Nous vous rappelons que, afin de garantir l'accès de tous les inscrits aux salles de réunion, l'inscription aux réunions est gratuite mais obligatoire.
Inscriptions closes à cette réunion.
15 personnes membres du GdR ISIS, et 32 personnes non membres du GdR, sont inscrits à cette réunion.
Capacité de la salle : 156 personnes.
L'objectif de cette journée GdR ISIS est de faire une introduction générale aux outils issus de la théorie des jeux en se concentrant sur les liens avec les outils plus populaires comme : l'optimisation distribuée, l'optimisation multicritères, l'apprentissage statistique et la théorie du contrôle. Ces concepts seront illustrés par des applications au traitement du signal.
La théorie des jeux est une branche des mathématiques qui permet de modéliser et d'analyser les interactions stratégiques entre plusieurs preneurs de décision (ou joueurs). Un jeu est une situation dans laquelle typiquement le bénéfice ou le coût de chaque joueur dépend non seulement de ses propres actions mais aussi des actions des autres joueurs : les actions et les objectifs des joueurs sont couplés.
La théorie des jeux est de plus en plus utilisée grâce aux applications dans le contexte des réseaux distribués, dans lequel les interactions entre les noeuds peuvent être modélisées sous forme d'un jeu. Les noeuds peuvent être en compétition ou peuvent former des coalitions pour améliorer leur qualité de service. L'interdépendance entre les actions des noeuds est peut être typiquement due à l'exploitation de ressources communes (e.g., ressources de calcul ou de stockage, ressources spectrales, ...). Des exemples d'utilisation de la théorie des jeux dans le traitement du signal pour les communications sont : contrôle de puissance dans les réseaux sans fils, formation de voie, précodage dans les systèmes multi-antenne, sondage spectral pour la radio cognitive, gestion de spectre et d'interférence, gestion de ressources dans les systèmes multimédia, segmentation d'images, estimation distribuée dans les réseaux des capteurs, brouillage dans les réseaux sans fils et applications radar multi-antennes.
Enfin, les interconnections entre la théorie des jeux et le traitement du signal sont renforcées à travers les outils comme l'optimisation distribuée et l'apprentissage statistique.
Les chercheurs et chercheuses intéressé(e)s à présenter leurs travaux peuvent envoyer par mail aux organisateurs le titre et le résumé de leur intervention souhaitée avant le 25 Octobre 2017. La participation des doctorant(e)s est fortement encouragée.
Cette journée est également soutenue financièrement par l'ANR ORACLESS - Online resource allocation for unpredictable large-scale wireless systems (2016-2021, PI Panayotis Mertikopoulos - CNRS, LIG).
Nous remercions l'ESIEE Paris pour l'accueil de cette journée GdR ISIS et en particulier nous remercions Romain Negrel (enseignant-chercheur, ESIEE) pour son aide précieuse dans la réservation d'amphi.
http://www.esiee.fr/fr/acces
This GdR ISIS meeting targets a wide signal processing audience. The main goal is to provide a comprehensive introduction to game theoretic tools with a clear focus on its connections to more popular tools such as: distributed optimization, multi-objective optimization, machine learning and control theory. Furthermore, these concepts will be illustrated via signal processing applications.
Game theory (GT) is a branch of mathematics that enables the modeling and analysis of the interactions between several decision-makers (or players) that can have either conflicting or common objectives. A game is an interactive situation in which the benefit or cost achieved by each player depends on its own decisions and also on those taken by the other players. For example, the time a car driver needs to get back home generally depends not only on the route he/she chooses but also on the decisions taken by the other drivers. Therefore, in a game, the actions and objectives of the players are tightly coupled.
