Portail HAL Ensta
Not a member yet
11080 research outputs found
Sort by
Hybridizing two linear relaxation techniques in interval-based solvers
No full textInternational audienceIn deterministic global optimization, techniques for linear relaxation of a non-convex program are used in the lower bound calculation phase. To achieve this phase, most deterministic global optimization codes use reformulation-linearization techniques. However, there exist also two interval-based polyhedral relaxation techniques which produce reliable bounds without adding new auxiliary variables, and which can take into account mathematical operations and most transcendental functions: (i) the affine relaxation technique, used in the IBBA code, based on affine forms and affine arithmetic, and (ii) the extremal Taylor technique, used in the Ibex-Opt code, which is based on a specific interval-based Taylor form. In this paper, we describe how these two interval-based linear relaxation techniques can be hybridized. These two approaches appear to be complementary, and such a hybrid method performs well on a representative sample of constrained global optimization instances
Change Detection in Polarimetric and Multilook SAR imagery using Stochastic Distance
No full textInternational audienceChange detection is useful for monitoring environmental evolution or tracking anthropogenic changes over time. One of the promotion of satellite monitoring is advocated for its ability to revisit the same sites periodically and gives us a growth data from different sensors like Radars. Synthetic Aperture Radar (SAR) data is preferred because weather conditions (e.g. cloud cover) do not hamper the ability to perform soil measurements. However, traditional change detection methods are sensitive to speckle noise. We propose to compare several statistical change detection methods in order to reduce false alarm. In this paper, the data used are polarimetric and multilook which are simulated by their covariance matrices following a complex Wishart law. Change detection methods rely on the pixel patches whose size we fix. After estimating the parameters of the Wishart law on patches of pixels from the same area but acquired on two different dates, we compare these two patches of pixels with three stochastic divergences: Kullback-Leibler, Bhattacharyya, and Hellinger. The decision threshold is then analyzed by comparing two methods, one by the Otsu segmentation method, the other by the Chi-square distribution. We evaluate these different methods by the number of true positives, false positives, and the number of targets identified. On simulation data, Bhattacharyya distance with the Chi-square threshold gives a higher true positive rate and a lower false alarm rate than other methods. Moreover, we note that the smaller the size of the pixel patches, the more the number of detected targets increases, but with a greater number of false alarms. Due to the strong influence of speckle, which causes false alarms, a post-processing method based on the concentration index of Getis and Franklin is proposed
Jump Conditions for Boussinesq Equations Due to an Abrupt Depth Transition
No full textInternational audienc
Highly-efficient electron ponderomotive acceleration in underdense plasmas
No full textInternational audienceLaser-plasma accelerators represent a promising technology for future compact accelerating systems, enabling the acceleration of tens of pC to above GeV over just a few centimeters. Nonetheless, these devices currently lack the stability, beam quality and average current of conventional systems. While many efforts have focused on improving acceleration stability and quality, little progress has been made in increasing the beam's average current, which is essential for future laser-plasma-based applications. In this paper, we investigate a laser-plasma acceleration regime aimed at increasing the beam average current with energies up to few-MeVs, efficiently enhancing the beam charge. We present experimental results on configurations that allow reaching charges of nC and a maximum conversion efficiency of around %. Through comprehensive Particle-In-Cell simulations, we interpret the experimental results and present a detailed study on electron dynamics. From our analysis, we show that most electrons are not trapped in a plasma wave; rather, they experience ponderomotive acceleration. Thus, we prove the laser pulse as the main driver of the particles' energy gain process
AP/HTPB heterogeneous combustion with revised kinetics
