Facebook develops a poker AI solution that can take on the pros
ReBeL can teach itself how to play poker with no prompting
Artificial intelligence (AI) and poker are highly related. For some experts, the game of poker, a game of imperfection, can provide a lot of teaching ability for AI machines. Case in point, the latest release by Facebook, a poker-playing AI but that uses a different approach than just relying on domain knowledge. This AI creation is capable of defeat the best poker players out there by using Recursive Belief-based Learning (ReBeL) that operates two different AI models to learn about making correct decisions based on given and hidden information.
This ReBeL tool also uses reinforcement learning to learn how to play the game rapidly. However, for the model make better decisions, this software also uses two AI models – one for value, another for policy – to create public belief states. This works in a similar way in which a poker player tries to figure out what’s on his or her opponent’s mind. Therefore, ReBeL does a lot more than just gathering game data, such as cards, bet size or hand range; it also goes deeper into information that was not given and which is not visible. By gathering this data, the AI machine comes up with subgames alternatives to figure out all the probabilities and all the possible actions from opponents and make a decision taking all these aspects into consideration.
This AI poker player was already tested against one of the top heads-up Texas Hold’em players, Dong Kim, and it was able to defeat him while playing faster than other human opponents. The results were even more impressive than the Libratus system, which scored lower than ReBeL. As poker is an imperfect game, this advancement can be taken to other areas in which decisions are made in a similar way – considering available and unavailable information – like in negotiations, auctions, cybersecurity, and self-driving trucks and cars.