The AI ​​breaks the game with a hole that no one has ever found

"Without AI control, the consequences can be unpredictable" - many will say.

and video games like a perfect couple. Thanks to the game, researchers get a virtual environment with fully defined and fully controllable targets, and AIs can freely run without causing any harm. However, sometimes AI also causes damage to some things.

In a study published this week by three machine researchers from the University of Freiburg in Germany, an AI "broke" a video game. Three researchers are learning a method to teach AI to move in the game. They tested a game on Atari from the 80s named Q * bert. While studying, they discovered an unusual incident: the software they tested detected an in-game error, allowing the AI ​​to achieve infinite scores.

Descriptive report, while AI is learning how to play Q * bert game , it discovered "an interesting solution". Usually, in this game, the player must jump from one box to the other, and for each jump, the color of the box will change. If you can change all the blocks, the player will be rewarded with points and will be able to go to the next table. However, this AI has found a much better way.

Game interface Q * bert.

Initially, the AI ​​completed the first round, and it started jumping from one box to another in a seemingly random way. For some reason, the game didn't go to the next game, and the boxes started blinking non-stop. AI has since been able to earn infinite points.

This is a hole that no one has discovered. If the AI ​​is thinking, maybe it will think, why should we suffer from jumping from box to box, when there is a good way to score like this.

AI "break" game.

An important point to note is, AI does not approach this issue in a way that people often do. It has no intention of finding a hole. The study of these three scientists is actually a wide-ranging test of AI studies known as "evolutionary algorithms." In the test, they will give the competition algorithms to see which algorithms complete the best task, and then add a few small improvements to the winners to see if they can work well. more not That way, the algorithms will become better and better.

This is not the most powerful or most commonly used AI form, but this form is gradually becoming popular again. The ability to "break" the game Q * bert may be a good omen, suggesting that evolutionary algorithms will become very useful in the future.