Pause
Lecture
Moteur de recherche d'offres d'emploi CEA

Learning Hybrid Automata Models H/F


Détail de l'offre

Informations générales

Entité de rattachement

Le Commissariat à l'énergie atomique et aux énergies alternatives (CEA) est un organisme public de recherche.

Acteur majeur de la recherche, du développement et de l'innovation, le CEA intervient dans le cadre de ses quatre missions :
. la défense et la sécurité
. l'énergie nucléaire (fission et fusion)
. la recherche technologique pour l'industrie
. la recherche fondamentale (sciences de la matière et sciences de la vie).

Avec ses 16000 salariés -techniciens, ingénieurs, chercheurs, et personnel en soutien à la recherche- le CEA participe à de nombreux projets de collaboration aux côtés de ses partenaires académiques et industriels.  

Référence

2021-18454  

Description de l'unité

The host unit in CEA will be the LECS team whose main mission is the development of methods and tools for the engineering of system requirements and compliance.

Description du poste

Domaine

Mathématiques, information  scientifique, logiciel

Contrat

Stage

Intitulé de l'offre

Learning Hybrid Automata Models H/F

Sujet de stage

The objective of the internship is to develop algorithms for learning models of hybrid automata

Durée du contrat (en mois)

6

Description de l'offre

Models play an important role in Cyber-Physical Systems (CPS). They constitute a primary vector of information for design engineers and can help them identify design issues very early. Moreover, models constitute a reference for formal analyses such as proofs, model-checking, test case generation, etc. Building efficient models capable of representing systems whose behavior results from continuous and discrete processes is challenging, in particular when adequate abstractions are not systematically applicable. The literature [1, 2, 3] shows great interest in learning hybrid models from execution traces in the form of time series. In this work, we are interested in applying and adapting machine learning techniques to the context of learning hybrid dynamical models.

We will consider a hybrid system as a black box and wish to address the task of learning an accurate model by:

  • Using machine learning methods such as [4] combined with black box testing and allowing to acquire faithful models while ensuring coverage of the system behaviors
  • Simplifying, optimizing learned models using genetic algorithms based on fitness functions

 The objective will be to learn automata (simple, timed or hybrid) that are abstractions of observable behaviors of the black box system. Current techniques of the literature consist of learning models while minimizing the estimation error obtained by computing the distance between training data traces and the model traces.

 

The candidate will be asked to perform the following tasks:

  • Study the existing methods for learning hybrid automata models and identify weak and strong hypotheses in these techniques
  • Elaborate and apply learning methods of hybrid models that could leverage some of the identified weaknesses
  • Implement necessary code to justify empirically the benefits of the proposed approach

 

[1] Oded Maler, Grégory Batt, Approximating Continuous Systems by Timed Automata, FMSB 2008.

[2] Daniel L. Ly and Hod Lipson. 2012. Learning symbolic representations of hybrid dynamical systems. J. Mach. Learn. Res. 13, 1 (January 2012), 3585–3618. 

[3] Santana, P. et al. “Learning Hybrid Models with Guarded Transitions.” AAAI (2015).

[4] Chen et al. “Neural Ordinary Differential Equations”

Profil du candidat

The ideal candidate will be pursuing a masters or engineering degree in computer science with a focus on machine learning and have good knowledge in automata, differential equations and unsupervised learning methods (Expectation-Maximization, Clustering algorithms and so on) and great programming skills.

Additional knowledge regarding formal methods, symbolic execution is desired.

Localisation du poste

Site

Saclay

Localisation du poste

France

Ville

Palaiseau (site de Nano Innov)