By Super nerve
Scene Description: Microsoft has released a mahjong AI model, in professional sports platform to achieve the highest success dan. They swept the country in the global entertainment and even the birth of what God AI bird to overcome the difficulties, the birth and what deeper meaning of this technology.
Keywords: Mahjong Suphx depth of reinforcement learning
Author: nerve knife neurotic
Edit: nerve stars
In the World Conference on Artificial Intelligence held today, Microsoft released a "bird god AI" - Suphx, on the professional mahjong competitive platform, the strength is better than the average of the top human players.
Suphx, stands for Super Phoenix (Phoenix super), in March 2019 landed in Japan Mahjong professional sports platform, "Fung days."
Executive vice president of Microsoft's global Shum site presentation Suphx
On this most famous of mahjong platform, AI can participate in open competitive "special never ending" in, Suphx with more than 5,000 human players started four games mahjong game, and gradually played their own strength and level.
To June, Suphx reached ten special section highest Dan never ending, but did not limit the aspirations of the most Suphx eleven Dan "Tianfeng bit", but the platform also does not allow access to the most advanced AI system war room.
Tianfeng platform since its launch in 2006, four player mahjong reached Ten sections of about 180, while the active duty period ten human players, but also a dozen. But in stable Dan measure the strength of the level, Suphx reached 8.7 section 7.4 section human players far ten segment higher.
Suphx is all day on the Phoenix platform in the highest levels of AI
Previously, Tianfeng mahjong platform also active in two other AI systems, respectively, in 2015 published by the University of Tokyo "blast beat", and in 2018 issued by Dwango "NAGA25" but stable Dan both are below 6.5 , is Suphx far behind them.
Mahjong thousand years of history: the slow evolution of leisure
Mahjong, also known as "sparrow" or "bird card", is authentic quintessence.
Mahjong statement about the origin of a wide variety, as it is also difficult to trace the truth, but it is certain that, since the advent of Mahjong, a national pastime project, mahjong has been popular in folk, enduring.
Mahjong predecessor, dating back to the Han Dynasty leaves play cards
The tiles made of symbols and also after a number of changes, the earliest tiles bamboo and animal bones production, after which time there had cards.
In the dignitaries, once used rhino horn, ivory, gold, silver, blue and white porcelain to create, then as the tiles by the Seiko Qiao will have a picture carved from.
LV, Prada, Hermes these international brands, have launched high-level customization Mahjong
Until after 1960, the popularity of plastic goods, and the development of mechanization, making mass production to gradually mahjong material.
But in addition to changes in the production process, mahjong in the highest technology , in addition to AI, probably automatic mahjong machine up.
AI points to win, thanks reasoning
Prior to AI research, a lot of people think that once in mahjong, luck is the decisive factor. But in fact, the competitive rules of mahjong, in fact, is a very complex issue.
136 mahjong permutations and combinations are many kinds of results , between the same player two cards, mixed up with the other three players the cards, as well as their own are drawn, but also "eat", "touch", "bar "let Licensing Board will produce dynamic changes.
Secondly, this is an imperfect information problems . Each player can only know their own 13 hand, and the card is played, while others cards and the remaining cards are unknown, these hidden information led to a number of variables.
The complexity of the comparison of several card games
Even very experienced players, it is difficult to sort out the logical relationship between the surface and the best known brand of play , a wealth of hidden information will lead to increased complexity of the game.
It takes in the whole process, good planning strategy, such as in an unfavorable situation, the strategic "blasting" allows players to win fourth place, in order to prevent the second go-ahead score.
So, you want to create a master mahjong AI, only the strong force is not enough to count, but also required that the AI has intuition, prediction, reasoning and fuzzy decision-making capabilities.
A generation of bird god, reinforcement learning by depth
In response to these difficulties, Microsoft take advantage of the depth of reinforcement learning to build Suphx, by the latest algorithm, step by step in learning and debugging, the promotion of competitive mahjong has become the strongest bird God.
Bird is doing to God Suphx
The first is the "initialization" phase, the use of "Tianfeng" platform of public data, researchers supervised learning , get an initial model, and the model is based on a way of reinforcing the lessons of self-training game.
Then, against imperfect information game challenges, Suphx innovative way to try to improve the prophet coaching techniques to strengthen the effectiveness of learning.
During the training phase, the use of invisible hidden information to guide the direction of AI training model, to make it more clear learning paths, close to the optimal path of meaning perfect information, thus contributing to the AI model-depth understanding of visual information, to find effective strategies .
The classic search tree structure, AI want to minimize the maximum benefit opponent
But not for mahjong games
In addition, complex licensing mahjong surface expression and scoring mechanisms they use to predict the overall technology, build a bridge between the end game and each round after the results of eight.
By compact design predictor model to understand the impact on the final results of each round, so there is the perspective of global decision-making.
The research team also introduced a new mechanism to regulate the dynamic process of the Board, let Suphx adjust the policy according to the latest information in the inference stage, making adaptive decisions.
The final step is to enter combat, through continuous human players to participate in the game, let the AI continuous learning to upgrade their skills.
In the council, Suphx not only the first player or Big Three
Since March Tianfeng into the platform, Suphx been in constant self-evolution. At present, the balance of attack and defense, Suphx enables a more sensible strategy than the top human players, strategically complete trade-off between short-term and long-term loss of income, and rapid decision-making based on existing vague information.
Mahjong AI: not just the outcome of the poker table
Thanks to its new training techniques and algorithms, Suphx unique in style and play.
Top human players on the Tianfeng platform, it praised Suphx on social media, he thought he saw a lot of Suphx game, learned a lot of techniques never seen before.
In addition, there are a lot of players say that in Suphx battle, the Battle learn practical skills, and therefore have to be referred to as "Mahjong textbook", "Suphx teacher."
Technical revelation brought 136 Mahjong
For winning or losing mahjong, urban people are enjoying the thrill of luck and experience , and master enjoy the intellectual contest .
Such a "bird god" AI, in addition to create a invincible mahjong coach, but also open a new perspective, let us from the dimension data algorithms to resolve this entertainment.
No longer as dependent on luck as a gambler, but under the mental aura, gradually leaving behind those random uncertain things, to explore a set of laws victory.
This is not exactly the way AI development, the most fascinating beam of light it?
Content Reference: Microsoft Research AI headline "Microsoft super mahjong AI Suphx, crack imperfect information game" ( https://mp.weixin.qq.com/ S / S- axCx41WKD JG . 2B iGGTZfg)
-- Finish--
—————————————
Past Wonderful: