For decades, video games have been criticized for purportedly wasting time, stifling creativity, and even influencing violent behaviors. Now, it seems that video games have become an unlikely tool for AI researchers to improve their systems.
Seeing stop signs
Take, for example, Artur Filipowicz, an AI researcher at Princeton University who’s been trying to develop software for autonomous vehicles. This software needs to be able to properly identify a stop sign — which can vary in appearance due to surroundings, conditions, and individual differences — no matter what. Failure here could cost a human life, so it’s important for the algorithm to view lifelike, variant images of stop signs to “understand” what a stop sign is like.
The solution to this problem? Grand Theft Auto V. Seriously.
In a game criticized for its violent and adult themes, stop signs are depicted somewhat realistically, and Filipowicz has been able to make modifications to the game so that his autonomous vehicle software can navigate the graphically rendered streets and respond to stop signs as if it were in a live environment.
A couple of years ago, DeepMind — the algorithm that recently bested a human Go champion in a feat once believed impossible — started training itself by playing a suite of Atari games. The catch is, developers didn’t tell the algorithm how to play the games. Instead, DeepMind had to learn on its own — and it did, by playing the games over and over, until it perfected the art. Now, DeepMind can beat just about any top score on any Atari video game.
Of course, high scores weren’t the end goal. The goal was to measure and improve DeepMind’s ability to learn using only external inputs, with no central programming to tell it “how” to play.
Privately funded organization OpenAI has taken the world of video game-based AI development to new levels, with a piece of software it calls Universe. With permission from the individual publishers, Universe has collected and modified a library of thousands of games, from basic games like those found on the Atari to major recent titles like Portal 2.
Each game in the library has been tweaked to allow an appropriate AI algorithm to explore it immediately. Best of all, the software is completely free to use — making it available to any AI researcher who wants to make use of this new trend.
So why is it that video games are so good at helping AI researchers solve problems?
Ultimately, the benefits can be reduced to four main areas:
- Training for the real world. Like with Grand Theft Auto V, some video games can be used as simulations for a real-world environment. Because these systems aren’t tangible, they’re cost-effective and can test new AI programs safely with limited repercussions.
- Problem-solving reduction. The big problem with developing AI is figuring out how to solve complex problems. Video games often take big, complex problems and reduce them to smaller, more manageable chunks. This allows researchers to not only optimize their algorithms to master problems in chunks, but also helps them understand how machine learning can stitch those chunks together to solve a bigger, more pressing problem.
- Repeatable learning environments. Video games are also predictable to some degree, even when they’re specifically engineered to randomly develop environments. They can be played and simulated at speeds faster than a human would use, and can therefore serve as practically infinitely repeatable learning environments. For example, DeepMind was able to pick up the “tunneling” trick to beating Breakout after 600 repeated games. For a human being, that could take weeks of training. For a machine, it takes hours.
- Transferring lessons. Machine learning and human learning are very different, and AI researchers are trying to bridge that gap. Currently, machines are very good at executing series of repeatable tasks, over and over, at an astonishing rate that human brains simply can’t match. However, human brains are very good at grasping high-level concepts, coordinating learning areas to work together and transferring lessons learned in one area to another area. Training a machine to take lessons learned in a game like Breakout and apply them to a game like Portal 2 is a massive jump, but it’s helping researchers to understand how machines could transfer lessons and understand high-level concepts in the future.
AI researchers relying on video games has a number of potential benefits and side effects. Video games are proving to be a safe, reliable, effective, and, most of all, cost-efficient way to test and develop new algorithms. On top of that, video game developers and publishers are recognizing the opportunities inherent in designing playground-like worlds for AI systems.
The future of video games in AI development is rich with potential, and we’re just starting to explore its full capabilities.
Larry Alton is a freelance writer covering artificial intelligence.