This article is from: Qubit
As expected of a Chinese AI, playing mahjong "easily" reached the top.
According to the latest news from Tencent, its chess and card AI "Fine Art" LuckyJ has refreshed the best result of AI in the field of mahjong on the Japanese mahjong platform "Tianfeng" .
Only 1,321 games were needed for Fantastic Art LuckyJ to reach the highest level of AI, 4,052 games less than the second place Suphx.
And its ability is stable at the tenth rank. You must know that the average stable rank of the top human players is only 7.4.
On the Japanese Mahjong Tianfeng platform, Fantastic Art LuckyJ has also become one of only 27 players (including AI) who have reached the tenth rank, and the average number of active players on this platform is 238,000.
Not only Japanese mahjong, but in national standard mahjong, Fantastic Art LuckyJ has also defeated six professional players, becoming the first mahjong AI to defeat professional national standard mahjong players .
One of the national standard mahjong players made such an evaluation.
What we usually call clever hands, a flash of inspiration, or even a choice based on experience and feeling to die and survive may be a routine operation for AI.
After Jueyi LuckyJ won the tenth dan of Japanese mahjong, many netizens came to congratulate him.
After all, behind AI playing mahjong, the improvement is actually the ability of AI in problem decision-making, which is also conducive to allowing AI to solve more complex problems in real life.
So how did Fine Art LuckyJ do it?
Playing mahjong AI needs to learn a more balanced strategy
Let's first look at the difficulty of AI playing mahjong.
For AI, playing games is an excellent way to test its ability, common forms such as chess, go, and glory of kings.
Among them, Go and Chess are perfect information games . That is to say, both players in the game can see the overall information before each move, that is, they can see each other.
This is not very difficult for AI, because it can solve it violently through powerful computing power and find an optimal solution.
But playing mahjong is more complicated.
Not only can players not see the cards of multiple other players, but also a lot of information is hidden in the cards that have not been revealed. This is a typical imperfect information game .
That is to say, both AI and human players can only see 13 cards in their hands at the beginning, and more than 100 cards are unknown. And every time you play cards, you have to make a series of complicated decisions, such as whether to eat cards, touch cards, win cards, etc., and the decisions you make must take into account both offense and defense.
At the same time, the operation of other players to eat the bar will change the order of the next card drawing, making the decision more complicated.
In such an icon, where the abscissa represents the amount of observable information and the ordinate represents the amount of hidden information, it can be seen that mahjong contains far more hidden information than other board games .
So what to do?
Tencent AI Lab proposed a self-playing technique based on reinforcement learning and regret value minimization .
This enables AI to self-learn and improve its capabilities from scratch, and finally converge to a strongest mixed strategy, which can have a more balanced strategy ability in the actual battle process.
At the same time, considering that the traditional imperfect information search algorithm is difficult to play a big role in the face of mahjong, the researchers also proposed an efficient imperfect search method based on the idea of optimistic value estimation , so that AI can play a large number of games with hidden information. In the state, adjust the current strategy in real time to better cope with the changing battle situation.
According to the researcher, compared with humans, AI has a more balanced strategy in the game of mahjong, and the calculation of the situation is very accurate, including the expected profit of playing each card, which types of possible future games, and so on. Under such "strategic" training, AI can enter other industries more quickly in the future.
In the actual combat test, Fantastic Art LuckyJ faced off in "Tianfeng".
This is an old Japanese mahjong game platform, founded in 2006.
Judging from the bootstrap distribution, Jueyi LuckyJ is significantly stronger than the other two Japanese mahjong AIs (Suphx, NAGA): LuckyJ vs Suphx p value=0.02883; LuckyJ vs NAGA p value=3e-05.
In addition, in the nearly 2,000 games of national standard mahjong, the average win rate of Fantastic LuckyJ reached 1.76.
(here means the settlement unit of the national standard mahjong, the larger the value, the more you win)
Game AI has been applied across industries
However, so much effort to improve AI's ability to play mahjong is certainly not just to let it play mahjong.
According to a researcher at Tencent AI Lab, the ability to promote decision-making AI in the game environment ultimately hopes that AI can move from virtual to reality and solve complex problems in the real world .
The real world is full of scenarios that require decision-making under imperfect information, such as financial transactions, autonomous driving, transportation and logistics, auction systems, etc.
Moreover, Tencent AI Lab already has actual cases.
Another of its decision-making AI , Juewu , has learned to identify the location of lesions in the pathological full scan image, and the efficiency is 400% of the traditional method.
This AI method of finding the optimal viewing path is based on reinforcement learning.
It avoids the traditional exhaustive method to analyze local image slices, but first decides to find areas of observation value, and obtains representative features across multiple resolution levels to accelerate the completion of full-film interpretation.
By imitating the human way of thinking, it not only improves the efficiency of watching movies, but also saves costs.
To sum up, Juewu who can play "Minecraft" can already help the world, and people look forward to what can Juewu who can play mahjong do across industries?
What do you think are the applicable directions for LuckyJ?