Until very recently, GT has been used only marginally in signal processing (SP). However, the real catalyzer of the application of GT to SP has been the blooming of all issues related to networking in general, and distributed networks, in particular. The interactions that take place in a network can often be modeled as a game, in which the network nodes are the players that compete or form coalitions to get some advantage and enhance their quality-of-service. The large interdependence between the actions of the network nodes due to factors such as the use of common resources, e.g., computational, storage, or spectral resources (interference across wireless networks). Examples of this approach can be found in the broad field of SP for communication networks in which GT is used to address fundamental networking issues such as: controlling the power of radiated signals in wireless networks; beamforming for smart antennas; precoding in multi-antenna radio transmission systems; data security; spectrum sensing in cognitive radio; spectrum and interference management; multimedia resource management; and image segmentation.
Motivated by the applications above, GT is also pervading many other branches of SP, and has very recently been used for modeling and analyzing the following "classical" SP problems: distributed estimation in sensor networks; adaptive filtering; waveform design for multiple-input multiple-output (MIMO) radar estimation; jamming in wireless communications and MIMO radar applications; and finding the position of network nodes. In addition to the examples above, we must eventually point out the important connection that is building up between GT and SP through the fields of machine learning and distributed optimization.
Interested researchers that wish to present their work should send a title and an abstract of their prospective talk to the organizers via email before October 25th, 2017. The participation of Ph.D. students is strongly encouraged.
This meeting is also funded by the ANR ORACLESS - Online resource allocation for unpredictable large-scale wireless systems project (2016-2021, PI Panayotis Mertikopoulos - CNRS, LIG).
We are truly thankful to ESIEE Paris graduate school for hosting this event and in particular we thank Romain Negrel (Assistant Professor in ESIEE) for his precious help in the amphi reservation process.
http://www.esiee.fr/en/address-and-getting-here
9h00 - 9h30 *** Welcoming coffee ***
9h30 - 9h45 : Introduction
9h45 - 10h30 *** Keynote speech ***
Eduard A. Jorswieck, "Wyner Meets Walras on an Interference Channel"
10h30 - 11h00 : Panayotis Mertikopoulos, "Multi-agent online learning: Game theory meets signal processing (and feels no regret)"
11h00 - 11h30 : Richard Combes, "Multi-armed bandits: a generic, asymptotically optimal algorithm and applications"
11h30 - 12h00 : Constantin Morarescu, "Clustering and multi time-scales in multi-agent systems"
12h00 - 13h30 *** Lunch Break ***
13h30 - 14h00 : Pascal Bianchi, "Distributed optimization on graphs using operator splitting methods"
14h00 - 14h30 : Alessio Zappone, "Game Theory for distributed and energy-efficient resource allocation in future cellular networks"
14h30 - 15h00 : Kenza Hamidouche, "Popular Matching Games for Correlation-aware Resource Allocation in the Internet of Things"
15h00 - 15h30 *** Coffee Break ***
15h30 - 16h00 : Olivier Beaude, "Flexibility of electricity consumption: study as a game, and link with learning and signal processing"
16h00 - 16h30 : Inaki Estella Aguerri, "Statistical Learning via Information Bottleneck"
16h30 - 17h00 : Nikolaos Liakopoulos, "Robust User Association for Ultra Dense Networks"
Speaker: Eduard A. Jorswieck (TUD)
Abstract: In modern wireless communication networks, technology layer parameters and economical aspects interact closely. Adaptive and flexible resource allocation in coexisting wireless networks requires to balance carefully achieved Quality of Service and costs in terms of energy, bandwidth and time. The talk presents a framework to connect the interference channel with confidential messages with an exchange economy based on a fictional meeting of Wyner and Walras. By identifying correspondencies between the market model and the interference channel, we explain how to find novel cooperative and distributed resource allocation algorithms for the optimization of various system objectives, including the efficient rate, secrecy rate, and secret key generation rate in multi-antenna interference channels. A comparison to the non-cooperative case illustrates the gain by cooperation even if the costs of cooperation are taken into account.
