Identification of Physical Parameters Using Change-Point Kernels

Ankit Chiplunkar (1, 2) , Emmanuel Rachelson (3) , Michele Colombo (1) , Joseph Morlier (2)

1 AIRBUS Operations Ltd.
2 ICA - Institut Clément Ader
3 ISAE-SUPAERO - Institut Supérieur de l'Aéronautique et de l'Espace

Engineering design is a costly exercise, primarily because gathering data for various design cases requires constructing and experimenting on that design point. Hence engineers learn basic principles of the system and construct more detailed models based on the initial principles and simplifying assumptions. Eg. FEM in structural design or CFD in fluid simulations. Gaussian Process is a flexible, non-linear, prior over functions. It enables to tractably compute the posterior distribution which is consistent with prior belief and observed data. The prior can be easily manipulated to encode a hypothesis space or family of functions. The algorithm tries to put a changepoint at every observation point. Osborne et al perform changepoint estimation by placing a prior over them and performing a MAP estimate. We would like use this method in the future for changepoint estimation.

Mots clés: Gaussian Process Regression; Change point

Domaines: Machine Learning [stat.ML]; Mécanique statistique [cond-mat.stat-mech]; Mécanique [physics.med-ph]

Ankit Chiplunkar, Emmanuel Rachelson, Michele Colombo, Joseph Morlier. Identification of Physical Parameters Using Change-Point Kernels. Society for Industrial and Applied Mathematics, Uncertainty Quantification, 2016, Apr 2016, Lausanne, Switzerland. 2016. ⟨hal-01555401⟩