DeepMind's upgrade to the AlphaGo algorithm, AlphaGo Zero, beat its predecessor in a 100-game Go match, acquiring skills without human data by playing millions of games against itself.
"By not using human data or human expertise, we've actually removed the constraints of human knowledge," says University College London (U.K.) professor David Silver. "It's able to create knowledge for itself from first principles."
DeepMind CEO Demis Hassabis says the methods for building AlphaGo Zero can be used in real-world scenarios in which vast possibilities must be explored.
Both Silver and Hassabis note many situations may lack a sufficiently large body of examples to learn from, making self-learning a requirement.
AlphaGo Zero, like AlphaGo, used a deep neural network and a search algorithm to determine its next move, but it differed from the original AlphaGo program in its use of a single network to execute both functions.
From Technology Review
View Full Article - May Require Free Registration
Abstracts Copyright © 2017 Information Inc., Bethesda, Maryland, USA
No entries found