Man vs. Machine

When I wrote Spin to Win, I reminded my readers about the 1983 science fiction film WarGames and of man’s unintended struggles with Artificial Intelligence in his ever more powerful machines. In that movie the computer inadvertently starts the countdown to World War III, threatening all of humanity. In the end disaster is averted when it  learns the lesson the film teaches us all: the only way to win was not to play.

Well, thirty or so years later, that may still be true. And yet, despite the warnings in our sci-fi cinema and literature, we continue our efforts to build smarter and smarter machines. And we program these intelligent machines with abilities that could enable them someday to turn against and defeat their makers. You see it in sci-fi movies all the time! Ex Machina, for instance, is a modern day reimagining of Frankenstein’s Monster and AI gone awry. In the Terminator and myriad of other sci-fi movies over the years, we often become the hapless victim of our own creation. Are there no lessons to be learned here, nothing to be afraid of? Why are we building these super-smart machines anyway? And why are we still playing games with them? 

And who would win that competition if the game was played today: man or machine? 

We have the answer to the man vs. machine match-up, at least when it comes to playing poker. This spring from April 24 through May 8, 2015, Artificial Intelligence (AI) once again squared off against world class human gamers. And, for now, in most people’s opinions, the humans won. 

Other past matches haven’t turned out so well for the machines’ human challengers. In 1997 the IBM Supercomputer Deep Blue defeated chess master Gary Kasparov. In 2011 the Watson program defeated Jeopardy champions Ken Jennings and Brad Rutter. But chess and trivia aren’t the only games in town. This time the Carnegie Mellon University’s computer Claudico went up against four of the world’s best professional poker players: Jason Les, Dong Kim, Bjorn Li, and Doug Polk. The tournament was held at the Rivers Casino in Pittsburgh in an epic 14 day and 80,000 hand game of no-limit Texas hold-em. And, by a narrow margin, the humans won.

The game is of particular interest to Artificial Intelligence researchers because poker is a game of incomplete information, and of all the poker variations, no-limit hold-em is one of the most sophisticated. Each player is dealt two cards only he can see. After a round of betting, the dealer presents five cards available to all the players —three cards (the flop), one card (the turn), and the last card (the river) with a round of betting after each. In no-limit anyone can bet anything from one chip to going “all in.” You can leverage a strong hand or bluff with a weak one. 

Researchers have been looking at the game since the 1990’s. An algorithm capable of determining optimal strategy for incomplete information scenarios could have far-reaching applications for cyber security, business, medicine, and military strategy, they say, because most world settings are imperfect information games. No limit hold-em is the last big challenge according to  professor Tuomas Sandholm whose team designed Claudico because the number of unique situations that can arise in the game is so big they can’t even all fit into memory. He claims there are 10 (to the 161 power) possible situations a player could face - more than all the atoms in the universe.  That’s 1 followed by 161 zeros. That's a lot.

There were three stages to the computer’s game theory development according to articles on the web. First the programmers fed the game rules into an abstraction algorithm, reducing it to something smaller in scope and more comprehensible. Then they customized algorithms to come as close as possible to Nash Equilibrium, a game theory concept involving optimal strategy. Finally they used reverse mapping techniques to input that strategy back into the algorithms for the games original parameters. Don’t worry if you don’t know what any of that means — I don’t either. Most of us have little real understanding of mathematics and statistical analysis. What I think we’re supposed to gather, from what I read, is that there are not algorithms to solve every possible problem, and the machine merely approximates ideal rational play, whatever the circumstances. And sometimes it doesn’t do too good a job of this.

Anyway, Microsoft Research and Rivers Casino put up $100,000 to cover the four players’ appearance fees for the grueling 13 hours of play each day. Claudico ran on Pittsburgh Supercomputing Center’s Blacklight computer. The humans used their own laptops to connect. with the supermachine. The challenge was structured so that the computer would simultaneously play each human one-on-one over a large sample of 20,000 hands, with the winner decided by who had the most virtual chips after the 80,000 hands. After two weeks of play, the ability to exploit Claudico’s departure from optimal play led the humans to victory. When the final hand was played, the players had wagered almost $170 million and the team of human professional gamblers was ahead by $732,713. 

Yes, the humans won — unless you’re a statistician. In that case, the match was statistically a tie. 

In the end both sides won the Texas hold-em challenge. The humans declared victory (despite the statistics), and the machine got 80,000 poker hands to study before the rematch next year and AI’s eventual dominance.