Abstract
Parabolic optimal control problems with control constraints present significant challenges in both theoretical analysis and algorithmic design. This paper explores the application of the Alternating Direction Method of Multipliers (ADMM) to these problems, leveraging its ability to decouple control constraints from parabolic partial differential equation (PDE) constraints. We propose a novel, mesh-independent inexactness criterion for the internal iterations of the ADMM, allowing for efficient, low-to-medium accuracy solves of unconstrained subproblems while maintaining rigorous global convergence. Numerical results demonstrate that this approach significantly reduces computational overhead compared to traditional exact methods and state-of-the-art solvers like Semi-Smooth Newton (SSN).
Introduction
Optimal control problems (OCPs) constrained by PDEs are fundamental to modern engineering and scientific modeling, ranging from thermal management to financial forecasting. These problems require finding a control variable that minimizes a cost functional while satisfying the underlying physics (the PDE) and physical limitations (e.g., box constraints on the control).
The primary hurdle in solving these problems is the "curse of dimensionality." Upon discretization in space and time, OCPs yield massive algebraic systems. Traditional methods often struggle with these systems because they require high-precision solutions at every step, which becomes increasingly expensive as the mesh is refined and the system's condition number deteriorates.
Methodology: The Inexact ADMM Framework
Decoupling Constraints
The core strategy of this research is the application of the Alternating Direction Method of Multipliers (ADMM). ADMM is particularly effective here because it "unties" the complex PDE constraint from the simpler box constraints on the control. This allows the algorithm to iterate between two distinct steps:
- The Physics Step: Solving an unconstrained parabolic optimal control problem (the PDE constraint).
- The Constraint Step: Projecting the control variable onto the feasible set (the box constraints), which is a computationally trivial operation.
The Inexactness Criterion
Standard ADMM implementations typically require the "Physics Step" to be solved to high precision, which is computationally prohibitive for large-scale PDEs. The authors introduce a specific inexactness criterion: the internal PDE problem is solved only until the residual falls below a fraction ($\sigma$) of the residual from the previous iteration.
This criterion offers several advantages:
- Mesh Independence: The stopping rule is independent of the spatial ($h$) or temporal ($\tau$) discretization parameters.
- Computational Efficiency: By allowing "lazy" or low-accuracy solves during early iterations, the algorithm avoids unnecessary CPU cycles when the global solution is still far from convergence.
- Automatic Tuning: The criterion adjusts dynamically based on the algorithm's progress, removing the need for users to pre-set arbitrary accuracy constants.
Algorithmic Implementation
The internal solves are performed using the Conjugate Gradient (CG) method. Because the "Physics Step" always utilizes the same coefficient matrix, a high-quality preconditioner can be computed once and reused throughout the entire iterative process, providing a distinct advantage over Newton-type methods where the system matrix changes at every iteration.
Numerical Experiments and Results
The methodology was validated using several synthetic benchmarks where analytical solutions are known, allowing for precise error measurement.
- Parabolic OCPs: In a 2D square domain (Example 1), the inexact ADMM (Algorithm 3) demonstrated superior speed. For a mesh size of $2^{-8}$, the inexact version converged in approximately 3,840 seconds, whereas versions requiring high-precision internal solves took up to 37,100 seconds.
- Comparison with Semi-Smooth Newton (SSN): In elliptic PDE tests (Example 4), the ADMM-PCG approach outperformed the established SSN method as the mesh became finer. At a mesh size of $h=2^{-9}$, ADMM-PCG required only 32.5 seconds, compared to 54.7 seconds for SSN.
- Versatility: The framework was successfully extended to the Wave Equation (hyperbolic) and Elliptic equations, proving its robustness across different classes of linear PDEs.
Discussion and Limitations
The proposed inexact ADMM framework offers a robust and "software-friendly" approach to PDE-constrained optimization, showing stability across various regularization ($\alpha$) and penalty ($\beta$) parameters. However, certain limitations remain:
- Linearity: The current theory and implementation are tailored for linear PDEs. Nonlinear problems, such as those in fluid dynamics, present additional complexities.
- Memory Requirements: 3D time-dependent problems require storing state variables at every time step for adjoint calculations, which can challenge available RAM. Integration with memory-saving techniques is a necessary next step.
- Convexity: The framework assumes convex optimization; extending these convergence guarantees to non-convex control problems remains an open area of research.
Conclusion
By introducing a mesh-independent inexactness criterion, this work transforms the ADMM into a highly efficient tool for parabolic optimal control. The ability to solve internal subproblems to low accuracy without sacrificing global convergence provides a significant performance boost, making it a viable alternative to traditional high-precision solvers in large-scale scientific computing.
Summary updated