No full textInternational audienceThis study first proposes a revised combustion model for composite propellants containing ammonium perchlorate (AP) and hydroxyl-terminated polybutadiene (HTPB). It comprises a detailed gas-phase kinetic mechanism, assembled leveraging recent work from the chemical kinetics community. A model for homogenized AP/HTPB decomposition at the burning surface is formulated to perform coupled flame/solid simulations with the proposed gas-phase mechanism. The obtained combustion model is validated against available experimental data on AP/HTPB combustion for propellants containing 60% to 80% mass fraction. Employing this combustion model, 2D heterogeneous combustion of an AP particle surrounded by a layer of homogenized AP/HTPB binder is then studied. These 2D simulations allow to investigate the flame structure and conditions on the propellant surface, such as burning rate. The effect of AP particle size and ambient pressure on the combustion regime is evidenced: conditions leading to a premixed regime or a diffusion-limited regime are identified. Production of chemical species within the flame is finally studied. Significant differences are observed, with respect to a widely used model from the literature, mainly in the formation of end products and NOx species. In particular, it is shown that C4H6 isomerization and NO formation are important chemical steps in the flame
Analyse expérimentale et théorique des algorithmes d'apprentissage par renforcement
No full textIn Reinforcement Learning (RL), an agent learns how to act in an unknown environment in order to maximize its reward in the long run. In recent years, the use of neural networks has led to breakthroughs, e.g., in scalability. However, there are still gaps in our understanding of how to best employ neural networks in RL. In this thesis, we improve the usability of neural networks in RL in two ways, presented in two separate parts. First, we present a theoretical analysis of the influence of the number of parameters on learning performance. Second, we propose a simple feature preprocessing based on the Fourier series, which empirically improves performance in several ways.In the first part of this thesis, we study how the number of parameters influences performance. While in supervised learning, the regime of over-parameterization and its benefits are well understood, the situation in RL is much less clear. We present a theoretical analysis of the influence of network size and L2 regularization on performance. We identify the ratio between the number of parameters and the number of visited states as a crucial factor and define over-parameterization as the regime when this ratio is larger than one. Furthermore, we observe a double descent phenomenon, i.e., a sudden drop in performance around the parameter/state ratio of one. Our analysis is based on the regularized Least-Squared Temporal Difference (LSTD) algorithm with random features in an asymptotic regime, as both the number of parameters and states go to infinity while maintaining a constant ratio. We derive deterministic limits of the empirical, the true Mean-Squared Bellman Error (MSBE), and the true Mean-Squared Value Error (MSVE) that feature correction terms responsible for the double descent. We show that those correction terms vanish when the L2 regularization increases or the number of unvisited states goes to zero.In the second part of this thesis, we study the preprocessing of features through a Fourier series. In addition to the number of parameters, the amount of optimization that can be achieved in practice remains limited. Neural networks behave thus as under-parameterized models that are also regularized through early stopping. This regularization induces a spectral bias since fitting high-frequency components of the value function requires exponentially more gradient update steps than the low-frequency ones. We propose a simple Fourier mapping for preprocessing, which improves the learning of high-frequency components and thus helps to overcome the spectral bias in RL. We present experiments indicating that this can lead to significant performance gains in terms of rewards and sample efficiency. Furthermore, we observe that this preprocessing increases the robustness with respect to hyperparameters, leads to smoother policies, and benefits the training process by reducing learning interference, encouraging sparsity, and increasing the expressiveness of the learned features.En apprentissage par renforcement (RL), un agent apprend comment agir dans un environnement inconnu de façon à maximiser sa récompense sur le long terme.Ces dernières années, l'utilisation de réseaux de neurones artificiels a conduit à de nombreuses avancées, notamment en termes de scalabilité.Cependant, de nombreuses lacunes subsistent dans notre compréhension de la meilleure manière d'employer les réseaux de neurones en RL.Dans cette thèse, nous proposons d'améliorer l'utilisation des réseaux de neurones en RL de deux manières, présentées dans deux parties distinctes.La