ADMM Penalty Parameter Selection by Residual Balancing


Appropriate selection of the penalty parameter is crucial to obtaining good performance from the Alternating Direction Method of Multipliers (ADMM). While analytic results for optimal selection of this parameter are very limited, there is a heuristic method that appears to be relatively successful in a number of different problems. The contribution of this paper is to demonstrate that their is a potentially serious flaw in this heuristic approach, and to propose a modification that at least partially addresses it.
Submitted 20 Apr 2017 to Optimization and Control [math.OC]
Published 21 Apr 2017