HAL Portal AU (University of Avignon)
Not a member yet
35713 research outputs found
Sort by
Les trois dames qui troverent l'anel 2 - Édition, traduction et notes d'après le ms. Berlin, Staatsbibliothek und Preussischer Kulturbesitz, Hamilton 257
No full textÉdition, traduction et notes des Trois dames et l'anel 2 d'après le ms. Berlin, Staatsbibliothek und Preussischer Kulturbesitz, Hamilton 25
Dispositifs d'aide préventifs et résolution de problèmes mathématiques: Comment outiller de manière préventive les élèves en difficulté et faire de l'erreur un levier d'apprentissage en situation de résolution de problèmes mathématiques?
No full textInternational audienc
Deep Learning and Natural Language Processing in the Field of Construction.
No full textThis article presents a complete process to extract hypernym relationships in the field of construction using two main steps: terminology extraction and detection of hypernyms from these terms. We first describe the corpus analysis method to extract terminology from a collection of technical specifications in the field of construction. Using statistics and word n-grams analysis, we extract the domain's terminology and then perform pruning steps with linguistic patterns and internet queries to improve the quality of the final terminology. Second, we present a machine-learning approach based on various words embedding models and combinations to deal with the detection of hypernyms from the extracted terminology. Extracted terminology is evaluated using a manual evaluation carried out by 6 experts in the domain, and the hypernym identification method is evaluated with different datasets. The global approach provides relevant and promising results
Risk factors and influence of surgical technique on the risk of caesarean scar defect formation: A systematic review of the literature
No full textInternational audienceObjectiveTo determine the factors associated with an increased risk of cesarean scar defect formation.MethodsA systematic literature search was performed up to July 2022 in PubMed databases following the PRISMA recommendations. All available English-language clinical studies presenting one or more factors that may affect the risk of cesarean scar defect were included.Results39 studies meeting the selection criteria were identified. An association was found between the number of previous cesarean sections and a significant increase in the risk of cesarean scar defect formation. Regarding patient age, gestational age at cesarean section, birth weight and emergency context did not appear to influence the risk of cesarean scar defect. However, cesarean sections performed during labor advanced stages of labor, may increase the risk. The data remain too limited to conclude on the impact of BMI, flexion uterine, and pregnancy pathologies (gestational diabetes, preeclampsia, premature rupture of membranes), the use of oxytocic, or infectious and hemorrhagic complications. Regarding the surgical technique, the literature suggested that a lower hysterotomy is associated with an increased risk of scar defect. However, the single- or double-layer suture technique did not provide a change in risk, and the data were too limited to conclude on the impact of the type of thread or suture used.ConclusionThe present systematic review of the literature suggests that several factors may increase the risk of developing a cesarean scar defect, such as the number of previous cesarean sections, a cesarean section performed during advanced labor, and a lower hysterotomy. However, the current state of the literature does not allow definitive conclusions to be drawn on most other factors
Linking ecotoxicological effects on biodiversity and ecosystem functions to impairment of ecosystem services is a challenge: an illustration with the case of plant protection products
No full textInternational audienceThere is growing interest in using the ecosystem services framework for environmental risk assessments of chemicals, including plant protection products (PPPs). Although this topic is increasingly discussed in the recent scientific literature, there is still a substantial gap between most ecotoxicological studies and a solid evaluation of potential ecotoxicological consequences on ecosystem services. This was recently highlighted by a collective scientific assessment (CSA) performed by 46 scientific experts who analyzed the international science on the impacts of PPPs on biodiversity, ecosystem functions, and ecosystem services. Here, we first point out the main obstacles to better linking knowledge on the ecotoxicological effects of PPPs on biodiversity and ecological processes with ecosystem functions and services. Then, we go on to propose and discuss possible pathways for related improvements. We describe the main processes governing the relationships between biodiversity, ecological processes, and ecosystem functions in response to effects of PPP, and we define categories of ecosystem functions that could be directly linked with the