Doctoral presentations

Registration for poster presentation for the Doctoral presentation on Monday 12th June should be made through  https://goo.gl/forms/Vx57jjmsCK4PC6Rq1 before 6th June 2017. 

All presentations related to the summer school are accepted, if you have any doubt do not hesitate to contact us by email: 

The presentation will be a short slide + a poster during the session.

More information on the slide will be given to the selected presentations. 

Accepted poster presentations 

Brethomé Lucile

Modelling and optimization of a passenger railway transportation plan with regards to mobility

"For historical reasons, the design of the SNCF railway services are, as elsewhere, mainly driven by the incremental optimization of the use of production resources (and primarily the available train-paths and material resources): timetable, infrastructure maintenance and rolling stock. Moreover, each year, marginal modifications are executed on the previous timetable, but a totally new timetable is rarely designed.
SNCF Transilien is SNCF passenger offer for Paris area. In this region, the traffic has increased by 30% over the last decade, reaching 3 million passengers a day on the network. Furthermore, timetables are saturated in term of capacity. To answer this mobility growth, it becomes more and more important to have decision support tools to build an optimized transportation plan with regard to passenger flows, service quality and its adaptation. This approach breaks up with the historical paradigm of optimization of the production resources. Proposing an offer consistent with actual needs is crucial for SNCF, to ensure reliability of public transport for future generations. The passengers will experience improvements concerning average travel time, service reliability, connections and comfort.
To this aim, several models have been developed to solve two stages of the transportation plan design. First, an ILP model with two distincts objective functions determines the stopping patterns and the frequencies of the trains. To manage with the two objective functions, we use a multiobjective method, called epsilon-constraint method. Then, a quadratic model has been designed to define the timetable. As the Timetabling Problem will manage large instances, we designed an iterative method, combining an heuristic and a MILP."

Laura Calzada-Infante

Dominance Network Analysis of Multiobjective Evolutionary Algorithms

"My proposal consists in applying a methodology developed in my on-going Ph.D, called Dominance Network Analysis (DNA) in the realm of Multiobjective Optimization. Originally, DNA was developed to assess the efficiency of input-output systems commonly studied using Data Envelopment Analysis (DEA). DNA consists in building a weighted direct acyclic graph that captures the dominance relationships between a sample of multidimensional points. This allows the visualisation of multidimensional data and the application of Complex Network Analysis (CNA) tools. Two specific applications of DNA to MOEAs are the following:
1. DNA of MOEA Evolution; It consists in using as sample points all the solutions generated by a MOEA along its search space exploration. This would lead to large network in which the most nodes will be dominated except the solutions that form the final Pareto Front (PF) approximation. This is temporal DN in which the nodes have attached the iteration in which that solution was generated. I believe that in this way DNA can help study the evolution of a MOEA and gain insight into its effectiveness
2. DNA for MOEA Benchmarking: It consists in using as sample points the PF approximations computed by different MOEAs. Although, by definition, each of those PF are formed by non-dominated solutions, when they are merged dominance relationships among those solutions can arise. I believe that DNA can thus be used to benchmark multiple MOEAs and develop new CNA-based performance indicators.
"

Cindy Calderón-Arce

Adapted Dictionary Learning for on-line estimation of surrogate models

"Expensive multiobjective optimization usually uses surrogate models to reduce the amount of objective function evaluations.  This is particularly necessary when the objective function has a high evaluation cost. Several methods have been proposed for such estimation based on artificial neural networks or gaussian processes. In this work we propose a novel approach which explores key modifications to the dictionary learning methods in order to estimate surrogate models.  These models are based on sparse linear combinations of \emph{words}, composed of smooth parametric functions. Hence, the surrogate models guide which locations of the objective function should be evaluated next and avoid unnecessary evaluations."

Simon Anderer

Evolutionary Algorithms for the Optimization of Mobility Systems in Real-Time (Working Title)

To tackle real-world optimization problems in transport logistics, I examine the introduction of a so-called Decision Maker as an interface between real-time dynamic optimization problems and evolutionary optimization algorithms. One the one hand, this decision maker is responsible for event-handling and information updates. On the other hand, it decides upon the realization of new (better) solutions suggested by the optimization algorithms. At the current stage of my PhD, a first version of this concept is implemented by considering a dynamic VRP in real-time, which is generated from the Solomon/Gehring & Homberger benchmark files. 

Hani Zbib

A multi-move decent algorithm for the No-Split Multi-Compartment Capacitated Arc Routing Problem

The No-Split Multi-Compartment Capacitated Arc Routing Problem (MC-CARP) is an extension of the CARP arising in different applications such as waste collection where different waste types exist at each edge and the fleet of vehicle is multi-compartmented. The aim of the MC-CARP is to find a set of least cost routes that service the demand for each waste type at all required edges without exceeding compartment capacities, with the condition that when a vehicle visits an edge it either picks the totality or none of its demands. We propose a heuristic approach to solve the MC-CARP, consisting of a multi-move descent algorithm that combines traditional CARP local search moves with optimal edge orientation choice and ejection chains. The algorithm iteratively chooses at random a set of moves and computes all the savings obtained from the search of the whole move neighborhood. The savings are then plotted on an alternate moves graph where each MC-CARP edge is transformed into a set of nodes and an edge is added between every two nodes presenting a positive saving. A shortest path algorithm is subsequently run on that graph to determine the set of multi-moves that bring the largest saving to the current solution. The algorithm has been tested on a large set of instances obtained from real-life waste collection companies across Denmark, varying in sizes between 25 to 11656 nodes and 18 to 8584 required edges, and in the number of waste types on each edge (2 to 4).

Glize Estèle

Branch-and-price algorithm for Multi-Objective Vehicle Routing Problems

"We propose an extension of the single objective exact algorithm of Baldacci et al. (2008) to generate the complete Pareto front of a bi-objective VRP. The resulting algorithm combines the column generation and the branch-and-bound methods. The algorithm requires the set of non-dominated points of the LP-relaxation to define a lower bound of the problem. Each point is combined with a direction in which the solution optimizes the problem. The algorithm also needs an upper bound representing feasible solutions of the initial problem. For each direction, the associated gap between the corresponding lower bound point and the closest upper bound point is computed. Then, all routes having their parametrized cost within this gap are generated and solving the resulting VRP provides a Pareto optimal solution. Afterwards, every pair of consecutive Pareto optimal solutions previously found defines a reduced area where other optimal solutions could be. Those are generated successively by calculating their most efficient direction, lower bound and gap. The algorithm is applied to a bi-objective VRP with time-windows to minimize two different costs on each route. "

Aymeric Blot

Reacting and Adapting to the Environment - Designing autonomous methods for multi-objective combinatorial optimisation

Multi-objective combinatorial problems are usually tackled by metaheuristics. These metaheuristics expose from few to large numbers of parameters in order to be efficient on many types of problems and instances. However, their performance on a given problem instance are closely related to their parameters values that are generally not optimal, either kept to their default value or laboriously set by hand.

Optimised values of these parameters can be inferred from a set of training instances (Algorithm Configuration / Parameter Tuning) and kept fixed during the algorithm execution. However, in order to better adapt to the instance to solve and its particular structure, these values can be modified during the algorithm execution (Parameter Control).

In this thesis, we study both aspects of automatic algorithm design on Multi-Objective Local Search (MOLS) algorithms, efficient multi-objective metaheuristics for which many parameterisable strategies and components have been proposed.