Engineers DeepMind - the London division of Google, engaged in developments in the field of artificial intelligence, taught computer algorithms "to plan ahead on the basis of imagination". As reported by the company, AI agents I2As use the internal "imaginative converter" to better predict how they should act in this or that situation.

The converter allows neural networks to present "in the head" various scenarios for the development of the situation and choose the most favorable. Predicative capabilities of AI were tested, in particular, on the logic puzzle game Sokoban, in which the player needs to place boxes within the labyrinth. If an error occurs and the object is in the wrong position, it is no longer possible to move it, and the game ends.

DeepMind scientists programmed Sokoban so that the new level was generated anew each time, and instructed the AI ??agent "with imagination" to solve the puzzle. The process of "thinking" I2As is shown on the right side of the video.

Before each move, algorithms "in the head" thought out which way to best push the box to win. As a result, I2As with internal planning coped with the task faster and for fewer moves than the AI ??agents without it.