Search

International audience; Les lasers à cascade quantique sont des sources semiconductrices basées sur des transitions intersous-bandesau sein de la bande de conduction. Pouvant émettre sur une large plage de longueurs d’onde allantdu moyen infra-rouge au terahertz, ils sont devenus une source privilégiée pour des applications telles que laspectroscopie de gaz, les communications en espace libre ou les contre-mesures optiques [1]. Une analyse temporelledes fluctuations de la puissance optique émise a récemment révélé un comportement chaotique du laser à cascadequantique soumis à une force de rappel optique [3]. Cette étude expérimentale a mis en évidence un scénariode bifurcation impliquant une déstabilisation à la fréquence de cavité externe, puis un chaos sous la forme defluctuations basses fréquences. Ce scénario est typique des lasers à gaz de type classe A [4] mais n’avait jamaisété observé jusqu’alors dans un laser semiconducteur, où la première déstabilisation est basée sur la fréquence desoscillations de relaxation [5]. Ces observations expérimentales ont été validées par une étude numérique exploitantles équations de Lang et Kobayashi [5], en y incorporant les paramètres spécifiques du laser à cascade quantiqueétudié. Cependant, il est important de comprendre si ce scénario de bifurcation peut se reproduire sur d’autresstructures ou dans d’autres conditions expérimentales.L’objectif de ce travail est d’étudier numériquement la dynamique à retard du laser à cascade quantique en fonctionde trois paramètres clés : le courant de pompe, la longueur de la cavité externe et le facteur d’élargissement spectraldu laser (facteur α), intrinsèque à la structure. Les simulations obtenues montrent des déstabilisations proches decelle observée expérimentalement, avec l’apparition d’oscillations à la fréquence de cavité externe. Cependant, sil’on augmente le courant de pompe ou que l’on diminue la valeur du facteur α, la première bifurcation de Hopf, àlaquelle ces oscillations apparaissent, se produit à des taux de réinjection significativement plus élevés, et la zonechaotique tend à disparaître. De plus, si l’on augmente la longueur de cavité externe, la bulle de chaos apparaîtà de plus faibles taux de réinjection, bien que la première bifurcation de Hopf reste à des taux de réinjectionsimilaires, mais cette zone se réduit fortement, et finalement disparait.En conclusion, cette étude numérique montre que le scénario de déstabilisation caractérisée par des fluctuationsbasses fréquences est bien reproductible dans d’autres conditions expérimentales ou sur d’autres types de structuresà cascade quantique. Cependant, le taux de réinjection auquel a lieu la première bifurcation de Hopf ou l’étenduede la zone chaotique peuvent varier, et il est donc fondamental d’étudier chaque cas afin d’éviter une déstabilisationou émission chaotique parasite.

This paper deals with the improvment of the Teager-Kaiser energy algorithm (TKEA), designed to track a frequency component corrupted by an additive noise. Methods based on the original TKEA are not robust at all when corrupted by such perturbations. It is proposed here a method based on the Noise Assisted Data Analysis (NADA) paradigm. It is shown that it allows to significantly increase this robustness., Cet article propose une technique pour améliorer l'opérateur de Teager-Kaiser (OTK) utilisé pour le suivi d'une composante fré-quentielle perturbée par un bruit additif. Les méthodes basées sur l'OTK classiques sont très peu robustes aux diverses perturbations possibles. Une méthode d'assistance par le bruit est mise en place. Nous montrons qu'elle permet de statistiquement améliorer cette robustesse. Abstract – This paper deals with the improvment of the Teager-Kaiser energy algorithm (TKEA), designed to track a frequency component corrupted by an additive noise. Methods based on the original TKEA are not robust at all when corrupted by such perturbations. It is proposed here a method based on the Noise Assisted Data Analysis (NADA) paradigm. It is shown that it allows to significantly increase this robustness.

