Interactive design of experiments: A priori global versus sequential optimization, revised under changing states of knowledge

Abstract

Predicting hydro(geo)logical or environmental systems is subject to high levels of uncertainties, especially if appropriate data for model calibration are lacking. For subsurface systems, where data acquisition is cost intensive and time demanding, it is especially important to collect only those data that provide the largest amount of relevant information. The high expenses call for optimal experimental design, which is widely recognized for maximizing the efficiency of experiments. In model-based design of experiments, the analysis of the design efficiency and the resulting optimal design are based on the initial state of knowledge about the modeled system. Joint optimization of multimeasurement designs is a well-known challenge, and the usefulness of global optimization approaches is widely recognized in this context. However, we will show that the benefit for such global optimization becomes questionable when measurement data become available sequentially. Instead, the optimization effort should be invested within an interactive design approach. Today's fast telecommunication, global connectivity, and high-performance computing allow to consider such interactive coupling. This study will use a synthetic case study to compare the standard en-bloc global optimization approach to two interactive design approaches. The approaches are implemented in a Bayesian framework and are compared based on their complexity and overall performance. The key conclusion confirms a previously untested presumption: for models that trigger nonlinear parameter inference problems, interaction (which may come at a loss of global optimization) is more beneficial than global optimization based on the initial state of knowledge (which typically implies the impossibility of interactivity).