Unlike Chess and Go, where Google AI manifested their supremacy, Libratus had to prove itself in the game of no-limit Texas Hold’em, a popular form of Poker.
Held in January this year, the 20-day competition against four human poker professionals involved 120,000 poker hands. All of this was to win the grand prize of $200,000. The researchers again proved their AI’s poker skills several months later during a poker competition held in China where they used a new version of the AI called Lengpudashi.
The reason why the AI’s victory in Texas Hold’em Poker is different is because the game is an example of imperfect-information game solving. Here, not a single player is aware of all the game elements at any point of time during the game. Games like Chess and Go are perfect-information games where all the related information is available in front of the players.
The most challenging thing about no-limit Texas Hold’em is it requires the players to sense entirely unpredictable moves of other players and make sure they don’t get deceived by other players’ bluffing skills. For an artificial intelligence system, it’s a tough task to master such games which involves taking small defeats for a bigger gain in the end.
The researchers have now revealed the inner workings of their poker playing AI in a new research paper published in Science journal.
According to the researchers, the techniques used by Libratus to conquer the game aren’t based on expert human knowledge. In fact, they aren’t designed specifically for poker and can be used for a host of imperfect-information games.