After 5,000 games, Microsoft’s Suphx AI can defeat top Mahjong players

Bolstered by algorithmic breakthroughs and an abundance of cheap computing power, AI has become a veritable games champion, notching wins against the world’s top-ranked players of chess, the Chinese board game Go, Montezuma’s Revenge, and countless others. In yet another victory for the machines this week, Microsoft detailed an AI system — Super Phoenix, or Suphx for short — capable of besting champions at Mahjong.

“For as long as researchers have studied AI, they have worked to build agents capable of accomplishing game missions,” said corporate vice president of Microsoft Asia Pacific R&D Group and MSR Asia Dr. Hsiao-Wuen Hon. “Mahjong is more complex than other board games, [so] playing becomes an art as well as a science. Good Mahjong players rely on a combination of observation, intuition, strategy, calculation, and chance that presents unique challenges for an AI system.”

As Dr. Hon and colleagues explained further, Mahjong is what’s known as an imperfect information game, meaning a number of factors remain unknown to players throughout matches. For instance, they must account for opponents’ unseen tiles and decide whether to fold, leading to bluffs.

To overcome this, the team had Suphx play against thousands of people on Tenhou, a Japan-based global online Mahjong competition platform with more than 300,000 members, to self-learn common strategies. It soon developed its own play style and learned to balance attack and defense moves, strategically weighing short-term losses against long-term gains and making decisions with ambiguous data.

Suphx demonstrated expert-level play after 5,000 games over the course of four months, and it recently became the first AI system to compete at Tenhou’s 10th dan ranking, something only 180 humans have ever done.

Microsoft’s AI gains come after San Francisco-based OpenAI’s Dota 2 bot defeated 99.4% of players in public matches and a team of professional players twice, and after DeepMind’s AlphaStar AI swept StarCraft II pros in head-to-head match. Tencent’s Honor of Kings AI earlier this month beat a team of professional players. And beyond the video game arena, Facebook AI Research and Carnegie Mellon University in July detailed Pluribus, a poker-playing AI that Facebook claims is the first to best 15 human professionals in Texas Hold’em.

The advancements aren’t merely pushing forward game design. Rather, they’re informing the development of systems that might one day diagnose illnesses, predict complicated protein structures, and segment CT scans. “[Our game-playing AI] is a stepping stone for us all the way to general AI,” Alphabet’s DeepMind cofounder Demis Hassabis told VentureBeat in a previous interview. “The reason we test ourselves and all these games is … that [they’re] a very convenient proving ground for us to develop our algorithms. … Ultimately, [we’re developing algorithms that can be] translate[ed] into the real world to work on really challenging problems … and help experts in those areas.”