Atari 2600 game system (Photo : Reuters)
A new AI (Artificial Intelligence) system can beat its human opponents in some old-school Atari 2600 games like clockwork. It does this after learning the rules like humans do, step by step.
The creators believe that the AI system will be able to do even more spectacular actions in the future. One day the system might be able to drive your vehicle!
The research performed by Google DeepMind was published in the latest issue of the Nature magazine. Human brains were the inspiration for the research, according to NBC News.
The learning system proceeds from "pixels to actions" according to Demis Hassabis from DeepMind. It can even do tasks that are difficult for humans, such as Atari video games from the 1970s and 1980s.
In the research, the AI defeated humans in 23 out of 49 games, including Breakout, Space Invaders, and Video Pinball.
In fact, the AI system scored close to the humans' scores in most games, according to Wired. It used two AI methods: deep learning, and deep reinforcement learning.
The AIs generally performed better in action games. That is because they were not required to experiment or explore.
The AIs are different from systems that are programmed to train to beat chess grandmasters. Instead, they learn and adapt based on their experiences.
The AI system, the Deep Q-network agent, learns like a human does. It learns how to play by observing the game, and then learns how to play the game better.
In fact, the system even found unique strategies that humans have never tried. They include earning easy points or finding certain safe areas.
Besides that, Deep-Q can also apply its lessons to different situations in different games. Usually this is a skill that humans do much better than AI systems can.
