EVALUATION OF GRADIENTS FOR PIECEWISE CONSTANT OPTIMAL-CONTROL

To solve nonlinear optimal control problems, the problems are here transformed into finite dimensional optimization problems, which can then be solved using nonlinear programming (NLP). Since for the rapidly converging NLP methods it is necessary to supply the gradients of the performance index and of the problem constraints, three approaches to obtain the gradients are examined. By using two nonlinear optimal control problems which are typical chemical engineering systems, it is shown that in computers that allow vectorization, the trajectory sensitivity and finite-difference approaches can be computationally more efficient than the more commonly used adjoint system approach.