Search

Railway operations in dense traffic areas are very sensitive even to small disturbances, and thus require careful planning and real-time management. Dwell times in stations are in particular subject to a high variability and are hard to predict; this is mostly due to the interactions between passengers and the railway system during the dwelling process. This paper presents a data-driven approach for assessing the influence of the numbers of alighting, boarding and on board passengers on the dwell time. We propose to split the dwell time into a deterministic component depending on the passenger flow, called the Minimum Dwell Time, and a random component. A method for estimating the minimum dwell time is provided. Based on the knowledge of this value, observations can be classified according to the main determinant of dwell time, namely timetable constraints or passenger exchange. The latter observations are used for estimating the conditional distribution of dwell time given passenger flows. Numerical experiments are carried out on stations located inside the dense traffic area of Paris suburban network. The obtained results indicate that the presented method can be used for a variety of applications, such as capacity assessment or stochastic simulation.

International audience; Laboratory of Mechanics and Acoustics (LMA) worked for several years with the French railway company (SNCF) and the telecommunication operator Orange. These collaborations had different consequences in terms of applications. LMA and SNCF worked together on annoyance, loudness, and on detection of emerging signals in noise. From the knowledge provided by this cooperation, but also from the state of the art, SNCF tried to integrate loudness in the standard used in the railway domain (ISO-EN-3381). But it was not accepted for various reasons: suppliers are used to use 1/3 octave bands and dB(A) in their models for interior prediction, loudness is too complicated to calculate. However, SNCF included in its own specification loudness density and loudness for the comfort, but without penalty. Concerning emergent signals, a method measuring tonality was included in the standard, but it is a simplified method which does not work for multicomponent signals. In telecommunications, loudness of speech is one of the main parameters for a good listening quality. For 30 years, the Loudness Rating (LR) has been used to evaluate loudness at sending and receiving terminals. This indicator was first developed for narrow-band (0.3-3.4kHz) handset terminals, then extended to wideband (0.1-8kHz). But it failed in predicting loudness in wideband and has not been extended to larger bandwidths. Orange Labs highlighted the problem several years ago and proposed collaboration with LMA to study how it would be possible to use loudness instead of LR, through a PhD supervised jointly. Orange Labs proposed a model to ITU-T, where SG12 eventually adopted a new standard (P.700) in which loudness is calculated using the ISO-532B model adapted for telephonometry. This standard is valid for all bandwidths and all conditions (handset, handfree, teleconference). Now Orange Labs use ITU-T P. 700 and ask their suppliers to characterize their products using loudness.

International audience; Based on the analysis of the railway system in the Paris region in France, this paper presents a rescheduling problem in which stops on train lines can be skipped and services are retimed to recover when limited disturbances occur. Indeed, in such mass transit systems, minor disturbances tend to propagate and generate larger delays through the shared use of resources, if no action is quickly taken. An integrated Integer Linear Programming model is presented whose objective function minimizes both the recovery time and the waiting time of passengers. Additional criteria related to the weighted number of train stops that are skipped are included in the objective function. Rolling-stock constraints are also taken into account to propose a feasible plan. Computational experiments on real data are conducted to show the impact of rescheduling decisions depending on key parameters such as the duration of the disturbances and the minimal turning time between trains. The trade-off between the different criteria in the objective function is also illustrated and discussed.

International audience; This paper presents a novel method for the state health monitoring of high-speed train suspensions from in-line acceleration measurements by embedded sensors, for maintenance purposes. We propose a model-based method relying on a multibody simulation code. It performs the simultaneous identification of several suspension mechanical parameters. It is adapted to the introduction of uncertainties in the system and to the exploitation of numerous high-dimensional measurements. The novel method consists in a Bayesian calibration approach using a Gaussian process surro-gate model of the likelihood function. The method has been validated on numerical experiments. We demonstrate its ability to detect evolutions of the health state of suspension elements. It has then been tested on actual acceleration measurements to study the time evolution of the suspension parameters.

Brands are today looking for new ways to convey their identity. Currently, brands sound identity is mostly based on music for ads or for selling places, and communication elements like voice or jingles. In this thesis, we focus on products sound design as a way to convey brand identity. Is it possible to take into account brand values in the writing of the requirements for a sound designer? For a given object, what are the relevant design features to consider in order to communicate a desired identity? In our works, we propose a sound design methodology for the translation of brand values into sound attributes, and for conveying the resulting sound identity through a series of products. Based on the analysis of several studies on sound verbal description, we elicited a sound lexicon of 35 words adapted to the description of sound properties. We propose to use this lexicon along with two communication tools, moodboards and a “sound design deck”, to translate brand identity into words easily understandable by both experts and non-experts. We also propose a methodology for linking the physical constraints of the objects to design to words of the lexicon. This method is based on a sound indexing experiment, inspired from sensory analysis. The methodology we propose is applied to the design of SNCF sound identity through different elements (alarms, validating machines, split-flap displays…). The originality of this work resides in its interdisciplinarity combining different approaches such as marketing, design, sensory analysis and auditory perception. This work proposes both methodologies and tools that facilitate communication in sound design, and shows that it is possible to convey brand identity through products sound design.; Les marques cherchent à communiquer leur identité à travers de nouveaux vecteurs d’émissions. Aujourd’hui, l’identité sonore des marques concerne principalement les musiques associées aux publicités ou diffusées sur les points de vente, ainsi que des éléments de communication comme la voix ou les jingles. Dans cette thèse, nous nous intéressons au design sonore des produits et équipements en tant que vecteur de l’identité de marque. Comment prendre en compte l’identité de la marque dans les spécifications communiquées à un designer sonore ? De quelle marge de manœuvre dispose-t-on pour véhiculer une identité par un son produit par un objet mécanique ou diffusé sur un support numérique ? Nous proposons une méthodologie de design sonore identitaire permettant de traduire les valeurs d’une marque en termes sonores, et de véhiculer l’identité sonore ainsi définie sur une série d’objets caractéristiques de la marque. A partir d'une analyse d'un grand nombre d'études sur la perception des sons, nous avons établi un lexique sonore de 35 termes adapté à la description des différentes caractéristiques du son. Nous proposons d’utiliser ce lexique ainsi que deux outils de communication, des planches de tendances et un « jeu de cartes sonores », pour traduire les valeurs identitaires d’une marque en termes sonores compréhensibles par tous les acteurs du projet en design sonore, qu’ils aient ou non une expertise dans le domaine du son. Nous proposons également une méthode d’analyse fonctionnelle permettant de traduire à l’aide du lexique les contraintes techniques liées à la nature ou à la fonction de l’objet choisi pour véhiculer l’identité. Cette méthode repose sur une expérience originale, l'indexation sonore, inspirée de l'analyse sensorielle. La méthodologie proposée a été appliquée au design de l’identité sonore de la marque SNCF véhiculée à travers différentes sources (alarmes, composteurs, tableau d’affichage, validations électroniques…). L’originalité de nos recherches réside dans leur caractère interdisciplinaire, reliant les domaines du marketing, du design, de l’analyse sensorielle et de la perception auditive. Ces travaux ont permis de construire et de tester des méthodologies et outils facilitant la communication dans le design sonore, et ont montré que l’identité des marques pouvait s’exprimer à travers la conception des sons de leurs produits et équipements.

