Eric Fabre Senior Research Scientist SUMO team Address       IRISA  Campus de Beaulieu  F-35042 Rennes Cedex   France  Tel  +33 2 99 84 73 26  Fax  +33 2 99 84 71 71 Email  eric.fabre (at) inria.fr Team assistant  Laurence Dinh,   ldinh(at)inria.fr   +33 2 99 84 74 84

Research Interests

• large scale stochastic systems : learning, approximation, verification, control
• machine learning techniques for dynamic systems
  
• distributed management of distributed systems (observation, control, diagnosis, optimization, planning, ... ) 
• distributed discrete events systems : Petri nets, concurrency, unfolding techniques
• Bayesian networks, graphical models of interactions, Markov fields, iterative/turbo algorithms
• information theory and its distributed (multi-user) aspects
  
• Application domains :
  
• autonomic management of telecommunication networks, diagnosis of software defined networks
• distributed regulation of trains in subway lines
• bio informatics : modeling and verification of the dynamics of cell populations
  

Open positions !

Master : Machine learning approaches to learn automata from their traces  (in french)

PhD : several PhD topics to open soon (contact us)

- robustness verification of large neural networks
- verification methods for reinforcement learning
- learning dynamic systems from their traces

   
Some recent (closed) topics, as an indication of our research interests       

•     PhD :    Approximation methods for the soundness of control laws derived by machine learning (on going)
  
  
•     PhD :    Fault management in software defined networks - CIFRE thesis, in collaboration with Alcatel-Lucent (Bell Labs France)
•     PhD :    Monitoring and diagnosis of large scale systems
•     PhD :    Verification of large scale stochastic systems; application to diagnosis and opacity 
•     PhD :    Distributed self-management methods: optimization and planning for resource allocation in telecommunication networks   
•                  This topic was in cooperation with Alcatel-Lucent Bell Labs
•     PhD :    Distributed Optimal Planning for Large Distributed Systems

•     Master : Estimation of the opacity degree of a system holding a secret   (version française)
•     Master : Diagnosis of large systems by progressive model refinement    (version française)
  
•     Master : Robust supervision and diagnosis of discrete event systems    (version française) 
  
•     Master : Optimal control of a complete urban train system   (version française)

Good candidates are always welcome !
Ideal profile : master in EE or CS, with a taste for formal methods.
Please send a resume (CV) and a motivation letter.

Teaching

• TI:  Information Theory and Coding,
  Ecole Normale Superieure de Cachan (Ker Lann), 2nd year (= Master 1 level)
  Télécharger les  notes de cours  (poly inachevé, en Français)
  Here is also a conference given to new students at ENS Cachan, about classical and more recent aspects of information theory.
  
  
• MAD : Models and Algorithms for Distributed systems
  Master in Computer Science, 2nd year.
  
  
• OPT: Introduction to Numerical and Combinatorial Optimization,
  Master in Computer Science, 2nd year
  No longer taught, but the material below could be of interest.
  Download the Lecture Notes, and the slide sets of  lectures 1-2   and  lectures 3-4-5     
  
  

Projects

Publications

                  You may try this list, automatically generated from our team data base.

Patents

• Method and equipment for adjusting power amplification in an optical network - European patents EP 09290408 and EP 09290409 (also US Patents) -  June 2009
  
• Detection and optimal correction of routing loops in a network using OSPF-like routing protocols - European patent EP 12305306.8 - March 2012
  

Demos : urban train regulation

• objective here is to balance headways (time distance to previous train)
• on the outer ring, each colored dot is a train
• train speed remains constant on each track segment, and changes at junctions
• train speeds suffer random perturbations
• regulation is performed through a distributed control algorithm, with local information (positions and speeds of previous and next trains)
• spikes from the center are proportional to the headway
• the curves in the middle represent train speeds, compared to their target value for each segment
• this algorithm prevents the bunching phenomenon
  

• two branches meeting at a junction
• bottom = train positions
                 trains arrive randomly from the two branches to the left, and compete to access the branch to the right
  
• top = train speeds
           trains have constant speed on each track segment, but speed requirements suffer random perturbations
• there should be at most one train in the green zone (junction) to leave time for the needle to switch track
• headways in all 3 branches should be balanced, and trains correctly interleaved
• regulation is performed through an optimal planning algorithm
• and, yes, trains tend to speed up in the critical zone... (still to be fixed)