Neo is an Inria project-team whose members are located in Sophia Antipolis (S. Alouf, K. Avrachenkov,
G. Neglia, S. M. Perlaza), in Avignon (E. Altman) at Lia (Lab. of Informatics of Avignon) and in Montpellier (A. Jean-Marie).
E. Altman is also with the LINCS (Lab. for Information, Networking and Communication Sciences).
S. M. Perlaza is also with the ECE department at Princeton Univ., N.J. USA; and the Mathematics Department of the Univ. de la Polynésie française (Laboratoire GAATI), Faaa, Tahiti.
The team is positioned at the intersection of Operations Research and Network Science. By using the tools of Stochastic Operations Research, we model situations arising in several application domains, involving networking in one way or the other. The aim is to understand the rules and the effects in order to influence and control them so as to engineer the creation and the evolution of complex networks.
The problems studied in Neo involve generally optimization, dynamic systems or randomness, and often all at the same time. The techniques we use to tackle these problems are those of Stochastic Operations Research, Applied Probabilities and Information Theory.
Stochastic Operations Research is a collection of modeling, optimization and numerical computation techniques, aimed at assessing the behavior of man-made systems driven by random phenomena, and at helping to make decisions in such a context.
The discipline is based on applied probability and focuses on
effective computations and algorithms. Its core theory is that of
Markov chains over discrete state spaces. This family of stochastic
processes has, at the same time, a very large modeling capability and
the potential of efficient solutions. By “solution” is meant the
calculation of some performance metric, usually the
distribution of some random variable of interest, or its average,
variance, etc. This solution is obtained either through exact
“analytic” formulas, or numerically through linear algebra
methods. Even when not analytically or numerically tractable,
Markovian models are always amenable to “Monte-Carlo” simulations
with which the metrics can be statistically measured.
An example of this is the success of classical Queueing Theory,
with its numerous analytical formulas. Another important derived
theory is that of the Markov Decision Processes, which allows to
formalize optimal decision problems in a random environment.
This theory allows to characterize the optimal decisions, and provides
algorithms for calculating them.
Strong trends of Operations Research are: a) an increasing importance of multi-criteria multi-agent optimization, and the correlated introduction of Game Theory in the standard methodology; b) an increasing concern of (deterministic) Operations Research with randomness and risk, and the consequent introduction of topics like Chance Constrained Programming and Stochastic Optimization. Data analysis is also more and more present in Operations Research: techniques from statistics, like filtering and estimation, or Artificial Intelligence like clustering, are coupled with modeling in Machine Learning techniques like Q-Learning.
Network Science is a multidisciplinary body of knowledge, principally concerned with the emergence of global properties in a network of individual agents. These global properties emerge from “local” properties of the network, namely, the way agents interact with each other. The central model of “networks” is the graph (of Graph Theory/Operations Research). Nodes represent the different entities managing information and taking decisions, whereas, links represent the fact that entities interact, or not. Links are usually equipped with a “weight” that measures the intensity of such interaction. Adding evolution rules to this quite elementary representation leads to dynamic network models, the properties of which Network Science tries to analyze.
A classical example of properties sought in networks is the famous “six degrees of separation” (or “small world”) property: how and why does it happen so frequently? Another ubiquitous property of real-life networks is the Zipf or “scale-free” distribution for degrees. Some of these properties, when properly exploited, lead to successful business opportunities: just consider the PageRank algorithm of Google, which miraculously connects the relevance of some Web information with the relevance of the other information that points to it.
In its primary acceptation, Network Science involves little or no
engineering: phenomena are assumed to be “natural” and emerge
without external interventions. However, the idea comes fast to intervene in
order to modify the outcome of the phenomena.
This is where Neo is positioned.
Beyond the mostly descriptive approach of Network Science, we aim at
using the techniques of Operations Research so as to engineer complex
networks.
To quote two examples: controlling the spread of diseases through a “network” of people is of primarily interest for mankind. Similarly, controlling the spread of information or reputation through a social network is of great interest in the Internet. Precisely, given the impact of web visibility on business income, it is tempting (and quite common) to manipulate the graph of the web by adding links so as to drive the PageRank algorithm to a desired outcome.
Another interesting example is the engineering of community structures.
Recently, thousands of papers have been written on the topic of community
detection problem.
In most of the works, the researchers propose methods,
most of the time, heuristics, for detecting communities or dense subgraphs
inside a large network. Much less effort has been put in the understanding
of community formation process and even much less effort has been
dedicated to the question of how one can influence the process of community
formation, e.g. in order to increase overlap among communities and reverse
the fragmentation of the society.
Our ambition for the medium term is to reach an understanding of the behavior of complex networks that will make us capable of influencing or producing a certain property in a given network. For this purpose, we will develop families of models to capture the essential structure, dynamics, and uncertainty of complex networks. The “solution” of these models will provide the correspondence between metrics of interest and model parameters, thus opening the way to the synthesis of effective control techniques.
In the process of tackling real, very large size networks, we increasingly deal with large graph data analysis and the development of decision techniques with low algorithmic complexity, apt at providing answers from large datasets in reasonable time.
Let us describe new/updated software.
marmoteCore is a C++ environment for modeling with Markov chains. It consists in a reduced set of high-level abstractions for constructing state spaces, transition structures and Markov chains (discrete-time and continuous-time). It provides the ability of constructing hierarchies of Markov models, from the most general to the particular, and equip each level with specifically optimized solution methods.