Brands are today looking for new ways to convey their identity. Currently, brands sound identity is mostly based on music for ads or for selling places, and communication elements like voice or jingles. In this thesis, we focus on products sound design as a way to convey brand identity. Is it possible to take into account brand values in the writing of the requirements for a sound designer? For a given object, what are the relevant design features to consider in order to communicate a desired identity? In our works, we propose a sound design methodology for the translation of brand values into sound attributes, and for conveying the resulting sound identity through a series of products. Based on the analysis of several studies on sound verbal description, we elicited a sound lexicon of 35 words adapted to the description of sound properties. We propose to use this lexicon along with two communication tools, moodboards and a “sound design deck”, to translate brand identity into words easily understandable by both experts and non-experts. We also propose a methodology for linking the physical constraints of the objects to design to words of the lexicon. This method is based on a sound indexing experiment, inspired from sensory analysis. The methodology we propose is applied to the design of SNCF sound identity through different elements (alarms, validating machines, split-flap displays…). The originality of this work resides in its interdisciplinarity combining different approaches such as marketing, design, sensory analysis and auditory perception. This work proposes both methodologies and tools that facilitate communication in sound design, and shows that it is possible to convey brand identity through products sound design.; Les marques cherchent à communiquer leur identité à travers de nouveaux vecteurs d’émissions. Aujourd’hui, l’identité sonore des marques concerne principalement les musiques associées aux publicités ou diffusées sur les points de vente, ainsi que des éléments de communication comme la voix ou les jingles. Dans cette thèse, nous nous intéressons au design sonore des produits et équipements en tant que vecteur de l’identité de marque. Comment prendre en compte l’identité de la marque dans les spécifications communiquées à un designer sonore ? De quelle marge de manœuvre dispose-t-on pour véhiculer une identité par un son produit par un objet mécanique ou diffusé sur un support numérique ? Nous proposons une méthodologie de design sonore identitaire permettant de traduire les valeurs d’une marque en termes sonores, et de véhiculer l’identité sonore ainsi définie sur une série d’objets caractéristiques de la marque. A partir d'une analyse d'un grand nombre d'études sur la perception des sons, nous avons établi un lexique sonore de 35 termes adapté à la description des différentes caractéristiques du son. Nous proposons d’utiliser ce lexique ainsi que deux outils de communication, des planches de tendances et un « jeu de cartes sonores », pour traduire les valeurs identitaires d’une marque en termes sonores compréhensibles par tous les acteurs du projet en design sonore, qu’ils aient ou non une expertise dans le domaine du son. Nous proposons également une méthode d’analyse fonctionnelle permettant de traduire à l’aide du lexique les contraintes techniques liées à la nature ou à la fonction de l’objet choisi pour véhiculer l’identité. Cette méthode repose sur une expérience originale, l'indexation sonore, inspirée de l'analyse sensorielle. La méthodologie proposée a été appliquée au design de l’identité sonore de la marque SNCF véhiculée à travers différentes sources (alarmes, composteurs, tableau d’affichage, validations électroniques…). L’originalité de nos recherches réside dans leur caractère interdisciplinaire, reliant les domaines du marketing, du design, de l’analyse sensorielle et de la perception auditive. Ces travaux ont permis de construire et de tester des méthodologies et outils facilitant la communication dans le design sonore, et ont montré que l’identité des marques pouvait s’exprimer à travers la conception des sons de leurs produits et équipements.

Deploying deep learning models in real-world certified systems requires the ability to provide confidence estimates that accurately reflect their uncertainty. In this paper, we demonstrate the use of the conformal prediction framework to construct reliable and trustworthy predictors for detecting railway signals. Our approach is based on a novel dataset that includes images taken from the perspective of a train operator and state-of-the-art object detectors. We test several conformal approaches and introduce a new method based on conformal risk control. Our findings demonstrate the potential of the conformal prediction framework to evaluate model performance and provide practical guidance for achieving formally guaranteed uncertainty bounds.

International audience; This paper studies the influence of an energy storage system (ESS) on the fuel consumption of a diesel-electric locomotive. First, an energetic model of a diesel-electric locomotive is established using energetic macroscopic representation (EMR). An inversion-based control is deduced, and the model is validated by experimental results on a real locomotive. Second, from this validated model, a battery/supercapacitor ESS is added in simulation to study the benefit of hybridization before integration on the real vehicle. The simulations show that simple energy management based on a frequency approach allows for the reduction of 25% on fuel consumption on a real drive cycle.

Thanks to its new connected trains, Transilien is now able to measure the number of passengers boardings and alightings per train door in real time.Our research uses this unique dataset to contribute to a better synchronization of train and passenger flows during the operational stage of railway operations.We first evaluate several statistical learning models to estimate dwell time as a function of railway operation and passenger flows variables. These models allow us to isolate critical situations where passenger flows significantly impact dwell time. Our research forecasts each of the railway operation and passenger flows variables, one stop ahead, from bidirectional autoregressive models exploiting their recent past. We then simplify those models’ using patterns derived from timetables. Finally, we estimate the occupancy rate by zone of the open gangways rolling stocks in order to inform passengers about the comfort on board and build two models of on-board passenger movements., Grâce à ses rames connectées, Transilien mesure en temps réel le nombre de montées et de descentes par porte du train. Nous contribuons à une meilleure synchronisation en phase opérationnelle des flux de trains et de voyageurs à l’aide de ces données uniques. Nous évaluons plusieurs modèles d’apprentissage statistique afin d’estimer les temps de stationnement en fonction des variables d’exploitation ferroviaire et des flux de voyageurs. Ces modèles permettent d’isoler des situations critiques où les flux de voyageurs impactent les temps de stationnement. Nous prévoyons chacune des variables, à l’horizon d’un arrêt, à partir de modèles autorégressifs bidirectionnels exploitant leur passé proche. Ces modèles se simplifient grâce aux motifs issus de la grille horaire. Nous estimons enfin des taux d’occupation par zone des rames traversantes, afin d’informer les voyageurs sur le confort à bord et proposons deux modèles de déplacement des voyageurs à bord.

International audience; The use of reduced basis has spread to many scientific fields for the last fifty years to condense the statistical properties of stochastic processes. Among these basis, the classical Karhunen-Loève basis corresponds to the Hilbertian basis that is constructed as the eigenfunctions of the covariance operator of the stochastic process of interest. The importance of this basis stems from its optimality in the sense that it minimizes the total mean square error. When the available information about this stochastic process is characterized by a limited set of independent realizations, the covariance operator is not perfectly known. In this case, there is no reason for the Karhunen-Loève basis associated with any estimator of the covariance that are not converged to be still optimal. This paper presents therefore an adaptation of the Karhunen-Loève expansion in order to characterize optimal basis for projection of stochastic processes that are only characterized by a relatively small set of independent realizations.

