Stay on Top of Enterprise Technology Trends
Get updates impacting your industry from our GigaOm Research Community
Deep learning is seemingly everywhere these days, although most of the time we’re reading (or writing) about it recognizing objects in images, understanding our language or recognizing our voices. But new research from North Carolina State University suggests that one type of deep learning model might also help make our video games a lot smarter.
The researchers used a type of deep learning model (a stacked denoising autoencoder, to be precise) in order to predict players’ goals in video games that don’t have a set series of challenges, specified order of events or set-in-stone solution paths. You can imagine the artificial intelligence engine in a game ramping up the obstacles, or perhaps guidance, as it figures out what a player is actually trying to accomplish at any given time.
The researchers looked at five types of data generated during gameplay — action state, action argument, location, narrative state and previously achieved goals — and their system was able to classify with 62.3 percent accuracy what goals a player was trying to achieve. According to their paper, that’s a nearly 29 percent improvement over previous approaches. The researchers gathered their data by analyzing the gameplay of 137 eight-graders on Crystal Island, an educational gaming environment developed at NCSU.
And better yet, while previous approaches required researchers to hand code the features their models needed to analyze, the deep learning approach was able to learn representations of the pertinent features on its own.
While such a limited study is far from definitive proof that deep learning could revolutionize artificial intelligence in different types of games (many of today’s commercial titles are probably much more advanced in terms of the number of possibilities and the number of things going on) it is a good example of the possible applications of deep-learning-based pattern recognition beyond what it’s currently known for. In fact, DeepMind, the startup Google bought for $400 million in January, had already trained a deep learning system to learn the rules of video game just by analyzing gameplay data.
And I’d be willing to bet there are a handful of gaming companies already applying deep learning, at least in their labs, to see if they can do just what the NCSU researchers did, but on a bigger scale and with real games.