Bulletin of Computational Applied Mathematics (Bull CompAMa)
Modified Spectral Projected Subgradient Method: Convergence Analysis and Momentum Parameter Heuristics
Milagros Loreto, Samantha Clapp, Charles Cratty, Breeanna Page
The Modified Spectral Projected Subgradient (MSPS) was proposed to solve Langrangen Dual Problems, and its convergence was shown when the momentum term was zero. The MSPS uses a momentum term in order to speed up its convergence. The momentum term is built on the multiplication of a momentum parameter and the direction of the previous iterate. In this work, we show convergence when the momentum parameter is a non-zero constant. We also propose heuristics to choose the momentum parameter intended to avoid the Zigzagging Phenomenon of Kind I. This phenomenon is present in the $MSPS$ when at an iterate the subgradient forms an obtuse angle with the previous direction. We identify and diminish the Zigzagging Phenomenon of Kind I on Setcovering problems, and compare our numerical results to those of the original MSPS algorithm.
Keywords: spectral projected gradient; subgradient methods; momentum term.