This software was started within the ANR MARMOTE project: ANR-12-MONU-00019.
The calculation of cumulative distribution functions (CDFs) of sums of random vectors is omnipresent in the realm of engineering. Approximations to these CDFs, e.g., Gaussian approximations and saddlepoint approximations have gained remarkable popularity. In the case of Gaussian approximations, multi-dimensional Berry-Esseen-type theorems provide upper bounds on the approximation errors. These bounds are particularly precise around the mean. Alternatively, saddlepoint approximations are known to be more precise than Gaussian approximations far apart from the mean. Unfortunately, this claim is often justified by numerical analysis as formal upper bounds on the error induced by saddlepoint approximations are rather inexistent.
In 29 and 54, S. M. Perlaza, together with J.-M. Gorce and D. Anade (both from Inria, MARACAS) and P. Mary (INSA de Rennes), proposed a real-valued function that approximates the CDF of a finite sum of real-valued independent and identically distributed random vectors. The approximation error is upper bounded by an expression that is easy to calculate. As a byproduct, an upper bound and a lower bound on the CDF are obtained. Finally, in the case of lattice and absolutely continuous random variables, the proposed approximation is shown to be identical to the saddlepoint approximation of the CDF.
Many systems require frequent and regular updates of certain information. These updates have to be transferred regularly from the 5G(s) to a common destination. In 39, K. Veeraruna (IIT Bombay) and E. Altman consider scenarios in which an old packet (entire information unit) becomes completely obsolete, in the presence of a new packet. We consider transmission channels with unit storage capacity; upon arrival of a new packet, if another packet is being transmitted then one of the packets is lost. We consider the control problem that consists of deciding which packet to discard so as to maximise the average age of information (AAoI). We derive drop policies that optimize the AAoI. We show that the state independent (static) policies like dropping always the old packets or dropping always the new packets are optimal in many scenarios, among an appropriate set of stationary Markov policies.
Prefetching is a basic technique used to reduce the latency of diverse computer services. In 19, K. Keshava, A. Jean-Marie and S. Alouf propose and analyze a model for optimizing the prefetching of documents, in the situation where the connection between documents is discovered progressively. A random surfer moves along the edges of a random tree representing possible sequences of documents, which is known to a controller only up to depth
Count-Min Sketch with Conservative Updates (CMS-CU) is a popular algorithm to approximately count items’appearances in a data stream. Despite CMS-CU’s widespread adoption, the theoretical analysis of its performance is still wanting because of its inherent difficulty. Y. Ben Mazziane, under the supervision of S. Alouf and G. Neglia, has proposed a novel approach to study CMS-CU and derive new upper bounds on the expected value and the CCDF of the count estimation error under an i.i.d. request process. The formulas obtained can be successfully employed to derive improved estimates for the precision of heavy-hitter detection methods and improved configuration rules for CMS-CU. The bounds have been evaluated both on synthetic and real traces.
In 33, K. Avrachenkov together with V. Gaitsgory and L. Gamertsfelder (both from Macquarie Univ., Australia) study asymptotic properties of problems of control of stochastic discrete time systems with time averaging and time discounting optimality criteria and with general compact state and action spaces (equivalently Markov Decision Processes, MDPs), and they establish that the Cesàro and Abel limits of the optimal values in such problems can be estimated with the help of a certain infinite-dimensional (ID) linear programming (LP) problem and its dual.
A search engine maintains local copies of different web pages to provide quick search results. This local cache is kept up-to-date by a web crawler that frequently visits these different pages to track changes in them. Ideally, the local copy should be updated as soon as a page changes on the web. However, finite bandwidth availability and server restrictions limit how frequently different pages can be crawled. This brings forth the following optimization problem: maximize the freshness of the local cache subject to the crawling frequencies being within prescribed bounds. While tractable algorithms do exist to solve this problem, these either assume the knowledge of exact page change rates or use inefficient methods such as MLE for estimating the same. We address this issue here. In 15 K. Avrachenkov and K. Patil together with G. Thoppe (Indian Institute of Science) provide three novel schemes for online estimation of page change rates, all of which have extremely low running times per iteration. The first is based on the law of large numbers and the second on stochastic approximation. The third is an extension of the second and includes a heavy-ball momentum term. Our main theoretical results concern asymptotic convergence and convergence rates of these three schemes. In fact, our work is the first to show convergence of the original stochastic heavy-ball method when neither the gradient nor the noise variance is uniformly bounded. We also provide some numerical experiments (based on real and synthetic data) to demonstrate the superiority of our proposed estimators over existing ones such as MLE.
Monitoring and control processes in power systems are supported by supervisory control and data acquisition (SCADA) systems, and more recently, by advanced communication systems that acquire and transmit observations to a state estimator. This new sensing and communication infrastructure, which possesses security vulnerabilities, exposes the power system to malicious attacks. In this context, one of the main threats faced by modern power systems are data injection attacks (DIAs). A DIA can alter the state estimate obtained by the power-system operator by tampering with the observations (measurements) without triggering the attack detection system. In 61, S. M. Perlaza together with X. Ye, I. Esnaola, and R. F. Harrison (all from Univ. of Sheffield, UK) presented sparse attacks that minimize simultaneously the information obtained by the operator and the probability of detection. When the assumption on the sparsity is dropped, S. M. Perlaza together with S. Ke and I. Esnaola (both from Univ. of Sheffield) have studied these attacks in 51.