première partie présente une analyse théorique de l'impact du nombre de paramètres sur la performance d'apprentissage.La seconde partie propose un prétraitement simple des données, basé sur la série de Fourier, qui améliore empiriquement les performances des réseaux de neurones de plusieurs façons.Dans la première partie de cette thèse, nous étudions l'influence du nombre de paramètres sur la performance.Alors que dans l'apprentissage supervisé, le régime de surparamétrisation et ses avantages sont bien compris, la situation en RL est beaucoup moins claire.Nous présentons donc une analyse théorique de l'influence du nombre de paramètres et de la régularisation L2 sur la performance.Nous identifions le rapport entre le nombre de paramètres et le nombre d'états visités comme un facteur crucial et définissons la surparamétrisation comme le régime où ce rapport est supérieur à un.De plus, nous observons un phénomène de double descente, caractérisé par une chute soudaine de performance au-delà d'un rapport paramètres/états visités de un.Notre analyse est basée sur l'algorithme de Least-Squares Temporal Difference learning (LSTD) régularisé avec des caractéristiques aléatoires dans un régime asymptotique, où le nombre de paramètres et d'états tendent vers l'infini tout en maintenant un rapport constant.Nous dérivons des limites déterministes de l'erreur quadratique moyenne de Bellman (MSBE) basée sur des échantillons collectés, de la vraie MSBE, et de l'erreur quadratique moyenne de la fonction de valeur (MSVE) qui comportent des termes de correction responsables du phénomène de double descente.Nous démontrons que l'influence des termes de correction diminue avec l'augmentation de la régularisation L2, ou avec le nombre d'états non visités.Dans la seconde partie de cette thèse, nous proposons l'étude d'un prétraitement des données basé sur la série de Fourier.En effet, outre le nombre de paramètres, le nombre d'optimisations réalisé en pratique reste souvent limité.Par conséquent, les réseaux de neurones tendent souvent à se comporter comme des modèles sous-paramétrisés régularisés par un arrêt prématuré.Cette forme de régularisation induit notamment un biais spectral, puisque l'apprentissage des composantes à haute fréquence de la fonction cible requiert exponentiellement plus d'itérations dans la descente de gradient stochastique que pour les composantes à basse fréquence.Pour pallier à ce problème, nous proposons un prétraitement des données basé sur la série de Fourier afin d'améliorer l'apprentissage des composantes à haute fréquence et surmonter le biais spectral en RL.Nous présentons des expériences indiquant que ce prétraitement peut conduire à des améliorations significatives des performances, en termes de récompenses obtenues et de données utilisées.De plus, nous observons que ce prétraitement favorise une plus grande robustesse face aux hyperparamètres, conduit à l'élaboration de politiques plus régulières, et bénéficie au processus d'entraînement en réduisant l'interférence d'apprentissage, en encourageant l'apprentissage de caractéristiques distinctes et sparses (ou creuses), et en augmentant l'expressivité des caractéristiques apprises
Open the Chests: An Environment for Activity Recognition and Sequential Decision Problems Using Temporal Logic
No full textInternational audienceThis article presents Open the Chests, a novel benchmark environment designed for simulating and testing activity recognition and reactive decision-making algorithms. By leveraging temporal logic, Open the Chests offers a dynamic, event-driven simulation platform that illustrates the complexities of real-world systems. The environment contains multiple chests, each representing an activity pattern that an interacting agent must identify and respond to by pressing a corresponding button. The agent must analyze sequences of asynchronous events generated by the environment to recognize these patterns and make informed decisions. With the aim of theoretically grounding the environment, the Activity-Based Markov Decision Process (AB-MDP) is defined, allowing to model the context-dependent interaction with activities. Our goal is to propose a robust tool for the development, testing, and bench-marking of algorithms that is illustrative of realistic scenarios and allows for the isolation of specific complexities in event-driven environments
Characterization on thermal hysteresis of shape memory alloys via macroscopic interface propagation
