Abstract:

There has been significant interest in developing methods that bridge between classical optimization and modern deep learning, under the broad theme of learning to optimize (L2O), for improved optimizers. In this paper, we propose Switch-L2O – a new primal-only method for learning a constraint optimizer. Empirically, our method is shown to enjoy a better optimality gap and reduces constraint violations against prior methods on convex and nonconvex optimization problems with possibly nonconvex constraints.