In 13, K. Avrachenkov, A. Bobu (former postdoctoral fellow in Neo) and M. Dreveton study clustering random geometric graphs. While the standard spectral clustering is often not effective for geometric graphs, they present an effective generalization, which they call higher-order spectral clustering (HOSC). It resembles in concept the classical spectral clustering method but uses for partitioning the eigenvector associated with a higher-order eigenvalue. They establish the weak consistency of this algorithm for a wide class of random geometric graphs which is called Soft Geometric Block Model. A small adjustment of the algorithm provides strong consistency. They also show that the HOSC method is effective in numerical experiments even for graphs of modest size.
In 32, K. Avrachenkov and M. Dreveton together with L. Leskelä (Aalto Univ., Finland) study the recovery of static communities in a temporal network. They introduce a temporal stochastic block model where dynamic interaction patterns between node pairs follow a Markov chain. They render this model versatile by adding degree correction parameters, describing the tendency of each node to start new interactions. It is shown that in some cases the likelihood of this model is approximated by the regularized modularity of a time-aggregated graph. This time aggregated graph involves a trade-off between new edges and persistent edges. A continuous relaxation reduces the regularized modularity maximization to a normalized spectral clustering. The authors illustrate by numerical experiments the importance of edge persistence, both on simulated and real data sets.
In 63, A. Hodgkinson (Univ. Exter, UK) and A. Jean-Marie explore the hypothesis that the growth of a population may be limited by the social constraints on pair-bonding (necessary for reproduction in some societies) rather than by ecological constraints. They propose a novel nonlinear Markovian mechanism and embed it within a dynamic graph model, called the Dynamic Markovian Ecological Network (DyME-Net). They study in particular how subdivision in the population may overcome these nonlinearities.
In 50 K. Avrachenkov and K. Patil together with V.S. Borkar and H. Dolhare (both from Indian Institute of Technology Bombay) analyze the DQN reinforcement learning algorithm, reinforcement learning algorithm with a neural network, as a stochastic approximation scheme using the o.d.e. (for 'ordinary differential equation') approach and point out certain theoretical issues. They then propose a modified scheme called Full Gradient DQN (FG-DQN, for short) that has a sound theoretical basis and compare it with the original scheme on sample problems. They establish a rigorous convergence proof and observe a better performance for FG-DQN on the sample problems.
In 31 K. Avrachenkov and K. Patil together with V.S. Borkar from (Indian Institute of Technology Bombay) apply FG-DQN reinforcement learning algorithm to web crawling. Specifically, a search engine uses a web crawler to crawl the pages from the world wide web (WWW) and aims to maintain its local cache as fresh as possible. Unfortunately, the rates at which different pages change in WWW are highly nonuniform and also, unknown in many real-life scenarios. In addition, the finite available bandwidth and possible server restrictions on crawling frequency make it very difficult for the crawler to find the optimal scheduling policy that maximizes the freshness of the local cache. We model this problem in a multi-armed restless bandits framework, where each arm represents a web page or an aggregate of statistically identical web pages. The objective is to find the scheduling policy that gives the exact indices of the pages to be crawled at a particular instance. We provide an online learning scheme using FG-DQN algorithm which learns the unknown page change dynamics on the fly along with the optimal crawling policy. Finally, we run numerical simulations to compare our approach with state-of-the-art algorithms such as static optimisation and Thompson sampling. We observe better performance for our approach.
In 30, K. Avrachenkov and M. Kamalov together with A. Boisbunon (MyDataModels) propose a novel framework called Graph Diffusion & PCA (GDPCA) in the context of semi-supervised learning on graph structured data. It combines a modified Principal Component Analysis with the classical supervised loss and Laplacian regularization, thus handling the case where the adjacency matrix is sparse and avoiding the curse of dimensionality. Our framework can be applied to non-graph datasets as well, such as images by constructing similarity graph. GDPCA improves node classification by enriching the local graph structure by node covariance. We demonstrate the performance of GDPCA in experiments on citation networks and images, and we show that GDPCA compares favourably with the best state-of-the-art algorithms and has significantly lower computational complexity.
The increasing size of data generated by smartphones and IoT devices motivated the development of
Federated Learning (FL), a framework for on-device collaborative training of machine learning models. First efforts in FL focused on learning a single global model with good average performance across clients, but the global model may be arbitrarily bad for a given client, due to the inherent heterogeneity of local data distributions.
Federated multi-task learning (MTL) approaches can learn personalized models by formulating an opportune penalized optimization problem.
The penalization term can capture complex relations among personalized models, but eschews clear statistical assumptions about local data distributions.
In 40, O. Marfoq together with G. Neglia, A. Bellet (Inria team MAGNET), R. Vidal (Accenture Labs) and L. Kameni (Accenture Labs) propose to study federated MTL under the flexible assumption that each local data distribution is a mixture of unknown underlying distributions.
This assumption encompasses most of the existing personalized FL approaches and leads to federated EM-like algorithms for both client-server and fully decentralized settings.
Moreover, it provides a principled way to serve personalized models to clients not seen at training time. The algorithms' convergence is analyzed through a novel federated surrogate optimization framework, which can be of general interest.
Experimental results on FL benchmarks show that this approach provides models with higher accuracy and fairness than state-of-the-art methods.
