Add another activity to the list of things that computers are now better than us at.
Computers and gaming go together like peanut butter and bananas, and the struggle between the human mind versus the cold intelligence of computers is a constant back and forth. One of artificial intelligence’s first great victories came from a computer called Deep Blue (from IBM), which beat the world’s best chess player at the game in 1966.
Since then, computers have bested the greatest champions of mankind in other games, including the Chinese game “Go” and the popular game show “Jeopardy!” Now in 2017, a new artificial champion has risen to take its place among the pantheon of computerized heroes.
Libratus (latin for balanced) is the name of the machine that swept through the competition at the end of January. Built by a team of Carnegie Mellon researchers, Libratus is designed to play poker. Specifically, Libratus was used to no limit Texas Hold’em.
During the latter half of January, a poker tournament was held in Pittsburg to test the computer against some of the country’s best players. Over the course of 20 days, Libratus played its way through a high stakes gambling event against several human professionals in one-on-one matches. It ended up winning $1.7 million, with its competitors being left in the hole (though it is important to note that the players involved weren’t playing for real money).
One of the biggest challenges facing Libratus and the team behind it is that unlike chess, where all the information you need is always in front of you, or “Jeopardy!,” which is as much about what you know as how you play, poker is a different creature that heavily leans on imperfect information. In other words, when you play poker, you’re never privy to all the information. Opponents cards are hidden away, and at best, can only be guessed at. Since computers rely on information input, it makes games like poker much harder to conquer.
The other big challenge a computer faces in a game of poker is the bluff. Bluffing is when you try to convince your opponents that what you have in front of you is much better than it actually is, and is a huge aspect of poker. To succeed in poker means that Libratus had to learn how to bluff, and how to do it well.
To overcome these obstacles, Libratus is designed to learn and adapt. At the beginning of the tournament, Libratus was more prone to make mistakes, but as it made them, it observed how its opponents exploited it, and came up with a patch for it, so it never made the same mistake twice. Halfway through the tournament, LIbratus became much more difficult to read, and its missteps became even rarer.
The ability to bluff (and by extension lie) is a behavior that as of now hasn’t been explored in artificial intelligence. But Libratus has shown that when the goal is clear, computers are now capable of deliberately and strategically handling misinformation. Combined with a better way of working with imperfect information, and you have the majority of human interaction.
Who knows? Maybe an AI will win “The Bachelor” soon…