No full textInternational audienceA specimen of Shape Memory Alloy (SMA) starts martensitic phase transformation normally by the nucleation of a macroscopic austenite-martensite interface at the specimen's boundary (i.e., the nucleation is sensitive to boundary conditions). By contrast, the interface propagation only needs to overcome the energy barrier of the incompatibility between the austenite and martensite phases (i.e., the interface-propagation driving force reflects material's intrinsic properties). In this paper, we observe the thermally induced forward and reverse quasi-static propagation of the macroscopic austenite-martensite interface in a Ni-Mn-Ga single-crystal bar. It is found that the temperature difference between the forward and reverse propagation is only 2.8 °C, which is obviously less than the thermal hysteresis evaluated from Differential Scanning Calorimetry measurement (6.3 °C) on the same material. This result not only gives a better characterization of the thermal hysteresis of the phase transformation, but also helps deeply understand the relation between the phase-transformation kinetics and the microstructures of the macroscopic austenite-martensite interface
Synthèse totale de produits naturels cyclotripeptidiques de la série des aurantiomides, dans des conditions de fonctionnalisation oxydative
No full textThe quinazoline heterocycle, as a pharmacophore, is significant in recent anticancer drug discovery and in natural products containing the piperazino[2,1-b]quinazolin-3,6-dione core. Compounds like anacine and its derivatives, found in marine Sponge-derived Penicillium species, show promising biological activities. However, their synthesis presents numerous challenges. In this study, we devised a three-step strategy, leveraging microwave-assisted condensation of tripeptides, to directly synthesize compounds bearing the piperazino[2,1-b]quinazolin-3,6-dione heterocyclic structure. Additionally, we successfully completed the full synthesis of anacine and polonimide C. Recognizing that aurantiomides A-C and verrucine F are oxidatized variants of anacine and verrucine A at the C-1 position, we employed DDQ to achieve the selective and efficient C-1 oxidation of the piperazino[2,1-b]quinazolin-3,6-dione structure. This approach demonstrated high efficacy across several substrates and allowed the conversion of anacine to aurantiomide C. Furthermore, we observed the spontaneous installation of the alkylidene group via autooxidation during the cyclocondensation process. Additionally, the successful application of nickel-catalyzed aminocarbonylation of a bromoethyl intermediate enabled us to achieve the first total synthesis of aurantiomide C, polonimide A, and verrucine F.Le cycle hétérocyclique quinazoline, en tant que pharmacophore, revêt une importance significative dans la découverte récente de médicaments anticancéreux et au sein des produits naturels de type piperazino[2,1-b]quinazolin-3,6-dione. Des composés tels que l'anacine et ses dérivés, trouvés dans les espèces de Penicillium issues d’Eponges marines, présentent des activités biologiques prometteuses. Cependant, leur synthèse présente de nombreux défis. Dans cette étude, nous avons élaboré une stratégie en trois étapes, en exploitant la condensation de tripeptides assistée par micro-ondes, pour synthétiser directement des composés portant la structure hétérocyclique piperazino[2,1-b]quinazolin-3,6-dione. Nous avons ainsi réussi à synthétiser l'anacine et le polonimide C.Les aurantiomides A-C et la verrucine F étant des variants oxydés sur la position C-1 de l'anacine et de la verrucine A, nous avons développé une méthode utilisant la DDQ pour réaliser l'oxydation sélective et efficace en C-1 de la structure piperazino[2,1-b]quinazolin-3,6-dione. Cette approche a démontré une efficacité élevée sur plusieurs substrats et a permis la conversion de l'anacine en aurantiomide C. De plus, nous avons observé l'installation spontanée du groupe alkylidène par auto-oxydation lors du processus de cyclocondensation. De plus, l'application réussie de l'aminocarbonylation catalysée au nickel d'un intermédiaire bromoéthyle nous a permis de réaliser la première synthèse totale de l'aurantiomide C, du polonimide A et de la verrucine F
Whither a Planetary Turn? A Roundtable
No full textInternational audienceIn recent years the concept of planetarity has seen a meteoric rise amongcritical theorists. But they often leave its exact signi cance opaque or evenactively obscure the meaning of ‘planet’. Can space history unearth a moreaccurate conceptualization by de-metaphorizing planetarity and fostering aspatial understanding of planet Earth instead? In this roundtable, a philosopher,a media theorist, a cultural historian, a historian of technology and a religiousstudies scholar discuss the uses of planetarity in their respective disciplines.They ask what all that planetarity talk is about, if proclaiming a planetary turncan be justified, and whether understanding the transformation of Earth into aplanet over the course of the twentieth century necessitates an extra-terrestrialperspective