![\"\"](\"https:\/\/images.newscientist.com\/wp-content\/uploads\/2024\/02\/07171355\/SEI_190206783.jpg?width=1200\")
<\/div>Minecraft is a game for humans, but it could help AI too<\/p>\n
Minecraft<\/p>\n<\/div>\n<\/figcaption><\/figure>\n<\/p>\n
Minecraft<\/em> is not only the best-selling video game in history, it could also be key to creating adaptable artificial intelligence models that can pick up a variety of tasks the way humans do.<\/p>\n Steven James<\/a> at the University of the Witwatersrand in South Africa and his colleagues developed a benchmark test within Minecraft<\/em> to measure the general intelligence of AI models. MinePlanner assesses an AI\u2019s ability to ignore unimportant details while solving a complex problem with multiple steps.<\/p>\n Lots of AI training \u201ccheats\u201d by giving a model all the data it needs to learn how to do a job and nothing extraneous, says James. That is a fruitful approach if you want create software to accomplish a specific task \u2013 such as predicting the weather<\/a> or folding proteins<\/a> \u2013 but not if you are attempting to create artificial general intelligence, or AGI.<\/p>\n James says that future AI models will need to tackle messy problems, and he hopes that MinePlanner will guide that research. AI working to solve a problem in the game will see the landscape, extraneous objects and other detail that isn\u2019t necessarily needed to solve a problem and must be ignored. It will have to survey its surroundings and work out by itself what is and is not needed.<\/p>\n <\/p>\n MinePlanner consists of 15 construction problems, each with an easy, medium and hard setting, for a total of 45 tasks. To complete each task, the AI may need to take intermediate steps \u2013 building a set of stairs in order to place blocks at a certain height, for instance. That demands that the AI can zoom out of the problem and plan ahead in order to achieve a goal.<\/p>\n 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. Fast Downward was only able to solve one of the medium problems, and five of the easy problems, while ENHSP performed slightly better by completing all but one of the easy problems and all but two of the medium problems.<\/p>\n \u201cWe can\u2019t require a human designer to come in and tell the AI exactly what it should and shouldn\u2019t care about for each and every task it might have to solve,\u201d says James. \u201cThat\u2019s the problem we\u2019re trying to address.\u201d<\/p>\n Topics:<\/p>\n<\/section><\/div>\n