AI opponents are usually looked down upon by players because of their limited number of moves and rather predictable decision-making. It usually doesn’t take long for even beginners to learn how to perfectly counter most patterns a computer opponent can come up with. Plenty of popular gaming formats are well-known for introducing a level of difficulty by simply having a better head start than human players or artificially inflating stats.

There’s a different case for the Mahjong Game, with its many layers of possibilities that cannot be resolved by brute-forcing all possible combinations at once. Between 4 and 3 years ago, mahjong players from all over the world were taken by surprise at the development of Super Phoenix. This learning AI was able to outcompete 99% of all officially-ranked human players of the Tenhou platform.

The introduction of deep reinforcement learning along with newly-introduced techniques like global reward prediction, oracle guiding, and run-time policy adaptation brings an exciting new phase in mahjong AI. The most recent product of these methods, Suphx, is out to surprise players with its stunning mastery and sheer ability to outplay most human opponents.

But what are we expecting with this 2019 technological marvel of the gaming sphere? What are the aspects that make Suphx the toughest mahjong AI in the world at present?

Adaptive gameplay

Suphx is made to learn from every game it plays. Opponents will directly influence how it plays future games, incorporating, predicting, and most likely, actively countering these strategies in case it runs into them more than once. There’s a reason Suphx has outplayed many of the world’s top mahjong players. So far, most reviews from opponents who took on Super Phoenix are that it leans more on defensive strategies.

A fusion of old and new

This mahjong automation is designed to combine traditional AI methods like search algorithms and pattern recognition with new approaches such as machine learning and neural networks. Gone are the days when AI typically lists all possible combinations and attempts to win games through brute force. Suphx can choose between several approaches based on the situation imposed by its opponents. Winning strategies are going to be constantly refined and updated to give players an added sense of challenge.

Constant development

This AI is constantly learning and can develop new ways to win games over time. It gains experience regardless of outcome and can combine various data to formulate a winning strategy. The more games it plays, the more combinations it absorbs from its opponents. Future updates are looking into giving it more strategic depth and additional playing behavior.

Applications outside of gaming

While this may be a game to some, what has been learned from Suphx can be transferred to various fields beyond mahjong. This is a stepping stone to developing specialized AI that deals with decision-making in complex environments where information is scarce or incomplete. Fields of work like logistics and supply chain management benefit easily from Suphx’s methods in achieving maximum efficiency.

Some people remember Deep Blue’s victory over Garry Kasparov in 1997. This IBM creation led to the development of even more sophisticated chess computers and AI assistants that help players develop existing strategies and create twists in executing them. Suphx follows the same path but with applications beyond mahjong. The future of developing more versatile AI platforms that perform more sophisticated tasks with near-flawless efficiency is on the horizon.