Ph.D
Group : Learning and Optimization
Multiobjective parallel evolutionary algorithms : application on diesel combustion.
Starts on 09/02/2009
Advisor : SCHOENAUER, Marc
Funding : Convention industrielle de formation par la recherche
Affiliation : Université Paris-Saclay
Laboratory : PSA Vélizy & LRI
Defended on 03/07/2012, committee :
- Mohamed Masmoudi, professeur, université de Toulouse (rapporteur)
- Frédéric Saubion, professeur , université d'Angers (rapporteur)
- Marc Schoenauer, directeur de recherche, INRIA Saclay (directeur de thèse)
- Ludovic Thobois, Anciennement PSA PEUGEOT CITROEN (co-encadrant)
- Laurent Dumas, professeur, UVSQ (examinateur)
- Noredine Melab, professeur, INRIA Lille (examinateur)
- Abdel Lisser, professeur, LRI (examinateur)
- Laurent Duchamps Delageneste, PSA PEUGEOT CITROEN (Invité)
Research activities :
Abstract :
In order to comply with environmental regulations, automotive manufacturers have to develop efficient engines with low fuel consumption and low emissions. Thus, development of engine combustion systems (chamber, injector, air loop) becomes a hard task since many parameters have to be defined in order to optimize many objectives in conflict. Evolutionary Multi-objective optimization algorithms (EMOAs) represent an efficient tool to explore the search space and find promising engine combustion systems. Unfortunately, the main drawback of Evolutionary Algorithms (EAs) in general, and EMOAs in particular, is their high cost in terms of number of function evaluations required to reach a satisfactory solution. And this drawback can become prohibitive for those real-world problems where the computation of the objectives is made through heavy numerical simulations that can take hours or even days to complete.
The main objective of this work is to reduce the global cost of real-world optimization, using the parallelization of EMOAs and surrogate models.
Motivated by the heterogeneity of the evaluation costs observed on real-world applications, we study asynchronous steady-state selection schemes in a master-slave parallel configuration. This approach allows an efficient use of the available processors on the grid computing system, and consequently reduces the global optimization cost.
In the first part of this work, this problem has been studied in an algorithmical point of view, through an artificial adaptation of EMOAs to the context of real-world optimizations characterized by a heterogeneous evaluation cost.
In the second part, the proposed approaches, already validated on analytical functions, have been applied on the Diesel combustion problem, which represents the industrial context of this thesis. Two modelling approaches have been used: phenomenological modelling (0D model) and multi-dimensional modelling (3D model).
The 0D model allowed us, thanks to its reasonable evaluation cost (few hours per evaluation) to compare the asynchronous steady-state approach with the standard generational one by performing two distinct optimizations. A gain of 42 % was observed with the asynchronous steady-state approach.
Given the very high evaluation cost of the full 3D model, the asynchronous steady-state approach already validated has been applied directly. The physical analysis of results allowed us to identify an interesting concept of combustion bowl with a gain in terms of pollutant emissions.