International audience

International audience; Five-period AlN/BAlN heterostructure containing boron as high as 11% has been successfully grown by MOVPE. Good periodicity of two alternative layers has been observed by both SIMS profile and Z-contrast HAADF-STEM images. The BAlN layers demonstrate columnar polycrystalline feature. The BAlN layers exhibit an emission peak by CL and absorption edge in transmission spectra at around 260 nm. The results enable the development of BAlGaN based multi-layered heterostructure for UV and deep UV applications.

International audience

International audience

International audience; Boron containing III-nitrides are attractive system for deep-UV LEDs and LDs because of their wide bandgaps and flexible lattice. However the crystallinity and boron content have been limited due to large mismatch between BN and other nitrides. In this work, BAlN layers with boron composition from 1% to 5% were successfully grown on AlN template substrates by low-pressure organometallic vapor phase epitaxy. The samples were grown at 650˚C and then annealed at 1020˚C for recrystallization. Growth techniques such as temperature, growth time and TEB/III ratio in the gas phase were investigated. High quality BAlN layers were grown using flow-modulate epitaxy method that allows to enhance surface migration of boron atoms. 70 nm-thick layers show a good surface morphology. For the first time, clear XRD peak relating to 5% boron for this new material was observed, which suggests the formation of single-phase solid solution. Adding more boron to the AlN produced a shift in the peak positions to greater angles. Further results by TEM and optical characterizations will be presented. This new material is promising for deep-UV applications and gives more freedom for bandgap engineering of multi-structure devices.

International audience

National audience

International audience; We present our preliminary work developing a reservoir computing platform using field-programmablegate arrays (FPGAs), with the ultimate goal of detecting features of complex systems. We show thatthe three basic properties required for reservoir computing, namely that different input states aremapped to different reservoir states, input states that are close together are mapped to identicalreservoir states, and a fading memory, can be realized even in moderately-sized Boolean networkssynthesized on an FPGA. The networks are realized as ring oscillators consisting of multiple-inputXOR gates that accept an input (Boolean) voltage and time-delayed feedback. These oscillators exhibitlong chaotic transients when the input voltage is flipped, and it is found that these transientsoffer promising dynamics for reservoir computing.

International audience; We propose a new experimental realization of a periodic oscillator that we term Boolean phase oscillator (BPO) [1]. One of the important features of our design is that it allows us to tune the coupling strength using only Boolean signals. Our BPO is realized on a field-programmable gate array(FPGA) in its autonomous mode of operation. It consists of a ring oscillator with state-dependent delay taking two discrete values depending on an external coupling signal. We demonstrate experimentally and numerically that two uni- and bidirectionally coupled BPOs can exhibit phase synchronization. We characterize the synchronization regions as a function of the frequency detuning between two BPOs and a quantity related to the variation of state-dependent delay. We notice that the synchronization regions become larger when the second quantity is increased. This allows us to conclude that this quantity is an effective coupling strength in BPOs. We gratefully acknowledge the financial support of the U.S. Army Research Office Grant W911NF12-1-0099 and National Science Foundation IGERT Grant DGE-1068871.

International audience; Chimera states in dynamical networks have been found recently in several experiments using chemical, optical, and mechanical systems. However, most of these experiments still involve a computer to mediate the network coupling or rely on rather simple network topologies. Here, we overcome these limitations with novel Boolean realizations of coupled oscillators termed Boolean phase oscillators that are realized experimentally with embedded logic circuits. They include state-dependent delay, which allows weak phase adjustments when subject to coupling from neighboring nodes [1]. Using a field-programmable gate array (FPGA), we implement a ring of 100 non-locally coupled Boolean phase oscillators and measure experimentally frequency and phase in the network. We observe typical experimental signatures of chimera states, where parts of the network are synchronized and other parts are desynchronized.

