In the past, artificial intelligence programs looked at how we play video games, and then tried to repeat after the person and also the games were played. So AI learned to play go and did a good job with different video games for the Atari prefix. Therefore, researchers from the American Institute of Technology in Georgia decided to go further and teach artificial intelligence to recreate videogames themselves.
The team published the results of the work in the article "Game Engine Learning from Video", which described in detail the process of recreating the game of artificial intelligence, which was shown to him earlier. It's interesting that AI could only watch the game process, but did not have access to the game code. While the program is not very good at the task: the game is buggy, but nevertheless, they can still be passed.
To teach the program to recreate the previously seen games, the developers provided AI with a visual dictionary that contains all the game sprites, and also provided the program with a set of basic concepts about the position of all objects and their speed. So AI was able to navigate through the video, analyzing what he "sees". Watching the game from the outside, artificial intelligence shatters frame by frame, marks the seen by markers and, comparing the pictures with its documentation, recreates the game independently.
While AI is able to work with 2D-platformers. On his account several successfully recreated games, among which there is Super Mario Bros. and Megaman, but in the future, developers are planning to expand the program's capabilities by teaching it to do 3D games on video.
The article is based on materials .
Comments
Post a Comment