Constraint Boundary Wandering Framework: Enhancing Constrained Optimization With Deep Neural Networks.

Area of Science:

• Computational Mathematics
• Machine Learning
• Operations Research

Background:

• Constrained optimization problems are widespread but challenging for traditional scalable methods.
• Deep neural networks (DNNs) offer potential for advanced optimization techniques.

Purpose of the Study:

• To develop a novel learning-based framework for scalable constrained optimization.
• To address limitations of conventional methods in handling complex optimization tasks.

Main Methods:

• Introduced the Constraint Boundary Wandering Framework (CBWF).
• Incorporated a boundary wandering strategy inspired by active-set methods.
• Treated the Lipschitz constant as a learnable parameter and evaluated regularization terms, favoring the nonsmooth L2 norm.

Main Results:

• CBWF demonstrated superior performance on synthetic and ACOPT datasets.
• Outperformed existing deep learning-based solvers in objective and constraint loss.
• Showcased enhanced equality constraint feasibility.

Conclusions:

• The Constraint Boundary Wandering Framework (CBWF) offers a powerful new approach for scalable constrained optimization.
• This deep learning strategy effectively handles complex optimization challenges, improving upon existing methods.