Israeli developers were able to teach AI to defeat a man in Mortal Kombat

The 3D environment is still difficult for the perception of a weak form of AI, which entails computer problems during the passage of such games.

With the help of video games, modern artificial intelligence specialists are going to teach AI methods of overcoming obstacles and solving emerging problems “on the go”. For example, DeepMind employees together with Blizzard turned StarCraft II into a medium for teaching a weak form of AI. Last year, Google’s artificial intelligence system independently mastered 49 old Atari games.

And this is not about a system integrated into the game (like AI opponents in fighting games, football simulations or racing simulators), which is well aware of the conditions and rules. AI, which is taught by developers in computer games now, put in equal conditions with man. The system monitors the image on the screen, learning from trial and error. And such a program is able to find a solution not only in games, it is suitable for finding a solution in the widest range of tasks, regardless of rules or conditions.


A group of students from the Israel University of Technology recently announced its development, the Retro Learning Environment (RLE) system. This is a software platform that allows you to teach AI on the example of many games of the 90s, including those that came out for the Nintendo and Sega consoles. These are, for example, many well-known F-Zero, Wolfenstein, and Mortal Kombat. According to the developers, for AI, many games turned out to be difficult, some of the system never learned to understand and pass. But RLE perfectly learned to play Mortal Kombat. The results of their work, the experts stated in an article on arXiv . The AI ​​has repeatedly managed to outright win against a human opponent. And this opponent was by no means a novice. The article states that the computer was opposed by an experienced player in Mortal Kombat.
')
In Wolfenstein, where the volume levels, plus you need to navigate during the passage of the maze and identify a number of objects, the system showed not very good results. In Gradius III, RLE was able to study the technical aspects of the game, which include the need to destroy encountered enemies with subsequent actions. But the system could not show a better result than the human player. Here it is necessary to improve the character's abilities by meeting artifacts. The more artifacts a player misses - the harder it is to go through the game. The computer almost did not pay attention to the power-up objects, which significantly complicated the process of passing.

The fact that the program was able to learn how to play a computer game is so good that it began to win from a person - the undoubted merit of the developer. For a computer, learning how to play a game through trial and error is not so easy, it is a difficult task that few software platforms overcome. “If algorithms can play complex games, then we can start working on the introduction of such systems in the real world to solve real problems,” said Shai Rosenberg, one of the authors of the study. “Just as a child learns to play games, a computer also sees only information on the screen. They (both the child and the computer) learn to avoid obstacles and solve problems in order to get the maximum reward, ”he continues.


The AI ​​learned quite well how to play both Boxing on Atari and Mortal Kombat simply by “looking at the screen” and assessing the consequences of their actions in the gaming environment.

In the real world, the ability of computer systems to learn from their mistakes and predict the consequences of certain “deeds” can be useful in many areas. Robots can move through complex spaces (corridors of rooms, for example) with a large number of obstacles, without colliding with them. Any small mistake made by the computer will be taken into account by them the next time, when performing the same or similar task.

According to Rosenberg, RLE can learn to go through more complex gaming systems, and not just play SNES games. The next stage of the project will be mastering the games of the PlayStation platform. True, so far Israeli developers have focused on teaching their system how to play most of the games they master. The fact that the computer has learned to play Mortal Kombat is good, but not enough - yet a significant part of the games was left behind, RLE could not master them.


The results of the passage of different games by the RLE system using different passing algorithms

“In the subsequent stages, we consider it possible and even relatively easy to adapt our training system to more complex games, including, for example, Grand Theft Auto,” the developers said. Now, unfortunately, games like Grand Theft Auto V AI are not available - they are too complicated.

The developers made the source code of their system open and posted it on Github. Get the source here .