Massive data are surrounding us in our daily lives. Urban mobility generates a very high number of complex data reflecting the mobility of people, vehicles and objects. Transport operators are primary users who strive to discover the meaning of phenomena behind traffic data, aiming at regulation and transport planning. This paper tackles the question "How to design a supportive tool for visual exploration of digital mobility data to help a transport analyst in decision making?” The objective is to support an analyst to conduct an ex post analysis of train circulation and passenger flows, notably in disrupted situations. We propose a problem-solution process combined with data visualisation. It relies on the observation of operational agents, creativity sessions and the development of user scenarios. The process is illustrated for a case study on one of the commuter line of the Paris metropolitan area. Results encompass three different layers and multiple interlinked views to explore spatial patterns, spatio-temporal clusters and passenger flows. We join several transport network indicators whether are measured, forecasted, or estimated. A user scenario is developed to investigate disrupted situations in public transport.

International audience; Today, in Electric Railways, the old DC supply systems are reaching their limits. To improve their efficiency and increase railroad traffic, a new DC Electrification has recently been proposed at 9-kV. It is now necessary to prepare the migration of infrastructure and rolling stock, using Power-Electronic-Transformers (PETs), for adaptation to this voltage level. For this application, high efficiency and reduced volume are essential. This paper clearly demonstrates that it is now possible to achieve compact, high-power, isolated DC-DC converters using 3.3-kV SiC-MOSFET power modules with high efficiency. After a preliminary study based on simulations, the paper focuses on characterization and implementation of elementary isolated DC-DC converters. The proposed topology is a series-resonant converter rated for a nominal power of 300-kW at 1.8-kV. First, laboratory testing using an “opposition-method” is used to evaluate the elementary converters up to their nominal power using both electrical and thermal measurements to accurately determine losses and efficiency. At the nominal output power, an efficiency of 98.93% is obtained. This is quite remarkable for an isolated DC-DC converter operating under 1.8-kV with a switching frequency of 15-kHz. Finally, two elementary isolated DC/DC converters are associated in ISOP configuration (Input Seriesed/Output Paralleled) in order to achieve a 3-kV/1.5-kV PET with a nominal power of 600-kW as a prelude to the final 9-kV/1.5-kV power conversion.

International audience

International audience; We consider a suburban railway network line in the greater Paris area, with sub-branches. Trains of the line are equipped both with automatic vehicle localization (AVL) and automatic passenger counting (APC) devices, leading to a rich data set with simultaneous measurements of variables related to railway operations (arrival delay, dwell time) and passenger flows (numbers of passengers alighting and boarding, total load at departure). We aim for one-station-ahead forecasting of each of these five variables independently from each other. To do so, we build a bi-auto-regressive approach consisting of using the past values of the variable of interest along a first dimension, given by past stations along the train ride, and along a second dimension, given by past trains at the station. A building block of this approach is a train-station representation that accommodates different types of train services. We identify repeated patterns in this representation and exploit this fact. Indeed, the proposed bi-auto-regressive models are based on linear regressions whose coefficients depend on the stations and possibly only on the location of the train ride within a repeated pattern. This results in models that have a smaller complexity than extremely local models tailored to the timetables, with no significant decrease in accuracy.

International audience; This work focuses on the detection of anomalies in the message traces of the communication infrastructure of the information system of the French national railway company (SNCF). It shows that for this task a method can be designed by combining two recent and independent techniques. The first is the storage and indexation of time series in a tree called iSAX tree, and the second is an anomaly detection score named CFOF, that has been proven to resist to the concentration phenomenon in high dimension. We show in this paper, that it is possible to use the information structure of the iSAX tree to quickly determine an approximation of good quality of the CFOF score. We show that the approximate score is close to the exact score on both synthetic and real industrial datasets, and the first feedback indicates that this score is relevant for triggering alarms related to the anomalies observed in the activity of the SNCF information system.

International audience; Controlling the energy consumption is an important stake in today’s world. In the railway field, the energy consumed by high-speed trains depends on many variables such as the driver behaviour. Significant variations have been noticed for different drivers on the same journey. To help drivers, crossing points are defined along the journey, but differences still exist. The industrial objective of this work is to define a model, able to describe the train dynamics and to propose an optimization method, which aims to minimize the energy consumption. This work is composed of two parts. First, a deterministic model is defined to describe the train longitudinal dynamics based on a Lagrangian approach [1]. This model is calibrated based on commercial trains measurements. Afterwards, the optimization of the command is performed using the CMA-ES method [2] to minimize the energy consumed while punctuality, security, and comfort constraints are respected. Nevertheless, the high-speed train system is complex, and taking into account the uncertainties of the model parameters is necessary. Therefore, a Bayesian inference method [3] is applied in order to include uncertainties in the previous deterministic model. Finally, an optimization under uncertainty method is used to find the optimal command. The originality of this work lies on its transposability to real train systems. Indeed, pneumatic braking is distinguished from dynamic braking (able to recover a part of the energy consumed). The optimization method is applied to the driver command and it combines both punctuality andphysical constraints. Many energy measurements are used to calibrate and validate the models and verify the quality of the optimal solution. Finally, the rolling environment of the train is determined carefully by the use of wind predictions, track declivity and curvature measurements.REFERENCES[1] Nespoulous J., Soize C., Funfschilling C., and Perrin G. Optimization of train speed to limit energy consumption. Vehicle System Dynamics. 2021.[2] Hansen N. The CMA evolution strategy: a tutorial. 2016. ArXiv e-prints, arXiv:1604.00772v1 [cs.LG], 4 April 2016 and https://hal.inria.fr/hal-01297037.[3] Box G.E.P. and Tiao G.C. Bayesian inference in statistical analysis. John Wiley & Sons. 2011 (40).

A multitude of explainability methods and associated fidelity performance metrics have been proposed to help better understand how modern AI systems make decisions. However, much of the current work has remained theoretical -- without much consideration for the human end-user. In particular, it is not yet known (1) how useful current explainability methods are in practice for more real-world scenarios and (2) how well associated performance metrics accurately predict how much knowledge individual explanations contribute to a human end-user trying to understand the inner-workings of the system. To fill this gap, we conducted psychophysics experiments at scale to evaluate the ability of human participants to leverage representative attribution methods for understanding the behavior of different image classifiers representing three real-world scenarios: identifying bias in an AI system, characterizing the visual strategy it uses for tasks that are too difficult for an untrained non-expert human observer as well as understanding its failure cases. Our results demonstrate that the degree to which individual attribution methods help human participants better understand an AI system varied widely across these scenarios. This suggests a critical need for the field to move past quantitative improvements of current attribution methods towards the development of complementary approaches that provide qualitatively different sources of information to human end-users.

