Playing Atari with Deep Reinforcement Learningcache is certainly the most catchy paper I've read in years. But before I start talking about the paper, you should definitely watch this video from it's presentation:
The basic idea is very simple - to train and evaluate the AI, let it play a variety of computer games. This has several advantages:
- Training and Evaluation can happen faster than real time
- Using several very different games can prevent over-fitting
- Games are usually created to be challenging for humans, thus containing an implicit requirement for intelligence.
DeepMind achieves this by letting it's Deep Reinforcement Learning algorithms play Atari games, using only the pixel values and the current score as input. These are simple enough to represent the whole world on a single visible screen - no off-screen action, are computationally very cheap to execute and still difficult for humans.
I won't go into details of their algorithm (go read the paper already!), but I expect they will move to more modern 3D games once they've reached the limits of simple 2D games. This will present several new challenges, most importantly stereo vision and object permanence.