Batch-sequential design and heteroskedastic surrogate modeling for delta smelt conservation

Delta smelt is an endangered fish species in the San Francisco estuary that have shown overall population decline over past 30 years. Researchers developed a stochastic, agent-based simulator to virtualize system with goal of understanding relative contribution natural and anthropogenic factors might play role their decline. However, input configuration space high dimensional, running time-consuming, its noisy outputs change nonlinearly both mean variance. Getting enough runs effectively learn input–output dynamics requires nimble modeling strategy parallel evaluation. Recent advances heteroskedastic Gaussian process (HetGP) surrogate helps, but little known about how appropriately plan experiments for highly distributed simulation. We propose batch sequential design scheme, generalizing one-at-a-time variance-based active learning HetGP, as means keeping multicore cluster nodes fully engaged runs. Our acquisition carefully engineered favor selection replicates which boost statistical computational efficiency when training surrogates isolate signal from noise. Design are illustrated on range toy examples before embarking large-scale simulation campaign downstream high-fidelity sensitivity analysis.