Measurements can be used in an optimization framework to compensate the effects
of uncertainty in the form of model mismatch or process disturbances. The main
bottleneck when trying to reformulate a dynamic optimization problem as a
control problem is the definition of the inputs, of the outputs and of the
corresponding setpoints. Among the various options for input update, a promising
approach consists of directly enforcing the Necessary Conditions of Optimality
(NCO) (that include two parts, the active constraints and the sensitivities), by
manipulating the appropriate parameters of the input profile.

In this talk, after a general presentation of the methodology, it will be shown
that the variations of the NCO due to parametric uncertainty can be used to: (i)
quantify the negative impact on uncertainty on the NCO, (ii) separate input
parameters to properly control the two parts of the NCO, and (iii) design
appropriate update laws for the inputs. To implement these control laws, two
control algorithms will be proposed, for which global convergence can be
established for a class of nonlinear systems. Finally, the theoretical concepts
will be illustrated through the run-to-run optimization of an industrial batch
inverse-emulsion copolymerization reactor.