AI Can Now Play Minecraft Just Like You – That’s Why Matters
OpenAI experts have trained a neural network to play minecraft To the same high level as human players.
The neural network was trained on 70,000 hours of varied in-game footage, which was a small . was supplemented with database Number of videos in which contractors performed specific in-game tasks, including keyboard And rat Enter the input as well.
After fine-tuning, OpenAI found that the model was capable of performing all manner of complex skills, from swimming to hunting animals and consuming their meat. It also captured the “pillar jump”, a move whereby the player places a block of material under mid-jump to gain height.
Perhaps most impressive, the AI was able to craft diamond tools (requiring a long string of actions in sequence), which OpenAI described as an “unprecedented” achievement for a computer agent.
An AI breakthrough?
The significance of the Minecraft project is that it demonstrates the efficacy of a new technology deployed by OpenAI in the training of AI models – called video pretraining (VPT) – which the company says is designed for “general computer-using agents”. can accelerate the development of
Historically, the difficulty of using raw video as a source for training AI models has been that what Easy enough to understand, but not necessarily How, In fact, the AI model will absorb the desired results, but have no understanding of the input combinations required to reach them.
However, with VPT, OpenAI combines a large video dataset drawn from public web sources with a carefully curated pool of labeled footage with relevant keyboard and mouse movements to establish a basic model.
To fine-tune the base model, the team then plugs in smaller datasets designed to teach specific tasks. In this context, OpenAI used footage of players performing early-game tasks, such as chopping down trees and building crafting tables, which is said to have led to a “massive improvement” in reliability with which The model was able to perform these tasks.
Another technique involves “rewarding” an AI model for achieving each step in a sequence of tasks, an exercise known as reinforcement learning. This process allowed the neural network to collect all the material for a diamond pickaxe with a human-scale success rate.
“VPT paves the way for allowing agents to learn how to act by watching a large number of videos on the Internet. Compared to generative video modeling or paradoxical methods that would only achieve representational priors, VPT offers the exciting prospect of directly learning behavioral priors at a large scale in domains more than just language,” OpenAI explained in a statement. blog post (opens in new tab),
“While we only experiment in Minecraft, the game is very open and the basic human interface (mouse and keyboard) is very generic, so we believe our results hold good for other similar domains, such as computer use.”
To encourage further experimentation in space, OpenAI has partnered with Minor NeurIPS Competition, donating your contractor data and model code to competitors attempting to use AI to solve complex Minecraft tasks. Grand Prize: $100,000.