Real-world applications in Optimisation (Vehicle Routing Problem) and Machine Learning (Image recognition) are very challenging due to their large-scale structure. To cope with this issue, decomposition techniques have been historically introduced in mathematical programming. These techniques split a given problem into smaller and more affordable subproblems which are subsequently solved. Column Generation is a reference on this matter but suffers from certain pathologies that we first propose to tackled using Learning approaches. In the same vein, Machine Learning (ML) models keeps growing and become time-consuming and energy-hungry models that can only be tackled with High-Performance Computing platforms. Nonethless, ML experts and academics did not inspired from mathematical programming and the decompostion techniques developed in the same context but in a different field. In this project, we are going to transpose and adapt decomposition techniques to ML models which are intrisically otpimisation problems.
The SERENITY project aims to design and develop a brokering system dedicated to space data that brings optimised costs/revenues and QoS by advancing the state of the art in the field of data lakes, HPC, and brokering optimization.
The UltraBO project aims to design and implement novel parallel hybrid optimization algorithms for modern supercomputers. Many experiments reported in the literature demonstrate that high-quality results are obtained through hybridization. However, finding efficient and effective hybridization schemes is challenging and tedious. On the other hand, according to Top500 modern supercomputers are increasingly large (millions of cores), heterogeneous (various CPUs, GPU, FPGA, …) and less and less reliable (Mean Time Between Failures - MTBF<1h) making their programming increasingly complex. The design and implementation of parallel algorithms for these ultra-scale supercomputers is still in its infancy especially in combinatorial optimization.
The ADARS (Automating the Design of Autonomous Robot Swarms) project aims to propose a unique approach to automatically generate behaviours for distributed aerospace and space systems (DASS) thanks to a cross-fertilisation between multi-objective optimisation and machine learning techniques. ADARS will demonstrate through specifically designed software simulations and real field tests with multi-rotor drones, that state-of-the-art results can be obtained on two challenging DASS applications: swarm formation for a counter UAV system and swarm formation of small satellites for asteroid observation.