Google's DeepMind artificial intelligence (AI) is learning to navigate thee-dimensional (3D) environments and puzzle-solving games, including a soccer game in which the AI plays as a virtual ant, according to DeepMind's David Silver.
Silver says the AI controls the ant's four-legged movement to chase down the virtual soccer ball, dribble, and score a goal; previously, the AI mastered two-dimensional Atari games. The ant soccer challenge was solved via reinforcement learning and DeepMind's Deep-Q Network algorithm, which stores a bot's experiences and estimates actionable rewards.
DeepMind also developed an asynchronous actor-critic algorithm, called A3C, to help the AI learn how to play soccer without prior experiences. Silver says A3C uses standard multicore central-processing units instead of graphics processing unit-based algorithms to solve motor-control challenges and 3D navigation using visual input in a fraction of the training time.
He also notes DeepMind is testing its AI with Labyrinth, a 3D maze and puzzle game using only visual cues.
DeepMind also has created Gorila, a reinforcement learning system with quick training times, which has been applied to Google recommender systems, according to Silver.
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA