Douglas Wiens: Topics in Robustness of Design

Abstract: I will discuss techniques of robustness of design, for experiments meant to give reliable predictions and inferences in the face of model discrepancies. Particular attention will be paid to (i) robustness against a misspecified linear response function; and (ii) robustness against violations of the assumed independence of the observations. The solutions to these problems involve a blend of mathematical and numerical optimization methods, which will be outlined. The methods are extended in a number of directions which will be discussed: (i) sampling schemes for a stochastic process whose autocorrelation function is possibly misspecified; (ii) active learning, as a further extension of the techniques from experimental design to sampling design; and (iii) discrimination problems in nonlinear regression, in which the competing models are only partially and perhaps incorrectly specified.