Graphics processing units that were originally intended just to pump out polygons and pixels

happen to be the perfect systems to train neural networks.

But artificial intelligence might repay the favor, and make video games even better.

And I'm not just talking about AI's applications for rendering graphics, like by taking lower

quality images and through the magic of computerized guesswork turning them into high-resolution

images.

Or by helping render realistic reflections through a process called ray tracing.

I'm talking about AI that could design entirely new experiences for players.

One area AI could be great for is level design.

Every game genre has different needs.

Maps in multiplayer games need to be balanced so one team doesn't have an obvious advantage.

Single player games need to be interesting to explore but easy to navigate so you don't

get stuck going in circles for hours when you can't find a stupid door.

Horror games have to build tension and atmosphere so when monsters pop out of vents you poop

your pants.

As you might imagine then, crafting an exquisite game world takes a lot of time and tinkering.

Could AI speed up the process?

Some scientists in Italy thought so, and they developed two AI programs to craft new levels

for the fast-paced first-person shooter Doom.

The original one from 1993, not the 2016 remake… or the movie with the Rock.

Working with the old game has a lot of upsides for their research.

Because it's old it's much simpler than games today, and because it's a classic

it has tons of levels made by the original creators and fans alike.

To train their AI they fed it thousands of human-made levels and asked it to create its

own.

Then the researchers used a second AI that was trained to spot a computer-generated map

from a human made one.

They repeated the process until eventually the AI-made maps were indistinguishable from

those made by hand.

Doom is not the most realistic game ever though, what with its sharp edges and the player's

weapon seeming to come out of a third arm in the middle of their chest.

But AI is not limited to boxes and odd anatomy.

Some researchers have used AI to produce a lifelike game world.

Using thousands of hours of footage taken from cars driving around cities in the meatspace,

an AI was able to generate its own city blocks a player could drive around in.

AI could also improve the non-player characters in video games.

These poor bots, which are also casually referred to as AI, are famously pretty dumb when compared

to human players.

In the single-player campaigns of combat games, most people will have no trouble mowing down

hordes of them, giving the player a false sense of confidence.

Confidence that is then crushed when they try the online multiplayer and 13- year-olds

halfway around the world destroy them and tell them things they never knew about their

mom.

AI that behaves more human-like could offer a better challenge, enhancing immersion and

eliminating the need to have so many bot opponents.

Or to play online multiplayer for that matter.

Video games designers just have to make sure AIs aren't too good though, or that can

ruin the game too.

Where does AI go from here?

Who knows, AI's uses are limited only by your imagination.

Perhaps an AI advanced enough could process whole libraries of games and generate their

own.

Maybe the next Doom that we're still playing 25 years after launch will be thought up,

designed, and coded entirely by a machine.

This topic is near and dear to me as a science nerd and a gamer.

Am I the only one out there or are there a lot of other gamers lurking in the comments

of our science channel?

Let us know!

And if you want to an idea of how we're shaping AI, check out our video on neural

networks here and don't forget to subscribe.

Thanks for watching Seeker!