Outdated games, “re-employment” in the field of artificial intelligence

I don’t know how many people still have impressions of the WeChat mini game “Airplane Wars” ——

WeChat version 5.0 was launched in 2013

Briefly repeat the gameplay: the player operates a small plane to dodge around, the small plane can automatically fire bullets, and two “biubiu” can kill a small plane on the opposite side. The screen is very simple, and there is no strategy to speak of. As long as you are focused enough and have hand-eye coordination, it will not be particularly difficult to score high.

With the update of the WeChat version, this little game has disappeared without a trace in WeChat. However, there are some Python tutorials that take this as a more basic exercise. After all, being able to write a “running, 0 error, 0 warning, and playable” program from start to finish is much more exciting to a novice than a simple “hello world”.

But “Airplane Wars” can teach users more than that. In the field of artificial intelligence deep learning, a series of small games represented by “Airplane Wars” are still shining brightly. There are a lot of related videos on station B. For example, the user whose ID is “vs little monster Monster” specially uploads various content of “mini game VS artificial intelligence”, of which the most played is this “ Punishment Artificial Intelligence ” , Forcing AI to Grow ” (BV12h411X7yY), interested readers can go and see. In this video, artificial intelligence realized the final solution of “Airplane Wars”: shoot down a small plane occasionally, and curl up in a corner most of the time – as long as I don’t take the initiative to attack, I will never lose when I am the king of good life .

Of course, thanks to artificial intelligence and algorithm recommendations, I quickly found more artificial intelligence on the scene. The funniest of them all is this one: “ About My Reinforcement Learning Model Not Converging ” (BV1e5411R7rF). The video tells the whole process of how a young lady who studies natural language processing (NLP) enters the pit of reinforcement learning. The game that Miss Li Rumor uses is “Tennis” on the Atari platform. She originally thought that she could train the strongest tennis prince on the surface——

But in fact, after 50,000 simulations, what was obtained was a rotten AI that would never serve and never start the game——

Well, “…as long as I never end, I’m invincible!”

There is also a more well-known rotten scene: “artificial retarded wolf”. The rule is the basic rule of “wolves eat sheep”. The wolf AI has 20 seconds of action time. The more sheep it eats, the better. The longer it takes, the more points will be deducted. At the same time, some stones will be generated on the ground to try to simulate the real environment. But after 200,000 simulations, the magical result appeared: the wolf chose to start badly, and did everything possible to sprint towards the stone at full speed, killing it . In most cases, wolves can’t eat sheep at all, and points will be deducted for wasting time in the process of catching sheep.

The logic is unbreakable and simple, and the absolutely rational AI has given its own answer.

“Wait, it’s weird,” my DNA as a gamer moved: “Airplane wars, wolves eat sheep, Atari mini-game resource library – these artificial intelligences that are at the forefront of the industry, how come they’ve been playing too much? A little game? “

Detroit: Become Human: Do androids dream of electric sheep?

Artificial intelligence (hereinafter referred to as AI), algorithm distribution, has formed part of today’s daily life and is still booming. But in fact this is not a new thing. Since the myths of ancient times, people have hoped to use their own hands and minds to build an artificial product similar to their own intelligence, which can also be understood as an “artificial human” fantasy. But the relevant technology and theoretical basis did not appear until the 1990s. Around 1950, some computer scientists in the United States wrote chess and checkers programs that could challenge amateur players of a certain level. Since then, game AI has been considered a yardstick for evaluating AI progress .

After going through repeated highs and lows, AI has finally achieved rapid development and application in the past 10 years. Among them, the most vigorously developed branch is the big category of “deep learning”. Deep learning is already a relatively cutting-edge application field in artificial intelligence applications. According to the definition of some novice introductory tutorials, deep learning refers to: “a learning route of machine learning, building and simulating a neural network for analysis of the human brain.”

Neural networks and reinforcement learning are deep learning solutions. They mainly simulate the exploration and strategy of a single agent (agent). The advantage is that they can simulate more times at the same time, and it is not difficult to run tens of thousands of simulations at the same time. In other words, in the absence of model input and a blank situation, how to stimulate the autonomous learning of artificial intelligence by controlling variables, and finally get a solution to the problem.

Building an effective neural network requires not only massive amounts of information for input and output, but also continuous training, adjusting the rules and its weights. Tell it which behaviors are allowed and which behaviors will be punished. After continuous adjustment and iteration, finally, it can get a weak artificial intelligence that barely participates in the game and obeys the rules, that is, “focuses on solving problems in specific areas” of artificial intelligence.

This is just a rough summary and description, let’s go back to the topic of games – why are these outdated games able to be laid off and re-employed in the field of artificial intelligence?

Publicly released Gym documentation debugging toolbox, built-in a large number of Atari mini-games

On the one hand, these little games are really small and take up very little storage space. We just mentioned that the neural network will perform concurrent simulations. The smaller the game, the smaller the storage and bandwidth occupied, the lighter the program runs, and the less money it will spend (renting a server or something). On the other hand, these small games have relatively simple rules, which are convenient for programmers to adjust parameters according to the data results. Finally, these mini-games are not used for profit in commercial projects, but only for research purposes, which is also reasonable from the perspective of copyright. Coupled with the field of artificial intelligence, the practice of using game projects to verify the progress of AI, in the end, those games you got tired of playing as a child are now the best teaching materials for artificial intelligence.

At the end of the “Prince of Tennis” video, Miss Sister’s rotten model finally began to score points. She said that the fun of learning algorithms is here, and she constantly refreshes her cognitive limits. There is a user’s comment in the comment area, which I think reflects the mentality of some algorithm engineers: “ Reinforcement learning to adjust parameters is like teaching your stupid son to read. You can’t learn even if you are tired. “

I also learned some algorithm basics when I was a child, but obviously, the outdated knowledge prepared for the competition has not kept up with the latest version environment. The development in the field of artificial intelligence is amazing. The translation tool DeepL, which is close to the level of human translation, the Disco Diffusion, which can automatically generate pictures by inputting pictures and instructions, and the neural network ceiling GPT-3… The content they generate can be said to be Mix the real with the fake. In terms of representation alone, these procedurally produced contents are no different from some relatively mediocre human works.

Disco Diffusion auto-generated images based on keywords

In the field of artificial intelligence, “whether it can pass the Turing test” is the touchstone to test the development level of artificial intelligence. The rules of the test are as follows: first assume that there are two entities of object AB, one of which is a real person and the other is an artificial intelligence. They are all hidden in places where they cannot be seen by others. At this time, an uninformed third person took the same question and went to seek the opinions of the object AB. After repeated inquiries, he could not distinguish the difference between AB according to the content of the answer. Then we can think that the artificial intelligence participating in the test can pass the test.

To achieve this goal, it is naturally inseparable from the joint efforts of multiple disciplines. It is still uncertain whether I will see such a powerful artificial intelligence in my lifetime. From a philosophical point of view, when an artificial intelligence really passes the Turing test, it will inevitably involve “representation and entity in the end which is more important” “As long as the appearance looks the same, then what is the motivation, can it be ignored?” in endless arguments.

But fortunately, we have not yet entered this stage, and artificial intelligence is not so smart, and it is still “rotten” in various algorithm models. No matter how powerful artificial intelligence is, it also needs human beings as the first driving force to put forward demands and find difficulties. And after watching countless “mentally retarded” scenes, I realized what is the most precious thing for human beings – the ability to ask questions, the courage to keep trying, the gentle eyes watching you, and a warm heart.

*This article is reproduced with permission from the public account “taptap discovers a good game”, author @星永