Federated learning offers naturally a certain level of privacy, as clients’ data is not collected at a third party. However, maintaining the data locally does not provide itself formal privacy guarantees. An (honest-but-curious) adversary can still infer some sensitive client information just by eavesdropping the exchanged messages (e.g., gradients).
In 48, C. Xu, and G. Neglia, initiate the study of local model reconstruction attacks for FL, where a honest-but-curious adversary eavesdrops the messages exchanged between the client and the server and reconstructs the local model of the client. The success of this attack enables better performance of other known attacks, such as the membership attack, attribute inference attacks, etc. They provide analytical guarantees for the success of this attack when training a linear least squares problem with full batch size and arbitrary number of local steps. One heuristic is proposed to generalize the attack to other machine learning problems. Experiments are conducted on logistic regression tasks, showing high reconstruction quality, especially when clients’ datasets are highly heterogeneous (as it is common in FL).
In 49, O. Zari, C. Xu and G. Neglia, study the membership inference attack, where the adversary can infer whether the client owns a specific data instance. They propose a new passive inference attack that requires much less computation power and memory than existing methods. Empirical results show that this attack achieves a higher accuracy on CIFAR100 dataset (more than 4 percentage points) with three orders of magnitude less memory space and five orders of magnitude less calculations.
In 26, C. Xu, G. Neglia, in collaboration with N. Sebastianelli (previous intern in Neo), study how to speed up the distributed training of machine learning models in the framework of synchronous parameters server.
This paradigm consists of n workers, which iteratively compute updates of the model
parameters, and a stateful PS, which waits and aggregates all updates to generate a new estimate of model
parameters and sends it back to the workers for a new iteration. Transient computation slowdowns or
transmission delays can intolerably lengthen the time of each iteration. An efficient way to mitigate this
problem is to let the PS wait only for the fastest
An increasing number of applications rely on complex inference tasks that are based on machine learning (ML). Currently, there are two options to run such tasks: either they are served directly by the end device (e.g., smartphones, IoT equipment, smart vehicles), or offloaded to a remote cloud. Both options may be unsatisfactory for many applications: local models may have inadequate accuracy, while the cloud may fail to meet delay constraints. In 44, T. Si Salem, G. Castellano, and G. Neglia, in collaboration with F. Pianese (Nokia Bell Labs), and A. Araldo (Télécom SudParis) present the novel idea of Inference Delivery Networks (IDNs), networks of computing nodes that coordinate to satisfy ML inference requests achieving the best trade-off between latency and accuracy. IDNs bridge the dichotomy between device and cloud execution by integrating inference delivery at the various tiers of the infrastructure continuum (access, edge, regional data center, cloud). Authors propose a distributed dynamic policy for ML model allocation in an IDN by which each node dynamically updates its local set of inference models based on requests observed during the recent past plus limited information exchange with its neighboring nodes. Their policy offers strong performance guarantees in an adversarial setting and shows improvements over greedy heuristics with similar complexity in realistic scenarios.
In the context of the exploratory action IDEM (Information and Decision Making), S. M. Perlaza and A. Jean-Marie, together with G. Bisson (Univ. de la Polynésie française), I. Esnaola (Univ. of Sheffield), and S. Rini (National Chiao Tung Univ.), studied the problem of empirical risk minimization (ERM) with relative entropy regularization (ERM-RER) in the context of supervised learning from the perspective of measure theory. In particular, the relative entropy is assumed to be with respect to a given measure, the reference measure, and not necessarily a probability mesure. This provides a unified treatment of two relevant problems in supervised learning. First, when the reference measure is a probability measure, the ERM-RER problem is shown to be a risk-information minimization problem. Alternatively, when the reference measure is the Lebesgue measure or a counting measure, the solution of the ERM-RER problem is shown to be identical to the solution to an entropy minimization problem with linear constraints, as the one typically induced by Jayne’s maximum entropy principle. The main result consists in a number of properties for the solution to the ERM-RER in terms of the reference measure, the regularization factor, and the training data. Some of the most important properties are described hereunder:
(i) The optimal solution to the ERM-RER is a probability measure that is unique and concentrates into a set arbitrarily small containing the minimizers of the empirical risk with arbitrarily high probability. The tradeoff between the cardinality of such set and the probability is governed by the regularization factor.
(ii) The expected value of the empirical risk, with respect to the ERM-RER optimal measure, is decreasing with the regularization factor. Nonetheless, via simple examples, monotonicity of the variance and higher cumulants is shown to be subject to conditions.
(iii) The transport of the ERM-RER optimal measure through the loss function is a sub-Gaussian probability measure. This property is central to study the sensitivity of the empirical risk for a particular data set with respect to changes around the optimal ERM-RER optimal measure.
A broad view at some advanced topics in data analytics for power systems is presented in the book 52, edited by S. M. Perlaza together with Ali Tajer (Rensselaer Polytechnic Institute, NY, USA) and H. Vincent Poor (Princeton Univ., USA). Therein, experts in data analytics and power engineering present techniques addressing the needs of modern power systems, covering theory and applications related to power system reliability, efficiency, and security. The book explores topics spanning large-scale and distributed optimization, statistical learning, big data analytics, graph theory, and game theory.
