Scientists have trained a neural network to proficiently play Minecraft like humans
Open AI researchers have formed a neural network to play Minecraft by having it watch videos for thousands of hours and using a small amount of tagged contractor data.
Also Read: The World Watched Minecraft Videos 1 Trillion Times on YouTube in 12 Years
They used a new method called Video PreTraining where they first started by collecting a small data set from contractors where they would record not only the video footage of the game but also the footage of their mouse movements and their keys..
Using this data, they trained an inverse dynamics model that predicts the action taken at each stage of the video. The IDM can use past and future information to guess the action at each step.
This was a more efficient alternative that required considerably less data than the behavioral cloning task of predicting actions based on past video frames only, which requires guessing and trying out what the person did. to achieve this. This can be used to train a larger dataset of online videos and learn to act using behavioral cloning.
The neural network was trained on 70,000 hours of IDM-labeled online video and the end result was a neural network capable of performing tasks in Minecraft not possible in reinforcement learning from scratch.
The AI learned how to chop down trees to collect logs, turn them into plants, and then craft the plants to make a crafting table. A competent human can do this in about 50 seconds or 1,000 consecutive game actions.
The model is also capable of performing other complex human skills such as swimming, hunting animals for food, etc. He even executed the “pillar jump” perfectly.
They then refined the neural network using behavioral cloning by asking entrepreneurs to play for 10 minutes in a brand new Minecraft Worlds and build a house out of basic Minecraft materials. By refining this dataset, they not only saw a massive improvement in reliable performance in early game skills from basic mode, but it allowed them to go further, crafting both tools wooden and wooden and even stone tools.
Also read: 13-year-old girl impresses Satya Nadella: she uses Minecraft to teach history and engineering
The researchers concluded in their study: “VPT paves the way for agents to learn how to act by watching a large number of videos on the Internet. Compared to generative video modeling or contrastive methods that would only produce representational priors, VPT offers the exciting possibility of directly learning large-scale behavioral priors in more areas than just language.
They added: “Although we are only experimenting in Minecraft, the game is very open and the native human interface (mouse and keyboard) is very generic, so we think our results bode well for other areas. similar, for example the use of a computer.”
To learn more about the world of technology and science, keep reading Indiatimes.com