Posts tagged "apprentissage"

Context Presentation

Anomaly detection in real fluid distribution applications is a difficult task, especially when we seek to accurately detect different types of anomalies and possible sensor failures. Our case study is based on a real context: sensor data from the SGE (Rangueil campus management and operation service in Toulouse).

We propose an automatic pattern-based method for anomaly detection in time-series called Composition-based Decision Tree (CDT). We use a modified decision tree and Bayesian optimization to avoid manual tuning of hyper-parameters. Our method uses sequences of patterns to identify remarkable points corresponding to multiple anomalies. The compositions of patterns existing into time-series are learned through an internally generated decision tree and then simplified using Boolean algebra to produce intelligible rules.

Our approach automatically generates decision rules for anomaly detection. All our experiments were carried out on real and synthetic data. We show that our method is precise for classifying anomalies compared to other methods. It also generates rules that can be interpreted and understood by experts and analysts, who they can adjust and modify.

Keywords

Anomaly detection, Time-series, Machine learning, Classification rules

Scientific goals

•    To detect different types of anomalies observed in real deployment

•    To generate interpretable rules for anomaly detection

•    To use learning methods for anomaly detection on static and continuous data

Contact

ines.ben-kraiem@irit.fr

Context Presentation

The 5G evolution is the key driving factor that provides promising support for efficient V2X (Vehicle to everything) communications. Various applications with different requirements have been designed on vehicular networks to improve the driving experience by offering multiple services. In this thesis, we are interested by safety-critical applications and focus on collision avoidance systems between vehicles and pedestrians. This type of applications imposes strict reliability, wide connectivity, and a minimum end to end delay requirements. The high dynamic topology and the different types of communication technologies raise up challenges such as scalability, heterogeneity, and high traffic load to handle.

We have chosen MEC (Multiple Access/Mobile Edge Computing) that offers direct communication exchange between the mobile nodes (vehicles and pedestrians) and the network Infrastructure and used the Network Slicing mechanism to separate the critical vehicular traffic with high priority and strict requirements from other traffic. In our scenario a network congestion risk while vehicles and pedestrians send their BSM and CAM messages to the network infrastructure. Indeed, periodical transmissions could be unnecessary or harmful especially when the network density is high. Therefore, an intelligent scheme should be developed to adapt the transmission frequency of these messages to a network server without overloading the network, while considering the dynamic network state.

The main contribution of our work is to exploit the rich environment Information and analyze those data to take future decisions and to predict the network state to decrease the traffic load and finally to meet the objective requirements. Thus, we argue for the use of advanced Machine Learning techniques to learn from the available network data and take the appropriate action by choosing the best parameters' configuration.

Keywords

Autonomous vehicles, V2X Communications, Resource Allocation, Network Slicing, MEC, Machine Learning

Scientific goal

•    Improve the performance of V2X Communications in the context of critical road safety applications.

Contacts

Chaima.Zoghlami@irit.fr, Rahim.Kacimi@irit.fr, Riadh.Dhaou@irit.fr

Context Presentation

Connected autonomous vehicles of level 3 are vehicles in which the human driver delegates driving control in specific situations. During these situations, it may be necessary for the human to regain control of the driving activity.

The main objective of this thesis is to develop a system for the safe and efficient transition of two-way control between the human and the autonomous vehicle.For this, the system must identify by self-observation and in real time situations in which the current driver will no longer be able to ensure driving. He must also provide a context for assessing the criticality of the situation as quickly as possible in order to anticipate and react to it as best as possible. The driving context is composed of indicators that characterize the elements that describes part of the driving process. The system is based on self-adaptive multi-agent learning systems.

Keywords

self-adaptive multi-agent systems, autonomous vehicle

Scientific goals

•    Dynamic learning using multi-agent systems

•    Generic approach to supervise the activity of a system

•    Insure the acceptability of the system by human driver

Contacts

kristell.aguilar-alarcon@irit.fr, marie-pierre.gleizes@irit.fr, loic.caroux@univ-tlse2.fr

Context Presentation

Cyber-physical and ambient systems surround the human user with services at her/his disposal. From these services, complex composites services, tailored to the user preferences and the current situation, can be composed automatically and on the fly.

To produce the knowledge necessary for automatic composition in the absence of both prior expression of the user's needs and specification of a process or a composition model, we develop a generic solution based on online reinforcement learning. It is decentralized within a multi-agent system in charge of the administration and composition of the services, which learns incrementally from and for the user.

