The furor over artificial intelligence reached the world of Mario Bros. Thanks to a group of engineers from the University of Copenhagen A tool was created that allows you to generate levels for this video game through text.
The researchers relied on GPT-2 to train the system with information on the levels of Super Mario Bros. and Super Mario Bros.: The Lost Levels and in this way the technology is capable of replicating the scenarios with high fidelity.
It may interest you: Image created with artificial intelligence cheated in a photography contest and won the prize
How Mario Bros levels are created with AI
MarioGPT it is the product of research and not a commercial tool, so it still has bugs. However, the engineers assure that they managed to make 80% of the levels created playable.
For their operation they preferred to use GPT-2 which is an easier technology to train and they were able to control it better with a small data set.
“Honestly, we went with the smaller one to see if it would work. I think that in general, with small data sets, GPT2 is more suitable than GPT3, while also being much lighter and easier to train. However, in the future, with larger data sets and more complicated prompts, we may need to use a more sophisticated model like GPT3,” he said. Shyam Sudhakaran lead author of the study, in an interview with TechCrunch.
It may interest you: Linkin Park used artificial intelligence for the video of one of his best songs
With artificial intelligence they create 250 levels of the classic video game.
Additionally, for the learning of intelligence they chose to give a specific character to each element of the scenario. For example, enemies is ‘E’, this way it was easier to write the requests to generate the levels.
Although this caused one of its most frequent errors and that is the creation of enemies, which is not so varied, since the system groups everyone under the same character.
The way in which the scenarios were generated was thanks to the indications that the artificial intelligence was given through text, with requests such as that it make levels with “many pipes and many enemies” or “many blocks, high elevation, no enemies”. ”. However, the scenarios in the water were not represented because they were not included in the learning.
It may interest you: An artificial intelligence bot answers why it is necessary to incorporate it into education
In total they created 250 different levels inspired by these first two games in the saga of Super Mario Brosof which 9 out of 10 could be finished by the research team.
Given the results, they highlight that using a more complex system would help to migrate the data to a more advanced model and that would become an important step for the automation of video games, in a more controlled and varied way.
With artificial intelligence they create 250 levels of the classic video game.
It may interest you: Hands with 6 fingers, teeth and more image errors made with artificial intelligence
Not the first time
MarioGPT is one more example of how AI is being adapted in the video game industry. The new york university developed a similar project with the game Sokoban, in which they automatically generated levels through the description of texts.
The difference with the other project is that they did opt for GPT-3, because from the experience of the researchers it works better with a small data set, reducing the information base required at the start. Although they rescued that what was done by the University of Copenhagen had a higher level of complexity in the training of artificial intelligence.
