NEW NEW NEW NEW The problems presentation can be found here:

And the final presentations can be found here:

Cloud is Mine

Clould is mine - Mise en place d'un systeme de recomendation avec traitement de données incomplètes

L'intuitivité est l’une des préoccupations de Cloud is Mine, société experte en conseil dans le domaine du cloud computing. Ayant mis en place un comparateur de logiciels SaaS à la disposition des dirigeants des TPE et PME visant à développer leurs performances opérationnelles à l'aide de solutions logicielles proposées par des éditeurs et dont les fonctionnalités renseignées par un expert de la société sont adaptés à leurs besoins, celle-ci envisage envisage d'injecter dans son application des outils d'analyse et d’exploitation de données dans le but de développer un système automatisé de recommandation. Ce système exploitant des connaissances métier correspondant à des métriques dont l’étude permettrait de déduire des relations de corrélation statistiques entre les logiciels ou les profils utilisateurs permettrait de proposer un contenu plus adapté en se basant sur une mesure de similarité. La premier problématique qui s’impose est de trouver un moyen d’effectuer un clustering des logiciels ou des utilisateurs en se fiant à cette mesure de similarité pour pouvoir organiser les solutions proposées en fonction de leur pertinence individuelle ou en fonction de leur interopérabilité par rapport aux attentes des utilisateurs.

Toutefois lors du processus de recommandation, il s’avère que certaines sources de données présentent des incertitudes ou des imprécisions (profil utilisateur incomplet, besoin mal renseigné, ...). La deuxième problématique sera donc de répondre à la questions d’organisation des connaissances imparfaites des profils utilisateurs en exploitant les différentes sources d’informations les concernant (formulaire directement renseigné, connexion avec un compte sur un réseau social professionnel) et en utilisant des techniques de clustering suivant les notions de proximité et de similarité sémantiques des termes utilisés par des utilisateurs non experts.



Orange labs -  Mobile networks migration optimization

Most telecommunication operators deploy and run several generations of mobile network in parallel (2G, 3G, 4G and 5G in the next decade). In such context, the design of effective 5-year master plans for mobile networks transformation is a key strategic activity for the French telco Orange. This activity consists in deciding which network technologies to invest on (including yearly network deployment and dimensioning investments as well as marketing incentives to users) versus which ones are to be decommissioned, so as to handle bandwidth upgrade expectations at minimum cost. From a mathematical point of view, this decision problem leads to a specific multi-period network design problem.



n-side - Vehicle routing for transporting people to their medical centers

This problem is about scheduling transportation of people from their home to their medical centres and bring them back at the end of the visit. We have a heterogenous fleet of vehicle. These vehicle may have different equipments and capacity. Some patients may be transported only with vehicles with given equipment. For each patient, we have their availability, the time they have to be on site and the estimated visit time. We also have the location of all patient’s home and medical centres.
One important and challenging flexibility is that patients may arrive at their medical center with one vehicle and leaving on another.
The goal is to maximise the number of patients that we can serve given our fleet of vehicles under service constraints, such as the maximum time a patient can be in the vehicle before (s)he arrives at destination. The patients that cannot be served with our internal fleet of vehicles will be transported by taxi or ambulances, which are more expensive. 
The solution proposed to solve this problem should be ready to take into account the uncertainty about the visit times. Indeed, visit times are highly stochastic, and we need to ensure that the solution proposed is robust with respect to what may happen the day after tomorrow.



EDF - Energy management in a decentralized setting

In classical energy management, the viewpoint is typically that of a central  decider (producer) of sufficient critical mass. The related optimization  problems model to some extent, the match between electrical load and generation. Perhaps in a goal to compute the optimal production schedule today for tomorrow ; to compute the optimal strategy of use of water for the coming year ; to compute an optimal maintenance schedule. Typically load is assumed inelastic. A balance needs to be found between a detailed representation of the power system dynamics and the representation of the inherent uncertainty. The resulting optimization problems are highly challenging and by no means well resolved. However, in the future, several actors will dispose of advanced technology to pilot locally their load and production (prosumers). To give an example, their interaction with the centralized actor could be through means of a contract. This could take the form of a price signal being transferred to the local actor against which his production is optimized. Other options are available too. In return the local actor signifies his net load. The global actor will need to account for these only partially controllable production levels in the global load balance. Since now the local actor optimizes his production while accounting for economic incentives. Therefore, these features change fundamentally any of three above given optimization problems: a new analysis needs to be made and new algorithms designed.
In order to obtain a better understanding of the inherent technical difficulties let us discuss some of technical constraints that traditional generation assets are subject to. Thermal plants are subject to limitations on variations of power over time, minimum up/down times, maximum numbers of starts per day, limitations on power modulations. Moreover their power output must fall between a minimal and maximal level. Hydro generation is part of a set of cascading reservoirs consisting of a network linked through turbines and pumps. The turbining power output is a non-linear function of the flow rate and may depend on the water 
head level. Several additional constraints force the turbining levels to have limited variations or even remain stable for some amount of time to avoid strain on the material. These traditional assets are completed with intermittent sources such as wind/solar generation. New demand management tools make consumer load partially elastic. One can think for instance of advanced technology allowing the operator to pilot electrical heating in houses, or hot water tank recharging. The availability of electric vehicles and their charging cycles as well as an action on the timing of this event provides even further load elasticity. Some, or part of the assets or decisions may incumb to a local actor the interaction of which with the global actor still needs to be defined. This interaction could be understood through a price incentive or contract. 
In the current environment the optimization problem dealing with finding the optimal production schedule today for tomorrow (also known as the unit-commitment problem) aims at finding the most cost-effective production schedule while satisfying the operational constraints of the units. The second optimization problem dealing with efficient annual use of constrained resources 
(say water) typically consider a less detailed model for the underlying assets, but incorporate some notion of uncertainty, perhaps a scenario tree. Similarly the last problem investigates the best moment to place the maintenance of a large power plant (cf. ROADEF/EURO challenge 2010). It is clear that the above sketched modifications of the environment call for changes in problem formulation, analysis and solution methodology. 



Amadeus - Estimation of air traffic per city pairs worldwide on Hadoop Infrastructure

The airline industry generates many large scale optimization problems that are today solved by decomposing them into small sub-problems. One of these large scale Amadeus optimization problems is the estimation of air traffic per city pairs worldwide. The problem has been currently modeled as a quadratic problem and its solution takes several days of computation. This could lead to suboptimal or even infeasible solutions that require numerous iterations to find a global solution.

On the other hand, there is a significant progress on distributed systems used mainly for data processing and more recently for predictive analytics (e.g., Hadoop, MapReduce, and Spark). However, to the best of our knowledge, the operations research community has not fully exploited yet this architecture.

Our main objective today is to reduce the time spent in the solution of the air traffic estimation problem without losing optimality. We would like to follow two strategies. In the first one would like to search for alternatives formulations of the same problem numerically competitive with the current one. For any formulation, our second goal would be to provide solutions approaches, based on Big-Data systems (in particular clusters running Hadoop (and Spark)), able to solve some real-life Amadeus optimization problem instances.






Online user: 1