A Single-Loop Bilevel Deep Learning Method for Optimal Control of Obstacle Problems
Researchers have developed a novel single-loop bilevel deep learning method for optimal control of obstacle problems. This mesh-free approach is designed to be scalable to high-dimensional and complex domains, avoiding the repeated solution of discretized subproblems inherent in classical methods. The method utilizes constraint-embedding neural networks and a Single-Loop Stochastic First-Order Bilevel Algorithm (S2-FOBA) for efficient training, demonstrating reduced computational costs and satisfactory accuracy in benchmark experiments. AI