Google-Logo-AH-1

Google Computer Learns And Masters 49 Atari 2600 Games Without Knowing The Rules

February 25, 2015 - Written By Christopher Dal Porto

 

Representing a huge leap in computers utilizing independent reasoning, reports of a computer which can both learn the rules and methods needed to win a game have surfaced. While this sort of artificial intelligence in the arena of playing games and devising strategies to win isn’t new, this instance takes existing methods to a whole new level. Programs written in the past at companies such as IBM which could learn and improve its methods to win at Jeaporady! are similar to this, however they were programmed with the knowledge of the rules from the start. This particular computer at Google is plunged headfirst into a game without awareness of the rules whatsoever.

This rule-learning method successfully masters video games, however the technology could prove groundbreaking in other contexts such as searching, voice recognition, and language translations. Researchers from the British journal Nature commented stating that the developed rule-learning software could generate great advancements in robots and even self-driving cars. The lead of the project at Google, Dennis Hassabis, says that the program is able to devise rules and winning strategies by combining sensory knowledge and learning in a dual capability method called “deep-Q network,” or D.Q.N. Dennis Hassabis finishes by explaining that D.Q.N. is how the program masters and gains understanding of the structure of tasks given to it. He was the founder of DeepMind, an artificial intelligence company that Google purchased last year for $400 million.

The program from Google was able to successfully learn and master over 49 Atari 2600 computer games. For certain games, such as Seaquest, the program was even able to devise tricks such as keeping the submarine near the surface to survive throughout the entire game. While this program shows new advancements in artificial intelligence, it also outlines the current limits the field is restrained to with today’s methods. For example, the program is able to devise rules to a game but will never be able to understand conceptually what a submarine is. For now, the ability to bring in prior knowledge to make conceptual inferences is still beyond the reach of artificial intelligence. The next step of the project will be to get the program to be able to navigate games such as Grand Theft Auto.