RailBeijing 2021, 9th International Conference on Railway Operations Modelling and Analysis (ICROMA), Pékin, CHINE, 03-/11/2021 - 07/11/2021; On-demand shared transport systems have become an efficient means of transport that addresses the challenges of urban transport in large cities, and provides a flexible transport service to users according to their needs. In this paper, a public transport system based on a fleet of autonomous electric vehicles, from the point of view of the SNCF company, the largest public transport operator in France. The service operates on a dedicated infrastructure, based on a fixed timetable at peak hours and an on-demand service the rest of the time. The problem amounts to scheduling autonomous heterogeneous electric vehicles ondemand during off-peak hours. To model the problem, we propose to split the infrastructure into sections and segments. By doing so, we can model the scheduling of vehicles and manage their maneuvers including overtaking, turnarounds, and the access to single-lane sections, by using railway-inspired techniques. In this paper, we propose a mathematical model of this problem.

International audience

WOS:000470074400013; International audience; In France, most of the trains are electrical and some are fed by 1500V DC voltage. When a train is travelling, the power consumption creates a voltage drop. Therefore, below a certain limit, the train has to slow down its speed. This limitation prevents the increase of traffic in some area which needs an electric reinforcement. This paper proposes to use embedded storage systems to support the voltage of the catenary. The main issue is: how to size such an energy storage system? In this way, several aspects (e.g.: the energy management strategy) have to be taken into account. The paper deals with an optimization-based sizing method for the embedded energy storage systems. The proposed method takes into account the energy management strategy in the sizing by optimization. A comparison is made between the use of batteries or supercapacitors.

WOS:000458127000025; International audience; The increasing railway traffic and the environmental issue demand to find new solutions to provide energy to autonomous train (train with embedded energy sources such as diesel power supply). Using hybrid diesel train with embedded storage elements is an interesting technical solution but this kind of multisource system presents new scientific and methodology challenges. Thus, in the paper, the problematic focus on the design and the energy management of the different sources for this system. Moreover, these two fields are linked to each other. Indeed, there is a strong influence of the sizing on the energy management but the reverse is also true. The paper deals with a new optimization approach to perform the design of both the hybrid sources sizing and their energy management. The multi-sources system is represented by a power flow model and the energy management strategy is based on a frequency approach and on dynamic programming. This direct optimization problem is solved by the Sequential Quadratic Programming (SQP) algorithm. Thus, in the paper, this optimization approach is applied to a real railway study case. A comparison is made between two cases with different energy management methods.

When using Finite Element methods to solve wave propagation problems, the spatial domain must be truncated at a finite distance. This can be done using absorbing boundary conditions or layers, which effectively inhibit the reflection of outgoing waves back into the computational domain. They cannot however handle moving loads coming from outside of the (truncated) computational domain, for which some incoming waves must be allowed. After illustrating this fact, this paper proposes a technique to introduce such moving loads in time simulations over finite size domains. This technique involves the introduction of appropriate initial conditions within the computational domain. In general, these initial conditions can be computed numerically, and possibly re-used for various configurations. The interest of the method is illustrated on industrial cases of interest for the railway community.

International audience; The speed profile of a train plays an important role in energy consumption and resulting costs. The industrial objective of this work is to develop a method, which optimizes the train speed under constraints, in order to reduce the energy consumed over a journey. Some articles have presented similar problems solved with other methods [1, 2]. In this work, we choose to describe the dynamic problem thanks to a rigid body approach (Lagrangian formalism), which is projected onto the longitudinal axis. The aerodynamic, traction, and braking forces applied to the train are taken into account. The pneumatic braking is dissociated from dynamic braking that can recover energy.The aim is to reduce the energy consumed (cost function) playing on the driver commands (traction and braking forces) respecting comfort, security, and punctuality constraints. As the track topology is an important parameter, we describe track slopes and curves to fit with industrial needs. The railway system depends on uncertainties derived from weather and from the power available at the catenary. As speed is a functional feature of the curvilinear abscissa of the track, a discretization strategy is used. The optimization problem under uncertainties is solved using a CMA-ES method where the constraints are implemented using an augmented Lagrangian. The method is applied to a real high-speed line and the model has been validated with experimental measurements. The optimal trajectory is compared to the in-line trains.REFERENCES[1] P. Wang, R. Goverde, Multiple-phase train trajectory optimization with signaling and operational constraints. Transportation Research Part C: Emerging Technologies, 192, 913 – 922, 2016.[2] G. Scheepmaker, R. Goverde, L. Kroon, Review of energy-efficient train control and timetabling. European Journal of Operational Research, 257, 355 – 376, 2017.

International audience; Dans cet article nous reproduisons un scénario d’apprentissage selon lequel les données cibles ne sont pas accessibles et seules des données connexes le sont. Nous utilisons une approche par méta-apprentissage afin de déterminer si les méta-informations apprises à partir de messages issus de médias sociaux, finement annotés en émotions, peuvent produire de bonnes performances une fois utilisées sur des messages issus de conversations, étiquetés en émotions avec une granularité différente. Nous mettons à profit l’apprentissage sur quelques exemples (few-shot learning) pour la mise en place de ce scénario. Cette approche se montre efficace pour capturer les méta-informations d’un jeu d’étiquettes émotionnelles pour prédire des étiquettes jusqu’alors inconnues au modèle. Bien que le fait de varier le type de données engendre une baisse de performance, notre approche par méta-apprentissage atteint des résultats décents comparés au référentiel d’apprentissage supervisé.

International audience; Non-fluorinated coatings have recently emerged as a new approach to avoid contamination on transparent surfaces. This research work concerns the synthesis of highly transparent, hydrophobic, durable, and self-cleaning coatings as an alternative to fluorine-based coatings used on glass. The developed coating is produced by a chemical route (sol–gel method) using two silica-based precursors, hexamethyldisilazane, and tetraethoxysilane (HMDS/TEOS). Then, the prepared Gel HMDS/TEOS was deposited on glass slides by the dip-coating technique. Afterward, the effects of mechanical durability, water-drop testing, thermal cycling testing, and aging times on the hydrophobicity of the gel were analyzed. The properties of the coating were characterized by optical microscopy, scanning electron microscopy, UV-VIS-NIR spectrophotometry, and contact angle measurement. The results show that the organic modification of the gel by trimethylsilyl (-Si-(CH3)3) groups leads to a homogeneous and a hydrophobic surface. Sufficient dilution of this gel with ethanol has been found to provide a highly transparent coating without losing its hydrophobic property. In addition, the robust and durable coating exhibits high self-cleaning performance suitable for various industrial applications.

