Over the past few years, we have seen various attempts to teach an artificial intelligence how to play a video game. Now, researchers at the Georgia Institute of Technology’s School of Interactive Computing published a paper outlining a method of teaching AI how to learn a game engine.
In previous projects, the AI has only been instructed on how to excel at a particular game. This study goes one step further, aiming to instill a deeper understanding of the mechanics at play, rather than just the fastest route to success.
The researchers decided to use Super Mario Bros. for the study, describing it in the paper as a “classic platformer.” The same game was at the core of previous researchers performed by the study’s authors, which developed an AI that was capable of creating new stages in the vein of those seen in the original.
For the purposes of this research, the AI was not given access to the game’s code, which would make it very easy for it to understand things like the height that the character could achieve with a jump or the effect various enemies had on the protagonist. Instead, it was fed video footage that it used to make assumptions how the game engine worked.
The footage was analyzed via a three-step process, starting with a scan that determines which objects were present in a given frame. A greedy matching algorithm was then run over a pair of adjacent frames to gauge changes made to those objects before each frame was parsed — if the second frame differed beyond a set amount from what the system expected to see in the next frame, an engine search would be performed.
“We anticipate this technique to aid in applications for automated game playing, explainable AI, gameplay transfer, and game design tasks such as automated game design,” the team states in the paper.
However, there are some limitations to the AI in its current form. For one, the technique that is currently in place doesn’t take player death or level transitions into account, but the team hopes to address this issue in a future study.