Battling AIs build new Doom levels

Even other machines think that the levels are man-made

Video games are now at an unparalleled level of realism, with extraordinary detail in their graphics, physics and storylines. We're a long way from the ZX Spectrum.

But those immersive, realistic games take a long time to create - multiple years for some of the biggest titles. With that in mind, researchers at Italy's Politecnico di Milano have set out to develop a machine learning/AI technique that could dramatically lower development time.

The basic idea is simple: offload the time-consuming process of level design to an AI programme, so that the humans have more time to spend on other areas of the game, like power balancing and high-level features such as specific map types.

The group didn't jump right in and start rewriting an MMO, though. Their starting point was the 1993 classic, Doom.

Widely considered a gaming milestone, Doom pioneered 3D graphics for MS-DOS PCs, introduced local play over a network and - this is the important part - allowed players to build their own game levels.

The number of purpose-built levels available for free online was critical to success. Edoardo Giacomello and his colleagues at the Politecnico di Milano used 1,000 such levels to train their AI, showing it the man-made environments before setting it free to build its own.

The team processed each level and generated a set of images that represented its important areas: walkable floors, walls, floor height and so on. They also built a vector that could capture other statistics, like the size and area of rooms, in numerical form.

Two neural networks - the generator and discriminator - were set against each other, with the first building levels and attempting to trick the second into believing that they were man-made. This deep learning technique is called a generative adversarial network (Gan).

It took 36,000 iterations to produce believable environments. The researchers who have tested them called them ‘interesting' to explore and play, with typical Doom features like narrow tunnels and large rooms.

Perfection is still a long way off - fine detail is especially elusive - but the implications are significant. Creating content like this is one of the most tedious parts of game design, but also one of the most important, with the potential to greatly impact the players' experience. Automating part of that process could greatly lower the cost of development, opening the field for smaller studios.