Keywords

Ambient intelligence; Service Discovery, Selection and Composition; Multi-agent System; Machine learning; Smart city; neOCampus

 Scientific goals

•    Design a decentralized and distributed system that learns and decides on compositions.

•    Consider user preferences and context.

Contacts

walid.younes@irit.fr, jean-paul.arcangeli@irit.fr, sylvie.trouilhet@irit.fr, francoise.adreit@irit.fr

Context Presentation

The current digital transformation requires the creation of autonomous applications that can be adapted to complex, dynamic, heterogeneous and unpredictable environments. These systems must be equipped with proactive learning capabilities. To this end, Self-Adaptive Multi-Agent principle allows the decentralization and self-observation of the learning process. Each knowledge granule is an autonomous agent that cooperates with its neighbors to improve learning from exogenous and endogenous feedbacks. Detecting and solving concurrences, conflicts and incompetencies leads to active endogenous learning.This work on an adaptive decentralized learning mechanism will be applied on application domains such as robotics, autonomous vehicles and smart cities.

Figure 1: « Schema of the Learning

Figure 2: « Implementation Example Multi-Agent System » on an Industrial Robot »

Scientific Goals

- Design a Self-Learning System

- Lifelong and Endogenous Learning

- Genericity and Scalability

Keywords

Self-Adaptive Learning, Endogenous Learning, Adaptive Multi-Agent Systems, Artificial Intelligence

Contacts

bruno.dato@irit.frfrederic.migeon@irit.fr marie-pierre.gleizes@irit.fr

Context Presentation

The number of sensors in buildings is constantly increasing, thanks to more accessible costs and the obvious interest of their use for optimized management. In this thesis we are interested in the use of data from these sensors to detect anomalies in buildings. These data, which are very numerous, can be of unknown and heterogeneous types.An anomaly is defined as unexpected and undesirable behaviour in a system and may depend on the context. In order to be able to deploy an anomaly detection system as widely as possible, it is necessary to create a decision support tool for energy experts. To address these issues, a system based on cooperative multi-agent systems implementing AMAS theory is being developed that allows anomalies to be detected by supervised learning. The anomaly detection system must take advantage of the feedback from one or more experts who label certain instances as anomalies or non-anomalies. These feedback are used for learning. The system we develop allows the addition or removal of new sensors without interrupting the detection of anomalies.

The system classifies situations

Scientific Goals

- Improve energy efficiency

- Detect anomalies in real time

- Learn continuously from the expert feedback

Keywords

Multi-Agent Systems, Smart Buildings, Internet of Things, Supervised Learning, Anomaly Detection

Contacts

Maxime.houssin@irit.fr

Présentation du contexte

La réalité d'un campus intelligent ou plus généralement d'une ville intelligente passe par une observation régulière de l'environnement par des capteurs ad-hoc, afin d’agir dans l’environnement avec des dispositifs automatiques pour améliorer le bien-être des usagers. Ces capteurs permettent d’obtenir une connaissance des activités humaines et des conditions dans lesquelles ces activités sont menées, mais le déploiement d'un grand nombre de capteurs peut être coûteux. Les coûts sont principalement liés à l'installation, la maintenance et les infrastructures de capteurs dans les bâtiments existants. Pour ces raisons, l’objectif de cette thèse vise à réduire ces coûts en utilisant quotidiennement des milliers d’informations partielles et intermittentes provenant de smartphones des usagers du campus de l’Université Toulouse III Paul Sabatier. Ces traitements sont fondés sur une technologie d’Intelligence Artificielle par systèmes multi-agents coopératifs.

Figure 1 : «On utilise les informations des dispositifs intermittents et mobiles pour fournir des estimations précises»

Objectifs scientifiques

- Apprendre à partir de données brutes, imprécises et intermittentes sans feedback.

- Fournir les informations en continu, même en l’absence de données de smartphone des usagers.

- Utiliser une approche hybride de l’Internet des objets qui mixe capteurs réels et capteurs virtuels.

Mots clés

Systèmes multi-agents auto-adaptatifs, fusion de données, apprentissage, smart campus

Contacts

Davide Andrea Guastella, Valérie Camps, Marie-Pierre Gleizes, {davide.guastella, camps, gleizes}@irit.fr

Context Presentation

The usual approach to ambient intelligence is an expert modeling of the devices present in the environment, describing what each does and what effect it will have. When seen as a dynamic and noisy complex systems, with the efficiency of devices changing and new devices appearing, this seems unrealistic. We propose a generic multi-agent (MAS) learning approach that can be deployed in any ambient environment and collectively self-models it. We illustrate the concept on the estimation of power consumption. The agents representing the devices adjust their estimations iteratively and in real time so as to result in a continuous collective problem solving. This approach will be extended to estimate the impact of each device on each comfort (noise, light, smell, heat...), making it possible for them to adjust their behaviour to satisfy the users in an integrative and systemic vision of an intelligent house we call QuaLAS: eco-friendly Quality of Life in Ambient Sociotechnical systems.

