New AI Can Develop Video Games Just by Watching Gameplay

Sunday, 17 September 2017 - 1:35PM
Technology
Artificial Intelligence
Sunday, 17 September 2017 - 1:35PM
New AI Can Develop Video Games Just by Watching Gameplay
Nintendo

As humans, we have both learned and innate behaviors. While some few, basic behaviors come to us instinctively, many more are learned through years of trial-and-error and watching those older than us. As it turns out, artificial intelligence systems can learn from watching as well.

Some machine learning mechanisms operate based off of an initial, large dataset, which can be both tricky and time-consuming to create. But this new AI, developed by researchers at Georgia Institute of Technology with the help of the original NES game Super Mario Bros., is capable of building a video game just by watching gameplay. In a new study, "Game Learning Engine from Video," researchers explained how the AI managed to recreate the engine of Mario's most famous outing.

Thankfully, this new tech isn't being designed to replace game developers in any way. Instead, the team hopes that this AI will be able to aid developers in an effort to "speed up game development and experiment with different styles of play," according to a recent press release.  And this AI does work quickly - the team recorded that the system only had to observe two minutes of Super Mario Bros. gameplay before it was able to build its own version of the game. 

Since then, they've experimented with showing the AI Mega Man and Sonic the Hedgehog as well, which it's also had some success in replicating. 
According to Joshua Preston, Georgia Institute's communications officer:

Opening quote
"To get their AI agent to create an accurate predictive model that could account for all the physics of a 2D platform-style game, the team trained the AI on a single 'speedrunner' video, where a player heads straight for the goal."
Closing quote


This method of studying frames directly and then predicting future movements and changes in the game, which Preston has said is potentially the most difficult for the program to "solve,"  has so far proven successful, more so than other systems.

Matthew Guzdial, the lead researcher on the team, assured in the same press release that the AI never had access to the original game's code, building the entire thing from scratch:

Opening quote
"Our AI creates the predictive model without ever accessing the game's code, and makes significantly more accurate future event predictions than those of convolutional neural networks... A single video won't produce a perfect clone of the game engine, but by training the AI on just a few additional videos you get something that's pretty close."
Closing quote


This AI could potentially revolutionize game development as we know it today. If you have an artificially intelligent assistant that can predict and model your own game, not only can you work faster, it could be possible to explore otherwise missed opportunities within development.

A system that not only intimately understands your development but has the capacity to predict what might come next could be an extremely useful tool for developers and completely change what games are capable of.


Science
Science News
Technology
Artificial Intelligence
No