Machine learning and competitions

The two games of Go played by the young Chinese genius of Go, Ke Jie, against AlphaGo took place in an unusual context: not only did they come after the successive defeats of Fan Hui and Lee Sedol, two of the best players in the world, but also, the Chinese authorities censored their broadcast, in what can only be assumed to be a gross misinterpretation of the national pride. For Ke Jie, who has recognized that Google’s algorithm is already “too strong for humans” and has become a kind of “Go god”, he finds himself being used as evidence, as the final proof of the superiority of machines at playing the game, a role previously experienced by a handful of humans in other games.

The importance of these types of challenges is relative. In reality, they are about letting the public know about the progress of machine learning by the companies sponsoring the challenge. Over time, we have seen how machine learning algorithms have taken over chess, Jeopardy, Go and, most recently, poker, without wondering what exactly each competition intended to prove.

Chess is a scenario game. Each move generates a new scenario, and good players are able to mentally see many moves ahead. When Kasparov lost against IBM’s Deep Blue in 1996, the only thing humans had to recognize was that a machine was already capable of calculating probabilistic scenarios better than we were. In other words, computational brute force. Given that we have been using calculators for decades, that is easy to assimilate: therefore, no offense taken, the human pride was still in a reasonable good position.

When, in 2011, another IBM product, Watson, beat the best ever players of Jeopardy, things were different. Here, a machine was better able to understand rhetorical questions expressed in human language, to look for its possible answers, to choose one of them, to press a button … and win. After Jeopardy, we knew that an algorithm is better than us at understanding our own language, opening up all kinds of innovation in conversational robots, chatbots, the law, medicine, and many more areas.

In 2015, a Google algorithm beat two of the best Go players in the world, a feat that has just been repeated few days ago. What’s going on? This challenge crowns the importance of a revolutionary technique: deep learning. After the algorithm was fed every game of Go ever played and registered, Google had a very good player, but not one that always won. It was as good as Go’s great masters, but not definitively better. Deep learning made the difference: the machine was programmed to invent new moves and to play against itself, to explore improbable scenarios. The result was that in some of matches, AlphaGo used moves that no human had ever made in any of the previous games ever played, moves with a probability of one in ten thousand, and managed to win. So we now live in a where an algorithm can develop intelligence for a task beyond anything humans are capable of.

Finally, in 2017, and an algorithm created by Carnegie Mellon University, Libratus, beat some of the world’s best poker players. Poker is an intrinsically human game, in which several cards remain face down, the identity of which can only be speculated on, while we get information about the rest of the cards on the table and from what the other players give us with their bets, which of course can be bluffs. After 120,000 poker hands played over 20 days, Libratus’ victory was absolute and unconditional: humans had no chance.

What does this mean? Simply that a machine is already capable of analyzing a situation with imperfect information, subject to uncertainty, and can make better decisions than a human could. That’s right: making decisions based on incomplete or imperfect information, including possibly false information, is what a manager does in a company; something I’ve been trying to teach my students for twenty-seven years. The approach taken by Tuomas Sandholm, creator of Libratus, is to create algorithm with a wide range of uses. He is not interested in an algorithm to play poker, but in one able to carry out cybersecurity analysis, medical diagnostics or business negotiations.

Competitions and business challenges should not be analyzed simply as the communication strategy they represent, but instead as what they really mean in terms of achievements, to show what is possible. If anybody reading this still doubts the possibilities of machine learning, let them come up with the next challenge.

(En español, aquí)