Minecraft videos helped OpenAI train a computer

nuclear fusion Minecraft maybe have more in common than all the countless hours you can invest in them. As MIT Technology Review As reported over the weekend, the artificial intelligence non-profit OpenAI recently built the most advanced in the world Minecraft-Playing bot by analyzing over 70,000 hours of human gameplay with a new training method. While researchers are currently relegated to crafting pixelated tools and buildings, researchers claim the bot’s achievements could one day help usher in breakthrough technologies like true self-driving vehicles and virtually limitless renewable energy resources.

To design the first bot capable of constructing “diamond tools”, Minecraft‘s in-game items, which take humans on average about 20 minutes and 24,000 actions to craft, the researchers used a technique known as imitation learning. As the name suggests, imitation learning requires an AI to observe and refine thousands of samples of human input to produce the intended results. Reinforcement learning, another popular and effective AI design method, instead focuses on an unfocused trial and error approach to its training.

[Related: This agile robot dog uses a video camera in place of senses.]

A major previous problem with imitation learning is that researchers typically have to hand-label “every step,” he explains technology review, i.e. “by this action this happens, by this action that happens and so on.” OpenAI managed to circumvent this immensely time-consuming process by building an entirely separate neural network capable of performing the labeling procedure in what it calls Video Pre-Training (VPT). The researchers initially hired gig workers to play Minecraftthen recorded 2,000 hours of their keystrokes, mouse clicks, and video games to use as a reference for training a future AI bot.

The addition of VPT allowed the new AI program to construct items Minecraft previously unattainable for bots that rely solely on reinforcement learning, like the estimated 970-step process of building a table from handcrafted planks. When imitation and reinforcement learning were combined, the bot could handle construction projects with over 20,000 consecutive actions.

[Related: An AI that lets cars communicate might reduce traffic jams.]

Though many years away, earlier reinforcement learning achievements such as support for nuclear fusion research and advances in self-driving could potentially benefit from additional support from imitation learning gains first revealed via video games as shown Minecraft. Until then, there are many ethical questions about what data treasures are used in methods like imitation and reinforcement learning and how effectively they can be applied.

OpenAI was co-founded in 2015 by a team including Elon Musk and Sam Altman, and counted Peter Thiel as an early investor. Musk resigned from the board in 2018.

We have reached out to OpenAI to clarify where it has collected its 70,000 hours of Minecraft games and if the authors of the videos are aware of its use, and will update accordingly.

Leave a Reply

Your email address will not be published. Required fields are marked *