In 22, S. M. Perlaza and collaborators at Edhec Business School (Nice, France), HEC Montreal (Montreal, QC, Canada), Cimiez Hospital (Nice, France), Univ. Hospital of Nice (Nice, France), Paul Brousse Hospital (Paris, France), and Centre Hospitalier Le Vinatier (Lyon, France), assessed the impact of the COVID-19 pandemic on hospital workers’ psychological parameters and attitudes toward substance use, before and during the French COVID-19 lockdown. An online survey was proposed to the staff of the University Hospital of Nice and Sainte-Marie psychiatric hospital in France assessing changes in daily habits, psychological distress and changes in substance use, including smoking. During the first lockdown, significant psychological changes (sadness, distress, irritability) associated with changes in tobacco use and physical activity were reported. Such results should encourage hospital leaders to implement dedicated policies to better accompany hospital workers’ psychological distress. Data analysis suggests that negative changes in health behaviors are associated with increased psychological distress in hospital workers during the COVID-19 lockdown. Tobacco consumers are having increased psychological distress than their non-consuming counterparts. Health promotion strategies aimed at embracing or preserving positive health behaviors should go toward reducing critical and chronic increases in psychological distress during these unparalleled times. Ongoing assessment of the impact of lockdown and social distancing on health behaviors is needed to shape targeted health promotion strategies.
In 14, K. Avrachenkov together with V.S. Borkar, S. Moharir and S. M. Shah (all from Indian Institute of Technology Bombay) introduce a model of graph-constrained dynamic social choice with reinforcement modeled by positively
The work 36 is the fourth in a series of papers that appeared in international conferences that came out of homework by students in a course by E. Altman on initiation to research on routing games. The students, S. Dalalli and C. Fontaine were in their 3rd year in Central-Supelec. They consider a splittable atomic game with lossy links on a ring in which the cost that each player
Epidemiology relies on models for propagation in populations through various types of propagation mechanisms. The control of such propagation may be useful in different contexts: in broadcasting information, in epidemic advertisement, in fighting against intrusion and denial of service and more generally in protection in context of cyberwar. On the other hand game theory is a powerful tool for protection against epidemics, and in particular pandemics such as the one due to SARS-CoV-2 (coronavirus). We have been studying during 2021 two game theoretic epidemic models for the study of cooperation versus free riding within various population types. This includes the mask game and the vaccination game.
Wearing a mask provides partial protection against epidemics but at some cost of discomfort. The players can be differentiated according to both their risk-state as well as their health-state (susceptible, infected and recovered). E. Altman, M. Datar, and S. M. Perlaza in cooperation with F. de Pellegrini (Univ. of Avignon) and D. Sadoc Menasche (Federal Univ. of Rio de Janeiro, Brazil) formulate the problem as a Bayesian game 11. A player takes decisions based on their own risk-state. Yet they have no information on the health and risk-state of the interacting players. They may not know their own health-state. Using ideas inspired from evolutionary games, we reduce the problem into a one shot equivalent game and derive the equilibrium.
In the second game, E. Altman studied 28 the case in which each player decides whether to purchase a vaccine or not, and if so, then they further decide which vaccine to purchase among a finite number of vaccine producers. The players need not be indistinguishable. A potential consumer belongs to a risk type that characterizes how important it is for them to be vaccinated. The cost of a vaccine may depend on the demand, on the cost of the production, and on the consumer's class. We prove the existence of an equilibrium within pure policies in the general multipopulation case. We further derive some properties of the equilibria in the case of a single risk-class.
In 18, A. Jean-Marie together
with M. Tidball (INRAe, France) and V. Bucarey (Univ. O'Higgins, Chile),
consider a discrete time, infinite horizon dynamic game of groundwater extraction. A Water Agency charges an extraction cost to water users, and controls the marginal extraction cost so that it depends, not only on the level of groundwater, but also on total water extraction (through a parameter
In 57, A. Jean-Marie, T. Jimenez (Univ. Avignon) and M. Tidball (INRAe, France), consider the application of conjectural learning to a water management model. They introduce the family of Conjectural Learning procedures, in which agents form conjectures about what the opponent will play, as a function of their action or some state variable. They prove general properties of these procedures, comparing them to two dynamic systems generated by nearsighted agents solving repeatedly a static game, either cooperatively, or noncooperatively. They then specify three simple functional forms of conjectures and analyze the five resulting dynamic systems, in terms of steady states and convergence conditions. They next further specify these procedures to the case of a game of groundwater management, using a quadratic functional form for the profit function of agents. They obtain then explicit expressions for steady states and convergence conditions for the dynamic systems. Finally, they conduct numerical experiments in order to assess the “performance” of pairs of agents adopting each of the five behaviors.
In 46, T. Si Salem and G. Neglia in collaboration with S. Ioannidis (Northeastern Univ., USA) study an online caching problem in which requests can be served by a local cache to avoid retrieval costs from a remote server. The cache can update its state after a batch of requests and store an arbitrarily small fraction of each content. They study no-regret algorithms based on Online Mirror Descent (OMD) strategies and show that the choice of OMD strategy depends on the request diversity present in a batch and that OMD caching policies may outperform traditional eviction based policies. In 20, T. Si Salem and G. Neglia in collaboration with Y. Li, and S. Ioannidis (Northeastern Univ., USA) study a cache network under arbitrary adversarial request arrivals. They propose a distributed online policy based on the online tabular greedy algorithm. Their distributed policy achieves sublinear
In 5G and beyond network architectures, operators and content providers base their content distribution strategies on Heterogeneous Networks, where macro and small(er) cells are combined to offer better Quality of Service (QoS) to wireless users. On top of such networks, edge caching and Coordinated Multi-Point (CoMP) transmissions are used to further improve performance. The problem of optimally utilizing the cache space in dense and heterogeneous cell networks has been extensively studied under the name of “FemtoCaching.” However, the literature usually assumes relatively simple physical layer (PHY) setups and known or stationary content popularity.