International audience

International audience

International audience

International audience; We first report on recently experimentally discovered two-mode dynamics in solitary Vertical-Cavity Surface-Emitting Lasers (VCSELs), which manifests the microscopic spin-flip processes taking place in the semiconductor quantum dots (QD)[1] We demonstrate that such dynamics leads to polarization chaos, a new phenomenon for class B semiconductor lasers. Our QD-VCSEL emits linearly polarized light at threshold and undergoes a pitchfork bifurcation when increasing injection current that creates two symmetrical elliptically polarized solutions ( Fig. 1a). Then, a sequence of bifurcations (Fig. 1b) brings it to two single-scroll chaotic attractors oscillating around the now unstable elliptically polarized steady-states (Fig.1c). Finally, the two chaotic attractors merge into a double-scroll "butterfly" chaotic attractor (Fig.1d). Experimentally, this "butterfly" attractor is observed as a mode hopping between two elliptically polarized modes (Fig.1e,f) [2]. The dwell time, i.e. the average time between two successive jumps exponentially decreases with injection current which is opposite to noise-induced polarization mode hopping in quantum well VCSELs [3]. Next, we review the features of VCSEL nonlinear dynamics induced by optical injection and optical feedback that are due to the lack of polarization anisotropy and to multiple transverse mode operation. In case of optical injection we demonstrate a new Hopf bifurcation on a two-polarization-mode solution that delimits the injection locking region and a new resonance tongue for large positive detunings that is due to polarization switching and injection locking of first-order transverse mode [4]. For the case of optical feedback in VCSELs[5] we show that for long and short external cavities the Lang-Kobayashi model is in good agreement with experiments on polarization switching, mode hopping and antiphase oscillatory dynamics at the delay time. These last dynamics modify polarization residence-time distribution making it oscillatory. We demonstrate numerically and experimentally coherence resonance in such bistable, time-delayed optical system. We present results on low frequency fluctuation and pulse-package dynamics in VCSELs and discuss the role of light polarization.

International audience

Inverse reinforcement learning (IRL) aims at estimating an unknown reward function optimized by some expert agent from interactions between this expert and the system to be controlled. One of its major application fields is imitation learning, where the goal is to imitate the expert, possibly in situations not encountered before. A classic and simple way to handle this problem is to see it as a classification problem, mapping states to actions. The potential issue with this approach is that classification does not take naturally into account the temporal structure of sequential decision making. Yet, many classification algorithms consist in learning a \textit {score function}, mapping state-action couples to values, such that the value of the action chosen by the expert is higher than the others. The \textit{decision rule} of the classifier maximizes the score over actions for a given state. This is curiously reminiscent of the \textit{state-action value function} in reinforcement learning, and of the associated \textit{greedy policy}. Based on this simple statement, we propose two IRL algorithms that incorporate the structure of the sequential decision making problem into some classifier in different ways. The first one, SCIRL (Structured Classification for IRL), starts from the observation that linearly parameterizing a reward function by some features imposes a linear parametrization of the Q-function by a so-called feature expectation. SCIRL simply uses (an estimate of) the expert feature expectation as the basis function of the score function. The second algorithm, CSI (Cascaded Supervised IRL), applies a reversed Bellman equation (expressing the reward as a function of the Q-function) to the score function outputted by any score-based classifier, which reduces to a simple (and generic) regression step. These two algorithms come with theoretical guarantees and perform competitively on toy problems.

International audience; Dialogue management optimisation has been cast into a plan- ning under uncertainty problem for long. Some methods such as Reinforcement Learning (RL) are now part of the state of the art. Whatever the solving method, strong assumptions are made about the dialogue system properties. For instance, RL assumes that the dialogue state space is Markovian. Such con- straints may involve important engineering work. This paper introduces a more general approach, based on fewer modelling assumptions. A Black Box Optimisation (BBO) method and more precisely a Particle Swarm Optimisation (PSO) is used to solve the control problem. In addition, PSO allows taking ad- vantage of the parallel aspect of the problem of optimising a system online with many users calling at the same time. Some preliminary results are presented.

