Things are moving fast in the world of artificial gaming intelligence. As a startup, DeepMind’s AI was able to learn (and eventually master) seven Atari 2600 games, getting good enough to beat humans at three of them. Now (and $628 million in Google investment dollars later), DeepMind is capable of schooling people in 22 out of 49 Atari games, and matching them at seven of them. But why the lag in performance on others? A new paper published in Nature explains.
DeepMind is unlike seemingly more impressive AI (like the famous chess-playing Deep-Blue) in that it’s not pre-programmed with moves and/or strategy – it has to learn as it goes. That makes it really good for games like “Space Invaders,” where it can crush humans with ease. But something as innocuous as “Mrs. Pac-Man?” Not so much. The authors explain why:
“The classic game Ms. Pac-Man neatly illustrates the software’s greatest limitation: that it is unable to make plans as far as even a few seconds ahead. That prevents the system from figuring out how to get across a the maze safely to eat the final pellets and complete a level. It is also unable to learn that eating certain magic pellets allows you to eat the ghosts that you must otherwise avoid at all costs.
DeepMind’s software is essentially stuck in the present. It only looks back at the last four video frames of game play (just a 15th of a second) to learn what moves pay off, or how to use its past experience to choose its next move. That means it can only master games where you can make progress using tactics that have very immediate payoffs.”
So in a game like “Space Invaders,” where the antagonists advance forward in regular, predictable patterns, DeepMind is more than capable of achieving God-like status. But in “Mrs. Pac-Man,” where eating magic pellets provides actual God status, things move a little too fast for DeepMind to keep up.
DeepMind’s strategies are complex, and its only real limitation is memory. Demis Hassabis, the leader of Google DeepMind, says his team is working to improve memory and attention span. It’s already capable of playing some Nintendo and PC games, which take place in a 3D environment. Hassabis also hints at what might be in store for DeepMind further into the future.
“Ultimately the idea is that if this algorithm can drive a car in a racing game, with a few tweaks it will be able to drive a real car,” he said at a press conference Tuesday.