A machine that can out-bluff a human being at poker is here. And the world of artificial intelligence is celebrating. Libratus, an AI robot might be the greatest advertisement yet for how far AI has advanced in the modern world. Do we welcome this development or do we view it with certain sense of dread about where this is all going? The primary fear Libratus stokes is robots may one day get so smart as to be able to defy the law of robotics — Isaac Asimov thought of and enunciated them before anyone else — and bring harm to human beings. Libratus is a sophisticated triple program that has learnt to play poker from scratch so well that it took on four of the world’s best poker players and won chips worth more than $1.76mn in virtual currency from them in a three-week challenge at a Pittsburgh casino.
Unlike chess, where all the pieces are on board and seen, poker is a game in which risk rewards players as they try to out-bluff each other with the ‘hand’ of cards or ‘hole’ that they hold, and five community cards that are open. Game events could not be more random. Libratus took on Dong Kim, Jason Les, Jimmy Chou and Daniel McCauley — all specialists in no-limit Texas Hold’em, which is an unlimited bid form of poker — and performed so well that, halfway through the competition, Kim began even to suspect if the program was somehow seeing his cards. Libratus did not have to peek. Melding three different systems, it sported intricacies of knowledge not through neural networks but through forms of AI known as ‘reinforcement learning’, said to be a method of extreme trial-and-error, and an algorithm called counterfactual regret minimisation.
Programmes have beaten the best humans in chess and Go but in poker, the AI had to be so good as to not only bluff in order to win but also interpret accurately the misleading information fed by other players’ bluffing. If all this sounds too complex, just relax in the knowledge that AI may take a lifetime more to become Albert Einstein. The only problem is AI is learning fast as Libratus showed with programmers sitting through each night after the game to build an algorithm that would identify the system’s patterns and remove them so that Libratus did not become predictable by bluffing in a linear manner.
Jason Les confessed, “Libratus turned out to be way better than we imagined. It’s slightly demoralising”. He may have been speaking for all of humanity as robots are getting better and better at outthinking humans. One day, a robot like this could make a billion dollars a day in the stock market for its owner, beating the Wall Street ticker of share price movements by the sheer weight of its knowledge on how mathematics works even in the most random of patterns.
The great sportsmen believe in taking positives out of even a losing situation. If we were to replicate that thinking, we can conclude that since the algorithms that power Libratus are not specific to poker, such programs could lead to a vast variety of applications as in medical treatment, cybersecurity So, we owe a debt of gratitude to the Carnegie Mellon professor Tuomas Sandholm and his student Noam Brown who built Libratus and breached the frontiers of AI. Forget for now, the fears that such an AI could one day run military applications in such sophisticated manner as to give its ‘parent’ a huge advantage over rest of humanity.
