- Computer Science Laboratory

XU Hao

PhD Student at Sorbonne University
Team : MoVe

Supervision : Souheib BAARIR
Co-supervision : ZIADI Tewfik

Hybridization of Constraint Compilation Methods, SAT Solvers, and Learning Techniques to Manage the Complexity of Renault’s Product Line

This thesis addresses the growing challenge of managing the variability modeling problem (VMP) in Renault's automotive product lines. The automotive industry, particularly at Renault, faces significant combinatorial complexity, as each vehicle can be configured with various options, features, and technical specifications. To illustrate the scale of this diversity, consider the example of the "Master 2" vehicle model. This heavy truck model can be configured in 10^21 different ways. Efficiently managing real-time requests regarding these configurations is the central problem to solve when dealing with such large-scale vehicle models. These requests come from multiple internal departments (e.g., supply chain, manufacturing, and marketing) as well as external customers who need to configure vehicles based on available options.

While it is necessary to respond quickly to these requests, enumerating all possible variants is computationally impractical due to the enormous configuration space. This makes managing such combinatorial diversity both a technical and operational challenge. Theoretically, this problem boils down to solving a constraint satisfaction problem (CSP).

Renault uses a knowledge compilation approach to manage product configuration. This approach precomputes the entire configuration space and stores it in a symbolic structure, allowing rapid query responses by searching within this precompiled space. This method avoids repeatedly solving NP-complete problems but introduces significant memory constraints, as the configuration space can occupy several gigabytes of memory.

However, with the increasing variability of vehicle models, memory requirements for these compiled structures have become problematic. Some models occupy up to 800 MB each, while operational systems handle multiple models simultaneously, leading to memory usage of several tens of gigabytes, putting pressure on system resources. The objective of this thesis is to optimize Renault's current configuration system by reducing the size of the configuration space without compromising response times. Specifically, two main contributions are proposed:

  1. Exploiting symmetries in the configuration space : The data structure used to encode the configuration space contains redundancies. By identifying and leveraging symmetries, where certain parts of the configuration space are structurally identical, the system can significantly reduce the size of the compiled data. Experimental results show a 52.13% reduction in the configuration space size and a 49.81% improvement in query response times.
  2. Automatic parameter tuning using machine learning : The system's performance heavily depends on appropriately tuning configuration parameters. Different variability models require different parameters for optimal performance. Machine learning models were developed to predict the best parameters for each variability model, thus improving system efficiency. Experiments demonstrate that machine-predicted parameters outperform default parameters, providing substantial performance gains.
Ultimately, this thesis aims to ensure that Renault's product configuration system remains efficient, scalable, and capable of managing the increasing complexity of future models.


Phd defence : 12/16/2024

Jury members :

Mme Élise VAREILLES, Professeure des universités, IMT Mines Albi, France [Rapporteur]
Mme Sana BEN-HAMIDA, Maître de conférences (HDR), Université Paris Nanterre [Rapporteur]
Mme Marie-Jo HUGUET, Professeure des universités, LAAS-CNRS
M. Chouki TIBERMACINE, Professeur des universités, University of Southern Brittany
M. Souheib BAARIR, Maître de conférences (HDR), Sorbonne Université SIM (Sciences, Ingénierie, Médecine)
M. Tewfik ZIADI, Maître de conférences (HDR), Sorbonne Université
Mme Siham ESSODAIGUI, Renault S.A.S.
M. Yves BOSSU, Renault S.A.S.

Departure date : 12/16/2024

2021-2024 Publications