Optimal control from theory to computer programs (Q1417206): Difference between revisions

This text provides a comprehensive introduction to several techniques for the numerical computation of control problems governed by (mainly elliptic) partial differential equations (PDE). The book's approach is to focus on classes of problems and various ways of discretization of the continuous variables involved. The solution of the control problem may then be calculated as a non-linear programming problem or via the solution of a variational inequality. The book develops convergence theory in some detail and then manages the transition from theory to algorithm and finally computer pseudo--code. The pseudo-code is clear and well documented but the final step of converting this into the code of a programming language is left to the reader (no CD is included). One of the satisfying aspects of the work is the detail given to methods of discetization of problems and the development of formulae which may be directly used in computer code. This book is most suited for graduate students in applied mathematics, numerical analysis or control engineering. Some of the earlier chapters could be used as part of an honours level course for mathematics students. The book's content is briefly as follows: The first chapter gives preliminary work on convex functions and functionals and their derivatives. The second chapter outlines some basic convex optimization theory and surveys various descent methods such as steepest descent with modified line searches, the conjugate gradient method and projective gradient methods. Chapter three studies elliptic variational inequalities (EVI) and convex optimization problems. The approximation of EVI via the Galerkin approximation and the finite element method (FEM) enables approximate solution to be computed. Penalty, Lagrange multiplier methods and success approximation is also studied. In Chapter four indirect methods are studied in which the state variables are eliminated and a convex optimization problem in the control variable is obtained. Descent methods applied to discretized problems allow approximate solution to be calculated. Here one and two phase Stefan problems are studied. Chapters five and six apply the methods developed in the preceding chapters to an epidemic model, simply supported clamped plate problems and structural optimization. Chapter seven studies control problems with quadratic objectives. FEM are used to obtain large scale but finite dimensional quadratic programming problems (QP) which are then solved using interior point methods. This approach leads to a large, indefinite and ill-conditioned system to be solved as a subproblem. Methods for this are developed and a code is presented. Finally, in Chapter eight control problems involving random variables are studied. In the main these reduce to the Bellman equations (in the discrete case) or the Hamilton-Jacobi-Bellman equations in the continuous time case. Here a finite difference method leads to the construction of approximating Markov chains. The resultant discrete equation allows the calculation of transition probabilities and interpolation intervals. Purely probabilistic methods are required as the underlying PDE may not be smooth enough for the FEM to be applicable.