Artificial intelligence is being explored in many industries, helping humans to focus on skilled work by taking care of time-consuming manual labour.
One of those industries is game design, where a huge amount of time is spent on building maps and levels. In an effort to solve this challenge, a team at Italy's Politecnico di Milano recently developed a system using competing AIs, or Generative Adversarial Networks (GANs). These networks are capable of autonomously building new levels for the 1993 classic video game, Doom, trained using fan-created maps. But what happens when that training data is missing?
Daniele Loiacono, an assistant professor at the Politecnico di Milano, told V3: "The approach we proposed...is based on the idea of generating novel content starting from a set of existing ones, that are used as examples… GANs need data to be trained (and the more data is available, the better), [but] other machine learning approaches to content generation have been proposed...to generate content without using examples."
The team chose Doom partly because of their fondness for it ("It was one of my favourite games in the mid '90s," Loiacono reminisced), but mostly because of the thousands of additional fan-made levels that exist.
"The open research question we tried to answer with this work is whether deep learning could be used to generate more complex game content such as maps of first person shooters, which involve a more complex spatial structure.
"Doom, beside being a milestone of the genre, is still played a lot today. There is a huge community and it is possible to find thousands of levels, which is very important for any project involving deep learning."
The Doom GAN used that data to learn how the maps for the game were put together and functioned. One AI then went on to build maps, while another evaluated the results.
Loiacono's team has previously worked on a similar GAN system that built maps without using training data. "The content generation [then] becomes an optimisation problem, where machine learning algorithms can be used to search the design space for good maps," he said.
When it was first covered, much was made of the team's system as a way to assist smaller studios with limited budgets. Is this actually possible without any base training data for the game - say, one in development? Yes, said Loiacono, "if we can train them on existing content created for a similar game." Think PUBG and Fortnite, or League of Legends and DOTA 2.
Isn't this just another No Man's Sky?
Procedural content generation (PCG) is nothing new in video games, where its use dates back to the 1980s in games like Rogue and Elite, or in more modern times universe-spanning creations like No Man's Sky. It uses an algorithm to generate specific types of content for a specific type of game.
PCG can add massive variety, while benefitting from small file sizes. The disadvantage is that the algorithms it relies on have to be built for each game, which itself is a time-consuming process. The neural networks in a GAN, on the other hand, learn by observing existing content.
"The benefit is that we don't need to explicitly develop a smart PCG algorithm to encode content design principles, because such knowledge is implicitly encoded in the examples of content used to train the networks," said Loiacono. "In addition, we can expect that this approach can be generalised to different type of game content, and also in different genres, while a specific PCG algorithm needs to be developed for any type of content."
The GAN went through 36,000 training iterations to build new Doom levels, although ‘reasonable' ones were being generated after about 20,000. The training process of a single network took around 30 hours on the team's server - which used a relatively underpowered (for GPU computing) Geforce GTX 1060.
Of course, no system is perfect first time. During the training the networks built maps that would have tested the ability of the best gamers, including some that were completely disconnected ("Like an irregular lattice"), or maze-like with excessive amounts of walls.
Will neural networks revolutionise the games industry in 2018? Probably not. The work being done on them, though, has huge potential, and that means investment is almost certainly close behind. We could see mainstream results in just a few years.
Findings made by reconstructing its orbit by numerical simulation
3D printer was specially adapted to build therapeutic biomaterials from multiple materials
Politicians in attendance complained that Zuckerberg skipped all the tough questions
Are you paying attention?