FYI.

This story is over 5 years old.

Tech

This Deep Learning Algorithm Can Predict Your Next Move in a Video Game

A new deep learning algorithm for video games can predict player actions with greater accuracy than ever before.
Image: NCSU

Video games have come a long way since my soda-fueled nights spent playing Sonic 2 with the volume just loud enough to hear the tinny refrain of Chemical Plant Zone without waking my parents. In those simpler times, artificial intelligence wasn't exactly a main concern for game developers making linear platformers.

But since games shifted to having open worlds, developers have had to figure out how to make narratives move forward consistently even when players don't. Designing areas to trigger new missions is one solution, but researchers at North Carolina State University have designed a completely different solution: a deep learning algorithm that can anticipate and adapt to unpredictable player actions.

Advertisement

"For games, the current state-of-the-art AI program for goal recognition has an accuracy rate of 48.4 percent," said Wookhee Min, lead author of the paper, in a statement. "The accuracy rate for our new program is 62.3 percent. That's a big jump."

"Goal recognition" is an approach to player modeling that guesses at what a player's goals are so that the game can respond accordingly by sending in enemies or adjusting difficulty levels. In open world games, player goals are "hidden" from the system and have to be inferred through low-level actions instead of being prescribed by well-defined parameters like set mission objectives.

A screenshot of Crystal Island from the paper, because you were surely curious what the game is like. (We couldn't find any gameplay videos in an admittedly-cursory YouTube search.)

The researchers' approach, described in a paper they're set to present at this year's Conference on Artificial Intelligence and Interactive Digital Entertainment, involved feeding their algorithm a massive data set culled from the actions of 137 payers recorded while playing through Crystal Island, an open world educational game.

The total number of goals completed by players in the training data was 893, and the average number of actions taken to achieve a goal was 86.4. The algorithm then analyzed the data set to tease out the statistical probability of of various combinations of game actions leading to certain goal outcomes with impressive accuracy.

Here's an example of how it works: Say I'm wandering around and blowing shit up in Grand Theft Auto: Vice City, as I'm wont to do. The game doesn't necessarily know what my plan is—maybe I'm working my way up to completing a mission that will drive the story forward, or maybe I don't even have a plan because I'm a loose cannon, baby.

In this scenario, the researchers' algorithm would be able to infer with reasonable accuracy that I'm really interested in heading up to Tommy Vercetti's mansion to save my game, and adjust the environment accordingly, as long as it's been trained on test data that I've provided.

According to the researchers, however, the most immediate application of their algorithm is in educational games that can facilitate the learning process by responding to player actions.

"We developed this software for use in educational gaming, but it has applications for all video game developers," said Dr. James Lester, a professor of computer science at North Carolina State. "This is a key step in developing player-adaptive games that can respond to player actions to improve the gaming experience, either for entertainment or—in our case—for education."