Cet article propose une étude comparative entre l'Apprentissage par Renforcement Inverse (ARI) et l'Apprentissage par Imitation (AI). L'ARI et l'AI sont deux cadres de travail qui utilisent le concept de Processus Décisionnel de Markov (PDM) et dans lesquels nous cherchons à résoudre le problème d'Apprentissage par Démonstrations (AD). L'AD est un problème où un agent appelé ap- prenti cherche à apprendre à partir de l'observation des démonstrations d'un autre agent appelé expert. Dans le cadre de travail de l'AI, l'apprenti essaie d'apprendre directement la politique de l'expert alors que dans le cadre de l'ARI, l'apprenti essaie d'apprendre la récompense qui explique la politique de l'expert. Cette récompense est ensuite optimisée pour imiter l'expert. On peut donc légitimement se demander s'il y a un intérêt à estimer une récompense qui devra ensuite être optimisée ou si l'estima- tion d'une politique est suffisante. Cette question assez naturelle n'a pas encore été réellement traitée dans la littérature pour l'instant. Ici, des réponses partielles à la fois d'un point de vue théorique et pra- tique sont produites. Mots-clés : Apprentissage par Renforcement Inverse, Apprentissage par Imitation, Apprentissage par Démonstrations.

International audience

Summary form only given. When using quantum dot (QD) as an active region for semiconductor laser, one can confine more than one energy level, thus allowing laser effect to take place in the ground state (GS), the excited state (ES) or both simultaneously [1]. In this contribution we theoretically consider the impact of optical feedback on the mode competition between GS and ES emission. Using an analytical approach, continuation techniques and numerical integration we provide a detailed analysis of the steady-states and bifurcation scenarios predicted, thus bringing new light on recent experiments [2].

Cet article traite le problème d'Apprentissage par Démonstrations (AD) dans lequel un agent appelé apprenti cherche à apprendre à partir des démonstrations d'un autre agent appelé expert. Pour aborder ce problème assez général, il est commun d'adopter le paradigme des Processus Décisionnels de Markov (PDM) qui est approprié pour les problèmes de prises de décisions séquentielles. Dans la littérature, il y a principalement deux façons de traiter ce problème (en faisant appel aux PDM) qui sont l'Apprentissage par Imitation (AI) où l'apprenti cherche directement à imiter la politique de l'expert et l'Apprentissage par Renforcement Inverse (ARI) où l'apprenti essaye d'apprendre une récompense qui pourrait expliquer la politique de l'expert. Ici, nous introduisons un paradigme inédit, appelé cadre de travail des politiques d'ensembles ( set-policy framework ), pour lequel il y a un lien naturel entre les méthodes d'AI et d'ARI. Ce paradigme permet de dériver des nouveaux algorithmes qui nécessitent uniquement la connaissance de couples état-action experts et d'exemples de transitions du PDM. Des expériences sont réalisés sur un problème qui fait référence (un simulateur de trafic routier) et sur une tâche plus générique (les Garnets) qui permet une comparaison plus générale des algorithmes. Mots-clés : Apprentissage par Renforcement Inverse, Apprentissage par Imitation.

International audience

International audience; This paper introduces the rllib as an original C++ template-based library oriented toward value function estimation. Generic programming is promoted here as a way of having a good t between the mathematics of reinforcement learning and their implementation in a library. Main concepts of rllib are presented, as well as a short example.

International audience; In this contribution, we demonstrate generation of polarization chaos, i.e. chaotic mode hopping between two elliptically polarized states, in a free-running quantum dot vertical-cavity surface-emitting laser and provide a theoretical framework of the phenomenon.

International audience

International audience; This paper presents a new, simple, and efficient data structure, namely, the \b{o}rdered \b{s}egment (OS): a mono dimensional string array that we have been using in our classification of big data. The essential idea in construction of OS is to make use of the redundancies that abound user-data. OS enables us to performs efficient retrieval, insertions and deletions of data. The theoretical and experimental observations show that the method presented is more practical than existing ones considering the use of dynamic string sets for the classifications of huge user-files.

