Artificial Intelligence is one of modern world’s most thrilling and maybe little disturbing projects. While all the truly major milestones with AI, as well as potential threats, seem to lay in the distant future, for the time being, we still have some accomplishments to be proud of. We successfully use some version of it in such many things, especially in IT sector, education, medical diagnosis, computer sciences, finance and a whole lot of other industries and aspects of life. We often put all kinds of robots and ‘thinking’ computers to different test and challenges against ourselves, both psychical and mental, just to see what happens and where we are as humans. Now the time has come for the game of poker.
AI has already gotten the better of our best performers in many disciplines. Go champion Lee Sodol lost to Google’s AlphaGo, chess master Garry Kasparov fell short of IBM’s Deep Blue, Elon Musk’s creations have beaten best e-sport players at Dota 2, and so on. We can now add poker to that list, and the events leading to such outcome are quite significant for science. The experiment has been conducted over the 20-day period and consisted of the total of 120 000 hands played between best players and Artificial Intelligence program. The system that faced poker pros is called Libratus and was developed by professors Tuomas Sandholm and Noam Brown from Carnegie Mellon University. Unlike the above-mention machines that managed to perfectly execute their knowledge of given set of rules, poker is much harder for AI to understand.
Not only the game has close to an infinite number of possible scenarios, which is greater than the number of atoms in the universe (10 to the 160th power), but also contains a lot of hidden information. You cannot see your opponent’s cards. Those of you familiar with texas holdem poker rules should not have any problem appreciating the importance of bluffing aspect throughout the course of the game. Having good hands is one thing but experienced players are well aware that you cannot win them all, as well as that you have to present something extra in the long run. Poker is recognized as Imperfect Information Game in contrast to Perfect Information Games, like chess or Go.
Therefore, the strategy had to take a different approach for the robot to gain an upper hand (pun intended). Libratus’ creators didn’t just teach their program intricacies of the game but allowed it to learn throughout interactions with human players. It was given instructions to win as much money as possible and told to optimize then these instructions. Through the course of the game, Libratus managed to learn subtle flaws in humans’ play and was able to capitalize on that. This may be considered as a seemingly-human way of bluffing for the AI. The implications of all this may be potentially huge for improvements in various areas. Computer programs might become better in helping humankind solving problems where not all the information is available and not all the facts are complete.
The Libratus experiment, other than being an entertaining show of AI’s superiority at poker, became much more significant. It may allow scientist to take the work on intelligent machines to the next level. They should be able to use this knowledge in AI development, hopefully making the world a better place to live for all of us.