Boosting the Scalability of Farm-Level Models: Efficient Surrogate Modeling of Compositional Simulation Output

Abstract

Surrogate modeling can overcome computational and data-privacy constraints of micro-scale economic models and support their incorporation into large-scale simulations and interactive simulation experiments. We compare four data-driven methods to reproduce the aggregated crop area response simulated by farm-level modeling in response to price variation. We use the isometric log-ratio transformation to accommodate the compositional nature of the output and sequential sampling with stability analysis for efficient model selection. Extreme gradient boosting outperforms multivariate adaptive regressions splines, random forest regression, and classical multinomial-logistic regression and achieves high goodness-of-fit from moderately sized samples. Explicitly including ratio terms between price input variables considerably improved prediction, even for highly automatic machine learning methods that should in principle be able to detect such input variable interaction automatically. The presented methodology provides a solid basis for the use of surrogate modeling to support the incorporation of micro-scale models into large-scale integrated simulations and interactive simulation experiments with stakeholders.