Figure 1: « eco-friendly Quality of Life in Ambient Sociotechnical systems »

Scientific Goals

- multi-learning in a highly dynamic environment,

- conditions for scaling up,

- sensitivity to disturbances and noisy signals

- convergence speed

Keywords

neOCampus, file, presentation, innovation, ambient Intelligence multi-agent systems, complex systems, collective learning.

Contacts

fabrice.crasnier@irit.frmarie-pierre.gleizes@irit.fr - jean-pierre.george@irit.fr

La transformation numérique actuelle demande la création d’applications autonomes et adaptables à des environnements ouverts, dynamiques, hétérogènes et imprédictibles. Ainsi, il faut doter ces systèmes de capacités d’apprentissage proactives. Pour cela, les Systèmes Multi-Agents Auto-Adaptatifs permettent de décentraliser le processus d’apprentissage en agentifiant les connaissances. Chaque granule de connaissance est alors autonome et coopère avec ses voisins afin de détecter des incohérences, des faiblesses ou des nouvelles zones à explorer pour perfectionner l’apprentissage. 

L’apprentissage proactif des granules autonomes conduit à créer de nouvelles connaissances par feedback endogène (sans feedback explicite de l’environnement). Les perspectives de ce travail sont de concevoir un mécanisme d’apprentissage adaptatif utilisable pour des applications diverses (robotique, véhicules autonomes, domotique …).

Agentification et auto-organisation des granules de connaissance (pavés verts et bleus), identification 

des incohérences et insuffisances (pavés rouges et zones blanches), transfert et généralisation

Objectifs scientifiques

Les objectifs de la thèse sont :

 Conception d’un système auto-apprenant

 Apprentissage tout au long de la vie

 Génération d’objectifs et de motivations propres

 Généricité et passage à l’échelle

Contacts

bruno.dato@irit.fr, frederic.migeon@irit.fr, marie-pierre.gleizes@irit.fr

Nous vivons dans un environnement qui regorge de systèmes artificiels dont le but est de nous assister dans notre quotidien. Toutes ces applications sont développées afin de servir un but précis défini avant leur conception. Cependant, il est impossible de prévoir à l’avance toutes les interactions qu’elles auront avec leur environnement. De plus, ces systèmes sont plongés dans un monde dynamique dans lequel divers dispositifs peuvent apparaître et disparaître. Face à ces besoins, il est légitime de penser que dans un futur proche, nous ne serons plus en mesure de concevoir et programmer tous ces systèmes, ils devront alors apprendre à être utiles de façon autonome et réactive.

Aujourd’hui, pour qu’un système puisse apprendre une tâche ou un service, il a besoin qu’un oracle lui dise ce qu’il est pertinent de retenir, on parle d’apprentissage par démonstrations. Cet apprentissage par démonstrations peut se traduire par une personne qui pilote un robot pour lui montrer une tâche à effectuer. Qu’en est-il si le système se retrouve dans une situation qu’il ne connaît pas ou s’il n’y a personne pour faire cette démonstration ?

Objectifs scientifiques

Ce stage a pour but d’aller au-delà de cet apprentissage par démonstrations et concevoir des méthodes d’auto-apprentissage utilisant seulement les actions et perceptions d’un système afin d’en apprendre un modèle de contrôle. C’est-à-dire de doter un système de capacités d’auto-observation lui permettant d’apprendre quelles conséquences ont ses actions sur ses perceptions sans faire d’hypothèses sur la nature de celles-ci. Ce modèle lui permettra alors d’évoluer dans la représentation qu’il a de son environnement, autrement dit de ses perceptions.

Un autre objectif de ce sujet de recherche est la généricité, les méthodes d’auto-apprentissage devront pouvoir s’appliquer à n’importe quel système doté d’actions et de perceptions. Enfin, le dernier objectif est le passage à l’échelle, cet auto-apprentissage doit pouvoir s’adapter à de grandes quantités d’actions et de perceptions.

Contacts

- Bruno Dato (IRIT) : bruno.dato@irit.fr

- Nicolas Verstaevel (IRIT) : nicolas.verstaevel@irit.fr 

- Marie-Pierre Gleizes (IRIT) : marie-pierre.gleizes@irit.fr