In 24, G. Neglia and G. Iecker Ricardo, together with E. Leonardi (Politechnic Univ. of Turin, Italy) and T. Spyropoulos (EURECOM) present
In 25, G. Iecker Ricardo and G. Neglia, together with T. Spyropoulos (EURECOM), address the average delay minimization problem by first formulating it as a static optimization problem. Even though the problem is NP-hard they are able to solve it via an efficient algorithm that guarantees a
In 16, O. Chuchuk, together with A. Di Girolamo, M. Lassnig, M. Schulz, A. Sciaba (all four from CERN, Geneva, Switzerland), T. Beermann (Bergische Univ. Wuppertal, Germany), M. Grigorieva (Lomonosov Moscow State Univ., Russian Federation), A.Klimentov (Brookhaven National Laboratory, USA), and E.Tretyakov (National Research Nuclear Univ. MEPhI, Russian Federation) have performed a multilateral evaluation of the complete data life cycle of the ATLAS Experiment at the Large Hadron Collider. This experiment generates petabytes of data that is distributed among 160 computing sites all over the world and is processed continuously by various central production and user analysis tasks. The popularity of data, which is typically measured as the number of accesses, plays an important role in resolving data management issues: deleting, replicating, moving between tapes, disks and caches. Previously, these data management procedures were carried out in a semi-manual mode and in this study the efforts have been focused on automating it, making use of the historical knowledge about existing data management strategies. The authors described the sources of information about data popularity and demonstrated their consistency. Based on the calculated popularity measurements, various distributions were obtained. Auxiliary information about replication and task processing allowed them to evaluate the correspondence between the number of tasks with popular data executed per site and the number of replicas per site. They also examine the popularity of user analysis data that is much less predictable than in the central production and requires more indicators than just the number of accesses.
In similarity caching systems, a user request for an object
In 23, G. Neglia, in collaboration with M. Garetto (Univ. of Turin, Italy), and E. Leonardi (Politechnic Univ. of Turin, Italy), provides a first comprehensive analysis of similarity caching in the offline, adversarial, and stochastic settings. The authors show that similarity caching raises significant new challenges, for which we propose the first dynamic policies with some optimality guarantees. They evaluate the performance of the proposed schemes under both synthetic and real request traces.
In 42, T. Si Salem and G. Neglia in collaboration with A. Sabnis, R. K. Sitaraman (Univ. of Massachusetts Amherst, USA), M. Garetto (Univ. of Turin, Italy), and E. Leonardi (Politechnic Univ. of Turin, Italy) propose a policy that uses gradient descent to navigate the continuous embedding space of objects and find the optimal objects to store in the cache.
In 45, T. Si Salem and G. Neglia in collaboration with D. Carra (Univ. of Verona, Italy) present AÇAI, a new similarity caching policy which improves on the state of the art by using (i) an (approximate) index for the whole catalog to decide which objects to serve locally and which to retrieve from the remote server, and (ii) a mirror ascent algorithm to update the set of local objects with strong guarantees even when the request process does not exhibit any statistical regularity. This paper received the best paper award at ITC 2021.
The k-Nearest Neighbors method aims at efficiently finding items close to a query in a large collection of objects, and it is used in different applications, from image retrieval to recommendation. These applications achieve high throughput combining two different elements: 1) approximate nearest neighbours searches that reduce the complexity at the cost of providing inexact answers and 2) caches that store the most popular items. In 35, G. Neglia, in collaboration with D. Carra (Univ. of Verona, Italy), proposes to combine the approximate index for the whole catalog with a more precise index for the items stored in the cache. Experiments on realistic traces show that this approach is doubly advantageous as it 1) improves the quality of the final answer provided to a query, 2) additionally reduces the service latency.
In 17, G. Neglia, in collaboration with M. Garetto (Univ. of Turin, Italy), and E. Leonardi (Politechnic Univ. of Turin, Italy), makes a first step into the novel area of similarity caching networks, where requests can be forwarded along a path of caches to get the best efficiency–accuracy tradeoff. The offline problem of content placement can be easily shown to be NP-hard, while different polynomial algorithms can be devised to approach the optimal solution in discrete cases. As the content space grows large, the authors propose a continuous problem formulation whose solution exhibits a simple structure in a class of tree topologies. They verify their findings using synthetic and realistic request traces.
Several networking applications studied in Neo are specifically
on 5G or 6G networks.
In 41, M. Masson, Z. Altman from Orange Labs and E. Altman consider a network with Massive Multiple Input Multiple Output (M-MIMO) base stations using a Grid of Beams (GoB) for data and control channels. 5G allows to establish interference relations between beams of neighboring cells. Such relations can be used to automatically generate a beam relation matrix, denoted as Automatic Neighbor Beam Relation (ANBR) matrix that can be very useful for optimizing different resource allocation processes. This paper shows how the ANBR matrix can be used to coordinate scheduling of neighboring cells with a small amount of information exchange. The coordination is performed by judicially muting or reducing the bandwidth of certain beams in the process of MultiUser (MU) Proportional Fair (PF) scheduling. Numerical results show how the coordination approach can bring about significant performance gain.
