Text / Microsoft Research Asia
Why in the field of artificial intelligence, scientists are always keen to let AI play chess and play games with humans? From simple checkers, backgammon, to more complex Chinese chess, chess, and recently very popular Go and Texas Hold'em. Every time AI successfully defeats the human player in a certain intellectual game, it will make everyone stunned and laments that AI will replace humans in the near future...
Fortunately, AI took over the earth has not yet happened. Not only do we not need to worry about such things, but we are also delighted to discover that technological advances in artificial intelligence have brought more convenience to life. An AI who plays chess is not the ultimate goal of scientists. The more positive significance is that the AI ​​algorithm continues to refine and improve in the process of studying chess art. It will bring more design innovations and thus fundamentally improve artificial intelligence. The ability and scope of the algorithm.
The reason why scientists are willing to choose chess games is because they have been regarded as the symbol of human intellectual activity since ancient times and the AI ​​that simulates human activities is naturally aiming at this. Successfully reaching humans or even humans can attract more attention and devote themselves to the research and application of artificial intelligence.
On the other hand, chess is also well suited as a benchmark for new AI algorithms. The rules of board games are concise and clear. Winning and losing are all on the disk and are suitable for computers. Theoretically, as long as there are new breakthroughs in computing power and algorithms, any new board game may be overcome.
In addition to board games, card games (such as Texas Hold'em, bridge cards, mahjong, landlords, etc.) have gradually become new directions for artificial intelligence research. In larger video games, such as StarCraft and Minecraft, scientists have also begun a new round of innovation in AI algorithms. What exactly is the difference between these different games in the eyes of researchers? What is the significance of these research results for our lives? Let's take a look at these two issues for everyone.
Chess AI family
To understand chess AI, we can start with its classification. This family can be divided into two contexts according to the brand's “candidnessâ€: one is good at “open skylight to speak bright words†and the other is “intelligence guessing†smart master.
Chess information such as Chess, Go, etc. is public, and the information received by both players is exactly the same. Therefore, it is also referred to as the “complete information†AI game; while Texas Hold'em, Bridge, Mahjong and other games cannot be seen by everyone. The opponent's hand, so called "incomplete information class" AI game.
Complete Information - Even if I see it
As the name implies, that is, everyone can see that the game information received by the two players is completely equal, such as chess and Go. In this type of game, the AI ​​only needs to search for the winning percentage in each case after calculation according to the current disk. In order to improve the search efficiency, it is generally necessary to extensively and deeply prune the "game tree" generated during the search process. This is how far we often say and how many we usually say. To be far, we generally need to make AI less likely to look at opponents and places where they are less likely to go, called strategy functions. In order to be accurate, we need to more accurately assess the winning rate of the multi-step disk and call it the value function. Finding a suitable function, coupled with the computational power of the computer, makes it possible for AI to reach or exceed humans. In the selection of the game tree and strategic value function, the "complete information" chess AI algorithm undergoes iterative updates from the "AlphaBeta Pruning Algorithm", "Monte Carlo Tree Search" to "Deep Neural Network", and the functions are also continuously updated. "evolution".
Checkers, Backgammon | Difficulty Index 燑/strong>★
The space complexity of checkers and backgammon is relatively low. Even without the need to prune the game tree, the computer can compute all the possibilities of the disk with powerful computing power. So in this relatively simple board game, humans have no chance of defeating AI.
Chinese Chess, Chess | Difficulty Index 牎铩铩铧/p>
Chess has a high degree of space complexity, and violent solution methods are not feasible. However, it is relatively easy to find a suitable value function. Taking chess as an example, an approximate score can be given based on the type and position of the remaining pieces on the board. For example, if there is a queen with 10 points on the board, there is a car plus 5 points, there are 3 points on the horse, based on which the calculation function is based. In order to improve efficiency, chess also has a huge start and end database to ensure the accuracy of the calculation of the endgame. Relying on these rules, "deep blue" defeated the human chess champion for the first time in 1997. Afterwards, the computer chess program can even run on the PC and beat top human players.
Go | Difficulty index 燑/strong>★★★★
The space complexity of Go is high, and it is estimated that Go's decision point is as many as 10 to 170 times. Finding the right strategy and value function has always been the core issue of Go AI. The Monte Carlo tree search algorithm uses a probabilistic approach to help Go AI find a more accurate value function and help the program reach the level of the amateur high segment. With deep neural networks, researchers have found better strategies and calculations of value functions. Through enhanced learning, AI can also infinitely simulate various game scenarios and generate hundreds of millions of data for training to generate more accurate functions. Dacheng’s “AlphaGo†won the world’s top chess player Li Shishi 4:1 in 2016. The ongoing battle between the new AlphaGo version and Ke Jie does not know what new algorithms and enlightenment we have brought.
>>>>Incomplete information class - three missing ones are not afraid of it
In the course of the game, if the information obtained by both parties is incomplete and unequal, it needs to guess the probabilities of the opponent's bottom cards. It is classified as incomplete information, such as Texas Hold'em, Bridge, and Mahjong.
