OpenAI trains an AI to play on Minecraft

Like a human, the AI ​​developed by the artificial intelligence company OpenAI managed to hunt, swim and build a house in Minecraft, owned by Microsoft. According to a blog post of the company, this was made possible by Video PreTraining (VPT). This method involves instilling skills into a neural network from a large amount of unlabeled videos and a few labeled videos. It then uses the countless video data available on the internet by semi-supervised imitation.

An AI that reaches the human level

The neural network model trained by OpenAI managed to fend for itself in the vast Minecraft universe, now accessible in augmented reality. To explain this feat, the company has followed new learning paths by Video PreTraining.

In the same category

The media still lacking in confidence in France and around the world

The VPT process begins with collecting gameplay footage with player actions, such as key presses and mouse movements. A model called “IDM” then processes the information collected to predict the action taken at each step. Finally, the AI ​​uses its knowledge to analyze and learn potential lessons from other unlabeled videos.

Process of the VPT method. Picture: OpenAI.

On Minecraft, the results were satisfactory in such a way as to obtain performance similar to that of a human, or even more. ” We’re the first to report computer agents capable of crafting diamond tools, which can take skilled humans over 20 minutes (24,000 environmental actions) of gameplay to accomplish. explains the OpenAI document.

OpenAI wants to profit from the millions of videos on the Internet

OpenAI chose Minecraft to test the VPT method, given that it is a high-profile title. It is one of the most played games in the world and, therefore, its videos shared on the internet are excellent sources of learning for algorithms.

The experiment has opened up new possibilities for learning by simply watching several videos on the web. ” Although we’re only experimenting in Minecraft, so we think our results bode well for other similar areas, such as computer use. concludes OpenAI.