Facebook develops poker AI that 'calculates the probability of every action of the other party', and will be applied to cyber security in the future

Facebook's AI development team has announced the AI ' ReBeL ' that shows superhuman poker play. By considering the probability of actions that players participating in the game can take, ReBeL will demonstrate high performance even in the types of games that AI has been weak in until now, and in the future, fraud detection and cyber security It is expected to be active in the real world, such as in fields.

ReBeL: A general game-playing AI bot that excels at poker and more

Games in game theory are ' perfect information games ' where each player can fully understand each other's decision contents and game development, such as shogi and chess, and 'incomplete information' such as poker where each player hides his or her hand. It is classified as 'game'.

In 2016, Google's AI AlphaGo defeated Go's top Go player Lee Seddle 9th Dan , and AI won in Go, which was considered to be 'the only board game that human beings have not yet lost to AI' at that time. , AI is very good at full information games.

In 2019, three years after AlphaGo's victory, AI `` Pluribus'' jointly developed by Facebook and Carnegie Mellon University defeats professional players in 6-player poker, AI is a human being even in an incomplete information game However, he is still not good at it compared to perfect information games.

AI finally beats professional poker players in 6-player poker-GIGAZINE

According to Facebook AI researchers Nome Brown and Anton Buffchin, AI's weakness in incomplete information games is due to AI's ' reinforcement learning (RL) + Search' algorithm. It doesn't work in imperfect information games.

The 'RL + search' algorithm, which is also used by AlphaGo and others, tends to calculate by assigning a fixed value to each action regardless of the probability that a specific action will be selected. In games like chess, this problem doesn't really surface because the good move is the good move and the bad move is the bad move, whether or not the player uses it frequently. However, in poker, the value of a particular action fluctuates depending on how often you use it, just as frequent use of ' bluffs ' can make you read a lot, so the probability that an action will be selected becomes very important.

Therefore, Facebook's newly announced AI 'ReBeL' calculates the probability distribution of various 'beliefs' that each player can have to determine the action. As a result, ReBeL will beat two poker AIs, including Baby Tartanian8, a poker AI developed by Carnegie Mellon University, in Texas Hold'em , the most popular rule of poker, to outperform top human players. It was successful. In addition, Liar's dice , a dice game that requires the opponent to lie, showed results close to Nash equilibrium , and showed strength in imperfect information games other than poker.

'ReBeL performed superhumanly, even with far less information than traditional AI,' Brown et al. Develop a universal and versatile technology. We believe this is a major step forward and a step towards the development of real-world AI such as fraud detection and cybersecurity. '