Two AIs Go Head-to-Head on Atari's 'Breakout' to Test Deep Learning

Thursday, 15 June 2017 - 12:09PM
Technology
Artificial Intelligence
Thursday, 15 June 2017 - 12:09PM
Two AIs Go Head-to-Head on Atari's 'Breakout' to Test Deep Learning
Image credit: Atari

It seems like every day brings a new AI more capable than the last. This was recently apparent with AlphaGo—it was pretty great at beating Breakout, then Google got involved and soon it was capable of beating the world's leading Go champion.

To do this, AlphaGo uses what is known as 'deep reinforcement learning'. For example, in Breakout, it will take raw image frames of the game as it's being played. Whether or not the ball is hitting the bricks in those frames will decide whether or not positive reinforcement is registered. Through this process, it can slowly get better at the game, "learning" what leads to positive reinforcement.

However, this reactionary style of learning is, according to one of its competitors, not close enough to how humans learn and is therefore limited. Vicarious, an AI company (their AI is called Schema Networks), does things a little bit differently. Their co-founder D. Scott Phoenix says that "To have AIs think the way you and I do, they need to move towards models that can reuse concepts, understand cause and effect." He asserts that deep reinforcement learning networks rely on trial and error, which hinders them. He also claims that their system of analyzing entire frames of the screen leaves details unnoticed.

To test whether or not Schema really is superior to deep reinforcement learning networks, Vicarious simply pitted them against each other in a game of Breakout:

Image Credit: Atari


Because Schema was able to adapt much faster to changes in the game, it beat even the highest scores of the deep reinforcement learning programs. Phoenix asserts that "The Schema Networks are all about actually learning the concepts of the game. What happens when a ball hits a paddle? It learns that concept, and then can generalize to different environments that it was never trained on."

It's clear that AI is moving closer and closer to human learning styles. But should we really run down this rabbit hole, until we finally create a truly human-like AI? What happens when they beat all our high scores on Pac-Man?

Science
Science News
Technology
Artificial Intelligence
No