FR |

mardi 18 juillet 2017.

I am a professor of Machine Learning and Optimization at ISAE-SUPAERO (the French Aeronautics and Space Institute) in Toulouse, France.

My research lies in the fields of Reinforcement Learning and Sequential Decision problems, with a general interest for Machine Learning and Operations Research. I work within the "Decision Systems" research group of the Department of Complex Systems Engineering (DISC).
I am in charge of the "Data & Decision Sciences" curriculum and am implied more generally on (most) related topics within the cursus.

Research interests

Reinforcement Learning for autonomous machines.

The "Learning to fly" (L2F) project stems from the inspiration of birds and unpowered flying machines such as paragliders, or sail planes, in their ability to soar in the convective layers of the atmosphere with little sensor data. The project tries to formalize the different learning and planning processes implied for the automated control of such flying machines. An overall goal of the project is thus to provide an AI-based solution to improve the energy efficiency of fixed-wing aircraft.
I have also a keen interest in two other projets : the iBoat autonomous sailboat and the energy extraction from wind gusts.

Examples of questions addressed :
- Online, light-weight Q-learning of an optimal setpoint controler for a fixed wing UAV in an unknown, time-dependent, convective environment, in order to maximize its long term energy gain.
- MCTS planning for a fixed wing UAV in a partially observed and time-dependent convective environment.
- Active exploration of an unknown, time-varying environment.
- Deep Reinforcement Learning for the control of a non-linear, non-observable system with a hysteresis loop.
- MCTS planning for a sailboat router across the Atlantic ocean.
- Identification of wind models in the atmostpheric convective layers.

Students involved :
Sebastian Rapp (undergrad 2016-17)
Marc Melo Oliver (undergrad 2016-2017)
Ludovic Becq (MS student 2016-2017)
Tam Le Minh (MS student 2016-2017)
Naïl Eljaafari (MS student 2016-2017)
Houssam Akhmouch (MS student 2016-2017)
Etienne Herlault (MS student 2016-2017)
Sreekanth Ganapathi Raju (MS 2016-2017)
Erwan Lecarpentier (PhD student 2016-2019)
Romain Madelaine (undergrad 2017)

Links :
L2F project website
The iBoat project
Latest L2F-sim simulator

Recurrent combinatorial optimization problems (ML contributions to).

Some combinatorial optimization problems are recurrent in nature. That is, they require the resolution of several instances over time. For example, the unit commitment problem in short-term power generation planning needs to be solved every day given the forecasted demand. The main question investigated is can we learn from past solutions ? Can we guide the resolution of future instances and get faster and better results ?

Contributions :
- NaiBX, an multi-label classification method that easily scales-up to a large number of labels.
- Applications to Unit Commitment and TSP problems.
- Learning of a heuristic for the optimization of the French hydro-electrical production

Students involved :
Matthieu Gruson (undergrad 2012-2013)
Quentin Deroubaix (MS student 2016)
Franck Wang (undergrad 2015)
Guillaume Plessix (undergrad 2015-2016)
Jules Dubois (undergrad 2015-2016)
Hugo Braguier (undergrad 2015-2016)
Ouali Tahar Chaouche (undergrad 2015-2016)
Luca Mossina (MS then PhD student 2016-2019)

Links :
The NaiBX code and experiment
The NaiBRec code and experiment

Other interests.

Spatio-temporal validity of optimal actions in sequential decision making.
This has been a topic of interest for a while : given a stochastic control problem (possibly time-varying), if you know an action in a given state is part of an optimal plan, how far (in time and state space) can you move away from this state and still guarantee your action remains optimal ?

Links :
The Localized Policy Iteration paper’s experiment

Games, puzzles, funny combinatorial problems.
I have an interest in toying with AI solutions for such problems. For instance :
- Creation of an AI for the "Pandemia" board game (2012), for Arimaa (2013), for Hanabi (2017).
- Optimal swaping problem during nuclear refueling operations.

Links :
The Optimal Swapping… paper’s experiment
A Hanabi simulator

Teaching activities

Classes taught :
- Algorithms in Machine Learning - Supervized and unsupervized (graduate, 2015-)
- Algorithms in Machine Learning - Reinforcement Learning (graduate, 2015-)
- ISAE-SUPAERO Hackathon (graduate, 2016-)
- Combinatorial Optimization (graduate, 2011-)
- Differentiable Optimization (graduate, 2017-)
- Being an Optimization engineer (undergrad, 2012-2014)
- Reinforcement Learning (undergrad, 2012-2015)
- Continuous Optimization (undergrad, 2011-2015)


Head of the Intelligent Decision Systems final year minor (180 hours).

