In many application areas, big optimization requires increasingly large-scale models to deal with a growing amount of decision variables and conflicting objectives, and subject to multiple sources of uncertainty. In this context, solving problems necessitates addressing the challenges of scalability and handling of uncertainty. This often requires revisiting not only the design and implementation of traditional optimization algorithms but also their parallelization on massively multi-core and heterogeneous (ultra-scale) supercomputers including in addition to multi-core processors, accelerators (e.g. GPU) and coprocessors (e.g. MIC).
The goal of BONUS is to come up with advanced approaches following three research topics constituting the roadmap of the project and which are the subject of several current and future collaborations of the team: decomposition-based optimization, optimization under uncertainty and ultra-scale optimization. The combined use of the three topics is rarely addressed in the literature. From a software point of view, the proposed approaches will mainly be integrated in the ParadisEO framework. In terms of application and industrial transfer, we target the scheduling of energy systems such as smart grids and engineering design.
Bilel Derbel
Optimisation Bayésienne Parallèle des Réseaux Neuromorphiques
COntributions to the Analysis and Design of Parallel Batched Bayesian Optimization Algorithms
Productivity-aware parallel cooperative combinatorial optimization for ultra-scale supercomputers
La conception automatique et optimisation multi-objectifs de réseaux de neurones pro-fonds parallèles pour la détection automatique des incidents en temps réel
Optimisation de modèles Deep Learning pour la prévision spatio-temporelle
Optimisation multi-critères et conception automatisée des réseaux de neurones profonds
Attaquer la "large échelle" : calcul haute performance pour l'intelligence computationnelle
Optimisation coûteuse multi-objectifs assistée par des modèles de substitution 09/11/2022