View a PDF of the paper titled A Semismooth Newton Stochastic Proximal Level Algorithm with Variance Discount, by Andre Milzarek and a pair of different authors

View PDF
HTML (experimental)

Summary:We develop an implementable stochastic proximal level (SPP) methodology for a category of weakly convex, composite optimization issues. The proposed stochastic proximal level algorithm incorporates a variance discount mechanism and the ensuing SPP updates are solved utilizing an inexact semismooth Newton framework. We set up detailed convergence outcomes that take the inexactness of the SPP steps into consideration and which might be in accordance with present convergence ensures of (proximal) stochastic variance-reduced gradient strategies. Numerical experiments present that the proposed algorithm competes favorably with different state-of-the-art strategies and achieves increased robustness with respect to the step dimension choice.