People generally recognize the restrictive relationship between "rock, paper and scissors". The characteristics of "fairness+randomness" make it not only a small game with active atmosphere, but also a relatively fair means to solve problems, which is widely used in resolving differences, deciding order, or determining ownership.
Needless to say, the "willing to gamble and admit defeat" that comes with guessing boxing can effectively maintain family harmony, and it can be called a family relationship mediator on call.
In most people's cognition, guessing boxing is a random event, and the probability of players winning should be the same, constant at one third, but this is not necessarily the case.
Recently, the research team of Professor Sailing He of Zhejiang University has developed an artificial intelligence model based on Markov chain, which is specially used to play guessing games. After 300 rounds of battles with 52 human players, AI defeated 95% of the players.
Figure | Changes in the number of net wins of AI model
For human players, the rule is to win +2 points, draw+1 point, and not lose points. Before the battle with AI, participants knew that winning would be rewarded with money, and the higher the total score, the more money they would win. Therefore, the probability of players deliberately releasing water or randomly choosing is extremely low.
Even so, AI defeated human beings. In the most unbalanced contest, Ai 198 won and 55 drew, only losing 47 times, and the winning rate was four times higher than that of human opponents. 15600 all the detailed original data of the game are given in the supplementary materials of the paper (see references for details).
If guessing boxing is really a random probability, statistically speaking, the probability that AI will gain such a big advantage after 15600 innings is very low.
In essence, guessing boxing is a game problem, and there is a classic Nash equilibrium behind it. Every individual's habit, cognition, strategy and strategy change will affect the actual winning rate. For example, if you are familiar with your opponent and may know that he/she often gives cloth, you can use scissors to restrain yourself.
The AI model put forward by Professor Sailing He of Zhejiang University also adopts a similar method, which proves that guessing boxing does have long-term winning strategies for different individuals, which can effectively improve the winning rate.
This AI model is based on the design of N-order Markov chain, which has memory and can trace back to N historical states at most and use them.
In order to cope with the different personalities and strategies of human players in actual combat, the research team also invented a multi-AI model.
"It is difficult to establish a single model that is effective for everyone, so we decided to combine single models so that they can distinguish and adapt to more different competitive strategies." The researchers explained in the paper.
In the first set of multi-AI models for human beings, they put 1-5 Markov chains, that is, five independent AI models, referring to the previous 1-5 actions respectively. Multi-AI refers to the decisions of five AI models as a whole. As for which one to choose, it depends on their performance in the last five times.
The "last five times" here is defined as a super parameter called focal length, which can be adjusted according to the situation to achieve further optimization. In the second set of multi-AI models for humans, this parameter is set to 10.
Figure | Decision Logic of Multiple Artificial Intelligence Models
For example, each N-order Markov chain model is like a strategist, and each model has different decision criteria. The multi-AI model is a think tank composed of commanders and many military advisers. When making a decision, each strategist will submit his own boxing advice, and the commander will adopt the advice of the person with the highest comprehensive score according to their performance in the past few times (focus length) to improve the long-term winning rate.
If human players win in succession, it will prompt Multi-AI to choose better solutions from other AI models. If the human player fails continuously, it is likely to change the strategy or break the previous punching rules, and then multiple AI can be adjusted accordingly.
The final social experiment results reflect the effectiveness of this idea. Of the 52 volunteers, less than 5 defeated Ai. Many people took the lead in the first 20-50 rounds, and then they were caught by AI and lost.
For those who play AI, the winning percentage is only slightly higher, and the gap is not big.
It is worth mentioning that when developing the algorithm behind the AI model, the research team read the research results of another Zhejiang University team six years ago, but used different game strategies.
Compared with the previous research on the data of all players in a statistical way, the multi-AI model here puts more emphasis on adjusting and controlling the personality differences and punching strategies of different players in time and choosing the most suitable game strategy at the moment.
20 14 May, many media reported a scientific research achievement about the game of "Rock, Paper, Scissors".
Actually, it is not. This research was also rated as one of the best results (preprint) in 20 14 by MIT Science and Technology Review.
Figure | MIT Science and Technology Review 20 14 Report
This paper reveals that there are different behavior patterns behind guessing games. For example, the winner will often make the same gesture in the next round, and the loser will often change. People are more willing to throw stones and so on. But the deeper purpose is to explore whether Nash equilibrium is established in the real game, study the framework of the real game model, and analyze the macro-cycle phenomenon and micro-behavior basis in the game. The basic theories used in this study cover many fields such as game theory, psychology and neuroscience.
Similarly, the latest research on "Rock, Scissors, Cloth" in 2020 turned out to be not only a very powerful guess AI, but also a very powerful cycle balance model analyst. It is expected to expand to other game scenarios in the future, such as predicting the next move of competitors, planning more effective campaign strategies, or formulating more favorable pricing schemes.
"(We found that) there are indeed laws to follow in human competitive behavior, and these laws can be utilized by using appropriate simple models," the researchers concluded in the paper. "The study of competitive behavior patterns and how to use them is expected to enable us to better model, predict and adapt to different competitive patterns."