Until now, we have mostly seen AI systems learning to play tough games like Go and DOTA 2 to defeat best players. There are systems which can see TV shows and automatically create subtitles. AI’s can also be capable of building games.
But what if there is an AI which doesn’t require anything to build a game, not even the source code. Researchers at the Georgia Institute of Technology are running an experiment where the AI analyzes Super Mario gameplay and recreate the game engine on its own.
AI doesn’t even need to play the game, just watch it and learn how it works. The output might not be the best, but still, it’s a great achievement which is first of its kind.
The researchers have published their findings in a paper titled, “Game Engine Learning From Video“. To analyze a given a gameplay frame-by-frame at the pixel level, the researchers supplemented the AI with two sets of information. A list of all the movable graphic objects on the screen, as well as, their position and velocity.
“Our technique relies on a relatively simple search algorithm that searches through possible sets of rules that can best predict a set of frame transitions,” the researchers conclude in the paper.
Over time, the system tries to learn how the game works. Meanwhile, it makes small rules based on the analysis which helps in the creation of the game engine. For instance, “imagine the case where Mario is just above a Goomba in one frame, and then the next frame the Goomba Is gone. From that it comes up with the rule that when Mario is just above the Goomba and his velocity is negative, the Goomba disappears” paper’s lead author Matthew Guzdial told The Verge.
The researchers chose Super Mario Bros. due to its 2-D nature and less complexity. Sourcing the supplement information in the case of a 3D game would be tedious and require advanced tools. Mega Man is another game considered by the researchers.
The system with limited capacity is nascent for now, but in the future, it might be able to handle 3D game engines. Moreover, the researchers want their AI to be efficient enough for a full game generation. And possibly, their AI could be improved to interpret real life situations. Probably, someday, to the extent humans understand it.
Guzdial has shared the video of the game level editor they created using the capabilities of their AI.