AI beats professionals at six-player Texas Hold ’Em poker

 AI defeats professionals in Texas Hold 'Em poker for six players

Artificial intelligence has overcome the greatest poker challenge: beating the top poker players in the popular Texas Hold 'Em game.

The AI has won 15 of the most successful online poker players worldwide, winning more than $1 million each.

Pluribus, the AI was tested against five human gamers and 10,000 rounds in which five Pluribus copies were played against one professional. The AI did better than both.

Noam Brown from Facebook AI Research and Tuomas Sandholm of Carnegie Mellon University in Pennsylvania created Pluribus. It is a significant improvement over Libratus, their poker-playing artificial intelligence, which was defeated by professionals at Heads up Texas Hold'Em in 2017.

Tristan Cazenave from the Paris Dauphine University says part of the reason poker is so hard for AI to master, is the sheer number of actions that can be taken. There are more possibilities in poker than there are atoms.

Hidden information also applies. Players have access to only the cards they see. An AI must take into consideration how different cards would interact so that it doesn't know when it has good hands.

Brown states that "most real-world interactions involve hidden data, multiple participants, or both." Pluribus's approach can be applied to situations in cybersecurity or having self-driving vehicles navigate traffic, he states.

Pluribus mastered the game playing against five copies. This approach has been adopted by other AIs in order to master games like StarCraft II, Dota 2, and Go. It began its journey as a beginner poker player, with no prior knowledge. Later it learned the rules and improved its strategy through reviewing each round.

You can play like a bot

Pluribus won against five professional poker players by winning an average of 48 mega-big blinds per round. This represents how many big blinds were each player won in a thousand hands of poker.

To deter players from collaborating against Pluribus, each player received an alias.

Brown said that the bot was not hiding its identity. Pluribus is able to play the first few actions of a round immediately because it has already planned its strategy. Human players typically take a few seconds to decide.

Jason Les, a professional player in poker, said that Pluribus knew which player was Pluribus so the AI could be tricked. He was one of five players who played against Pluribus in the rounds.

Read on: AI creativity is real and machines will see it.

Les states, "If you want to push AI to the limit, do all you can to find weakness." "Obviously, we weren’t able."

Les also played against Libratus during 2017. He said, "It was amazing that they had made this much progress in a matter of years." "What was especially impressive about this challenge? The AI played faster on less computing power."

Pluribus had to choose from fewer options, so the AI combined similar hands (e.g. king-high flush and queen high flush) and considered only a few sizes of bets.

Brown states, "Betting $150 is a lot like gambling $151 at the end of the night." Pluribus treats these bets together and makes them equal.

No guarantees

Brown states that the AI is produced using very limited computing resources. Training Pluribus took less than 500 gigabytes of memory. Cloud computing would have cost you less than $150.

Pluribus performed better than human players in poker, but according to the Nash equilibrium, there was no theoretical guarantee it will always win.

Nash equilibrium can be found in non-cooperative game where each player has their own strategy list and cannot improve their performance by switching to a new strategy. A Nash equilibrium strategy is unbeatable at Heads Up Texas Hol 'Em. However, there are no strategies that can be used for the six-player variant. Texas Hold Em Poker Tips - 2 Tips To Targeting Players