Developers from DeepMind and Carnegie Mellon University decided to use an interesting approach to learning. Learning AI languages is rather tedious, so scientists decided that it would be better if the artificial intelligence learns the basic principles of using the language on its own - for this, the researchers put two neural networks in an arcade game and made them chop until they turn blue, comparing individual objects with different characteristics to those until they begin to understand the language.
So, programs, trying to achieve the desired result in millions of different ways, learned to associate certain words with objects and their characteristics and began to operate with different concepts, such as "larger" or "smaller", for example, and began to distinguish similar externally objects from the game environment. So, the artificial intelligence was able to apply the knowledge gained in unfamiliar situations, encountering columns and other objects of the same color in the game, but, according to the assignment, it went to the columns, although it had not seen them before.
“The ultimate goal is to help the AI transfer the knowledge gained from the game to other areas of its activity,” says Devendra Chaplot, MA, Carnegie Mellon University, who gave a presentation on the team's work.
Video games are used by developers at DeepMind for other purposes as well. With the help of the Sokoban game, for example, AI is taught to imagine - before making the next move, the program itself calculates the possible options "in mind" and makes the right choice.
Viacheslav Larionov