AI can also play Minecraft!The Microsoft OpenAI team recently published a result, they let AI learn how to play wheat following watching 70,000 hours of YouTube videos. past, AI can create delicate and complex sentences, can produce non-existent faces, or a seemingly reasonable but unreasonable photo, such asan astronaut on a horse。But these are still relatively simple input and output calculations, which can be learned by feeding a large amount of data in advance. Open-world games contain complex operating instructions, so AI learning has always been a place where AI cannot break through.
AI in the past might only learn new things with simple instructions
Previous AI machine language algorithms,A game where only simple instructions can be learned and every step can be estimated and measured, a puzzle game like chess. But now AI can use a keyboard and mouse like a human and learn to play games.And isBy watching instructional videos, learn non-linear and random storyline games like Minecraft this time.
♦ TO Recommended reading:【VIDEO INCLUDED】Gamers build a complete computer in Minecraft!Tetris and Snake can really play
New advances in machine language algorithms that learn and execute multiple instructions
Microsoft development team OpenAI team explained that this timeSuccessfully teach AI the basics of Minecraftsuch as chopping trees, making tables by hand, or operating game characters to swim, hunt, cook, etc.
themMore than just basic tasksyou can also complete advanced tasks, such as making wooden tools, building basic settlements, exploring villages, etc., and evenYou can also craft a tool that humans take 20 minutes to craft: a diamond pickaxe。
AI can watch videos and learn to play games, just like humans
Video may be a good learning tool for humans, but not necessarily practical for AI. Because the film only demonstrates how people do it, but does not explain it.The structure of the film is complex, and the learning process of the machine language algorithm in the middle will be very time-consuming. Therefore, feeding a large number of films for AI to learn was very impractical in the past.
AI cannot learn actions directly, but must be further linked to actions through labeled segments.So the development team made annotations on 70,000 hours of film. This is a huge project, they finally decided to find another team to perform the recording and labeling of basic Minecraft operation instructions, and the outsourced team used the labelled 2000 hours of video to teach the second set of machine language algorithm IDM to label the previous 7 10,000 hours of videos (the team claims that the accuracy of IDM is as high as 90%).
A child learning the basics of wheat blocks may only need one or two videos to learn, and today’s AI needs more things to learn very “simple” skills.
♦ TO Recommended reading:OpenAI Releases Free AI You Ask Me Answer Playground! How will AI solve the trolley problem?
This article is open to partners for reprinting, reference materials:Singularityhubfirst image source: Unsplash