A Minecraft bot developed by OpenAI was successfully taught using 70,000 hours of gaming videos.

Why this matters: At first glance, Minecraft might not seem like a serious tool for cutting-edge AI study. After all, why would it be significant to train a computer to play a sandbox game that came out over a decade ago? Recent work by OpenAI suggests that a well-trained Minecraft bot has more to do with AI progress than is commonly believed.

To this day, OpenAI remains committed to advancing AI and ML for the greater good of society. Recently, the company used more than 70,000 hours of gameplay videos to teach a bot how to play Minecraft. This accomplishment goes well beyond the simple act of a computer program completing a task. It’s a huge step forward for machine learning techniques based on emulation and observation.

OpenAI
OpenAI

The OpenAI chatbot is a shining example of supervised learning (also known as imitation learning). Imitation learning is used to teach neural networks how to execute certain tasks by seeing how humans perform them, as opposed to reinforcement learning, which rewards a learning agent only once it has successfully reached a goal through trial and error.

AI and humans collaborate

OpenAI used publicly available gaming videos and instructions to teach their bot sophisticated in-game sequences that would normally take a human player around 24,000 separate actions to accomplish.

For imitation learning to work, videos must be labeled with information about the action and its result. However, due to the time and effort required, there may be insufficient data to use this method effectively. Since there aren’t enough datasets to go around, the agent can’t really learn anything by seeing its environment.

OpenAI
OpenAI

To increase the number of tagged movies without resorting to a laborious manual data tagging process, the OpenAI research team created a method called Video Pre-Training (VPT). An initial 2,000 hours of annotated Minecraft gaming were used to teach an agent to link player actions to in-game results. For the Minecraft bot to learn from, the resulting model was utilized to automatically produce labels for 70,000 hours of publicly available, unlabeled Minecraft content.

Read also:- 

About The Author

Leave a Comment

Your email address will not be published.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

Scroll to Top