[02325] A Hybrid Method for Solving Linear KKT Systems
Session Time & Room : 4D (Aug.24, 15:30-17:10) @E508
Type : Contributed Talk
Abstract : We propose an iterative method for solving linear systems arising from optimization problems with a separable objective function and dense constraints and suitable preconditioner for quick convergence. The method is implemented in Julia using Krylov.jl for the iterative solution, CUDA.jl to enable GPU capabilities, and a custom kernel for construction of the preconditioner. The method attains faster solution times than direct methods common in solvers such as Mosek and Gurobi in theory and in practice.
