Over-Relaxation Policies

Vector policies can exploit per-constraint information but may lose runtime gains because of per-row inference overhead. Scalar policies are cheaper to evaluate and often produce the best wall-clock runtime. The learned policies update the relaxation matrix every ten ADMM iterations without requiring matrix refactorization. The vector policy predicts per-constraint relaxation values using shared row-wise weights and per-constraint features. The paper learns low-dimensional solver hyperparameters rather than high-dimensional perturbations to the optimization iteration map. The scalar policy predicts one relaxation value from global residual and penalty features.