to be the case online ten years from now remains to be seen but, right now, the AIs are still demonstrating only theoretical potential. Some day, however, that theory could well be put into practice.
