Parcours Informatique et Interactions

Diplômés 2020

Les vendredi 4 septembre et lundi 7 septembre 2020, les treize étudiants du parcours Informatique et interactions ont soutenu leur mémoire de fin d'études. Félicitations à eux !

Stereo Event-Based Matching using Stereo PatchMatch with Slanted Support Window, and Time Surfaces, Simone Ballerio

Tutors: Jean Martinet, Université Côte d’Azur, CNRS, I3S, France.
Andrew Comport, CNRS, France.

Event Cameras are bio-inspired sensors that offer several advantages, such as low latency, high temporal resolution, high dynamic range and low power consumption. Depth estimation is one of the most important task in computer vision and robotics. This thesis presents a solution to the problem of Event-Based Stereo Matching from data captured by a stereo Event Camera system. The proposed method combines a Frame-Based Stereo Matching algorithm with Time Surfaces computed from a stereo stream of events, to produce multiresultion disparity maps of the scene. In this method, we improved the computation of the Time Surface exploiting the exponential properties, and we introduced a new method based on masking and merging to reduce errors coming from disparity maps. Finally, we set up a complete stereo recording environment with Event Cameras and we create our own dataset.

Extraction of knowledge from RDF graphs, Valeria Bellusci

Tutors: Andrea G.B. Tettamanzi, Université Côte d’Azur, CNRS, I3S, France.
Alberto Trombetta, Insubria University, Varese, Italy.

In the context of the Semantic Web, Linked Data (LD) is one of the main concepts andit is used to describe the practice of exposing, sharing, and linking data, information,and knowledge on the Semantic Web. A unique identifier (URI) is used for each resourceand a standard format is defined for the data (RDF).The Resource Description Framework (RDF) is an XML-based framework used to de-scribe resources of the Semantic Web in the form of subject-predicate-object triples andlink them together; RDF triples can be queried by SPARQL (SPARQL Protocol andRDF Query Language), which is a standard language for querying graph data. Queriesare performed through a set of SPARQL endpoints.In this work, we aim to analyze and refactor the works previously carried out in theknowledge extraction context, in particular the research work of Duc Minh Tran whichaims to discover new knowledge in the form of multi-relational association rules fromontology loaded as a file *.owl.Subsequently, the objective is to modify the algorithm, inserting a new approach thatallows to apply it directly to RDF graphs. Using a SPARQL endpoint, the modifiedalgorithm queries the RDF graph through SPARQL queries to obtain useful informationfor generating the association rules.Finally, the consistency of the rules with the knowledge base is checked querying theRDF graph.The refactored algorithms of Dun Minh Tran and the new algorithm were then publishedon an Open Source repository to make them available to the scientific community.

Agent-Based Modeling for Studying the Spontaneous Emergence of Money, Mattia Di Russo

Tutors: Celia da Costa Pereira, Université Côte d’Azur, CNRS, I3S, France.
Andrea G.B. Tettamanzi, Université Côte d’Azur, CNRS, I3S, France.
Paolo Massazza, Università degli studi dell’Insubria, Italy.

Money, defined as the media of exchange, is demonstrated to be an emergent property of multi-agent systems. We analyze an agent-based model of an economy in which individuals specialize in consumption and production and meet randomly over time, in a way that implies that trade must be bilateral and quid pro quo. We designed and developed a computational framework that represents this agent-based model, where the agents’ intelligence is guided by a learning classifier system that is demonstrated to be able to learn trading strategies (fundamental strategies) that involve short sequences of trades. Using this framework we show that it is possible to modify the classifier system in order to improve the agents’ behaviour when they are in charge of learning strategies that involve longer chains of trades (speculative strategies).

