News Categories

An A.I. is able to replicate video games simply by looking at its pixels

By Nickey Ross - on 11 Sep 2017, 12:32pm

An A.I. is able to replicate video games simply by looking at its pixels

The original Mega Man game is on the left, while the AI recreation of the game is on the right. <br> Image source: Georgia Tech via The Verge.

Scientists from the Georgia Institute of Technology are working to get artificial intelligence (A.I.) systems to learn how video games work. The team's research paper, Game Engine Learning From Video, details how an A.I. system is able to replicate the game engine of titles like Super Mario Bros and Mega Man, simply by looking at the pixels. The re-created versions of the games contain minor glitches, but are acceptable overall.

The system receives two sets of information, which are a visual dictionary containing all in-game sprites as well as a set of vital concepts including the placement and speed of in-game objects, which the A.I. system will use to analyze what it sees. The system deconstructs the gameplay frame-by-frame, identifies what it can see, and searches for rules that describe the action.

The A.I. system accumulates minor rules, notes them as a string of logic statements and merges the two to emulate the game engine. These rules are exportable, and may also be converted into several programming languages that can replicate the game itself. The A.I. system is working on 2D platform games for now, as it is dependent on humans to dictate the outcome of a certain game. This process would require more time and more developed machine vision tools, if a 3D game was used instead.

Georgia Tech believes that this technological approach could also be applied to real life situations in the future. However, A.I. systems would need to be advance enough to be able to perceive the world like how humans do. 

Source: The Verge.