A0426
Title: A novel active learning criterion for experiments with multiple responses
Authors: Luca Pegoraro - University of Padova (Italy) [presenting]
Rosa Arboretti - University of Padova (Italy)
Elena Barzizza - Università degli studi di Padova (Italy)
Nicolo Biasetton - Università degli Studi di Padova (Italy)
Riccardo Ceccato - University of Padova (Italy)
Marta Disegna - University of Padova (Italy)
Luigi Salmaso - University of Padova (Italy)
Abstract: Active Learning (AL) is a branch of Machine Learning (ML) in which the learner is in charge of the choice of data from which to learn. Its concept interweaves with the ones of adaptive sampling and sequential design from the experimental design literature, as the objective is to sample those data points that maximize information in terms of an acquisition criterion. We present a novel criterion that can drive adaptive batch sequential acquisition when the objective is to maximize the predictive accuracy of the models globally. The novel criterion is based on a ranking procedure that ranks candidate observations with respect to the uncertainty of predictions, a quantification of feature importance and a clustering approach for grouping candidates with respect to their proximity in the design space. The method can be applied when multiple responses are investigated in the same physical experiment and the data is noisy. We show the effectiveness of the proposed procedure through a simulation study and a case study application.