Graph decompositions and treelength, Thomas Dissaux

Tutor : Nicolas Nisse, CNRS, France

The treelength of a graph is the largest distance between two vertices of a bag of a tree-decomposition, over all tree-decompositions of the graph. Thanks to a tree-decomposition with small length, we can resolve some problem as the Traveling Salesman Problem in polynomial time, or approximate the treewidth of a graph. However, deciding whether a graph has a treelength at most 2 is NP-hard. In planar graph, the complexity of computing the treelength is an open problem. For some sub-classes of planar graphs, the treelength is already known as for Outerplanar graph and trees. The next sub-class of planar graphs to be studied is naturally the one of series-parallel (SP) graphs. Our main results are an 3/2-approximation algorithm, and the characterization of SP graphs with treelength at most 2 in terms of forbidden isometric subgraphs. To conclude, we also present some preliminary results for the characterization of SP graphs with treelength at most 3.

D́eveloppement d’une ǵeńerateur automatique de phrases standardiśees MNREAD, Arthur Doglio

Tutors: Pierre Kornprobst, INRIA, France.
Jean-Charles Régin, Université Côte d’Azur, CNRS, I3S, France.

A framework based on symptoms prediction to support medical assistance, Emmanuele Galasso

Tutors: Celia da Costa Pereira, Université Côte d’Azur, CNRS, I3S, France.
Cathy Escazut, Université Côte d’Azur, CNRS, I3S, France.
Binaghi Elisabetta, Università degli studi dell’Insubria.

The world of medicine is very vast, it always presents new complex challenges. Doctors are committed to saving lives and healing people every day. Whenever a doctor meets a new patient, he asks him some questions in order to understand the symptoms and find the disease the patient presents. This procedure must be done as soon as possible as sometimes a question makes a distinction between life and death. The purpose of this work is to create a framework that can help medical staff to find the right diagnosis as quickly as possible. The framework makes use of artificial neural networks trained through symptoms obtained through a scraper (a program capable of retrieving information from a site), in this case, the site is called SymCAT. We have also decided to test several classifiers to help in recognizing the disease that afflicts the patient. Finally, to help the doctor understand, educate himself, and learn what questions to ask the patient, we have constructed a graphical interface with which the doctor can interact in a user-friendly way. The result is a framework capable of predicting the next relevant symptom to ask the patient, thus helping to decrease the time needed by the doctor to guess the patient’s disease. However, we believe that this model can be further improved as it represents only a basis on which a powerful framework can be created.

Mailbox Half-Duplex Automata, Loïc Germerie Guizouarn

Tutors: Cinzia Di Giusto, Université Côte d’Azur, CNRS, I3S, France.
Etienne Lozes, Université Côte d’Azur, CNRS, I3S, France.

We introduce a new class of systems of communicating automata. The class represent half-duplex systems with mailbox communications: each process buffer stores all messages sent to a same automaton regardless of their senders. We prove that half-duplex mailbox systems are a subclass of 1-synchronizable systems. Such a characterisation allows to show that the membership problem is decidable as well as several verification problems such as: reachability, deadlock and boundedness. The algorithms we propose for reachability and deadlock are polynomial.

Intégration des données au produit HPE Unified Console, Violette Héron

Tutor: Jean-Charles PICARD, Hewlett Packard

Mon apprentissage se passe à Hewlett Packard Enterprise. J’ai intégré l’équipe travaillant sur le projet HPE Unified Console, un framework graphique permettant la création d’écrans de monitoring à partir de sources de données diverses. Lors de ma première année, j’ai développé un plugin rendant possible l’intégrationau projet de n’importe quelle base de données SQL en tant que source de données. Cette année a vu le début d’un nouveau projet: HPE Intelligence-Assurance. J’ai intégré une équipe internationale pour incorporer HPE Unified Console au projet en question.

Recherche de contraintes de séquences implicites, Victor Jung

Tutor: Jean-Charles Régin, Université Côte d’Azur, CNRS, I3S, France.

On introduit de nouveaux outils et de nouvelles approches basées sur les MDDs afin de déterminer si une contrainte est respectée par un ensemble de solutions. On se concentre essentiellement sur la contrainte de séquence, mais la portée des méthodes présentées peut s’étendre au-delà de cette contrainte. On présente tout d’abord une approche très générale s’appuyant sur l’introduction d’une nouvelle opération d’inclusion sur les MDDs, puis on s’intéresse à une approche plus spécifique, basée sur l’introduction de propriété(s) interne(s) aux noeuds du MDD. Les résultats montrent que l’introduction de l’opération d’inclusion semble prometteuse pour savoir si une contrainte est respectée, et l’ajout de paramètres aux noeuds permettent la mise en place de méthodes rapides exploitant la structure du MDD.

