Communications of the ACM
Collaborative Learning--For Robots
A new algorithm enables distributed agents like robots to collaborate on a machine-learning model without aggregating data.
Credit: Ogilvy South Africa
Researchers at the Massachusetts Institute of Technology (MIT) Laboratory for Information and Decision Systems have written an algorithm that enables independent agents to collaborate on a machine-learning model without aggregating data.
Distributed agents, such as robots surveying a building, gather and analyze data independently, and pairs of agents then exchange analyses. This analysis exchange is repeated many times, with the model growing increasingly refined.
The distributed algorithm outperformed a standard algorithm that works on data aggregated at a single location in experiments with various data sets. "A single computer has a very difficult optimization problem to solve in order to learn a model from a single giant batch of data, and it can get stuck at bad solutions," says MIT graduate student Trevor Campbell.
Although robot collaboration was the impetus for the algorithm, the work also could have applications in big data, enabling distributed servers to merge data analyses without aggregating data at a central location. In addition, the algorithm could be applied to learning problems such as topic modeling, in which a computer uses relative word frequency to classify documents according to topic. It also could enable scattered servers to independently work on documents and generate a collective topic model.
From MIT News
View Full Article
Abstracts Copyright © 2014 Information Inc., Bethesda, Maryland, USA
No entries found