Project 03: Box-Constrained Mixed Boundary Control
Goal¶
Add pointwise bounds to the mixed Dirichlet-Neumann boundary-control problem
of Project 02.
This project is the natural boundary-control analogue of the projected
gradient and PDAS laboratories.
Use the state equation from Project 02 and impose
uDa≤uD≤uDba.e. on ΓD, and
uNa≤uN≤uNba.e. on ΓN. The admissible set is
Uad={(uD,uN):uDa≤uD≤uDb,uNa≤uN≤uNb}. The cost functional is
J(y,uD,uN)=21∥y−yd∥L2(Ω)2+2αD∥uD−uD,ref∥2+2αN∥uN−uN,ref∥2. The variational inequality is
⟨gD,uD−uD⟩ΓD+(gN,uN−uN)L2(ΓN)≥0∀(uD,uN)∈Uad, where formally
gD=αD(uD−uD,ref)−∂np,gN=αN(uN−uN,ref)−p∣ΓN. For an L2 boundary discretization, a projected-gradient update is
uDk+1=P[uDa,uDb](uDk−sDgDk), uNk+1=P[uNa,uNb](uNk−sNgNk). Implementation Hints¶
Implement projected gradient first.
Add a projected-gradient residual for each boundary component.
As an extension, implement PDAS by freezing active boundary control DoFs.
Track lower-active, upper-active, and inactive boundary regions.
Relevant Examples¶
Course code: codes/dealii/execs/laplacian_box_constraints.cc,
codes/dealii/execs/kkt_box_constraints.cc.
Course lectures: lecture09.md, lecture11.md, lecture14.md.
deal.II: AffineConstraints<double> examples in many tutorials; step-7
for boundary terms.
Deliverables¶
Projected-gradient solver with Armijo backtracking.
Active-set visualization on ΓD and ΓN.
Numerical comparison with the unconstrained Project 02 solution.
A short discussion of how boundary control spaces affect projection.