Abstract
Optimal control of obstacle problems arises in a wide range of applications and is computationally challenging due to its nonsmoothness, nonlinearity, and bilevel structure. Classical numerical approaches rely on mesh-based discretization and typically require solving a sequence of costly subproblems. In this work, we propose a single-loop bilevel deep learning method, which is mesh-free, scalable to high-dimensional and complex domains, and avoids repeated solution of discretized subproblems. The method employs constraint-embedding neural networks to approximate the state and control and preserves the bilevel structure. To train the neural networks efficiently, we propose a Single-Loop Stochastic First-Order Bilevel Algorithm (S2-FOBA), which eliminates nested optimization and does not rely on restrictive lower-level uniqueness assumptions. We analyze the convergence behavior of S2-FOBA under mild assumptions. Numerical experiments on benchmark examples, including distributed and obstacle control problems with regular and irregular obstacles on complex domains, demonstrate that the proposed method achieves satisfactory accuracy while reducing computational cost compared to classical numerical methods.
Introduction
Obstacle problems are a fundamental class of mathematical models used to describe systems where a state variable must satisfy specific constraints—for instance, the deformation of an elastic membrane that cannot penetrate a rigid foundation or the pricing of American options in finance. When we introduce optimal control to these problems, the goal is to find an external force (the control) that steers the system toward a desired target state while ensuring the state remains physically consistent with the obstacle constraints.
This creates a complex bilevel optimization structure. The upper-level problem seeks to minimize the distance to a target, while the lower-level problem ensures the state is a valid solution to the obstacle problem given the current control. Traditional numerical methods, such as Finite Element Methods (FEM), struggle with these problems because the "free boundary"—the area where the state meets the obstacle—is nonsmooth and rarely aligns with a fixed grid. Furthermore, classical solvers often require "nested" loops, solving the entire lower-level problem to completion for every single update to the control, which is computationally prohibitive for high-dimensional or complex domains.
Methodology
To overcome these challenges, this research introduces a mesh-free framework that combines architectural innovation with a novel optimization algorithm.
Constraint-Embedding Neural Networks
Standard deep learning approaches for partial differential equations (PDEs) often use penalty terms to enforce constraints. However, this can lead to "soft" violations where the state slightly penetrates the obstacle. This method instead uses constraint-embedding neural networks. By designing the network architecture with specific mathematical transformations—such as ReLU-based shifts or squaring outputs—the model is physically incapable of producing an invalid state. This ensures that boundary conditions and obstacle constraints are strictly satisfied throughout the training process.
The S2-FOBA Algorithm
The primary contribution to efficiency is the Single-Loop Stochastic First-Order Bilevel Algorithm (S2-FOBA). Most bilevel solvers are "double-loop," meaning they are slow and resource-intensive. S2-FOBA utilizes a Moreau envelope-based reformulation to transform the bilevel problem into a single-level constrained problem. This allows the algorithm to update the control, the state, and auxiliary variables simultaneously using stochastic gradient descent. By eliminating the need for nested optimization, the method achieves significant speedups.
Two-Stage Refinement
The training process follows a two-stage strategy. After the initial bilevel training with S2-FOBA, a quick "Stage 2" refinement is performed. This step ensures the final state is perfectly consistent with the final control, further enhancing the physical accuracy of the solution.
Experimental Results
The framework was evaluated against several benchmark synthetic datasets designed to test edge cases in PDE research, including distributed control on square domains, irregular "flower-shaped" geometries, and problems involving non-symmetric operators.
Efficiency and Scalability
In head-to-head comparisons with the classical Active Set Method, the deep learning approach demonstrated superior scalability. For a high-resolution grid of $1024 \times 1024$, the classical method required approximately 178 seconds to converge. In contrast, the proposed method’s inference took only 0.044 seconds. Because the method is mesh-free, once the neural network is trained, it can be evaluated at any point in the domain nearly instantaneously.
Accuracy and Generalization
The method consistently achieved relative $L^2$ errors in the range of $10^{-2}$ to $10^{-3}$. This level of precision is considered highly satisfactory for nonsmooth, bilevel problems. Furthermore, the framework successfully handled irregular geometries where standard Physics-Informed Neural Networks (PINNs) often struggle, proving its robustness across different domain shapes and operator types.
Key Insights
- Hierarchy Preservation: Simply adding the upper and lower objectives together (objective combination) fails to capture the leader-follower relationship inherent in optimal control. This bilevel approach preserves that hierarchy, leading to more physically consistent results.
- Stochastic Advantage: By using mini-batch sampling to approximate integrals, the method avoids the "curse of dimensionality" and prevents the model from overfitting to specific grid points, a common issue in mesh-based methods.
Conclusion and Future Work
This work provides a scalable, mesh-free alternative to classical solvers for the optimal control of obstacle problems. While the results are promising, the authors note that the S2-FOBA algorithm requires careful hyperparameter tuning regarding step sizes and penalty sequences. Future research will focus on automating this tuning process and extending the framework to time-dependent obstacle problems and operator learning, where a single model could generalize across various obstacles and target states without retraining.
Summary updated