Network Slicing is one of the essential concepts that has been introduced in 5G networks design to support demand expressed by next generation services. Network slicing will also bring new business opportunities for service providers (SPs) and virtual network operators, allowing them to run their virtual, independent business operations on shared physical infrastructure. In 37, M. Datar and E. Altman consider a marketplace where service providers (SPs) i.e., slice tenants lease the resources from an infrastructure provider (InP) through a network slicing mechanism. They compete to offer a certain communication service to end-users. We show that the competition between SPs can be model using the multiresource Tullock contest (TC) framework, where SPs exert effort by expending costly resource to attract users. We study the competition between the SPs under a static and dynamic resource sharing scheme. In a dynamic resource sharing scheme, SPs are pre-assigned with fixed shares (budgets) of infrastructure, and they are allowed to redistribute their shares and customise their allocation to maximise their profit. The decision problem of SPs is analysed using non-cooperative game theory, and it is shown that the resultant game admits a unique Nash Equilibrium (NE). Furthermore, a distributed reinforcement algorithm is proposed that allows each SP to reach the game's unique Nash equilibrium. Finally, simulations results are conducted to analyse the interaction between market players and the economic efficacy of the network sharing mechanism.
Mobile phones rely on batteries to provide the power needed for transmission and for reception (up and downlink communications). Considering uplink, E. Altman, M. Datar and G. Ferrat (former intern in Neo) analyse in 27 how the characteristics of the battery affect the amount of information that one can draw out from the terminal. They focus in particular on the impact of the charge in the battery on the internal resistance which grows as the battery depletes.
Battery dependency is a critical issue when communications systems are deployed in hard-to-reach locations, e.g., remote geographical areas, concrete structures, human bodies, or disaster/war zones. In this case, the lifetime of the electronic devices or even the whole communications system is determined by the battery life. An effective remedy is using energy harvesting technologies. Specifically, energy can be harvested from different ambient sources such as light, vibrations, heat, chemical reactions, physiological processes, or the radio frequency (RF) signals produced by other communications systems. This observation rises the idea of simultaneous information and energy transmission (SIET) via RF. In 47 and 62, S. U. Zuhra, S. M. Perlaza, and E. Altman have studied the fundamental limits on the rates at which information and energy can be simultaneously transmitted over an additive white Gaussian noise channel. The underlying assumption is that the number of channel input symbols (constellation size) is finite. The main results are mathematical expressions of the achievable and converse information-energy rates as a function of the constellation size, number of channel uses, decoding error probability, and energy-outage probability. As a by product, guidelines for optimal constellation design for SIET are obtained in terms of all real-system implementation parameters.
In 43, R. Sambamoorthy and E. Altman in cooperation with J. Mandalapu, S. Peruru, B. Jain (IIT Madras and Kanpur) propose a Graph Convolutional Neural Networks (GCN) based scheduling algorithm for ad-hoc networks. In particular, we consider a generalized interference model called the k-tolerant conflict graph model and design an efficient approximation for the well-known Max-Weight scheduling algorithm. A notable feature of this work is that the proposed method do not require labeled data set (NP-hard to compute) for training the neural network. Instead, we design a loss function that utilises the existing greedy approaches and trains a GCN that improves the performance of greedy approaches. Our extensive numerical experiments illustrate that using our GCN approach, we can significantly (4-20 percent) improve the performance of the conventional greedy approach.
Botnets such as Mirai use insecure home devices to conduct Distributed Denial of Service (DDoS) attacks on the Internet infrastructure. Although some of those attacks involve large amounts of traffic, they are generated from a large number of homes, which hampers their early detection. In 12, P. Nain in collaboration with A. Ramtin and D. Towsley (both from Univ. Massachusetts), D. Sadoc Menasche and E. de Souza e Silva (both from Federal Univ. of Rio Janeiro), addresses the following question: what is the maximum amount of damage that a DDoS attacker can produce at the network edge without being detected? To that aim, the authors consider a statistical hypothesis testing approach for attack detection at the network edge. The proposed system assesses the goodness of fit of traffic models based on the ratio of their likelihoods. Under such a model, it is shown that the amount of traffic that can be generated by a covert attacker scales according to the square root of the number of compromised homes. Theoretical results are evaluated and validated using real data collected from thousands of home-routers connected to a mid-sized Internet service provider.
Neo members are involved in the
Neo has contracts with
Accenture (see §8.1.4),
MyDataModels (see §8.1.5),
Qu'Est-Ce Qui Tourne (see §8.1.6),
and NSP (see §8.1.7).
Over the last few years, research in computer science has shifted focus to machine learning methods for the analysis of increasingly large amounts of user data. As the research community has sought to optimize the methods for sparse data and high-dimensional data, more recently new problems have emerged, particularly from a networking perspective that had remained in the periphery.
The technical program of this ADR consists of three parts: Distributed machine learning, Multiobjective optimisation as a lexicographic problem, and Use cases / Applications. We address the challenges related to the first part by developing distributed optimization tools that reduce communication overhead, improve the rate of convergence and are scalable. Graph-theoretic tools including spectral analysis, graph partitioning and clustering will be developed. Further, stochastic approximation methods and D-iterations or their combinations will be applied in designing fast online unsupervised, supervised and semi-supervised learning methods.
