RESEARCH · RESEARCH · #2190
New paper introduces random-forest–supervised HSIC-ANOVA for hierarchical decision sensitivity in MAELIA
The arXiv preprint (arXiv:2610.22420v1) proposes treating decision-tree structured inputs in agent-based agro-ecological simulations as first-class objects and introduces a random-forest–supervised HSIC-ANOVA method to perform sensitivity analysis on hierarchical categorical and numerical design spaces; the authors demonstrate the approach on the MAELIA platform for outcomes such as nitrogen leaching, soil organic carbon change, and yields. The method uses surrogate modelling to rank main effects and interactions while handling parameters that are only conditionally active by higher-level decisions.
KEY POINTS
- The arXiv preprint (arXiv:2610.22420v1) proposes treating decision-tree structured inputs in agent-based agro-ecological simulations as first-class objects and introduces a random-forest–supervised HSIC-ANOVA method to perform sensitivity analysis on hierarchical categorical and numerical design spaces; the authors demonstrate the approach on the MAELIA platform for outcomes such as nitrogen leaching, soil organic carbon change, and yields.
- The method uses surrogate modelling to rank main effects and interactions while handling parameters that are only conditionally active by higher-level decisions.
- It matters because classical global sensitivity methods assume fixed, independent inputs and are biased on branched decision spaces, while this method enables interpretable, decision-oriented sensitivity analysis for hierarchical agent-based models.
WHY IT MATTERS
It matters because classical global sensitivity methods assume fixed, independent inputs and are biased on branched decision spaces, while this method enables interpretable, decision-oriented sensitivity analysis for hierarchical agent-based models.