Optimal sensitivity analysis under constraints: Application to fisheries
Résumé
We propose a new optimal (in some sense) approach to conduct a global sensitivity analysis (GSA) of numerical model outputs relatively to the model inputs when three specific constraints exist. The three constraints considered here are: i) A computation time too long to perform all simulations required when using the usual methods of GSA, typically those based on LHS , Sobol sequences,...; ii) The inputs are not independent because exist some structural correlations between them (or part of them), or functional relationships between them, or bounded combinations of inputs; iii) The presence of qualitative (categorical) inputs in addition to quantita- tive (continuous) inputs. The two main innovative aspects of the proposed approach are based upon the construction of a small size D optimal simulation design and the use of the PLS regression to manage the qualitative inputs and the correlated quantitative inputs. This new approach leads to compute specific Sensitivity Indexes, called SI-VIP, different from those of Sobol’ This new approach is applied here on a fisheries management problem using the ISIS-Fish modelling framework of the French Research Institute for Exploitation of the Sea.