International audience; This paper provides a comparative study between Inverse Reinforcement Learning (IRL) and Apprenticeship Learning (AL). IRL and AL are two frameworks, using Markov Decision Processes (MDP), which are used for the imitation learning problem where an agent tries to learn from demonstrations of an expert. In the AL Framework, the agent tries to learn the expert policy whereas in the IRL Framework, the agent tries to learn a reward which can explain the behavior of the expert. This reward is then optimized to imitate the expert. One can wonder if it is worth estimating such a reward, or if estimating a Policy is sufficient. This quite natural question has not really been addressed in the literature right now. We provide partial answers, both from a theoretical and empirical point of view.

International audience; This paper adresses the inverse reinforcement learning (IRL) problem, that is inferring a reward for which a demonstrated expert behavior is optimal. We introduce a new algorithm, SCIRL, whose principle is to use the so-called feature expectation of the expert as the parameterization of the score function of a multi-class classifier. This approach produces a reward function for which the expert policy is provably near-optimal. Contrary to most of existing IRL algorithms, SCIRL does not require solving the direct RL problem. Moreover, with an appropriate heuristic, it can succeed with only trajectories sampled according to the expert behavior. This is illustrated on a car driving simulator.

Data driven methods have long been used in Automatic Speech Recognition (ASR) and Text-To-Speech (TTS) synthesis and have more recently been introduced for dialogue management, spoken language understanding, and Natural Language Generation. Machine learning is now present end-to-end in Spoken Dialogue Systems (SDS). However, these techniques require data collection and annotation campaigns, which can be time-consuming and expensive, as well as dataset expansion by simulation. In this book, we provide an overview of the current state of the field and of recent advances, with a specific focus on adaptivity.

Statistical approaches to dialogue management have steadily increased inpopularity over the last decade. Recent evaluations of such dialogue managershave shown their feasibility for sizeable domains and their advantage in terms ofincreased robustness. Moreover, simulated users have shown to be highly beneficialin the development and testing of dialogue managers and in particular, fortraining statistical dialogue managers. Learning the optimal policy of aPOMDP dialogue manager is typically done using the reinforcement learning(RL), but with the RL algorithms that are commonly used today, thisprocess still relies on the use of a simulated user. Data-driven approaches touser simulation have been developed to train dialogue managers on morerealistic user behaviour. This chapter provides an overview of user simulationtechniques and evaluation methodologies. In particular, recent developments inagenda-based user simulation, dynamic Bayesian network-based simulations andinverse reinforcement learning-based user simulations are discussed indetail. Finally, we will discuss ongoing work and future challenges for usersimulation.

International audience

International audience; Spoken dialogue systems provide an opportunity for man machine interaction using spoken language as the medium of interaction. In recent years reinforcement learning-based dialogue policy optimization has evolved to be state of the art. In order to cope with the data requirement for policy optimization and also to evaluate dialogue policies user simulators are introduced. Almost all existing data driven methods for user modelling aims at simulating some generic user behavior from some reference dialogue corpus. However, this corpus consists of dialogues from multiple users and thus exhibit different user behaviors. In this paper we explore the possibility of identifying and simulating different user behaviors observed in the corpus. For this purpose inverse reinforcement learning-based user simulation method is employed. Using experimental results, we validate the effectiveness of the proposed method for building multiple behavior specific user simulators.

International audience

International audience

International audience; This paper addresses the Inverse Reinforcement Learning (IRL) problem which is a particular case of learning from demonstrations. The IRL framework assumes that an expert, demonstrating a task, is acting optimally with respect to an unknown reward function to be discovered. Unlike most of existing IRL algorithms, the proposed approach doesn't require any of the following: complete trajectories from the expert, a generative model of the environment, the knowledge of the transition probabilities, the ability to repeatedly solve the forward Reinforcement Learning (RL) problem, the expert's policy anywhere in the state space. Using a classi cation approach in which the structure of the underlying Markov Decision Process (MDP) is implicitly injected, we end-up with an e cient subgradient descent-based algorithm. In addition, only a small amount of expert demonstrations (not even in the form of trajectories but simple transitions) is required. Keywords: inverse reinforcement learning, structured multi-class classi cation

International audience; We make an in-depth analysis of a bifurcation scenario that leads to chaotic hopping between two elliptically polarized modes in VCSELs. Our work brings new light on recent experiments using quantum dot VCSELs.