International audience; The speed profile of a train plays an important role in energy consumption and resulting costs. The industrial objective of this work is thus to develop a method to reduce the energy consumed by a train over a journey by playing on the driver commands (traction and braking forces) while respecting punctuality constraints. A coupling between measured data and simulation is proposed to solve this optimization problem. First, a rigid body approach (Lagrangian formalism) is introduced to characterize the dynamics of each element of the train and their interactions with their environment. In particular, the aerodynamic (including the wind effect), traction, and braking forces are taken into account, and a special attention is paid to the vertical and lateral characteristics of the track as they play a key role on the train dynamics. Secondly, a model for energy consumption and recovery (thanks to dynamic braking) is introduced. Experimental measurements of a high-speed line are then used to estimate the parameters on which the two previous models are based and to validate their predictive capacities. The optimization problem under constraints is finally solved using an evolutionary algorithm where the constraints are implemented using an augmented Lagrangian formalism. The performance of the proposed method in terms of speed optimization and energy consumption reduction is compared to measurements associated with commercial trains.

International audience; Several recent studies on dyadic human-human interactions have been done on conversations without specific business objectives. However, many companies might benefit from studies dedicated to more precise environments such as after sales services or customer satisfaction surveys. In this work, we place ourselves in the scope of a live chat customer service in which we want to detect emotions and their evolution in the conversation flow. This context leads to multiple challenges that range from exploiting restricted, small and mostly unlabeled datasets to finding and adapting methods for such context.We tackle these challenges by using Few-Shot Learning while making the hypothesis it can serve conversational emotion classification for different languages and sparse labels. We contribute by proposing a variation of Prototypical Networks for sequence labeling in conversation that we name ProtoSeq. We test this method on two datasets with different languages: daily conversations in English and customer service chat conversations in French. When applied to emotion classification in conversations, our method proved to be competitive even when compared to other ones.

International audience; In the heart of a tense competitive context, railway transport has a major advantage through its low environmental impacts. Among these, particular attention is paid to acoustic emissions. For people living near railway infrastructures, noise emissions are a major concern. Noise barriers are an effective and traditionally implemented solution to reduce the acoustic impact of railway traffic. Nevertheless, these barriers are generally not well accepted by the population for aesthetic reasons. Sonic crystals are structures that provide significant acoustic attenuation, particularly around certain frequencies (so-called ?band gaps?). Using this new type of noise mitigation solution has advantages in terms of integration into the landscape and cost but remains a challenge given the plurality of railway acoustic sources, both in terms of their location, directivity and spectral content. This study presents results obtained by 2D Finite Element Method (FEM). This paper has two objectives. The first one is to study the sensitivity of acoustic performance to geometric parameters such as the filling factor or the shape of the elementary structure of the crystal. The second seeks to design a sonic crystal adapted to each type of railway source (rolling, equipment, aeroacoustics) - which differ from each other in their sound emission spectrum. A comparison with a full classic screen is also presented.

In this thesis, we propose methods to approximate an anomaly score in order to detect abnormal parts in data streams. Two main problems are considered in this context. Firstly, the handling of the high dimensionality of the objects describing the time series extracted from the raw streams, and secondly, the low computation cost required to perform the analysis on-the-fly. To tackle the curse of dimensionality, we have selected the CFOF anomaly score, that has been proposed recently and proven to be robust to the increase of the dimensionality. Our main contribution is then the proposition of two methods to quickly approximate the CFOF score of new objects in a stream. The first one is based on safe pruning and approximation during the exploration of object neighbourhood. The second one is an approximation obtained by the aggregation of scores computed in several subspaces. Both contributions complete each other and can be combined. We show on a reference benchmark that our proposals result in important reduction of the execution times, while providing approximations that preserve the quality of anomaly detection. Then, we present our application of these approaches within the SNCF information system. In this context, we have extended the existing monitoring modules by a new tool to help to detect abnormal behaviours in the real stream of messages within the SNCF communication system., Dans cette thèse, nous proposons des méthodes de calcul approchées d'un score d'anomalie, pouvant être mises en oeuvre sur des flux de données pour détecter des portions anormales. La difficulté du problème est de deux ordres. D'une part, la haute dimensionnalité des objets manipulés pour décrire les séries temporelles extraites d'un flux brut, et d'autre part la nécessité de limiter le coût de détection afin de pouvoir la réaliser en continu au fil du flux. Concernant le premier aspect du problème, notre étude bibliographique a permis de sélectionner un score de détection d'anomalies proposé récemment, le score CFOF, qui est le seul pour lequel il existe des garanties formelles quant à son adéquation pour les données en haute dimensionnalité. Nos contributions ont alors porté sur la proposition de deux méthodes d'approximation du score CFOF pour permettre son usage en continu sur des flux. La première est une approche combinant élagage et approximation lors du parcours des voisinages dans l'espace de description des objets. Notre second apport est une approximation par agrégation de scores obtenus sur des sous-espaces, qui complète la première contribution et se combine avec elle. Nous avons montré sur une collection de jeux de données, utilisés comme cadre d'évaluation de référence dans le domaine, que nos méthodes permettaient des gains importants en temps de calcul, tout en fournissant des approximations qui préservent la qualité des détections. Enfin, nous présentons également l'application de ces approches au sein du système d'information de la SNCF dans lequel de nombreux flux sont collectés en temps réel, transformés et rediffusés. Dans ce contexte, nous avons étendu la supervision de bout-en-bout existante par la mise en oeuvre d'un outil d'aide à la détection d'anomalies sur le flux de messages entrant d'une des principales plateformes de traitement.

International audience; Des expérimentations basées sur des technologies de traitement automatique du langage ont été menées au sein d'un programme de sécurité ferroviaire et de simplification documentaire afin d'améliorer la recherche d'information et la rédaction dans les textes de prescription SNCF.

IEEE ITSC 2020, 23rd IEEE International Conference on Intelligent Transportation Systems, Rhodes, GRECE, 20-/09/2020 - 23/09/2020; In order to provide a satisfying quality of service to passengers, companies operating suburban trains seek to implement timetables that perform well despite the occurrence of random disturbances. However, due to some specificities of Mass Transit, the standard approaches for robust train timetabling do not apply in that context. In this paper, we present a data-based stochastic simulation method for assessing the efficiency of train timetables for dense traffic areas. We describe models for the random variabilities that occur daily on such networks, and use them in a stochastic simulation algorithm. Results are presented on a saturated line of Paris suburban network.

This work is focused on the improving of the energy efficiency of electrical train in braking mode. The improvement is explored by integration and control of a reversible inverter in a DC power substation to inject the braking power from the DC railway electrical network to the AC transmission grid. This solution makes the power substation reversible. The objective of this study is to investigate a control strategy by adjustment of the catenary DC voltage by the inverter in order to recover the maximum of electric braking energy by keeping the DC voltage of the substation stable with a reference value. A comparative study of simulation results of the proposed control scheme with a droop control technique for the railway substation “Massena” in Paris is presented.

