It is common for scientists to use artificial intelligence to perform tasks that humans can complete, such as winning a game. What is not common is researchers using these technologies to accomplish another task that can be done by humans, such as creating a game.
Scientists at the Georgia Institute of Technology in the US want artificial intelligence to mimic game designers. That is, they put AIs to play, but not for the purpose of making them experts in the game, but rather for them to recreate it.
Artificial intelligence needs to watch only two minutes of someone throwing something – for the study, scientists used videos of speedrunners – and from the images, manages to create an accurate model with the physics and features of a platform game in two such as “Super Mario Bros.”, “Mega Man” or “Sonic the Hedgehog”, to stay in the games used during the experiment.
The result was a recreation of the phases of classic games with virtually all the features of the original game. “Our artificial intelligence creates a model without having to access the game’s code, and manages to predict future events more accurately than [artificial intelligence] with complex neural networks,” said lead researcher Matthew Guzdial, one of the researchers. “A single video will not produce a perfect clone of the original game, but by training artificial intelligence with a few more videos we get something pretty close,” he continued.
To test whether the cloned game was indeed similar to the original, the researchers placed a second artificial intelligence to traverse the stages re-created by the former. With this they observed that the differences between the original and cloned versions were minimal, and the game created by artificial intelligence might very well be the true one.
For the time being, the system is limited to 2D platform games, and it can never get out of it: developing an artificial intelligence that manages to recreate an entire world in three dimensions is considerably more complex and time-consuming and requires more advanced technology of what currently exists.
But researchers do not think this will be the end of technology. They want to use it in the real world, to help make machines understand some things the same way humans do.