National audience; Dans cet article, nous présentons une méthode générale pour optimiser un tuteur intelligent dans le domaine de l'acquisition d'une seconde langue. Plus particulièrement, le processus d'optimisation a pour but de trouver une stratégie qui propose la meilleure séquence de phases d'évaluation et d'enseignement afin de maximiser l'augmentation des connaissances de l'apprenant. La principale caractéristique de la méthode proposée est qu'elle est capable d'apprendre la meilleure stratégie à partir d'un jeu fixe de données, collectées à partir d'une stratégie définie à la main. Ainsi, aucun modèle, ni cognitif ni probabiliste de l'apprenant, n'est nécessaire. Seules sont requises des observations du comportement de l'apprenant alors qu'il interagit avec un système non-optimal. Pour ce faire, un algorithme de programmation dynamique approchée en mode hors-ligne est utilisé : l'algorithme LSPI (Least Square Policy Iteration). Des résultats obtenus avec des données simulées semblent prometteurs.

http://cap2012.loria.fr/pub/Papers/13.pdf; National audience; Cette contribution traite du problème de l'apprentissage par imitation par le biais de l'apprentissage par renforcement inverse (ARI). Dans ce contexte, un expert accomplit une tâche qu'un agent artificiel doit essayer de reproduire. L'ARI part du postulat que l'expert optimise avec succès une fonction de récompense ; le problème consiste à deviner cette fonction à partir de traces du comportement de l'expert. Les algorithmes d'ARI existants nécessitent une ou plusieurs des conditions suivantes pour fonctionner : trajectoires complètes de la part de l'expert, un modèle génératif pour les estimations de type Monte-Carlo, la connaissance des probabilités de transition, la capacité de résoudre le problème direct (celui de l'apprentissage par renforcement) de manière répétée ou l'accès à la strategie complète de l'expert. Notre contribution consiste en un nouvel algorithme d'ARI levant l'ensemble de ces contraintes. En utilisant une méthode supervisée dans laquelle nous introduisons implicitement la structure du processus décisionnel de Markov (PDM) sous-jacent, nous créons un algorithme basé sur une descente de sous-gradient, possèdant une faible complexité tant en échantillons que calculatoire et surtout ne nécessitant pas la résolution du problème direct.

http://cap2012.loria.fr/pub/Papers/10.pdf; National audience; Trouver des contrôleurs optimaux pour des systèmes stochastiques est un problème particulièrement difficile abordé dans les communautés d'apprentissage par renforcement et de contrôle optimal. Le paradigme classique employé pour résoudre ces problèmes est celui des processus décisionnel de Markov. Néanmoins, le problème d'optimisation qui en découle peut être difficile à résoudre. Dans ce papier, nous explorons l'utilisation de l'optimisation par essaim particulaire pour apprendre des contrôleurs optimaux. Nous l'appliquons en particulier à trois problèmes classiques : le pendule inversé, le mountain car et le double pendule.

International audience; Spoken Dialogue Systems (SDS) are natural language interfaces for human-computer interaction. User adaptive dialogue management strategies are essential to sustain the naturalness of interaction. In recent years data-driven methods for dialogue optimization have evolved to be a state of art approach. However these methods need vast amounts of corpora for dialogue optimization. In order to cope with the data requirement of these methods, but also to evaluate the dialogue strategies, user simulations are built. Dialogue corpora used to build user simulation are often not annotated in user's perspective and thus can only simulate some generic user behavior, perhaps not representative of any user. This paper aims at clustering dialogue corpora into various groups based on user behaviors observed in the form of full dialogues.