Speaker: Panayotis Mertikopoulos (CNRS, LIG)
Abstract: In many cases of practical interest, several interacting agents might be involved in a "game" without necessarily knowing its structure, rules, or even that they are playing a game - for instance, think of a set of wireless users obliviously modulating their transmit characteristics in response to the interference that they experience. In such cases, it is often beneficial for agents to follow a procedure leading to "no-regret", an important worst-case guarantee that has attracted significant interest in machine learning, signal processing, and theoretical computer science.
Motivated by its applications to signal processing (telecommunications, automated image generation, etc.), this talk focuses on the following question: does no-regret learning lead to an equilibrium of the game being played? I will present some recent contributions to this question (in both finite and continuous games), and I will discuss the impact of the feedback available to the agents.
Speaker: Richard Combes (CentraleSupelec)
Abstract: We present a novel and completely generic algorithm for multi-armed bandits with structure. Our framework covers all known bandit problems found in the literature such as: classical bandits, duelling bandits, combinatorial bandits, linear bandits, unimodal bandits, learning to rank etc. We prove that the proposed generic algorithm is asymptotically optimal for general structures. We then highlight how such a generic algorithm may be applied in practice to yield provably optimal algorithms to solve problems in communications networks such as: rate adaptation, channel assignment, routing and others.
Speaker: Constantin Morarescu (CRAN UMR 7039)
Abstract: Firstly, we present a consensus-like discrete time dynamics for the cluster detection in a graph representing the interaction network of a multi-agent system. Beside the decentralized nature of this cluster detection algorithm we also emphasize its effectiveness. Secondly, we propose a time scale modeling for consensus in large networks represented by time-varying directed graphs. For practical reasons we deal with large networks by collapsing the states of densely connected nodes in a single aggregate node. Here, we show that under suitable conditions, the states of the agents in each cluster converge fast toward a local agreement.
Speaker: Pascal Bianchi (Telecom ParisTech)
Abstract: Consider a network of N agents (computing units) having private objective functions and seeking to find a minimum of the aggregate objective. The aim is to design iterative algorithms where, at a each iteration, an agent updates a local estimate of the minimizer based on the sole knowledge of its private function and the information received from its neighbors. In this talk, i will first provide an overview of standard distributed optimization methods. Then, i will explain how recent and generic results in stochastic optimization can be used in order to design asynchronous and adaptive distributed optimization algorithms.
Speaker: Alessio Zappone (LANEAS, CentraleSupelec)
Abstract: Nowadays, the number of devices connected to the Internet is larger than the size of the world population, and it is increasing at an exponential rate. By 2020, a 1000x data- rate increase is required to serve so many connected devices, but this must be achieved while at the same time halving the network power consumption, due to ecological and economical reasons. This requires a 2000x increase of the network bit-per-Joule energy efficiency.
This talk will cover recent advances in the field of distributed resource allocation in future wireless interference networks. It will be shown how non-cooperative game theory can be successfully used to develop distributed algorithms for resource allocation in generic interference networks, with limited feedback overhead and low-complexity. The comparison between the distributed solutions and more sophisticated centralized schemes will also be discussed.
Speaker: Kenza Hamidouche (LANEAS, CentraleSupelec)
Abstract: In this talk, the problem of cell association is addressed in an Internet of things (IoT) system in which a set of devices deployed in a given area report ground information to a set of small base stations (SBSs) via uplink communication links. In this model, the key goal is to prevent multiple devices from reporting the same information to a given SBS by taking into account the spatial correlation between the IoT devices. In particular, the problem of correlation-aware cell association is formulated as a popular matching game in which the IoT devices are assigned to the SBSs to maximize the amount of information that is reported to the SBSs. To this end, the number of matched devices to every SBS must be maximized. For the formulated problem, a distributed two-level matching algorithm is proposed and the algorithm is proved to converge to a popular outcome. In that state, all the SBSs and devices prefer the matching that results from the proposed algorithm to any other possible matching. Simulation results show the performance of the proposed model.
