AI is Making Video Game Levels (Opinion Piece)
By: Nik S
Researchers at the Politecnico di Milano, the largest technical university in Italy, led by Edoardo Giacomello are creating an artificial intelligence (AI) that is creating levels for games like the classic Doom. There's a similar program being developed over at the University of California that is similar and makes Super Mario levels. The AI is, in a sense, trained to make new levels by showing it ones that have been designed by humans, both official and community-made. Doom alone has over 10,000 levels to choose from.
The researchers use a deep-learning technique known as a generative adversarial network (Gan). Basically, it's where two neural networks (a generator and a discriminator) are pitted against one-another; the first network tries to trick the second. The program is shown a set of images from each level so it understands where a character can move, walls, floor, room division and objects. It's also given information about the level like the perimeter and how many rooms it has. It takes around 36,000 iterations for the networks to generate something that is actually playable.
This is something that all gamers and developers should be ecstatic about. The process of making a video game is a long and tedious process, since to make a triple-A title can take a few years to make (2-5 years minimum). These larger games require a big team in multiple studios, and a lot of trial-and-error in finding ways to make their design fun. Modeling and decorating a map you can interact with is one of the hardest parts. Every single square inch of the map has to be presentable to the player.
Developers rely heavily on human expertise and some extensive testing to create the levels we know and love, but this can become very expensive and time consuming. The Gans is still only in the early stages and when it's ready for full release it's still going to need some human interaction to work on what the program can not. The AI can only focus on the bare-bones low-level parts of level design, so the monotonous part can be easily covered. If anything, it's going to help give time to the human designers to focus on the big-picture, like what type of maps or features they want to include.
Gans is still in the early prototype phase and isn't going to be seeing a full release for developers to use any time soon. Once it does come out there will probably be a few bugs and mistakes that human ingenuity will help fix. However, when the program starts to become perfected it's definitely going to have a rippling effect throughout the development community.
Source
The researchers use a deep-learning technique known as a generative adversarial network (Gan). Basically, it's where two neural networks (a generator and a discriminator) are pitted against one-another; the first network tries to trick the second. The program is shown a set of images from each level so it understands where a character can move, walls, floor, room division and objects. It's also given information about the level like the perimeter and how many rooms it has. It takes around 36,000 iterations for the networks to generate something that is actually playable.
This is something that all gamers and developers should be ecstatic about. The process of making a video game is a long and tedious process, since to make a triple-A title can take a few years to make (2-5 years minimum). These larger games require a big team in multiple studios, and a lot of trial-and-error in finding ways to make their design fun. Modeling and decorating a map you can interact with is one of the hardest parts. Every single square inch of the map has to be presentable to the player.
Developers rely heavily on human expertise and some extensive testing to create the levels we know and love, but this can become very expensive and time consuming. The Gans is still only in the early stages and when it's ready for full release it's still going to need some human interaction to work on what the program can not. The AI can only focus on the bare-bones low-level parts of level design, so the monotonous part can be easily covered. If anything, it's going to help give time to the human designers to focus on the big-picture, like what type of maps or features they want to include.
Gans is still in the early prototype phase and isn't going to be seeing a full release for developers to use any time soon. Once it does come out there will probably be a few bugs and mistakes that human ingenuity will help fix. However, when the program starts to become perfected it's definitely going to have a rippling effect throughout the development community.
Source
Comments
Post a Comment