Google’s scientific artificial intelligence (AI) research arm DeepMind is seeking to improve machine learning by giving computers the ability to dream.
In a paper published online, researchers at DeepMind said they had achieved a leap in the speed and performance of a machine learning system.
The paper revealed details about how DeepMind’s new system, dubbed Unsupervised Reinforcement and Auxiliary Learning agent (Unreal), learned to master a three-dimensional maze game called Labyrinth 10 times faster than the current best AI software.
Researchers said it can now play the Labyrinth game at 87% the performance of expert human players.
Bloomberg reported that DeepMind researchers Max Jaderberg and Volodymyr Mnih jointly wrote via e-mail, “Our agent is far quicker to train, and requires a lot less experience from the world to train, making it much more data efficient.
The agent was tested on a suite of 57 Atari games and Labyrinth with 13 levels.
In all the games, the same Unreal agent is trained in the similar way, on the raw image output from the game, to produce actions for maximising the score or to reward the agent.
The behavior needed to get game rewards changed, from picking up apples in 3D mazes to playing Space Invaders.
On Atari the agent now achieves on average nine times human performance. Researchers expect that this work will enable them to scale up their agents to ever more complex environments.
In another project, researchers from DeepMind and the University of California, Berkeley, have trained AI machines to learn the physical properties of objects by interacting with them virtually.
The research project drew inspiration from child development and sought to build agents that can learn to experiment so as to learn representations that are informative on physical properties of objects, using deep reinforcement learning.
