In this paper, we consider solving a class of nonconvex and nonsmoothproblems frequently appearing in signal processing and machine learningresearch. The traditional alternating direction method of multipliers encountertroubles in both mathematics and computations in solving the nonconvex andnonsmooth subproblem. In view of this, we propose a reweighted alternatingdirection method of multipliers. In this algorithm, all subproblems are convexand easy to calculate. We also provide several guarantees for the convergenceand prove that the algorithm globally converges to a critical point of anauxiliary function with the help of the Kurdyka- Lojasiewicz property. Severalnumerical results are presented to demonstrate the efficiency of the proposedalgorithm.
展开▼
