The team at Google DeepMind’s lab is working on models to emulate the ability to imagine the consequences of an action before taking it: it’s basically trying to figure out what intelligence and imagination are to convert them into agents, algorithms.
It is a crucial step for the future of artificial intelligence: will “systems” know and will be able to adapt to changing conditions and not specifically planned?
♪ When we put a glass on the edge of the table we probably consider the possibility that it falls, researchers write: ♪ On the basis of an event we might decide to move the glass. To achieve similar results, our algorithms must have the ability to imagine and reason about the future, and to make plans that use this ability.
Think before you do it! We have already seen a preview of artificial intelligence of “Deep Thought class,” the chess computers of IBM whose name is a tribute to Deep Thought (see the Galactic Guide for Hitchhikers). DeepMind has challenged and won the world champions of Go: unlike chess, the game of Go has a number of possible moves too high to be counted and each move leads to a number of possible scenarios that, it is said, is larger than the number of
In the Go, AI is therefore obliged to “play intuition.” But the rules of the real world are even more varied and complex than those of Go and Google’s team is working on a system that operates on another level.
Combined several approaches, including trial and error learning (reinforcement learning) and deep learning (through the processing of large amounts of data, in a similar way to the human brain), try to combine the method for attempts with the simulation capacity: in this
Playing learns. The method was tested with Sokoban (see below), a relatively simple 80s video game that offers a maze with paths that change by shifting obstacles. Some moves may make it unsolvable, so advanced planning is needed, but Deep Mind has not been given the rules of the game in advance: the AI with imagination has solved 85% of the levels, a remarkable step ahead compared to 60% previously obtained
13 PHOTOS Photogallery Welcome to 2020! GO TO THE GALLERY
The test highlighted the fact that these new AIs can better manage their gaps, collect more effectively useful information for their simulations and can learn different strategies to apply in subsequent games. It is not a “simple” early planning: it is something very similar to creative and fluid planning.
Despite its success, it is still the first days of this technology and the games it is subjected to are very far from representing the complexity of the real world. But it’s a promising start to the development of an artificial intelligence that doesn’t put a glass of water on the edge of the table.