COMPRAIL 2020 : 17th International Conference on Railway Engineering Design & Operation, Berlin, ALLEMAGNE, 01-/07/2020 - 03/09/2020; In this paper we propose an experimental study aimed to assess a new paradigm for railway traffic management. In current practice and academic state of the art, train priority decisions in case of perturbation are made centrally. In this work, train traffic is modelled as a multi-agent system in which trains are decision makers. We implement a microscopic simulator for assessing the viability of this decentralized traffic management and for comparing it with the centralized one. Our results show that it is indeed possible to let trains decide their own priority and it is possible to achieve results which are comparable to simple centralized approaches in terms of passenger travel time in the mass-transit area of Paris, in France.

Mode I + II fatigue crack growth tests are run, following sequential or pseudo-sequential loading paths, representative of those computed for ”squat type” cracks in rails. Stereo DIC provides the near-tip displacements, from which Δ K I effective and Δ K II effective are derived. Compression while shearing extends coplanar growth, by slowing down the wear-induced rise of the effective mode mixity ratio, up to the critical value for which bifurcation occurs. The consideration of plasticity and contact and friction stresses improves crack paths predictions, compared to the maximum tangential stress criterion. The coplanar crack growth rate correlates well with a combination of Δ K I effective and Δ K II effective .

Acronyme : ARPAN Le projet Approche Responsable des Passages à Niveau (ARPAN) entre dans le champ des recherches visant l'amélioration de la sécurisation des passages à niveau, en relation avec l'évolution des comportements des personnes amenées à les traverser, plus particulièrement les conducteurs de véhicules automobiles. Il pose comme hypothèse que l'amélioration attendue passera par des solutions innovantes concernant la lisibilité et la visibilité des infrastructures. Pour la valider, différentes situations d'approche et de traversée des passages à niveau ont été intégrées dans les scénarii et itinéraires de simulateurs de conduite. Ces situations s'inscrivaient dans des environnements correspondants : 1. aux infrastructures actuelles et aux infrastructures en évolution 2. aux infrastructures actuelles avec messages adressés aux conducteurs Un panel de 58 sujets a été recruté puis mis face aux situations choisies, à l'intérieur d'un parcours de conduite simulé d'une durée d'environ 20 à 30 minutes. Les données ont été recueillies à partir de l'observation de leurs comportements, de la mise en oeuvre d'entretiens visant à expliciter leurs actions mentales et physiques, d'entretiens semi-directifs sur leurs expériences. Des analyses quantitatives et qualitatives ont permis de valider l'hypothèse de départ par rapport à certaines des solutions choisies.

This paper highlights the potential of medium-voltage dc electrification for railway traction. A method is proposed for determining the dc voltage level in order to guarantee, in terms of traffic, equivalent performances to an ac-electrified railway line. Railroad traffic, type of transport service, and railway network topology being given, the analysis method takes into account the following constraints: voltage at the pantograph, heating of overhead line and rail-to-ground voltage. Once the dc voltage level is chosen, the dc electrification system is compared with a standard 25-kV ac electrification in the case of the French high-speed line from Paris to Strasbourg.

International audience; In this paper, we focus on the problem of robust rolling-stock planning for French passenger trains. First, we characterize robustness and define some indicators for the evaluation of rolling-stock rosters. We take a particular interest in homogenizing turning-times in a roster in order to absorb potential delays. Then, we propose a new approach to solve the problem of robust rolling-stock planning. The SNCF reference tool (PRESTO) calculates a solution to the rolling-stock planning problem. It consists of a multi-step approach to cover demand while minimizing operating costs, and to further add maintenance slots to the roster. We propose an integrated ILP model to add robustness to a roster while maintaining low operating costs compared to PRESTO. We have carried out tests on nine real French regional transport instances, and we use a simulation module to evaluate the results. We observe a significant improvement in robustness indicators while maintaining low operating costs and meeting maintenance requirements.

Les temps de stationnement en gare en Île-de-France peuvent avoir un impact significatif sur la régularité des trains transilien. L'optimisation des durées de stationnement est essentielle pour une exploitation efficace, en n'étant ni trop courtes, ni trop longues. Le déploiement d'un système de comptage automatique des voyageurs embarqué permet de comprendre les déterminants d'une telle durée optimale. Il en résulte une modélisation du temps de stationnement minimal, effectuée par gare et par sens, qui permet d'évaluer la probabilité que les temps de stationnement prévus dans la grille horaire soient respectés ou soient source de retards.

Forum Acusticum, LYON, FRANCE, 07-/12/2020 - 11/12/2020; On-board measurement of the acoustic performance of railway tracks is nowadays necessary to qualify networks on a large scale. The issue concerns not only the supply of prediction models for strategic noise mapping, but also the optimization of track maintenance. One of the key parameters to be measured is the rail acoustic roughness. Unlike direct measurements where sensors are directly applied to the rail surface, indirect measurements of rail roughness focus on quantities that result from wheel/rail interaction, such as noise or vibrations of axle-boxes or rail, and from which the effective wheel/rail combined roughness are estimated. In particular, on-board measurements make the qualification of long track lengths possible without major constraints on traffic. A number of improvements can be made to the existing methods, especially in estimating the transfer functions between the effective roughness and the signals provided by the sensors. This study is part of the MEEQUAI French project aiming to combine modelling and measurements to optimize the estimation of the transfer functions and the location of sensors while taking into account the variability of tracks. This paper concerns the measurement performed on a static vehicle/track configuration in order to validate and calibrate the numerical simulations.

International audience

International audience; Structures composed of repetitions of multiple identical cells are common and have been the object of a large body of literature on waveguides, periodic media, and cyclic symmetry. Starting from a cell model, possibly with a large number of Degrees of Freedom both inside the cell and on its edges, the objective of this paper is to propose a Ritz-Galerkin reduction procedure retaining the standard second order model form and periodicity properties of the original model, while controlling accuracy in terms of model bandwidth and reproduction of the forced response to applied loads. The procedure is classically decomposed in two phases: subspace learning and basis generation. For the learning phase, Wave Finite Element (WFE) and periodic computations are presented as alternatives. The latter are then preferred for their easier control on the reduced model accuracy using classical modal synthesis and the simple choice of few target wavelengths. For the basis generation phase, constraints needed to generate a periodic superelement are defined and numerical procedures to generate a basis verifying those constraints are proposed. The validity of the reduction is demonstrated for the case of a cell with random elastic properties presenting bandgaps, local modes and wavemode crossing. Accurate predictions of modes and damped forced response are given for both the infinite and finite cases, using frequency and time simulations. The proposed analysis illustrates tracking of waveshapes, evaluation of significant waves by computation of the forced response in the frequency/wavenumber domain and interpretation of the relation between the infinite and finite forced responses. The case of a railway track with edges and transitions between multiple periodic zones is finally used to illustrate scalability issues.

