video is a training resource with great potential.
Open AI recently announced their latest achievements in the field of artificial intelligence - AI, which is best at playing " Minecraft ". How powerful is
? Whether it is building complex buildings such as bunkers and pyramids, fighting against the evil dragons in the game, or "starting from scratch" and creating rare diamond tools, these tasks often require dozens of minutes of process and tens of thousands of clicks.
After training, this AI can independently complete this series of events without any external force assistance, which is something that other AIs could not do before. According to Open AI, this achievement can be achieved depends on new breakthroughs in imitation learning technology.
In this technology, AI will be trained to complete the corresponding tasks by observing and imitating human behavior. Many effects have previously applied similar principles, such as controlling robot arm or driving a car.
Since AI needs to observe and imitate, it is bound to require a certain amount of material. As for "proficient in operating "Minecraft"", there are actually a lot of learning materials on major video websites, which can fully meet the learning needs of AI.
But the problem is that another additional work is needed, namely "let the AI accurately understand the content in the video."
"Minecraft" is a game without a clear goal. Although there are a lot of online materials, it cannot be used directly as learning materials, because every action and operation in the video actually requires humans to put "labels" on it, that is, make corresponding annotations in order for AI to understand the meaning of operations.
If each action requires manual annotation, the efficiency will naturally be reduced. To solve this problem, the researchers came up with a new method - training AI that can automatically label video operations.
They first summoned a group of volunteer players to operate the game and kept the players' key record. After obtaining 2,000 hours of game data, the researchers trained a new model and manually corresponded the key operation to the effects in the game one by one, such as under what circumstances will the mouse click to make the character wave an axe.
The next step is to use this model to process the online "Minecraft" videos in Shanghai and label them as a tag that AI can understand. According to the published data, in order to achieve ideal results, the video materials provided by scientific researchers lasted 70,000 hours.
and this has indeed achieved certain results. For example, as mentioned at the beginning, AI has learned how to make diamond tools by itself. Bowen Baker, the head of Open AI, said: "We think Minecraft is a good field of research."
Since Minecraft does not have a clear winning or losing and has extremely high freedom, researchers can train AI to perform more complex tasks, and they believe that these tasks can eventually feed back to reality - if AI can design a table in the game, then it might be fine in reality.
