Google has developed a program that, for the first time, is able to learn to play on their own and independently, which is a very important step regarding the type of existing artificial intelligence so far. Deep Blue, to take a familiar example, won the chess to Kasparov in 1997 thanks to having been previously programmed by humans with the rules and specific strategies, which added to its computing power gave the necessary advantage. He did not learn alone.
In the case of this new “agent”, as they call it in Google, we are talking of leaving the program play for itself and learn independently what the best strategy to win is.
The new AI has learned to play alone at 49 retro titles
Last year Google bought a company called DeepMind which aims to build intelligent machines. This group has been responsible for developing the new agent, which has learned to play without help 49 different retro titles.
In this video we see how the AI learns to play the classic ‘Breakout’ Atari. The agent keeps training and learning and, after 600 rounds, gives a winning strategy. Decides that the best way to overcome levels is attacking the side of the barrier so that the ball is placed on the other side and finish with it from above:
When the agent starts playing for the first time a game is produced something similar to when a newborn opens his eyes and sees the world for the first time. In this case the IA notes the information on the screen and pressing buttons randomly to see what happens.
It uses a method called deep learning that allows to convert visual basic inputs into meaningful concepts, in the same way that the human brain is capable of transforming the raw sensory information in a rich understanding of the world. Thanks, on the other hand, Reinforcement Learning, the agent is able to detect what has value. And hence arises a series of basics like this: score points is good; losing is bad.
The agent is very good, but much remains to be perfect
In the study, figure that the agent had a yield of 75% of the level of a professional tester game, or even higher, in the middle of the tested games, among whom were from shooters side-scrolling Driving up games in 3D environments. In games like ‘Space Invaders’, ‘Pong’ or ‘Breakout’, the agent overcame humans, while in other games did much worse.
Researchers say that this is so mainly because of the lack of real memory by the agent, which does not allow you to establish long-term strategies that require planning. For this reason, the team DeepMind is trying to build a memory component to add to the system and apply it more realistic three-dimensional gaming environments.
Elon Musk, CEO of Tesla, was one of the first investors of DeepMind and ensures that advances in the field of Artificial Intelligence occur so rapidly that the risk of this happening is something really dangerous in a time frame located between the next five or ten years at most.
We have spoken of the agent in terms of game, but clearly this type of AI able to learn independently has many more applications: cars that drive themselves, personal assistants on smartphones, scientific research in various fields and much, much more.