ecological processes used as functional endpoints in investigations on the ecotoxicology of PPPs. We then explore perceptions on the possible links between these categories of ecosystem functions and ecosystem services among a sub-panel of the scientific experts from various fields of environmental science. We find that these direct and indirect linkages still need clarification. This paper, which reflects the difficulties faced by the multidisciplinary group of researchers involved in the CSA, suggests that the current gap between most ecotoxicological studies and a solid potential evaluation of ecotoxicological consequences on ecosystem services could be partially addressed if concepts and definitions related to ecological processes, ecosystem functions, and ecosystem services were more widely accepted and shared within the ecotoxicology community. Narrowing this gap would help harmonize and extend the science that informs decision-making and policy-making, and ultimately help to better address the trade-off between social benefits and environmental losses caused by the use of PPPs
Not every day is a sunny day: Synthetic cloud injection for deep land cover segmentation robustness evaluation across data sources
No full textInternational audienceSupervised deep learning for land cover semantic segmentation (LCS) relies on labeled satellite data. However, most existing Sentinel-2 datasets are cloud-free, which limits their usefulness in tropical regions where clouds are common. To properly evaluate the extent of this problem, we developed a cloud injection algorithm that simulates realistic cloud cover, allowing us to test how Sentinel-1 radar data can fill in the gaps caused by cloud-obstructed optical imagery. We also tackle the issue of losing spatial and/or spectral details during encoder downsampling in deep networks. To mitigate this loss, we propose a lightweight method that injects Normalized Difference Indices (NDIs) into the final decoding layers, enabling the model to retain key spatial features with minimal additional computation. Injecting NDIs enhanced land cover segmentation performance on the DFC2020 dataset, yielding improvements of 1.99% for U-Net and 2.78% for DeepLabV3 on cloud-free imagery. Under cloud-covered conditions, incorporating Sentinel-1 data led to significant performance gains across all models compared to using optical data alone, highlighting the effectiveness of radar-optical fusion in challenging atmospheric scenarios
Assessing Form Patterns for the Suburban 15-Minute City: The Case of Drap (France).
No full textInternational audienceThe 15-minute city is an urban model aimed at reducing motorized traffic and fostering more convivial, human-centred environments. However, transforming suburban areas into 15-minute cities presents a greater challenge than applying the model to compact central cities. Functionalist approaches that focus solely on distributing different functions within walking distance often fail if they do not consider the urban form requirements necessary for a truly walkable environment. To address this, the Evolutive Meshed Compact City (emc2) has been proposed as a more specific 15-minute city model for European suburban peripheries, emphasizing urban form. It features a mesh of vibrant central streets that integrate existing linear centralities in suburban areas.Suburban settlements structured around main streets are particularly apt to implement the emc2 model. Drap, a village on the outskirts of Nice, France, serves as a case study for this approach. This study codifies emc2 into a pattern language, mapping and assessing public space patterns in Drap. Urban morphometrics played a crucial role in generating detailed spatial maps, while fieldwork provided preliminary data and direct observations of human activities in public spaces.A statistical cross-analysis of human behaviour and urban form patterns identified key factors for Drap’s transformation: the articulation between main and ordinary streets, a web of activities around main streets, ample pedestrian space (which Drap lacks), and active façades along public spaces. While Drap has strengths in its urban structure, its main street remains largely utilitarian rather than vibrant.By integrating urban morphometrics, fieldwork, and statistical analysis, this innovative approach provides targeted planning recommendations to transform suburban settlements like Drap into integral components of a successful 15-minute city. It offers a replicable method for reshaping suburban environments to support walkability, accessibility, and local vitality
Apprentissage par renforcement pour l'affectation des ressources dans les systèmes Edge/Fog computing
