Uncertainty in optimal experiment design: comparing an online versus offline approaches

Abstract Model-based experiment design for parameter estimation is aimed at obtaining accurate parameter estimates with minimal variance. However, these experiment designs critically depend on the current best known parameter value. As the current best known values can differ from the true process, there can be a loss in information and this can lead to unwanted process behavior. In this paper we focus on the latter goal as we want to avoid constraint violations as much as possible. We review two offline approaches, namely a linearization and unscented transformation approach and we highlight the potential of the online receding horizon to avoid constraint violations. We illustrate these techniques on a benchmark bioreactor case study. For this case study, the online approach has a better potential for avoiding constraint violations even in view of parameter model/plant mismatches up to 50%.