After demonstrating the impressive capabilities and abilities that the google artificial intelligence, as you will surely remember capable of even being at the height of the best Go player on the planet or of learning to play StarCraft II, where surely after several months also demonstrating its immense qualities, the point has come that the work done by DeepMind of a new step so that now I am able to identify objects by their properties in order to decide the best possible way to grab them.
For this a team of developers and researchers formed by inengineers from Google and the University of California, has decided to start practicing with the algorithm so that he educates himself as a human being would do in his earliest childhood, that is, they will let him pull, push, break and generally experiment with the world within a virtual entrojo commanded by DeepMind.
The objective of this work is to make DeepMind capable of learn the properties of physical objects in order to interact with them. This type of teaching is known under the name of 'deep reinforcement learning'and will allow this platform to allow in real time to solve tasks without specific instructions, something very similar to our way of interacting with a certain object when we do not know what it is made of or how to use it, that is, instinctively.
Thanks to the use of deep reinforcement learning techniques DeepMind will be able to interact with any type of object.
To achieve this the researchers created two different environments So that DeepMind could experiment and learn from its mistakes, for this, in the first place, it faced the system with five blocks of the same size but with different weight, seeking to get the platform to identify which was the heaviest where it learned that the only way to guess it it was interacting with all objects.
Second, the platform was pitted against towers of different heights for DeepMind to calculate how many blocks were in each. In case of success, a series of rewards were offered while, if a failure occurred, negative feedback was given to the platform. With these tests the platform learned to discover new ways of acting based on ingenuity. Thanks to this DeepMind is now able to find solutions when there are no clear instructions or they are directly lacking.
Further information: Arxiv