Skip to main content
Back

UCF's HUANet Solves Optimization 44x Faster While Guaranteeing Constraint Satisfaction

HUANet turns each iteration of ADMM into a trainable neural layer, enforcing equality constraints exactly while running orders of magnitude faster than classical solvers.

  • HUANet unrolls ADMM iterations into shared-weight neural layers for parametric constrained convex optimization.
  • A differentiable correction stage guarantees equality constraint satisfaction to machine precision at every layer.
  • Self-supervised training uses KKT residuals as soft constraints, requiring no ground-truth optimal solutions.
  • Achieves 8x to 44x speedup over Clarabel and OSQP on 100-variable LASSO and QP problems.
  • Reaches up to 4700x speedup over vanilla ADMM on entropy maximization at higher dimensions.
  • Read the full paper on arXiv; no code release announced yet.

Solving the same convex optimization problem thousands of times with slightly different parameters is a daily reality in robotics, portfolio management, and power grid dispatch. Classical solvers like OSQP or Clarabel treat every instance as new, while end-to-end neural nets tend to produce solutions that violate the constraints. A team from the University of Central Florida has proposed HUANet, a network that unrolls the Alternating Direction Method of Multipliers into trainable layers while guaranteeing equality feasibility at every step.

Where learning-to-optimize breaks down

Learning-to-optimize methods learn the mapping from problem parameters to solutions, something classical solvers do not exploit, so they can handle new instances significantly faster. The catch is that purely end-to-end surrogate models struggle in constrained cases because the predicted outputs are not guaranteed to satisfy the constraints at test time. Predict-and-correct hybrids fix some of this but remain black boxes.

Algorithm unrolling turns each iteration of a solver into a neural layer, keeping the interpretability of the original method. Existing unrolling approaches still treat constraints as soft penalties with no feasibility guarantee, and prior unrolled ADMM work has focused only on unconstrained problems. HUANet targets exactly that gap.

Inside the architecture