Incomplete information game requires more complex reasoning ability, not only to look at what cards others have played, but also to guess what cards others have in their hands, and to calculate their optimal card exit method based on information implied by opponents' actions. . Since the opponent's behavior not only implies his information, but also depends on how much he knows about our private information, how much information our behavior reveals. Therefore, this kind of "circular reasoning" makes it difficult for a person to infer the state of the game in isolation.
Von Neumann, the founder of modern game theory and pioneer of computer science, has a well-known saying that it is appropriate to describe incomplete informational games: “There are many illusions and tricks in the real world, and you need to think about how others see through your strategy. This is the game involved in the theory I propose."
Texas Hold'em Poker | Difficulty Index 燑/strong>★★★★
The search complexity of Texas Hold'em is 10 to the 160th power, which is close to International Go. The principle of "Nash equilibrium" is mainly used in the game - at a particular moment, the optimal response to other participants is sought. Compared with Go, poker not only has to make complex decisions based on incomplete information, but also has to cope with opponents' bluffs, intentional weaknesses, and other moves. At the end of last year, DeepStack, a computer scientist from the University of Alberta, Charles University and Czech Technical University in Prague, defeated human professional players in two-man No-Limit Hold 'em; earlier this year, Libratus was developed by Carnegie Mellon University. It defeated four more outstanding professional players. This is AI's landmark breakthrough in incomplete information game. For artificial intelligence, the next challenge is to conquer multiplayer poker.
Mahjong | Difficulty index 燑/strong>★★★
At present, mahjong is mainly popular in Asia. Therefore, both the national standard mahjong and the Japanese mahjong have strong AI, which is higher than the average human level. However, there is still a long way to go from the level of the top human masters. Mahjong's search is much less complex than Go and Texas Hold'em, but because (generally) is a four-player game, its technical requirements are very different from the two-person zero-sum game (eg, one-on-one Texas Hold'em). The solution of the two-person zero-sum game is mainly to find the Nash equilibrium strategy or the approximate Nash equilibrium strategy. Because there are multiple equalization possibilities and multi-person interaction in the multiplayer game, the Nash equilibrium strategy has no guarantee of performance, and it is technically This means that everyone will almost start from scratch, which brings new technical challenges (similar to multiplayer poker).
StarCraft, My World | Difficulty index ★/strong>★★★★★
The complexity of games such as StarCraft and My World is not only in the information asymmetry, but also in its more open game rules. Such games are more similar to what people have encountered in the real world. The openness of the rules of the game will make the game world a new situation that many computers can hardly handle. Such as special features of terrain that have never appeared before, opponents of long-term plots and planning. Excluding the computer's advantage in operating speed, the computer has not really proved its ability in these games.
Tips: Compared to full-information games, Texas Hold'em and mahjong are sometimes lost to the game. Not all because of poor play. It is possible that the game is not good from the start, so the win is relatively low. The composition of luck is very heavy in this kind of chess game, which is different from chess and chess. In Go, the confrontation between professional players and non-professional players has never been lost because of luck.
What is the significance of chess and card AI?
From the perspective of social feedback, some people may worry that the success of machine-to-play humans will undermine the meaning of chess art itself, they will challenge the value of professional players, and even let more people give up learning chess; some people However, he thinks that such events can popularize various kinds of chess so that more people can become interested in these kinds of games and games. There are also people who exaggerate the threat posed by AI to humans.
It may actually cause some social problems in the process of technological advancement. However, it will be encountered at every historical stage in humanity. Humans will not slow down the pace of technological advancement, and some social problems that now seem to cause public discomfort. It will be gradually resolved. Einstein said: "Science, whether it brings happiness or catastrophe, depends entirely on people." After all, in a battle between the brain and the AI, not the machine has defeated humans. But mankind surpasses itself!
The wider use of AI in the future will certainly be in the real world of unmanned, smart security, and artificial intelligence assistants. In the real world, the problems encountered by AI are ever-changing. There will not be a unified rule and a unified function that will help explain the behavior. The chess and card AI is just a very early exercise of artificial intelligence.
Therefore, AI plays a variety of chess games against humans. Its significance lies not in winning or losing. More importantly, people are familiar with this kind of game and can learn about the latest progress of AI through competition. This has greatly promoted the development of AI. After all, the process of AI evolution is still quite long. Even watching the masses needs to understand the areas that will be closely related to everyone's life in the future.
We also look forward to the success and breakthrough of chess and card AI can inspire AI in other areas of research and application, and can apply innovation to more industries and fields, and encourage more people to devote themselves to the research and practical use of AI. Human life is more convenient, efficient and intelligent, enabling the entire human race and nature to benefit from AI. In the process of playing chess, the artificial intelligence research field technology and expert personnel training system have also been further improved, thereby promoting artificial intelligence to overcome the "highlands" of one technology and one application.
Author: Microsoft Research Asia, a senior researcher Yang Mao, director of researcher Qin Tao
燑br>
Zinc rod high purity of over 99.995% was widely used in vacuum coating and other purposes.
Zinc Rod,Pure Zinc Rod,Zinc Anode Rod,Zinc Welding Rod
Shaoxing Tianlong Tin Materials Co.,Ltd. , https://www.tianlongspray.com