Researchers use a famous game to train AI

Minecraft is the world’s best-selling video game, but not only that. Steven James and his colleagues at the University of the Witwatersrand in South Africa are using it to test artificial intelligence’s ability to adapt, reports New Scientist, relayed by Corey.

Not all AIs are the same. There are people who can create a specific task like predicting the weather or playing chess. They received all the data to complete their mission. But there are others who must sort and evaluate interactions, information using a “machine learning” process.

45 tasks to complete

Researchers were interested in this latter model. Thus they developed a reference test that made it possible to measure the ability to complete tasks while ignoring the redundant. With the MinePlanner model, they asked two AI programs to run Minecraft without telling them exactly how to do it. However, the programs had to complete 15 constructions with three levels of difficulty (easy, intermediate and difficult) or 45 different tasks.

mixed results

To complete certain tasks, the AI ​​must use steps, such as building stairs to raise blocks. This assumes that an AI can plan in advance to achieve a goal and discard anything unnecessary to accomplish its mission.

Results: The AI ​​model successfully completed fourteen easy, thirteen intermediate and zero difficult tasks. Other AIs complete only five simple tasks and one intermediate one. The researcher clarifies that “she had exhausted her memory before being able to do most things.” “There are still technical gaps to overcome in solving some of the problems. We hope our work will be used to develop new approaches in more complex areas,” concludes Steven James.