Export Reviews, Discussions, Author Feedback and Meta-Reviews

This paper considers the use of augmented Lagrangian or the method of multipliers to solve the linear program, attempting to address the scalability issue of classical linear program methods like interior point methods and simplex methods for large scale high dimensional machine learning problems. The resulting augmented Lagrangian method basically reduces to two steps which update primal and dual variables alternatingly. The primal variable update is to solve a bound-constrained quadratic problem which involves high computational effort. To solve the primal subproblem, the authors proposed to use another optimization algorithm like randomized block coordinate descent to solve it. Along with the outerloop, the proposed algorithm is a double-loop algorithm.