Artificial Intelligence Microsoft mastered mahjong





Scientists from Microsoft Research (MSR) Asia have created an artificial intelligence (AI) system that has learned the intricacies of mahjong. Mahjong is considered one of the most difficult games in terms of mastering algorithms, because there is a random factor in it.



Super Phoenix (Suphx) was the first AI system to receive 10th dan (rank) on Tenhou , an online riichi mahjong competition platform with over 300,000 participants from around the world. Only 180 people have such qualifications. This is the maximum discharge for the algorithm at the moment, only a few recognized professionals play at a higher level.



To create the system, deep reinforcement learning was used, a comprehensive principle of learning algorithms based on a combination of deep learning and reinforcement learning. Suphx mastered the intricacies of mahjong by participating in 5,000 games in 4 months. The system gradually increased the quality and speed of the game. Now the algorithm has its own strategy, which includes deliberately weak moves, allowing in the end to get a greater advantage.



Unlike chess and go, in which AI has already made significant progress, mahjong is a game with incomplete information. Not all factors are known to the participants, for example, it is impossible to see images on the opponent's knuckles. The game process resembles poker - participants need a good memory, observation, strategic thinking and the ability to bluff.



According to scientists from MSR, the presence of a random factor in mahjong is a unique challenge for AI systems, and the experience gained in teaching algorithms to nonlinear problems can find its application in real life situations involving a set of unknown factors.

Earlier, Microsoft Research and Maluuba trained artificial intelligence to play Ms. Pac-Man, which was released for one of the first Atari 2600 home consoles. The algorithm broke the world record with 999,990 points, while the best result for a person was 266,360 points.



All Articles