The poker winning machine is not a threat to humans
I wasn’t too worried when computers beat humans at checkers or chess or Go. It was, after all, only a matter of time before someone built a powerful computer with a vast base of data from known game situations. But now that the machines are beating the professionals at poker – a game of imperfect information – a question must be asked: is artificial intelligence beginning to threaten people in creative jobs?
In terms of game complexity – the number of allowed positions attainable in a game – no-limit Texas Hold’em poker is not an artificial intelligence researcher’s worst nightmare. The chess tree has 10 nodes up to the 120th degree. The one in Go has 10 at the 170th degree. Two-player No-Limit Texas Hold’em falls between the two with 10 to the power of 160 possible decision points. There are methods to make the complexity of the game tree manageable in a real-time game, often based on ignoring what led to a particular position and reducing the computational depth for future positions, and they have been successfully implemented.
Rise of the Machines: Keep an eye on the AI, experts say
But in poker, imperfect information creates an additional layer of complexity. As Matej Moravcik and his team of Canadian and Czech researchers wrote in a January 2017 article describing DeepStack, software they developed that outperforms professional poker players:
“The correct decision at any given time depends on the probability distribution on the private information the adversary holds, which is revealed by his past actions. However, how the adversary’s actions reveal this information depends on his knowledge of our private information and how our actions reveal This type of recursive reasoning is why one cannot easily reason about game situations in isolation, which is central to local search methods for games with perfect information.
In other words, it is difficult to reduce poker to a viable abstraction without compromising on the level of play. Two competing groups, however, seem to have overcome this problem lately: that of Moravcik and another, Carnegie Mellon University, which has yet to release a description of its winning program. The members of this group, however, have provided insight into what they have done in their previous work.
The language in which the creators of DeepStack describe their software disturbs anyone who fears being overtaken by machines. Moravcik and his team wrote that DeepStack had “intuition” – an ability to replace calculation with a “quick guess”. The machine developed it through “training” on many random poker situations. It worked well enough to beat 33 professional players from 17 countries.
Libratus, the product of the Carnegie Mellon team, is apparently based on different principles, using more precise calculations in the final part of the poker hand than in the beginning. He beat four of the best poker players, who came away amazed: The software managed to stay unpredictable and keep winning. Among other techniques, he varied the size of his bets to maximize profit in a way that even the best human player finds too difficult to imitate.
The good news for humans, however, is that even with all the complexity-reducing shortcuts researchers have developed, beating a good poker player requires enormous computing power. Deep Blue, the IBM machine that beat Gary Kasparov in chess, was a 32-node high-performance computer. Libratus used 600 supercomputer nodes, the equivalent of 3,330 high-end MacBooks. It would take a lot more, and probably more ingenious shortcuts, to create an artificial intelligence capable of winning real-world multiplayer poker games at a high level.
In AI research, it never pays to say something is impossible. The field is growing rapidly and boasting successes that seemed unattainable ten or even five years ago. But one can see how the introduction of even a small amount of uncertainty and information asymmetry immediately makes the work of AI developers much more difficult and resource-intensive. Poker, while extremely difficult to play well, is, after all, a game with well-defined rules. How much artificial intelligence and what unfathomable shortcuts will it take to excel in a game with few or no rules – like a trade negotiation or, at the extreme, a process like the Syria peace talks? Humans are used to situations in which rules develop in real time. No existing machine – and, judging by the state of the art, none that will be developed in the near future – can come close to our confidence in dealing with uncertainty and imperfect information.
Machines play an important role in eliminating routine tasks. What we’re seeing with recent developments in AI is expanding the way we define “routine” to most processes with clear rules. Even the most complex processes – like multiplayer poker – can be economically inefficient to automate. But processes without defined rules seem to be beyond the realm of practice. To be safe from machines, we humans must seek out such situations and learn to excel in them.
© 2017 Bloomberg L.P.
Leave a Reply