The artificial intelligence of a Hu family has come

  Artificial intelligence, Deep learning

Scenario description: Microsoft has released an AI model for playing mahjong, successfully reaching the highest level on a professional competition platform. What difficulties did the birth of AI sparrow god overcome and what deep significance did the birth of this technology have in this popular entertainment activity across the country and even around the world?

Key words: mahjong Suphx intensive learning

At the world conference on artificial intelligence held today, Microsoft released a “sparrow god AI”-suphx. on the professional mahjong competition platform, Microsoft’s strength surpasses the average level of top human players.

Suphx, whose full name is Super Phoenix, landed on Japan’s professional mahjong competition platform Tianfeng in March 2019.

图片描述

Microsoft Global Executive Vice President Shen Xiangyang Presents Suphx Live

On this most famous mahjong platform, in the open competition “special room” in which AI can participate, Suphx has played more than 5,000 four-person mahjong games with human players, gradually playing his own strength and standard.

By June, Suphx had reached the top ten levels of the special room. however, it was only the platform that did not allow AI system to enter the highest level room to play.

Since Tianfeng Platform was launched in 2006, there are about 180 players who have reached 10 stages of mahjong for four people, and there are only a dozen human players in the 10 stages of mahjong. However, on the stable level of strength, Suphx reached 8.7, far higher than the 7.4 of the ten human athletes.

图片描述
Suphx is the highest level in all AI on tianfeng platform.

Previously, Tianfeng platform was also active with two other mahjong AI systems, namely “burst play” released by the university of tokyo in 2015 and “NAGA25” released by Dwango in 2018. however, the stable bands of both were below 6.5 and were far behind by Suphx.

Mahjong with a History of 1,000 Years: Slowly Evolving Public Leisure

Mahjong, also known as “sparrow” or “sparrow brand”, is the quintessence of authentic Chinese culture.

There are various theories about the origin of Mahjong, and the truth cannot be verified. However, what is certain is that Mahjong has been popular among the people since its inception as an entertainment program for all.

图片描述
Mahjong’s predecessor can be traced back to the Ye Zixi brand in the Han Dynasty.

The symbols and making of mahjong tiles have also undergone many changes. The earliest mahjong tiles were made of bamboo and animal bones, and there was also a period of playing cards.

Among the dignitaries, rhinoceros horn, ivory, gold, silver and copper, and blue and white porcelain were also used to make mahjong tiles. At that time, the tiles were carved one by one.

图片描述
International brands such as LV, Prada and Hermes have all launched advanced customized mahjong

Until 1960, the popularity of plastic products and the development of mechanization led to the gradual mass production of mahjong with materials.

However, apart from the changes in production technology, the highest technology in mahjong, apart from AI, may be automatic mahjong machine.

AI Hu Card, All Depends on Reasoning

Before AI research, many people once believed that luck was the decisive factor in mahjong. But in fact, the rules of mahjong competition are very complicated.

There are many kinds of results in the arrangement and combination of 136 pieces of mahjong. Between the two cards played by the same player, the cards played by the other three players are mixed with their own cards. Moreover, the “eat”, “touch” and “bar” will make the game change dynamically.

Secondly, this is an imperfect information problem. Each player can only know his or her 13 cards and the cards played, while others’ cards and the remaining cards are unknown, which leads to many variables.

图片描述
Complexity Comparison of Several Card Games

Even experienced players are difficult to clarify the logical relationship between the known card face and the best play, and the rich hidden information will lead to increased complexity of the game.

This requires a good strategic plan throughout the process, such as strategically “shooting” the fourth player to win when the situation is unfavorable, in order to prevent being overtaken by the second overall score.

Therefore, in order to build a master mahjong AI, only strong calculation power is not enough. What is more needed is that AI has the ability of intuition, prediction, reasoning and fuzzy decision-making.

Becoming a generation of sparrow gods depends on intensive study

In view of the above difficulties, Microsoft has promoted Suphx, which is built by intensive learning, to become the strongest sparrow god of competitive mahjong through the latest algorithm in learning and debugging step by step.

图片描述

This is how the sparrow god Suphx did it.

First, in the “initialization” phase, researchers obtained an initial model through supervised learning by using the public data of the “Tianfeng” platform, and on the basis of the model, strengthened learning training was conducted by means of self-game.

Then, in response to the challenge of imperfect information game, Suphx innovatively tried the prophet coaching technology to improve the effect of intensive learning.

In the training phase, invisible hidden information is used to guide the training direction of the AI model so that its learning path is clearer and closer to the optimal path in the sense of perfect information, thus prompting the AI model to deeply understand visible information and find effective strategies.

图片描述

The classic search tree structure, AI hopes to minimize the opponent’s maximum profit, but is not suitable for mahjong games.

In addition, for mahjong’s complicated face expression and scoring mechanism, they use the overall prediction technology to build a bridge between the results of each round and the final set after 8 rounds.

Through the elaborately designed predictor, the model can understand the influence of each round of competition on the final result, thus having a global decision-making perspective.

The research team has also introduced a new mechanism, which can dynamically control the process of the match board, allowing Suphx to make policy adjustments and make adaptive decisions based on the latest information in the reasoning stage.

The last is to enter actual combat, through continuous participation in human players in the game, let AI continue to learn to improve their skills.

图片描述

In this bureau, Suphx is not only Hu brand but also da San yuan

Since entering tianfeng platform in march, Suphx has been continuously evolving itself. At present, Suphx can realize wiser strategies than top human players in balancing attacks and defenses, strategically balance short-term losses with long-term gains, and make quick decisions based on existing fuzzy information.

Mahjong AI: It’s not just the outcome at the table

Thanks to its new algorithm and training skills, Suphx is unique in its play and style.

Top human players on tianfeng platform praised Suphx highly on social media. he felt that he had watched many Suphx competitions and learned many technologies he had never seen before.

In addition, many players said that they had learned practical skills in the Suphx match, so they also called them “Mahjong Textbook” and “Suphx Teacher”.

图片描述

Technical Enlightenment from 136 Pieces of Mahjong

For winning or losing mahjong, the common people in the market enjoy the pleasure of luck and experience, while the superior enjoy the contest of intelligence.

Such a “sparrow god” AI, in addition to creating an invincible mahjong coach, can also open a new perspective, let’s analyze this entertainment from the dimension of data algorithm.

Instead of relying on luck like a gambler, under the halo of brain power, we gradually cast aside those random and uncertain things and explore a set of rules for victory.

Isn’t this the most fascinating light on AI’s development?
图片描述