My research aims at the study of **Adaptive algorithms**: overall,
my main concern over the last 15 years (and more) has been about the
study, design, implementation, ... of algorithms that adapt to their
time varying environment. The algorithm may be either dedicated to
control an agent, or predict some quantity, or whatever. In this
endeavor, I shifted from genetic algorithms to the field of machine
learning, and particularly the sub-field of reinforcement learning,
also known as approximate dynamic programming.

My mid-term objective is to study adaptive algorithms in time-varying
environments. There is a paradox in current main stream research in
reinforcement learning for instance: RL is fundamentally dedicated to
adatation to changing environements; however, the very large majority
of papers deals with non varying environments! This is clearly a bias
towards what we know how to study, rather than studying what really
matters.

A closely related theme is that of non asymptotic behavior of
algorithms; in a time varying environment, we are interested in the
adaptivity of an algorithm to its current environement, and before the
environment changes too much; so, asymptotically optimal algorithms
are really not what we care of in changing environments.

To reach this goal, an algorithm has to represent its environment and
I think that this representation should also adapt to the changes of
the environment. This led me to get interested in non parametric
function approximators, such as neural networks which architecture
changes over time, and kernel machines; these functions approximators
let us discover useful feature in a very natural way; thus, I got a
strong interest in sparse representation, l1 regularization, LARS,
kernelized-LARS, compressed sensing, ...

As I am a computer scientist, I am really interested in working
algorithms, effective, and efficient. Theorems are all nice, but an
algorithm that really does something useful, in a reasonnable amount
of time, is much better to my eyes. And if the algorithm rests on
sound principles, and a sound theory, and works efficiently in
practice, and is able to scale up with the amount of data, it is a
beautiful algorithm; there are not so many...

I have a strong interest towards web application, such as search
engine, recommendation systems, and computational advertizing. These
problems may easily be formulated as sequential decision making
problems.

Ok, now you have to shift to french, sorry about that(-(: