The developers invented an AI that is able to predict the future

Amazing design appeared in the United States.

The company DeepMind (owned by Google) has created a new type of artificial intelligence that can use their “imagination” to create an image of the near future and plan his actions. He proved to be more accurate and efficient than conventional AI.

In a posting on its website, DeepMind researchers provided a brief overview of the “new approach to planning based on imagination.” The so-called agents augmented by imagination or I2As (Imagination-Augmented Agents) use the built-in “code of the imagination”, which helps the AI to decide which assumptions about the future events is correct and what is not.

Recently, the founder of DeepMind, Demis Hassabis wrote an article, published in the journal Neuron about how the development of AI depends on the encoding of human abilities such as imagination, curiosity and memory. I2A is an attempt to copy these functions of the human brain.

The new algorithm was tested on the puzzle game Sokoban, where the little alien have to push blocks to the right place. They can not pull, so one wrong move leads to losing the round. The researchers gave artificial intelligence only one attempt that forced him to scroll “in the head” various options and the possible outcome of each movement. Only after such verification, the AI has made real progress in the game.

I2A was better than the AI without “imagination”. He has learned to quickly navigate the puzzles, extracting more information from the simulation gaming events. When the researchers added in AI “Manager” who was checking plan future action, the system began to solve problems faster and more efficiently.

Of course, the type of computer “imagination” too far from what people are capable of, however, the experiments prove that the AI might benefit from attempts to predict the future and plan their actions and decisions.

Swedish expert in the field of artificial intelligence Nick Bostrom believes that DeepMind more likely to create humanoid artificial intelligence. And it can happen in the coming decades.

© 2017, paradox. All rights reserved.