Reunion28112019

Programme:

 

Titre: Resilient state estimation in Linear Time-Discrete Systems

Orateur: Alexandre Kircher (Laboratoire Ampère)

Résumé : Being able to estimate the inner hidden state of a system can be a key component to accurately control it or monitor its health. If gaussian disturbances are well handled by estimation frameworks such as Kalman filters, many other types of disturbances, like injections attacks from an external operator or sensor failures, still constitue a challenge. In this presentation, we will focus on how we can analyze the performances of an estimator in order to state its resilience, i.e. the induced estimation error is bounded with a bound independent from the disturbance values or its statistical properties, as well as on a few different resilient 

estimators. We will also talk about how they can be implemented in both offline and online cases.

Titre: Caractérisation garantie de régions de confiances en régime non-asymptotique : Application à la localisation de sources 

Orateur: Michel Kieffer (L2S)  

Résumé: En estimation paramétrique, les régions de confiance associées à une estimée permettent de mesurer la qualité de cette estimée. Les approches SPS (Sign perturbed sums) et LSCR (leave out sign correlated regions) proposée par Campi et al., permettent d'obtenir des régions de confiance exactes en régime non-asymptotiques (peu de mesures disponibles) et ne requièrent que de faibles hypothèses sur la distribution des bruits de mesure. Grâce à l'analyse par intervalles, nous montrerons qu'il est possible de caractériser ces régions de confiance de manière garantie. Ces approches seront illustrées sur un problème de localisation d'émetteurs dans un réseau de capteurs à partir de mesures de puissances reçues. 



Titre: Design of satellite maneuvers for parameter estimation

Orateur: Carlo Nainer (CRAN)

Résumé: Accurate mathematical spacecraft models are necessary for the validation of satellite control algorithms and for the design of effective feedforward actions. Model estimation improvements can originate from design of experiment techniques. A good design of the experiment maneuver can affect the quality of satellite parameter estimates. It is therefore an important step of the overall identification process. In this presentation, the satellite experiment design problem is considered. First, a cubic B-spline representation of the maneuvers is proposed, then an improved maneuver is obtained by optimizing a cost function based on the Fisher information matrix. The proposed method has been applied for the maneuver design of a “satellite” platform that will be tested in a zero-G flight. The results show the effectiveness of the proposed excitation profile.



Titre: Improved PAC-Bayesian Bounds for Linear Regression

Orateur: Vera Shalaeva ( Inria Lille-Nord), Alireza FAKHRIZADEH ESFAHANI (CRIStAL)

Résumé: In this paper, we improve the PAC-Bayesian error bound for linear regression derived in Germain et al. (2016). The improvements are two-fold. First, the proposed error bound is tighter, and converges to the generalization loss with a well-chosen temperature parameter. Second, the error bound also holds for training data that are not independently sampled. In particular, the error bound applies to certain time series generated by well-known classes of dynamical models, such as ARX models.