International audience; Long-term averaged indicators are commonly used to dimension protections and satisfy existing regulations for railway infrastructures. These long-term indicator are generally used to assess the exposure of populations to environmental noise and, in turn, to evaluate health related effects. However, long-term indicators may not correlate well with noise perception by residents. In case of high speed train pass-by, for example, short-term indicators are more directly linked to the maximum noise levels and the duration of the event. For rolling stock characterization, indicators such as LA eq,Tpb>, SEL, TEL and LAmaxare commonly used. As an extension, such indicators, including LAeq,Te may also be used to assess the noise perceived by residents for single pass-by events. A new calculation scheme, compatible with existing normative methods, has been developed for predicting event-related noise indicators in complex environments, under varying meteorological conditions. The prediction methods are adapted to obtain the time-dependent sound pressure level time during the train pass-by and then to derive the different short-term indicators. This paper presents a validation study comparing measured and calculated short term indicators according to current engineering methods. Results are given for different trains and speeds as well as different sites and receiver configurations.

This work focuses on the detection of anomalies in a railway communication system. We present the detection method as well as the architecture of the system supporting it. The method is based on a preliminary stage to collect and aggregate the data. It then combines the use of a multidimensional indexation tree (i.e., iSax tree) to store time series, and the computation of an anomaly detection score (i.e., CFOF score). We propose a fast estimation of this anomaly score, based on the information stored in the indexation tree, so that the score can therefore be computed on the fly over the stream of messages in the communication system. We show by mean of experiments that this estimation is close to the exact score. We describe the platform that has been implemented, and we show that it effectively support abnormal behaviour detection in the real stream of messages within the communication system of the French Railway Company (SNCF).

WCRR 2019, 12th World Congress on Railway Research, TOKYO, JAPON, 28-/10/2019 - 01/11/2019; Passenger load forecasting can be valuable in transportation planning, operation management and for enriching the information available to passengers, particularly in high-density megacities. This paper investigates the long and short term forecasting of passenger loads in a transit network by using multiple sources of data (on-board headcount data and train timetables). With each passenger load being treated as a time series, one of the main challenges of this study is related to the dependence of the temporal dynamics of the time series to be predicted on the railway timetable. Machine learning models are proposed to predict the passenger load on each train passing each station. We will compare different models, including a random forest, and a gradient boosting tree. Different types of features (calendar, hour, last passenger load, train delay, and train route) will be considered to measure their contributions to the prediction task. The experiments are conducted on a real historical dataset covering the period from 2015 to 2016. The dataset was collected on a railway transit network line operated by SNCF in suburban Paris.

International audience

ECML/PKDD - The European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, Würtzburg, ALLEMAGNE, 16-/09/2019 - 20/09/2019; The increase in the amount of data collected in the transport domain can greatly benefit mobility studies and help to create high value-added mobility services for passengers as well as regulation tools for operators. The research detailed in this paper is related to the development of an advanced machine learning approach with the aim of forecasting the passenger load of trains in public transport. Predicting the crowding level on public transport can indeed be useful for enriching the information available to passengers to enable them to better plan their daily trips. Moreover, operators will increasingly need to assess and predict network passenger load to improve train regulation processes and service quality levels. The main issues to address in this forecasting task are the variability in the train load series induced by the train schedule and the influence of several contextual factors, such as calendar information. We propose a neural network LSTM encoder-predictor combined with a contextual representation learning to address this problem. Experiments are conducted on a real dataset provided by the French railway company SNCF and collected over a period of one and a half years. The prediction performance provided by the proposed model are compared to those given by historical models and by traditional machine learning models. The obtained results have demonstrated the potential of the proposed LSTM encoder-predictor to address both one-step-ahead and multi-step forecasting and to outperform other models by maintaining robustness in the quality of the forecasts throughout the time horizon.

5th International Workshop and Symposium TransitData 2019, Paris, France, 08-/07/2019 - 10/07/2019; On the basis of data collected by counting sensors deployed on trains, this paper deals with a forecasting of passenger load in public transport taking into account train operation. Providing passengers with train load forecasting, in addition to the expected arrival time of the next train, can indeed be useful for a better planning of their journeys, which can prevent over-crowding situations in the trains [6] [7]. The proposed approach is built on both a hierarchy of recurrent neural networks [8] and representation learning [9] with the aim to explore the ability of such mobility data processing to simultaneously perform a forecasting task and highlight the impact of events on the public transport operation and demand. An event refers here to an unexpected passenger transport activity or to a modification in transport operation compared to those corresponding to normal conditions. Two kind of historical data are used, namely train load data and automatic vehicle location (AVL) data. This latter source contains all information related to the train operation (delay, time of arrival/departure of vehicles ...). The proposed methodology is applied on a railway transit network line operated by the French railway company SNCF in the suburban of Paris. The historical dataset used in the experiments covers the period from 2015 to 2016.

CAP 2019, Conférence sur l'Apprentissage Automatique, Toulouse, France, 03-/07/2019 - 05/07/2019; Les possibilités offertes en termes de collecte et de stockage de données permettent de renouveler les approches de modélisation dans le domaine du transport. L'exploitation croisée de différentes sources de données a pour vocation la création de services à forte valeur ajoutée pour l'usager. Les travaux détaillées dans cet article portent sur le développement de modèles de prévision a base de méthodes d'apprentissage notamment profond, pour la prévision court-terme de la charge (nombre de passagers) des trains. Cette prévision de l'affluence dans les trains peut servir à enrichir l'information voyageur à destination des usagers des transports collectifs qui peuvent ainsi mieux planifier leur déplacement. Elle peut également servir aux opérateurs de transport pour une régulation "à la demande" de l'offre de transport. La principale difficulté dans la prévision est liée à la variabilité intrinsèque des séries temporelles des charges à prédire, induite par l'influence de plusieurs paramètres dont ceux liés à l'exploitation (horaire, retard, type de mission...) et au contexte (information calendaire, grand évènement, météo,...). Nous proposons un modèle LSTM encodeur-prédicteur pour résoudre cette tâche de prévision. Plusieurs expérimentations sont menées sur des données réelles du réseau Transilien de la SNCF sur une durée d'un an et demi. Les résultats de prévision sont détaillées en vue de comparer les performances d'un tel modèle à plusieurs horizons temporels avec celles d'autres modèles plus classiques utilisées en prévision.

