Surrogate assisted feature computation for continuous problems
The first method that enable the computation of problems features in a setting that is compatible with the
operation of search heuristics. Allows to limit the undersampling error of biased estimators, using a surrogate model.
A crucial tool to embed dynamic parameter setting within search heuristics.
Adaptive Learning Search, a New Tool to Help Comprehending Metaheuristics
An attempt at forcing metaheuristics out of metaphors. This is also the first occurrence of the idea that
simulated annealing can be viewed as a population algorithm that samples the objective function directly.
Still relevant, as hybrid sampling approaches are believed to exhibit better performances nowadays.
Pareto-Based Multiobjective AI Planning (2013).
One of the very first Pareto-optimal approach to automated AI planning. This work extends the D A E solver and
proposes a simple multi-objective benchmark with proven optimal solutions. It also outperforms the only
known metric-sensitive solver, competing on a —albeit simpler— aggregated objective problems.
Per instance algorithm configuration of CMA-ES with limited budget
The method that won the 2017 Black Box Optimization Competition in the single objective track. The seminal
work on learning parameters-features mapping that can be embedded within solvers. Classical CMA-ES
versions were leading the competition since years and this work enabled a performances breakthrough.
Divide-and-Evolve: the Marriage of Descartes and Darwin
This solver won the 2011 AI planning competition in the temporal track. It was the first time a stochastic
metaheuristic won the IPC and it was on its hardest problems. It's currently used in command and control
prototypes. This work with ONERA and INRIA also led to the solver that won the following IPC.
Operating Room Planning with Random Surgery Times
My first glance at automated planning. This work is an interesting combination of a classical Operations
Research approach with ideas that originates from the field of stochastic metaheuristics (most notably a
heuristic to solve the pricing sub-problem of a column generation approach).
Robust rigid registration of retinal angiograms through optimization
The application of my thesis. Back then, it was uncommon to use search heuristics in this field and high
resolutions images could not be handled because of inefficient multi-resolution algorithms. This work
showed that global optimization can outperform the windowed approaches generally used in imaging.