TY - JOUR
T1 - Tree-based ensemble methods for sensitivity analysis of environmental models
T2 - A performance comparison with Sobol and Morris techniques
AU - Jaxa-Rozen, Marc
AU - Kwakkel, Jan
PY - 2018
Y1 - 2018
N2 - Complex environmental models typically require global sensitivity analysis (GSA) to account for non-linearities and parametric interactions. However, variance-based GSA is highly computationally expensive. While different screening methods can estimate GSA results, these techniques typically impose restrictions on sampling methods and input types. As an alternative, this work evaluates two decision tree-based methods to approximate GSA results: random forests, and Extra-Trees. These techniques are applicable with common sampling methods, and continuous or categorical inputs. The tree-based methods are compared to reference Sobol GSA and Morris screening techniques, for three cases: an Ishigami-Homma function, a H1N1 pandemic model, and the CDICE integrated assessment model. The Extra-Trees algorithm performs favorably compared to Morris elementary effects, accurately approximating the relative importance of Sobol total effect indices. Furthermore, Extra-Trees can estimate variable interaction importances using a pairwise permutation measure. As such, this approach could offer a user-friendly option for screening in models with inputs of mixed types.
AB - Complex environmental models typically require global sensitivity analysis (GSA) to account for non-linearities and parametric interactions. However, variance-based GSA is highly computationally expensive. While different screening methods can estimate GSA results, these techniques typically impose restrictions on sampling methods and input types. As an alternative, this work evaluates two decision tree-based methods to approximate GSA results: random forests, and Extra-Trees. These techniques are applicable with common sampling methods, and continuous or categorical inputs. The tree-based methods are compared to reference Sobol GSA and Morris screening techniques, for three cases: an Ishigami-Homma function, a H1N1 pandemic model, and the CDICE integrated assessment model. The Extra-Trees algorithm performs favorably compared to Morris elementary effects, accurately approximating the relative importance of Sobol total effect indices. Furthermore, Extra-Trees can estimate variable interaction importances using a pairwise permutation measure. As such, this approach could offer a user-friendly option for screening in models with inputs of mixed types.
KW - Decision tree methods
KW - Ensemble learning methods
KW - Factor screening
KW - Global sensitivity analysis
UR - http://resolver.tudelft.nl/uuid:649d7279-157d-4a92-81b9-aa47b16e36cb
UR - http://www.scopus.com/inward/record.url?scp=85048952644&partnerID=8YFLogxK
U2 - 10.1016/j.envsoft.2018.06.011
DO - 10.1016/j.envsoft.2018.06.011
M3 - Article
AN - SCOPUS:85048952644
SN - 1364-8152
VL - 107
SP - 245
EP - 266
JO - Environmental Modelling and Software
JF - Environmental Modelling and Software
ER -