Computers and artificial intelligence to learn to play video games, that is the future that awaits us, along with non-player characters that are increasingly real in our games.
Now we have with us the best Minecraft player bot ever, an OpenAI creation that has watched 70,000 hours of video of people to learn the deepest secrets of this popular game.
The goal isn’t just to create a good player, it’s to test a powerful new technique that could be used to train machines to perform a wide range of tasks.
In this case, Minecraft AI learned to perform complicated sequences of mouse and keyboard clicks to complete in-game tasks like felling trees and crafting tools, all using a technique known as imitation learning, in which neural networks are trained to perform tasks by watching humans do them. Imitation learning can be used to train AI to control robotic arms, drive cars, or navigate web pages (which can be very dangerous, by the way).
In this case they have used a different approach to what currently exists, where video demos must be labeled at each step. With the new system, called Video Pre-Training (VPT), ​​it is possible to label the videos automatically, since the actions and consequences are automatically recorded.
They have chosen Minecraft because it is a game without a clear objective, we can do what we want, there is no winner or loser, being a good environment to train the AI.
The AI, so far, had already managed to create planks and turn them into a table, which implies around 970 consecutive actions. With this new method they are able to perform tasks that involve more than 20,000 consecutive actions.
We are still far from being able to say that this AI is better than a human playing, but that limit is getting closer.
You can read more details at MIT.