A growing number of network infrastructures are being presently considered for a software-based replacement: these range from fixed and wireless access functions to carrier-grade middle boxes and server functionalities. On the one hand, performance requirements of such applications call for an increased level of software optimization and hardware acceleration. On the other hand, customization and modularity at all layers of the protocol stack are required to support such a wide range of functions. In this scope the ADR focuses on two specific research axes: (1) the design, implementation and evaluation of a modular NFV architecture, and (2) the modelling and management of applications as virtualized network functions. Our interest is in low-latency machine learning prediction services and in particular how the quality of the predictions can be traded off with latency.
The considerable extent of the complexity of 5G networks and their operation is in contrast with the increasing demands in terms of simplicity and efficiency. This antagonism highlights the critical importance of network management. Self-Organizing Networks (SON), which cover self-configuration, self-optimization and self-repair, play a central role for 5G Radio Access Network (RAN).
This CIFRE thesis aims at innovating in the field of managing 5G RAN, with a special focus on the features of the SON-5G. Three objectives are identified: a) develop self-organizing features (SON in 5G-RAN), b) develop cognitive managing mechanisms for the SON-5G features developed, and c) demonstrate how do the self-organizing mechanisms fit in the virtual RAN.
IoT applications will become one of the main sources to train data-greedy machine learning models. Until now, IoT applications were mostly about collecting data from the physical world and sending them to the Cloud. Google’s federated learning already enables mobile phones, or other devices with limited computing capabilities, to collaboratively learn a machine learning model while keeping all training data locally, decoupling the ability to do machine learning from the need to store the data in the cloud. While Google envisions only users’ devices, it is possible that part of the computation is executed at other intermediate elements in the network. This new paradigm is sometimes referred to as Edge Computing or Fog Computing. Model training as well as serving (provide machine learning predictions) are going to be distributed between IoT devices, cloud services, and other intermediate computing elements like servers close to base stations as envisaged by the Multi-Access Edge Computing framework. The goal of this project is to propose distributed learning schemes for the IoT scenario, taking into account in particular its communication constraints. A first 12-month pre-PhD contract has been followed by a PhD grant.
Variational autoencoders are highly flexible machine learning techniques for learning latent dimension representation. This model is applicable for denoising data as well as for classification purposes. In this thesis we plan to add semi-supervision component to the variational autoencoder techniques. We plan to develop methods which are universally applicable to versatile data such as categorical data, images, texts, etc. Initially starting from static data we aim to extend the methods to time-varying data such as audio, video, time-series, etc. The proposed algorithms can be integrated into the internal engine of MyDataModels company and tested on use cases of MyDataModels.
In this project, together with InriaTech (Jean-Luc Szpyrka, Riham Nehmeh), we have designed, prototyped and tested a recommendation system for automatic suggestion of theatres to show producers. In particular, it was very interesting to observe that our Personalized PageRank, graph-based recommendation system performs much better than a more standard text-based recommendation system.
SmartProfile is a marketing platform that allows to collect, to enhance and to analyse marketing data. Typically the data is very heterogeneous (binary data, integer numbers, real numbers, functional data, images, etc) This contract was funding the internship of Ibtihal El Mimouni on the exploratory investigation of recommendation systems with heterogeneous data. Currently, we are following up with Cifre application.
The Embassy of France in the United States, via the progamme “make our planet great again”, has funded an initiative led by S. M. Perlaza and A. Tajer (RPI, USA) for addressing foundational questions pertinent to two emerging wireless communication technologies: (i) energy harvesting (EH) systems, and (ii) ultra low-latency systems for critical missions. This project explores two strongly symbiotic research directions for establishing the fundamental limits of (i) data transmission and, (ii) simultaneous energy and data transmission, in mission critical systems empowered by EH. The expected results have applications in, e.g., disaster relief, medical instruments, cyber-physical systems, and the Internet of things. This program was launched by the President of France Emmanuel Macron in June 2017 and handled by the Embassy of France in the United States of America. The aim of such fellowships is to reinforce the international engagements of the 2015 Paris Agreement on Climate Change by fostering collaborations between scholars in both US and France.
ANSWER is a joint project between Qwant and Inria, funded by the French Government's initiative PIA “Programme d'Investissement d'Avenir”.
The aim of the ANSWER project is to develop the new version of the search engine of Qwant by introducing radical innovations in terms of search criteria as well as indexed content and security. This initiative is a part of the Big Data Big Digital Challenges field, since a Web search engine deals with large volumes of heterogeneous and dynamic data.
Of the five characteristics of big data, the ANSWER project will focus more particularly on the aspects of Velocity in terms of near real-time processing of results, and Variety for the integration of new indicators (emotions, sociality, etc.) and meta-data. The Volume, Value and Veracity aspects will necessarily be addressed jointly with these first ones and will also be the subject of locks, especially on the topics of crawling and indexing.
This registration of the search engine in the Big Data domain will only be reinforced by developments in the Web such as the Web of data, and generally by the current trend to integrate the Web of increasingly diverse, rich and complex resources.
NEO members regularly perform reviews for journals such as
Discrete Applied Maths,
European Journal of Operations Research,
IEEE Trans. on Communications,
on Parallel and Distributed Systems,
on Information Theory,
on Communications,
on Signal Processing,
IEEE Communication Letters
and many others.
Note: UCA is the Univ Côte d'Azur.