Localisation de robots en utilisant la programmation par contraintes, Steve Malalel

Tutors: Guillaume Allibert, Université Côte d’Azur, CNRS, I3S, France.
Marie Pelleau, Université Côte d’Azur, CNRS, I3S, France.

La localisation en robotique dans le milieu sous-marin est soumise à différentes contraintes : les systèmes GPS sont difficilement utilisables, l’utilisation de balises peut être coûteux et dangereux, et l’environnement homogène des fonds marins permet peu d’observations utilisables pour les algorithmes de localisation populaires. Des recherches ont donc été menées afin d’utiliser la programmation par contraintes pour pouvoir faire de la localisation avec ces circonstances, et le but de ce stage était de comprendre le fonctionnement d’un tel procédé puis de l’adapter pour d’autres types de localisation.

Développement Blockchain, Matthias Percelay

Tutor: Sorin Manolache, Digital Factory, Orange, Sophia Antipolis, France.

La blockchain est une technologie récente et dont les applications potentielles continuent à être explorées. Le framework Hyperledger Fabric est une des plus importantes implémentations open source de blockchain privée. Sa grande flexibilité est un atout mais rend aussi son utilisation difficile. Dans le but de permettre à Orange de mettre en œuvredes solutionsbasées sur cette technologie, cette alternance vise à concevoir et implémenter un logiciel de «Blockchain as a Service»qui faciliterait grandement le développement, le déploiement et l’utilisation de solutions basées sur Fabric dans l’infrastructure d’Orange.

Étude d’algorithmes dedécomposition de graphes, Théo Qui

Tutor: Nicolas Nisse, INRIA, France.

Nous nous intéressons ici aux décompositions arborescentes de graphes. Savoir si un graphe a une treelength d’au plus 2 est un problèmeNP-Complet, c’est pourquoi nous cherchons à approximer ce paramètre. Pour cela, nous présentons 4 algorithmes de la littérature permettant d’obtenir des décompositions de graphes et d’en approximer la treelength :lex-M et Bfs-layering qui sont des 3-approximation de la treelength, uncle-trees pour qui nous ne connaissons pas de rapport d’approximation, et disk-tree qui est une 6-approximation mais que l’on suppose être en réalité une 2-approximation. Pour certains algorithmes, nous proposons une légère amélioration nechangeant pas le rapport d’approximation, mais permettant tout de même d’obtenir des décomposition de length inférieure dans certains cas. Après les avoir implémentés, nous lançons des tests sur plusieurs types de graphes afin de comparer leurs temps d’exécution, et la length des résultats. Enfin, nous présentons quelques résultats de nos travaux en cours qui, nous espérons, nous permettront de prouver que disk-tree est bien une 2-approximation de la treelength.

Anomaly detection in real time systems, Fabio Scantamburlo

Tutors: Enrico Formenti, Université Côte d’Azur, CNRS, I3S, France.
Ignazio Gallo, DISTA,Universit`a Degli Studi dell’Insubria, Italy

Anomaly detection by machine learning is a topic which received great attention in the domain of time series analysis. Anomaly detection is an applicable method to discover anomalies or irregular points in a time series. Due to the sensors in the Internet Of Things paradigm, the need for time series analysis by new techniques such as artificial intelligence will increase in both private and public sectors. In this work we carry out a state of the art of the common methods to find anomalies in time series domain with the practical results and experiments issued by a real world scenario. After realizing that the data-set provided has insufficient information for useful results, we experimented several data augmentation techniques in order to improve the quality of the anomaly detection. Finally, we propose our method to perform data augmentation and the classification task in order to find anomalies in our data-set. The case study is strictly related to an industrial problem, where data is unlabeled and a real time anomaly detection algorithm is needed.