Orateur / Speaker: Olivier Beaude (EDF Lab' Paris Saclay)
Abstrait : Dans les systèmes électriques du futur, qualifiés souvent de "réseaux électriques intelligents" ("smart grids"), une part importante de la consommation électrique va passer d'un statut "nonflexible", c'est-à-dire figé, non-contrôlable, au statut "flexible", c'est-à-dire contrôlable (par un déplacement dans le temps par exemple). Comme les consommateurs sont multiples, et souhaiteront garder la main sur leurs décisions de consommation pour une large part, ceci amène naturellement à des problèmes de coordination qui peuvent être analysés par la théorie des jeux. Vu d'un opérateur du système électrique (EDF par exemple), l'objectif est alors de concevoir des mécanismes de coordination efficaces des décisions individuelles de consommation flexible : efficacité pour le système électrique, mais aussi pour les consommateurs. En pratique, des boîtiers/compteurs intelligents ("smart meters") pourront par exemple piloter les usages flexibles résidentiels (chauffe-eau, véhicule électrique, machine à laver, etc.) sur la base d'incitations reçues des opérateurs du système électrique ; la décision et les informations restant locales. Quelques problèmes et méthodes analysés actuellement à la R&D d'EDF dans cette thématique seront présentés, en mettant en avant les liens forts avec l'apprentissage et le traitement du signal.
Abstract: In the future electricity systems, commonly named "smart grids", a large part of electricity consumption will become "flexible", i.e. controllable (e.g. by shifting it in time), while up until now it was "nonflexible", i.e. fixed, not controllable. As there are many customers whose decisions must be coordinated, this leads naturally to the framework of game theory. From the viewpoint of electricity system operators (e.g. EDF), the objective will be the design of efficient coordination mechanisms of individual flexible consumption decisions: efficient for both the system and the individual customers. In practice, this kind of mechanisms could be applied by smart meters to control flexible appliances (among them water-heater, electric vehicle, washing machine) based on incentives, or signals, sent by electricity system operators, information and decision being made locally. A few problems of this field, currently analyzed at EDF (R&D), will be presented with a focus on the strong link with learning and signal processing.
Speaker: Inaki Estella Aguerri (Mathematical and Algorithmic Sciences Lab - France Research Center, Huawei Technologies)
Abstract: We connect the information flow in a neural network to sufficient statistics; and show how techniques that are rooted in information theory, such as the source-coding based information bottleneck method can lead to improved architectures, as well as a better understanding of the theoretical foundation of neural networks, viewed as a cascade compression network. We illustrate our results and view through some numerical examples.
Speaker: Nikolaos Liakopoulos (Mathematical and Algorithmic Sciences Lab - France Research Center, Huawei Technologies)
Abstract: We study the user association problem in the context of dense networks, where standard adaptive algorithms become ineffective. The paper proposes a novel data-driven technique leveraging the theory of robust optimization. The main idea is to predict future traffic fluctuations, and use the predictions to design association maps before the actual arrival of traffic. Although the actual playout of the map is random due to prediction error, the map are robustly designed to handle uncertainty, preventing constraint violations, and maximizing the expectation of a convex utility function, which allows to accurately balance base station loads. We propose a generic iterative algorithm, referred to as GRMA, which is show to converge to the optimal robust map. The optimal maps have the intriguing property that they jointly optimize the predicted load and the variance of the prediction error. We validate our robust maps in Milano-area traces, with dense coverage and find that we can reduce violations from 25% (achieved by an adaptive algorithm) down to almost zero.
Date : 2017-11-08
Lieu : ESIEE Paris, Noisy le Grand (Amphi Marcel Dassault)
Thèmes scientifiques :
A - Méthodes et modèles en traitement de signal
Inscriptions closes à cette réunion.
Accéder au compte-rendu de cette réunion.
(c) GdR IASIS - CNRS - 2024.