Stabilization techniques for Column Generation: towards automated schemes.
Résumé
Stabilization procedures are critical feature to accelerate the convergence of column generation algorithms. Standard techniques include penalty functions that drive the dual solution close to the incumbent, or smoothing techniques that can be viewed as interior point cutting plane strategies in the dual. Our paper is a critical review of such methods, developing further convergence proof and proposing extensions. The efficiency of stabilization heavily rely on proper parametrization. To address this key issue, we develop auto-regulating strategies that aim to replace parameter tuning.