Minecraft could be the key to creating adaptable AI

Minecraft Not only is it the best-selling video game in history, but it could also be the key to creating applicable artificial intelligence models that can take on a variety of human-like tasks.

Steven James At the University of the Witwatersrand in South Africa, he and his colleagues developed a benchmark test in-house. Minecraft To measure the general intelligence of AI models. MinePlanner assesses AI’s ability to ignore unimportant details while solving a complex problem with multiple steps.

A lot of AI training “tricks” a model by giving it all the data it needs to learn how to operate and nothing else, James says. This is a productive approach if you want to build software to accomplish a specific task – eg Weather forecast or Folding proteins – but not if you’re trying to build artificial general intelligence, or AGI.

James says future AI models will need to tackle messy problems, and he hopes MinePlanner will lead that research. An AI working to solve a problem in a game will notice the landscape, external objects, and other details that are not necessary to solve a problem and should be ignored. He has to survey his surroundings and work out for himself what is needed and what is not.

MinePlanner consists of 15 construction problems, each with easy, medium and hard settings, for a total of 45 tasks. To complete each task, the AI ​​may need to take intermediate steps – for example, building a set of stairs to place blocks at a certain height. This requires that the AI ​​can step outside of the problem and plan ahead to achieve the goal.

In experiments with state-of-the-art planning AI models ENHSP and Fast Downward, open-source programs designed to handle sequential operations in pursuit of an overall goal, neither model was able to complete any of the hard problems. was Fast Downward was only able to solve one of the medium problems and five of the easy problems, while ENHSP performed slightly better, completing all but one of the easy problems and all but two of the medium problems. demonstrated.

“We can’t have a human designer come in and tell the AI ​​exactly what it shouldn’t care about and what it shouldn’t do,” says James. “That’s the problem we’re trying to solve.”