CASHOCS as Solver for Optimal Control Problems

Problem Formulation

As a model problem we again consider the one from Distributed Control of a Poisson Problem, but now we do not use cashocs to compute the adjoint equation and derivative of the (reduced) cost functional, but supply these terms ourselves. The optimization problem is given by

\[\begin{split}&\min\; J(y,u) = \frac{1}{2} \int_{\Omega} \left( y - y_d \right)^2 \text{ d}x + \frac{\alpha}{2} \int_{\Omega} u^2 \text{ d}x \\ &\text{ subject to } \quad \left\lbrace \quad \begin{alignedat}{2} -\Delta y &= u \quad &&\text{ in } \Omega,\\ y &= 0 \quad &&\text{ on } \Gamma. \end{alignedat} \right.\end{split}\]

(see, e.g., Tröltzsch, Optimal Control of Partial Differential Equations or Hinze, Pinnau, Ulbrich, and Ulbrich, Optimization with PDE constraints).

For the numerical solution of this problem we consider exactly the same setting as in Distributed Control of a Poisson Problem, and most of the code will be identical to this.

Implementation

The complete python code can be found in the file demo_control_solver.py, and the corresponding config can be found in config.ini.

Recapitulation of Distributed Control of a Poisson Problem

For using CASHOCS exclusively as solver, the procedure is very similar to regularly using it, with a few additions after defining the optimization problem. In particular, up to the initialization of the optimization problem, our code is exactly the same as in Distributed Control of a Poisson Problem, i.e., we use

from fenics import *
import cashocs


config = cashocs.create_config('config.ini')
mesh, subdomains, boundaries, dx, ds, dS = cashocs.regular_mesh(25)
V = FunctionSpace(mesh, 'CG', 1)

y = Function(V)
p = Function(V)
u = Function(V)

e = inner(grad(y), grad(p))*dx - u*p*dx

bcs = cashocs.create_dirichlet_bcs(V, Constant(0), boundaries, [1, 2, 3, 4])

y_d = Expression('sin(2*pi*x[0])*sin(2*pi*x[1])', degree=1)
alpha = 1e-6
J = Constant(0.5)*(y - y_d)*(y - y_d)*dx + Constant(0.5*alpha)*u*u*dx

ocp = cashocs.OptimalControlProblem(e, bcs, J, y, u, p, config)

Supplying Custom Adjoint Systems and Derivatives

When using cashocs as a solver, the user can specify their custom weak forms of the adjoint system and of the derivative of the reduced cost functional. For our optimization problem, the adjoint equation is given by

\[\begin{split}\begin{alignedat}{2} - \Delta p &= y - y_d \quad &&\text{ in } \Omega, \\ p &= 0 \quad &&\text{ on } \Gamma, \end{alignedat}\end{split}\]

and the derivative of the cost functional is then given by

\[dJ(u)[h] = \int_\Omega (\alpha u + p) h \text{ d}x\]

Note

For a detailed derivation and discussion of these objects we refer to Tröltzsch, Optimal Control of Partial Differential Equations or Hinze, Pinnau, Ulbrich, and Ulbrich, Optimization with PDE constraints.

To specify that cashocs should use these equations instead of the automatically computed ones, we have the following code. First, we specify the derivative of the reduced cost functional via

dJ = Constant(alpha)*u*TestFunction(V)*dx + p*TestFunction(V)*dx

Afterwards, the weak form for the adjoint system is given by

adjoint_form = inner(grad(p), grad(TestFunction(V)))*dx - (y - y_d)*TestFunction(V)*dx
adjoint_bcs = bcs

where we can “recycle” the homogeneous Dirichlet boundary conditions used for the state problem.

For both objects, one has to define them as a single UFL form for cashocs, as with the state system and cost functional. In particular, the adjoint weak form has to be in the form of a nonlinear variational problem, so that fenics.solve(adjoint_form == 0, p, adjoint_bcs) could be used to solve it. In particular, both forms have to include fenics.TestFunction objects from the control space and adjoint space, respectively, and must not contain fenics.TrialFunction objects.

These objects are then supplied to the OptimalControlProblem via

ocp.supply_custom_forms(dJ, adjoint_form, adjoint_bcs)

Note

One can also specify either the adjoint system or the derivative of the cost functional, using the methods supply_adjoint_forms or supply_derivatives. However, this is potentially dangerous, due to the following. The adjoint system is a linear system, and there is no fixed convention for the sign of the adjoint state. Hence, supplying, e.g., only the adjoint system, might not be compatible with the derivative of the cost functional which CASHOCS computes. In effect, the sign is specified by the choice of adding or subtracting the PDE constraint from the cost functional for the definition of a Lagrangian function, which is used to determine the adjoint system and derivative. CASHOCS internally uses the convention that the PDE constraint is added, so that, internally, it computes not the adjoint state \(p\) as defined by the equations given above, but \(-p\) instead. Hence, it is recommended to either specify all respective quantities with the supply_custom_forms method.

Finally, we can use the solve method to solve the problem with the line

ocp.solve()

as in Distributed Control of a Poisson Problem. The results are, of course, identical to Distributed Control of a Poisson Problem and look as follows

Note

In case we have multiple state equations as in Using Multiple Variables and PDEs, one has to supply ordered lists of adjoint equations and boundary conditions, analogously to the usual procedure for CASHOCS.

In the case of multiple control variables, the derivatives of the reduced cost functional w.r.t. each of these have to be specified, again using an ordered list.