AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-23
Learning SOCP Solver Dynamics with Graph Neural Networks
Second-order cone programs are widely used in engineering, but interior-point solvers can be costly when solved repeatedly. The authors propose SOCPNet, a recursive graph neural network that learns from the intermediate iterates of classical solvers and refines its solution step by step. Used to warm-start an interior-point solver, the trained model cuts solver iterations without loss of accuracy.