Randomized primal-dual methods based on stochastic quasi-Fejér monotonicity property.

A. Repetti, P. L. Combettes and J.-C. Pesquet.

European Conference on Stochastic Programming and Energy Applications (ECSP 2014),
Paris, IHP, France, 24 - 26 septembre 2014.

We propose a framework for designing fixed point algorithms relying on a notion of stochastic quasi-Fejér monotonicity. These methods rely on a sweep of blocks of variables activated at each iteration according to a random rule, and they allow stochastic errors in the evaluation of the involved operators. Using the proposed approach, we develop novel asynchronous distributed primal-dual methods in a multi-agent context. Then, we derive a class of proximal algorithms for solving composite convex variational problems.




Biomedical and Astronomical Signal Processing group
Institute of Sensors, Signals and Systems
Heriot-Watt University
Edinburgh EH14 4AS

mail: A.Repetti@hw.ac.uk