Deep-Q: An Artificial Intelligence That Learns with Atari 2600 Games

Testing an artificial intelligence by pitting it against video games is a common benchmark. In the past, we've seen a couple of examples involving Nintendo's top mascot, but this time the AI's 'victims' are 49 Atari 2600 games. Its name is Deep-Q, and all it needed was to train hundreds of times on each title.

While many have warned us about the evolution of AIs, that doesn't change the fact that many experts are working and advancing in this field. One of the main goals is for them to learn by themselves and make decisions without human intervention, which sounds fascinating and dangerous at the same time. If there is an activity that forces an artificial intelligence to learn from its mistakes, it's playing. At first, the AI doesn't even know how to move the characters, but with each new attempt it gets smarter. A little fast-forward through 600 or 700 games, and we have a fast, precise player, capable of exploiting programming errors, and ranking alongside a human player.

That brings us to Deep-Q, also known as Deep Convolutional Network, developed by scientists at DeepMind, Google's AI branch. In total, Deep-Q faced 49 Atari 2600 classics, including Breakout and Space Invaders. Its creators have highlighted that the only information available to Deep-Q is the pixels on the screen and the score counter. From there, Deep-Q must learn the controls, understand how to score points, and of course, try not to lose. When compared to Deep Blue, Demis Hassabis of DeepMind explains that they are two completely different creatures. Deep Blue was nothing more than a debugged and optimized program with millions of moves at its disposal, but Deep-Q is based on algorithms that learn from scratch through perceptual experience.

Some of the tricks Deep-Q recorded during its games include keeping a submarine safe just below sea level in SeaQuest, and creating small tunnels to hit blocks from the top in Breakout. For now, Deep-Q has managed to achieve more than 75 percent of the scores obtained by human players. The future plans for the algorithms available in Deep-Q include weather predictions and the stock market, but the general parameters they seek for Deep-Q are those of a two- or three-year-old child, and both Hassabis and his colleagues admit they are still far from that. As long as they don't teach it Missile Command, I think everything will be fine...

Source: