Tech companies like Google DeepMind have demonstrated how cutting-edge artificial intelligence can learn how to play classic video games better than most human players. But could an A.I. design classic video game stages as well as human designers? That’s a question posed by researchers from Italy, who have developed an artificial neural network that’s capable of generating a theoretically infinite number of new levels of the classic 1993 first-person-shooter Doom. If you’ve spent the past 25 years wishing Doom would never end, today is your lucky day!
To create the new levels, two deep-learning neural networks were shown 1,000 existing Doom levels. This gave them the ability to learn the features found in popular levels, and use these as the basis for generating new ones. The researchers didn’t directly have any input on the levels generated. However, their selection of the levels used to teach the network allowed them to exert a small amount of control — like the parent who tries to shape their kids’ music tastes by only playing them classic albums produced during the golden age of, say, 1988 to 1997.
“There are no explicit bias encoded in the networks that generate the levels, but we expected — and found in our analysis — the networks generate levels that share similarities with levels used for training,” Daniele Loiacono, an assistant professor at Italy’s Politecnico of Milano, told Digital Trends. “Accordingly, choosing the set of levels to use for training makes it possible to affect the quality and the characteristics of the levels generated.”
So what does this mean for future game design then? Are tomorrow’s AAA developer jobs going to snapped up by bots instead of human creators? Not necessarily.
“We think that this work, as well as several recent works in the game research literature, suggest that it would be possible very soon to develop better design tools, where A.I. could assist human designers with the game content creation,” Loiacono said. “Such ‘intelligent’ design tools could save time for human designers and, at the same time, allow them to work at a higher level of abstraction. Our other work deals with using A.I. to generate game content, including tracks for racing games, 3D assets, weapons and maps for FPS, and levels for platformers. In particular, this approach could be applied to generate maps also for real-time strategy, multiplayer online battle arena, and RPG games.”
Along with Loiacono, other researchers on the project included former Politecnico di Milano student Edoardo Giacomello and Pier Luca Lanzi, a full professor at the university. You can read their paper here. The project’s repository is also available for examination on Github.
- Microsoft’s friendly Xiaoice A.I can figure out what you want — before you ask
- What is an artificial neural network? Here’s everything you need to know
- Midrange phones can’t do A.I., but MediaTek’s P90 chip aims to change that
- G’day, Google: U.S. users can now give Assistant a British or Australian accent
- A.I. can do almost anything now, but here are 6 things machines still suck at