Abstract:

Complex numerical models, such as ocean models, reliability models of passive industrial components, climate models, etc. traditionally mix multi-physical aspects and a large number of approximations related to unknowns mathematics (e.g. approximate numerical diagrams) and computer aspects (e.g. errors rounding, completion of algorithms in finite time). It is wanted to obtain a certain readability, a separability and characterization of epistemic uncertainties (related to lack of knowledge of the reality represented) and random (related to the intrinsic behaviour of the quantities that feed the model). Arguments from algorithmic complexity and Cox-Jaynes-derived probabilistic logic make it possible to answer to the question of their representation and propagation by stochastic tools. Depending on time, the talk will attempt to provide some answers on the choice of probabilistic measures to be used in uncertainty propagation problems, highlight some typical issues and answers encountered by probabilistic sampling in such problems and will propose avenues for research.

Poster (Download here)