No full textThis thesis investigates how reinforcement learning (RL) can be applied to design intelligent decision-making systems for resource allocation in edge computing, focusing on algorithms tailored to the challenges of distributed, heterogeneous, and resource-limited environments.In the edge computing paradigm, data processing and decision-making are pushed closer to end devices, reducing latency and bandwidth usage but introducing strict resource constraints. Unlike the cloud, edge nodes operate under limited processing power, energy, and communication capacity, which makes traditional centralized optimization infeasible. Efficient orchestration therefore requires adaptive, decentralized control methods capable of satisfying performance and safety requirements under uncertainty. RL naturally fits this role: by learning through interaction, agents can adapt to dynamic environments and optimize long-term objectives even when system models are incomplete.The thesis is organized into three main parts. The first provides the theoretical foundations of RL and introduces an original contribution on multi-agent coordination; the second and third apply RL ideas to two classical problems of resource management, task offloading and load balancing, within the constrained edge-computing setting.The first part reviews the mathematical background of RL, including Markov decision processes, constrained RL, and multi-agent formulations. It then presents a novel framework for implicit coordination in sequential social dilemmas, where agents face conflicts between individual and collective interest. The proposed algorithm, C3-IPPO (Communication-free Constrained Coordination with Independent PPO), employs a three-timescale Lagrangian scheme in which each agent optimizes a local constrained policy while slowly adapting its internal penalty parameter. Coordination emerges implicitly through these updates, without explicit communication or shared rewards. Experiments on the Melting Pot benchmark show that C3-IPPO achieves cooperative behaviors comparable to communication-based baselines while remaining fully decentralized and scalable.The second part applies RL to task offloading in mobile edge computing: evices must decide whether to process data locally or offload it to an edge server, balancing latency, energy consumption, and information freshness measured by the Age of Information (AoI).A first study formulates the single-device problem as a Markov decision process combining offloading and energy harvesting. Exploiting the monotone structure of the optimal policy, the work introduces Ordered Q-Learning, a lightweight algorithm that preserves the convergence guarantees of Q-learning while accelerating learning in structured state spaces. The resulting policies closely match analytically derived optima and outperform standard deep RL baselines.The framework is then extended to a multi-agent offloading scenario, where several devices share the same edge server. Building on the coordination principles of C3-IPPO, a decentralized constrained-RL scheme enables agents to align local offloading actions with global resource constraints.The third part addresses load balancing under hard system constraints, distinguishing between (i) communication-link capacity limits and (ii) processing-capacity limits at servers. For the first case, classical queueing-theoretic methods are adapted to edge environments, yielding new safe policies that remain stable even in heavy-traffic regimes. For the second, the problem is cast as a constrained RL task, leading to the Decomposed Reward Constrained Policy Optimization (DRCPO) algorithm, which learns optimal admission and balancing decisions while provably satisfying safety constraints.Across all studies, the thesis demonstrates that combining RL with structural and constraint information yields solutions that are both adaptive and reliable.Cette thèse étudie comment le reinforcement learning (RL) peut être appliqué à la conception de systèmes de décision intelligents pour la gestion des ressources dans le cadre de l’informatique en périphérie de réseau (edge computing). Elle se concentre sur le développement d’algorithmes adaptés aux environnements distribués, hétérogènes et limités en ressources.Dans le paradigme de l’informatique en périphérie, le traitement et la prise de décision sont rapprochés des dispositifs terminaux, réduisant la latence et la charge réseau, mais introduisant de fortes contraintes en calcul, énergie et communication. Contrairement au cloud, ces nœuds de bord rendent l’optimisation centralisée inapplicable. Une orchestration efficace requiert donc des méthodes de contrôle adaptatives et décentralisées, capables de respecter des contraintes de performance et de sûreté dans des environnements incertains. Le RL s’inscrit naturellement dans ce cadre : en apprenant par interaction, les agents s’adaptent à des conditions dynamiques et optimisent des objectifs de long terme même en l’absence de modèle complet du système.La thèse est structurée en trois parties. La première présente les fondements du RL et propose une contribution originale sur la coordination multi-agent ; les deuxième et troisième parties appliquent ces idées à deux problèmes classiques de gestion des ressources — le déchargement de tâches (task offloading) et la répartition