On the 31st of January 2017, a poker tournament that lasted nearly 3 weeks ended in a casino in Pittsburgh. For 20 days, some of the best players in the world went toe to toe with an unusual adversary: an AI. Conceived by professor Tuomas Sandholm and his student Naom Brown from Carnegie Mellon University, the AI – nicknamed Libratus after the latin for “balance” – should fare better than its predecessor. Last year, during the first “Brain VS AI” tournament, the computer Claudico was not powerful enough to beat its human adversaries. However this time, Dong Kim, Daniel McAuley, Jimmy Chou and Jason Les were not able to beat Libratus, who seemed to see straight into their hand. But who was this possible?
Garry Kasparov thinks that machines will be unbeatable
Pitting human versus artificial intelligence didn’t just start yesterday. Remember the computer victory over chess champion Garry Kasparov in 1997? In 20 years, artificial intelligence has made leaps in terms of capabilities, and humans have had to cede intellectual territory in places they couldn’t even imagine. Watson, IBM’s AI, creamed the human competition in Jeopardy!, the american game show where participants are given an answer and they must define the question. And again in march 2016, Google’s AplhaGo AI defeated the South Korean Go champion Lee Sedol, classed 2nd in the world. Before the match, Kasparov had written an article in Le Monde stating that even if the AI would lose that it would eventually surpass human players.
Google’s AI is world champion Go player
Go is meant to be the hardest strategy game in the world, but this didn’t stop AlphaGo from beating Sedol 4-1. AlphaGo is classed as the “best player in the world” since this summer. Wired wrote an account of this historical match.
Libratus has beaten the world’s best poker players – by being the better bluffer
And now, here we are; an AI is able to beat the best human players in the world, a game where chance, strategy and bluffing all have their place. For 20 days, Libratus lined up Texas Hold’Em games against 4 champions, including 28 year old Dong Kim. Half way through teh contest, Kim though though that the AI was looking at its cards it was that good. “I’m not accusing it of cheating” he said. “But it’s just that good”. Wired also covered this fascinating tournament.
These victories are due in part to Neural networks.
If such spectacular AI victories are possible, it’s due to to neural network technologies. This ensemble of AI algorithms mimics human intelligence. Just like a young human learns to walk after many unsuccessful attempts, the machine learns from its errors to better itself. Thus, Libratus played billions of games against itself before it gained supremacy of the game. It’s what we know as reinforcement learning. It’s the same technique that allowed AI’s to learn how to drive by playing GTA V.