An active learning strategy for sequential designs of probabilistic systems with continuous and discrete input variables
Ravindu Udayantha Jayasekara, Sabarethinam Kameshwar · Engineering Optimization · 2025
This study proposes an active learning strategy for sequential experimental designs of probabilistic systems with mixed continuous and discrete input variables. The proposed method aims to sequentially explore local optima with high aleatory uncertainties along with the global optimum of the system. In complex probabilistic systems, the exploration of both of these regions is needed to build a more robust understanding across different conditions or scenarios. The proposed strategy employs a non-deterministic kriging (NDK) method, capable of handling mixed input variables, to account for aleatory uncertainties. A modified version of the expected improvement concept is used to identify the regions of interest for the sequential experiments. To avoid neglecting local optima with high aleatory uncertainties, subset simulations are used to identify these regions. The methodology was applied to a set of analytical examples and yielded accurate predictions for the optimum values and associated aleatory uncertainties.