Random Activations in Primal-Dual Splittings for Monotone Inclusions with a priori Information

TL;DR

This paper introduces a flexible stochastic primal-dual algorithm for monotone inclusions with a priori information, demonstrating convergence and improved performance in convex optimization problems with constraints.

Contribution

It proposes a novel random activation scheme for primal-dual algorithms that incorporates a priori information, extending existing methods and improving convergence speed.

Findings

Algorithm converges almost surely using stochastic Quasi-Fejér sequences.

Includes various activation schemes like Bernoulli and cyclic.

Numerical results show faster convergence in a transport network problem.

Abstract

In this paper, we propose a numerical approach for solving composite primal-dual monotone inclusions with a priori information. The underlying a priori information set is represented by the intersection of fixed point sets of a finite number of operators, and we propose and algorithm that activates the corresponding set by following a finite-valued random variable at each iteration. Our formulation is flexible and includes, for instance, deterministic and Bernoulli activations over cyclic schemes, and Kaczmarz-type random activations. The almost sure convergence of the algorithm is obtained by means of properties of stochastic Quasi-Fej\'er sequences. We also recover several primal-dual algorithms for monotone inclusions in the context without a priori information and classical algorithms for solving convex feasibility problems and linear systems. In the context of convex optimization with inequality constraints, any selection of the constraints defines the a priori information set, in which case the operators involved are simply projections onto half spaces. By incorporating random projections onto a selection of the constraints to classical primal-dual schemes, we obtain faster algorithms as we illustrate by means of a numerical application to a stochastic arc capacity expansion problem in a transport network.