Magical AI cloning technology, two minutes to build Super Mario

AI technology application is not a novelty in game development. AI technology is so powerful that we can't predict it. In the game world, game developers may soon have a new tool to speed up the game development of different styles of games, that is AI cloning technology.

Researchers at Georgia Tech have developed a new approach to using artificial intelligence to learn a complete game engine. The game's clone software can be used to control everything from character movement to graphic visualization.

The study was presented at the International Joint Conference on Artificial Intelligence, held in Melbourne, Australia in August.

Associate professor and project co-investigator Mark Riedl said: "The technology relies on a relatively simple search algorithm that can be searched to best predict the set of rules for a set of frame transitions."

To the best of our knowledge, this is the first AI technology to learn the game engine and use the game screen to simulate the game world.

The researchers' AI system only needs to watch less than two minutes of game video, then build their own game mode of operation by studying the framework and predicting future events, such as what path the character will choose or how the enemy reacts.

In order for AI agents to create an accurate predictive model that can explain all the physics of a 2D platform-style game, the team trained the AI ​​in a single "speedrunner" video, and the player can directly attack the target.

They first cloned Super Mario, first using AI and Mario Bros. game video to create a unique game-level design. Trying to experiment with the clone Rockman and the Hedgehog Sonic series.

Principal Investigator and Ph.D. Matthew Guzdial said: "Our AI has created predictive models that do not require access to the game's code and allow for more accurate prediction of future events than convolutional neural networks.

Pixel error comparison between previous frame, engine prediction and CNN

"A video doesn't produce a perfect game engine clone, but by doing artificial intelligence training in a few additional videos, you get a very close result."

They then tested the performance of the clone engine in the actual game. They hired two AI agents to play the game and ensure that the protagonist of the game does not fall from the solid floor, or if it is hit by an enemy, it will not be injured.

The result: artificial intelligence using clone engines cannot be distinguished from AI using the original game engine.

Part of the game video (left) is made by the original Super Marie engine, and the clone engine (right) demonstrates the ability to accurately predict the state of the animation.

First, each input video frame is scanned to determine the set of objects that exist, as shown: Visualization of the frame parsing process in the Super Marie Engine.

First a framework is parsed, find the spritesheet element in a frame, and then convert it into a list of fact frames.

Second, we run a greedy matching algorithm between two adjacent frames to determine how the object changes between frames.

Finally, when the second frame is different from the predicted next frame and exceeds some of the set thresholds, we parse each frame and run the engine search algorithm.

The current cloning technology is very effective for games that have a lot of motion on the screen. Guzdial said that the actions of Clash of Clans or other similar games may occur off-screen, which can be difficult for their system.

He said: "If smart agents are to fulfill their commitment to promote different technology applications, smart agents need to make predictions about their environment." "Our model can be used to train or educate various tasks in the scene, we I think this research will probably extend to all types of games.