A team of scientists from one of the companies owned by Google, Deep Mind, has managed to create an algorithm capable of learning to play classic games as well-known as Comecocos or “Space Invaders” (the old Martians) starting from of a trifle information about the game in question.
Until now, the learning method for artificial intelligence had been reinforcement based on stimulus and reward, but this technique has its limitations, such as, for example, that it is subject to controllable environments and simple situations. Now, thanks to the new algorithm, the machine can learn by itself, with hardly any prior information .
To create this innovative code, the experts combined the classic learning used in AI with an artificial neural network in the image and likeness of biological ones . The sum of both, called “Deep Q-network” , has achieved that the performance in video games “can be compared with that of a professional human video game tester”, explains Demis Hassabis, co-author of the study published in the journal Nature .
And is that after testing the performance of the new algorithm with a professional video game human, the machine achieved more than 75% of the score in more than half of the games tested, which were a total of 49 classic games from the Atari 2600 console from boxing games to car races.
“We use the same initial data for all the games, which shows that the agent successfully learns the procedures of each one based solely on the sensory inputs, which shows that using the same architecture , the machine can learn to optimize strategies in different environments ”, the researchers explain.