International audience; The management and selection of relevant information for a given speaking tour remains a problem for open-domain dialogue systems. For the latter, the possible interactions between a user and an agent are a priori infinite and indefinite. The possibility of an incorrect response from the agent to the user therefore remains high. For goal-oriented systems, the problem is considered solved, but in our experience no system shows remarkable robustness when evaluated in real situations. In this article, we present a state of the art of agent learning methods and different conversational agent models. In our opinion, one of the agent's avenues for improvement lies in his memory, because the latter (often represented by the triplet: current speaking time, dialogue history and knowledge base) is not yet modelled with enough precision. By providing the agent with a cognitively inspired memory model, we believe we can increase the performance of a goal-oriented dialogue system in real life situations, by using automatic learning algorithms with an antagonistic approach in support of a new memory model for the agent. ABSTRACT Goal-oriented dialogue systems: a recent overview and research prospects Determining which information is relevant for a given dialogue turn is still a problem for open-domain dialogue systems since the interaction between a user and the system is a priori infinite and indefinite. Thus, the possibility for error remains high for the dialog system. Although the problem is considered solved, to our experience, no system has shown outstanding performance when confronted to "real-world" situation. We review the state-of-the-art of machine learning approaches used for dialogue system development and present the various models of agent architecture. We claim that one of the main directions for improvement is the agent architecture. Often the agent's memory (represented by the triple: dialogue turn, dialogue history and knowledge base) is not modeled accurately enough. Thus we decided to investigate more cognitive oriented models of memory. We hypothesize that combining machine learning approaches in an adversarial setup with a new memory model for the agent would result in improved performances for goal-oriented dialogue systems. KEYWORDS: dialogue systems, learning by reinforcement, memory model, learning by antagonism, goal-oriented systems.; La gestion et la sélection des informations pertinentes pour un tour de parole donné restent un problème pour les systèmes de dialogue à domaine ouvert. Pour ces derniers, les interactions possibles entre un utilisateur et un agent sont a priori infinies et indéfinies. La possibilité d'une réponse erronée de l'agent à l'utilisateur demeure donc élevée. Pour les systèmes orientés-but, le problème est considéré comme résolu, mais d'après notre expérience aucun système ne montre une robustesse remarquable lorsqu'il est évalué en situation réelle. Dans cet article, nous dressons un état de l'art des méthodes d'apprentissage de l'agent et des différents modèles d'agent conversationnel. Selon nous, l'une des pistes d'amélioration de l'agent réside dans sa mémoire, car cette dernière (souvent représentée par le triplet : tour de parole courant, historique du dialogue et base de connaissances) n'est pas encore modélisée avec assez de précision. En dotant l'agent d'un modèle de mémoire d'inspiration cognitive, nous pensons pouvoir augmenter les performances d'un système de dialogue orienté-but en situation réelle, par l'emploi d'algorithmes d'apprentissage automatique avec une approche antagoniste en support d'un nouveau modèle de mémoire pour l'agent. ABSTRACT Goal-oriented dialog systems : a recent overview and research prospects Determining which information is relevant for a given dialog turn is still a problem for open-domain dialog systems since the interaction between a user and the system is a priori infinite and indefinite. Thus, the possibility for error remains high for the dialog system. Although the problem is considered solved, to our experience, no system has shown outstanding performance when confronted to "real-world" situation. We review the state-of-the-art of machine learning approaches used for dialog system development and present the various models of agent architecture. We claim that one of the main directions for improvement is the agent architecture. Often the agent's memory (represented by the triple : dialog turn, dialogue history and knowledge base) is not modeled accurately enough. Thus we decided to investigate more cognitive oriented models of memory. We hypothesize that combining machine learning approaches in an adversarial setup with a new memory model for the agent would result in improved performances for goal-oriented dialog systems. MOTS-CLÉS : systèmes de dialogue, apprentissage par renforcement, modèle de mémoire, appren-tissage par antagonisme, systèmes orientés-buts.

International audience; The management and selection of relevant information for a given speaking tour remains a problem for open-domain dialogue systems. For the latter, the possible interactions between a user and an agent are a priori infinite and indefinite. The possibility of an incorrect response from the agent to the user therefore remains high. For goal-oriented systems, the problem is considered solved, but in our experience no system shows remarkable robustness when evaluated in real situations. In this article, we present a state of the art of agent learning methods and different conversational agent models. In our opinion, one of the agent's avenues for improvement lies in his memory, because the latter (often represented by the triplet: current speaking time, dialogue history and knowledge base) is not yet modelled with enough precision. By providing the agent with a cognitively inspired memory model, we believe we can increase the performance of a goal-oriented dialogue system in real life situations, by using automatic learning algorithms with an antagonistic approach in support of a new memory model for the agent. ABSTRACT Goal-oriented dialogue systems: a recent overview and research prospects Determining which information is relevant for a given dialogue turn is still a problem for open-domain dialogue systems since the interaction between a user and the system is a priori infinite and indefinite. Thus, the possibility for error remains high for the dialog system. Although the problem is considered solved, to our experience, no system has shown outstanding performance when confronted to "real-world" situation. We review the state-of-the-art of machine learning approaches used for dialogue system development and present the various models of agent architecture. We claim that one of the main directions for improvement is the agent architecture. Often the agent's memory (represented by the triple: dialogue turn, dialogue history and knowledge base) is not modeled accurately enough. Thus we decided to investigate more cognitive oriented models of memory. We hypothesize that combining machine learning approaches in an adversarial setup with a new memory model for the agent would result in improved performances for goal-oriented dialogue systems. KEYWORDS: dialogue systems, learning by reinforcement, memory model, learning by antagonism, goal-oriented systems.; La gestion et la sélection des informations pertinentes pour un tour de parole donné restent un problème pour les systèmes de dialogue à domaine ouvert. Pour ces derniers, les interactions possibles entre un utilisateur et un agent sont a priori infinies et indéfinies. La possibilité d'une réponse erronée de l'agent à l'utilisateur demeure donc élevée. Pour les systèmes orientés-but, le problème est considéré comme résolu, mais d'après notre expérience aucun système ne montre une robustesse remarquable lorsqu'il est évalué en situation réelle. Dans cet article, nous dressons un état de l'art des méthodes d'apprentissage de l'agent et des différents modèles d'agent conversationnel. Selon nous, l'une des pistes d'amélioration de l'agent réside dans sa mémoire, car cette dernière (souvent représentée par le triplet : tour de parole courant, historique du dialogue et base de connaissances) n'est pas encore modélisée avec assez de précision. En dotant l'agent d'un modèle de mémoire d'inspiration cognitive, nous pensons pouvoir augmenter les performances d'un système de dialogue orienté-but en situation réelle, par l'emploi d'algorithmes d'apprentissage automatique avec une approche antagoniste en support d'un nouveau modèle de mémoire pour l'agent. ABSTRACT Goal-oriented dialog systems : a recent overview and research prospects Determining which information is relevant for a given dialog turn is still a problem for open-domain dialog systems since the interaction between a user and the system is a priori infinite and indefinite. Thus, the possibility for error remains high for the dialog system. Although the problem is considered solved, to our experience, no system has shown outstanding performance when confronted to "real-world" situation. We review the state-of-the-art of machine learning approaches used for dialog system development and present the various models of agent architecture. We claim that one of the main directions for improvement is the agent architecture. Often the agent's memory (represented by the triple : dialog turn, dialogue history and knowledge base) is not modeled accurately enough. Thus we decided to investigate more cognitive oriented models of memory. We hypothesize that combining machine learning approaches in an adversarial setup with a new memory model for the agent would result in improved performances for goal-oriented dialog systems. MOTS-CLÉS : systèmes de dialogue, apprentissage par renforcement, modèle de mémoire, appren-tissage par antagonisme, systèmes orientés-buts.