Missions

The successful candidate will conduct research in the context and towards the objectives of the ERC dynaBBO project. He/she/they will be required to interact and collaborate with team members as well as other partners involved in the project. The successful candidate will be required to produce deliverables and monitoring reports as part of project management.

Activités

When faced with an optimization problem, we often lack time, knowledge, or other resources to
develop a dedicated approach to solve it. In such situations, it is convenient to resort to black-box
optimization algorithms, approaches designed to provide high-quality solutions without requiring
manual adjustments nor expert knowledge about the problem. Given their ease of use, black-box
optimization algorithms are among the most widely applied optimization techniques, deployed to
solve numerous problems across broad ranges of industrial branches and academic disciplines every
day.

A plethora of different black-box optimization strategies exist, complementing each other in
strengths and weaknesses for different problem types and for different stages of the optimization
process. While this complementarity is widely acknowledged, we lack efficient approaches to
leverage it, resulting in sub-optimal solutions that cause an ineffective use of our limited resources.
With the dynaBBO project, we set out to fill this important gap. Relying on a hybrid approach
synergizing knowledge about black-box optimization algorithms with automated machine learning
techniques, we obtain an efficient system, capable of dynamically switching between different
black-box optimization algorithms “on the fly”.

The three main research questions that guide our project are which algorithm to select for the
initial phase, when to switch from one algorithm to another, and how to warm-start the selected
solver so that it can continue the search for high-quality solutions as effectively as possible. The
key novelty of our approach lies in (1) a revised modeling of the algorithms, better suited to control
their behavior, (2) the ability to switch between algorithms of fundamentally different types, and
in (3) an adaptive choice of the moment(s) when to switch.

As we have demonstrated in a series of recent works [GECCO 2019, GECCO 2022, FOGA 2023, GECCO 2025], theory-guided benchmarks can support the development of automated dynamic control poicies by providing benchmarks with proven ground truth. However, the variety of examples for which rigorously proven control policies are know is fairly limited. To fill this gap, The PostDoc will work with us on the design and analysis of a broader spectrum of benchmarks, by varying the complexity of the optimization problems, the algorithms, or the state space information that may be taken into account by the control policy.

Compétences

The candidate should bring solid experience in the formal analysis of black-box optimization algorithms.

Contexte de travail

The selected candidate will work at the LIP6 Computer Science department of Sorbonne Université, where they will be supervised by CNRS research director Carola Doerr. The PostDoc will integrate into the Operations Research team (RO) of LIP6.
The position is funded via the ERC Consolidator grant “dynaBBO: Dynamic Selection and Configuration of Black-box Optimization Algorithms”. Funding for traveling and conference attendance is available. The PostDoc will have access to the computing facilities of the LIP6 Computer Science lab and of Sorbonne University.
Our working language is English. No French skills are required.

Contraintes et risques

Not applicable