Complete re-design of the minor’s contents. Changed the focus to decision support disciplines : statistics, machine learning, optimization, AI, databases. Introduced student projects and conferences.
Growth : from 10 to 30+ students in two years.

Coordinator for the design of the Decision Sciences final year major (240 hours).

This major offers a common core on the scientific foundations in decision making (Statistics, Optimization, Decision Theory) and three distincts tracks in "Industrial Engineering", "Finance" and "Data and Decision Sciences".
Participation in the definition of the new global ISAE-SUPAERO training cursus. Animation and coordination of the Decision Science major’s design. Animation and coordination of the "Data and Decision Sciences" track’s design. Construction of classes and contents for the common core training and the "Data and Decision" track.

Head of the Decision Sciences final year major.

Academic team management (common core training), administrative tasks, curriculum evolutions (contents, teaching methods), networking, student advising.

Head of the Data and Decision Sciences program (within the Decision Sciences major)

Academic team management (almost all courses), administrative tasks, curriculum evolutions (contents, teaching methods), networking, student advising.

For an overview of the Data and Decision Sciences program, see this flyer :
PDF icon


2011-present Associate Professor, ISAE-SUPAERO (Toulouse, France).
2011 Research engineer, EDF R&D (Clamart, France)
2010 Post-doctoral researcher, University of Liège (Belgium)
2009 Invited researcher, EDF R&D (Clamart, France)
2009 Post-doctoral researcher, Technical University of Crete (Chania, Greece)
2005-2008 PhD candidate, ONERA Toulouse (France)

2008 PhD in Artificial Intelligence - University of Toulouse (France)
2005 MS in Control Theory - University of Toulouse (France)
2005 Generalist and Aeronautics Engineer – SUPAERO (Toulouse, France)

Paragliding instructor (and passionate pilot).

PDF icon

A few extra things

A few extra / funny / important things to know.
Most of them are the result of my adaptation to the unstructured world of academia and my personal experience.

e-mail solving

I receive more than a hundred e-mails each day (spam and mailing-lists excluded) and try to devote only around 1h to e-mail solving(*).
If you did not receive an answer from me (and expected one) :
- Resend your e-mail, it probably went away with the flow without me noticing.
- Reconsider.
Can it wait ? Is it so urgent that you can’t come see me or get an answer by phone ?
Can you solve your problem without me ? I’m honored that you thought I could help, sometimes I just don’t have the answer, or it is not within my prerogatives or duties, or I just don’t have the resources to help.
- Come to my office.
If your need is _really_ urgent and can only be solved by me I’ll do my best to help you. Even if I can’t help you’ll get an answer.

(*) e-mail solving : series of actions engaged to solve a problem submitted to me by e-mail, often ending with replying to said e-mail and triggering the reception of a new request.

Happy together (for students)

You want to work with me ? Great ! Here’s what you should expect.

  • be prepared to be :
    - autonomous : my schedule is often busy and I try to focus on where I’m most useful so you should sort out the rest by yourself as much as possible.
    - creative : suggest your own solutions to the problem we work on, imagine new ways of solving it, generalize from the current problem to a more formal statement.
    - problem-solving oriented : the goal is to solve a problem, not just think about a good way of doing it.
  • in return you get :
    - ideas, leads, constructive criticism from me
    - contacts with scientific experts and great people
    - lots of fun
  • get started in my topics of interest :
    - learn about Reinforcement Learning, AI planning, ML and related topics.
    Usual references : "RL an introduction" by Sutton and Barto, "Algorithms for RL" by Szepesvari, the bandits and MCTS survey by Munos, the Policy Search survey by Deisenroth et al., etc.
    - read about my current projects and interests (on my webpage for example)
  • Once all this is clear :
    Come in, let’s discuss your ideas and your interests. If I’m not the right person for you I’ll redirect you to more adequate people.
  • I look for solutions, not culprits.
  • When you ask a yes/no question, be prepared for both answers (that applies to me too).
SPIP | | Suivre la vie du site RSS 2.1