de charge (load balancing) — dans le contexte contraint de l’edge computing.La première partie revoit les bases du RL, notamment les Markov decision processes, le constrained RL et les formulations multi-agent. Elle introduit le cadre C3-IPPO (Communication-free Constrained Coordination with Independent PPO), un algorithme lagrangien à trois échelles de temps où chaque agent apprend une politique locale contrainte tout en ajustant un paramètre interne. La coordination émerge implicitement, sans communication explicite ni partage de récompenses. Les expériences sur le benchmark Melting Pot montrent que C3-IPPO atteint des comportements coopératifs comparables aux méthodes basées sur la communication, tout en restant entièrement décentralisé et évolutif.La deuxième partie applique le RL au déchargement de tâches dans le mobile edge computing. Les dispositifs doivent décider s’ils traitent localement ou envoient leurs données à un serveur de bord, en équilibrant latence, consommation d’énergie et fraîcheur de l’information (Age of Information, AoI). Une première étude formule le problème pour un seul dispositif et introduit Ordered Q-Learning, un algorithme léger exploitant la structure monotone de la politique optimale, qui accélère l’apprentissage tout en préservant les garanties de convergence du Q-learning. Le cadre est ensuite étendu à plusieurs dispositifs partageant un même serveur, à l’aide d’un schéma de RL contraint décentralisé inspiré de C3-IPPO.La troisième partie traite de la répartition de charge sous contraintes fortes du système, en distinguant (i) les limites de capacité des liens de communication et (ii) celles des serveurs. Dans le premier cas, des politiques sûres issues de la théorie des files d’attente sont adaptées à l’edge computing et demeurent stables même en régime de forte charge. Dans le second, le problème est formulé comme une tâche de RL contraint, donnant naissance à DRCPO (Decomposed Reward Constrained Policy Optimization), un algorithme apprenant des décisions d’admission et d’équilibrage tout en garantissant la satisfaction des contraintes de sécurité.Dans l’ensemble, la thèse montre que la combinaison du RL avec des informations structurelles et des mécanismes de contrainte permet de concevoir des solutions adaptatives, fiables et décentralisées. Elle met également en évidence la complémentarité entre les approches d’apprentissage et les méthodes analytiques classiques pour la gestion efficace des ressources dans les systèmes distribués modernes
Enjeux et apports d’une modélisation des structures paysagères par une approche ontologique spatiale hybride : le cas de la région de Klaipėda (Lituanie) depuis le XXe siècle
No full textInternational audienceL’évolution des structures paysagères en Lituanie depuis le XXe siècle résulte de transformations politiques, économiques, sociales et culturelles multiples, intenses et rapides. Pour en analyser les effets, une modélisation spatiale rétrospective des classes d’occupation et de couverture des sols à travers les périodes pré-soviétique, soviétique et post-soviétique s’avère nécessaire. Cette analyse repose sur l’intégration de données géospatiales libres récentes et de cartographies historiques, dont la combinaison génère une complexité liée à l’hétérogénéité des résolutions spatiales, des systèmes de représentation et des classifications utilisées.Dans cette étude, une approche ontologique spatiale hybride est adoptée. Elle s’appuie sur des nomenclatures prédéfinies d’occupation et de couverture des sols, déconstruites et enrichies par l’analyse de métriques configurationnelles, fonctionnelles et morphologiques. Ces métriques mobilisent des concepts paysagers classiques tels que la fragmentation, la connectivité, la proximité, etc. Un clustering multivarié par l’algorithme k-means, appliqué aux métriques extraites, permet d’identifier pour chaque période d’étude des structures paysagères caractérisées par leur distribution spatiale, leur organisation interne et leur profil d’utilisation et couverture des sols. Cette approche vise à réduire la complexité inhérente à la diversité des données sources.Dans la région de Klaipėda, cette méthode révèle une relative inertie des structures paysagères à travers les trois périodes d’étude, marquée par la persistance de grands ensembles tels que le système hydrologique littoral, lagunaire et fluvial ; les parcelles forestières ; les zones d’agriculture à intensité variable ; ainsi qu’un environnement deltaïque semi-naturel. Toutefois, leur structure interne évolue sensiblement, notamment sous l’effet des politiques agraires : développement de la petite propriété paysanne, réforme de collectivisation sous le régime soviétique, politiques de restitution foncière à partir des années 1990. Bien que la méthode soit largement contrainte par l’incomplétude des données sources, certaines lacunes peuvent être partiellement compensées par la reconstitution des structures paysagères associées à d’autres classes d’occupation et de couverture des sols