Neo is an Inria project-team whose members are located in Sophia Antipolis (S. Alouf, K. Avrachenkov,
G. Neglia, and 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.

G. Neglia was recognized one of the “top reviewers” of the 25th International Conference on Artificial Intelligence and Statistics (AISTATS) and TPC distinguished member for IEEE International Conference on Computer Communications (INFOCOM).

G. Neglia was awarded a Chair by the
Interdisciplinary Institute for Artificial Intelligence
3IA Côte d'Azur,
in the theme “Core Elements of AI”.

K. Avrachenkov and M. Dreveton have co-authored and published a new book 49: “Statistical Analysis of Networks”. Now Publishers, Oct 6, 2022.

S. Perlaza was re-appointed “Visiting Research Collaborator” in the Electrical and Computer Engineering Department at Princeton University for the academic year 2022-2023; and re-appointed “Associate Researcher” in the Laboratory of Algebraic Geometry and Applications to Information Theory (GAATI) at the Université de la Polynésie Française for the academic year 2022-2023.

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.

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. In 31, 19 Y. Ben Mazziane, S. Alouf, and G. Neglia propose 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 16, 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.

In 17, K. Avrachenkov
and K. Patil (former Neo member) with G. Thoppe (Indian Institute of Science, IISc, Bangalore, India) consider the following optimization problem: maximize the freshness of the local cache of a web crawler 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. The authors 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. All these schemes only need partial information about the page change process, i.e., they only need to know if the page has changed or not since the last crawled instance. Numerical experiments (based on real and synthetic data) are also provided that demonstrate the superiority of the proposed estimators over existing ones such as MLE. The algorithms are also readily applicable to the synchronization of databases and network inventory management.

Discrete-time discrete-state finite Markov chains are versatile mathematical models for a wide range of real-life stochastic processes. One of most common tasks in studies of Markov chains is computation of the stationary distribution. In 15, K. Avrachenkov and P. Brown, in collaboration with N. Litvak (Twente University, the Netherlands) propose a new controlled, easily distributed algorithm for this task, briefly summarized as follows: at the beginning, each node receives a fixed amount of cash (positive or negative), and at each iteration, some nodes receive 'green light' to distribute their wealth or debt proportionally to the transition probabilities of the Markov chain; the stationary probability of a node is computed as a ratio of the cash distributed by this node to the total cash distributed by all nodes together. The proposed method includes as special cases a wide range of known, very different, and previously disconnected methods including power iterations, Gauss-Southwell, and online distributed algorithms. The authors prove exponential convergence of our method, demonstrate its high efficiency, and derive scheduling strategies for the green-light, that achieve convergence rate faster than state-of-the-art algorithms.

In most standard queueing systems, in the case of overload, customers are either lost or experience very long or infinite delays. However, in many real systems, rejected or significantly delayed customers leave the queue and then return later. This natural phenomenon motivates the development of retrial queues. In retrial queueing models, once a customer is rejected, he or she goes into the orbit and retries from there. This significantly increases the complexity of the system and makes the analysis of retrial queues challenging. There are two main types of retrial: independent retrials and constant rate retrials. In the former case the retrying customers are in “competition” and retry from the orbit independently. In the latter case, retrying customers wait in the orbit, according to the FIFO principle, and retry only when they are at the front of the orbit queue. In 18, K. Avrachenkov overviews the stability conditions for multi-class retrial queues with constant retrial rates. Also, a very interesting phenomenon of partial (or local) stability is discussed when some orbit sizes are stable (tight) whereas the other orbit sizes grow to infinity. This type of queueing models can find application in access control systems, and in particular, in access control system in overload.

Prioritizing genes for their role in drug sensitivity, is an important step in understanding drugs mechanisms of action and discovering new molecular targets for co-treatment. In 29, A. Jean-Marie, together with F. Cazals and D. Mazauric (Inria Abs team), J. Roux, A. Sales de Quieroz and G. Sales Santa Cruz (all from Univ. Côte d'Azur), introduce Genetrank, a method to prioritize the genes in some source set for their likelihood
to regulate the genes in some target set.
Genetrank uses asymmetric random walks with restarts, absorbing states, and a suitable
renormalization scheme. Neo's software Marmote, successor of marmoteCore (Section 6.1.1) was used for the intensive numerical experiments.

Since 2020, S. Perlaza in collaboration with X. Ye, I. Esnaola, and R. Harrison (Univ. of Sheffield) have studied sparse stealth attack constructions that minimize the mutual information between the state variables and the observations. In 62, the attack construction is formulated as the design of a multivariate Gaussian distribution that aims to minimize the mutual information while limiting the Kullback-Leibler divergence between the distribution of the observations under attack and the distribution of the observations without attack. The sparsity constraint is incorporated as a support constraint of the attack distribution. Two heuristic greedy algorithms for the attack construction are proposed. The first algorithm assumes that the attack vector consists of independent entries, and therefore, requires no communication between different attacked locations. The second algorithm considers correlation between the attack vector entries and achieves a better disruption to stealth tradeoff at the cost of requiring communication between different locations. Numerical evaluations show that it is feasible to construct stealth attacks that generate significant disruption with a low number of compromised sensors.

K. Avrachenkov and M. Dreveton publish a book “Statistical Analysis of Networks” 49 where various random graph models are applied to network inference problems such as community detection also known as graph clustering, graph-based semi-supervised learning and network sampling.

In 24 K. Avrachenkov together with M. Hamibouche (Eurecom, France) and L. Cottatellucci (FAU University, Germany) study the spectrum of the normalized Laplacian and its regularized version for random geometric graphs (RGGs) in various scaling regimes. Two scaling regimes are of special interest, the connectivity and the thermodynamic regime. In the connectivity regime, the average vertex degree grows logarithmically in the graph size or faster. In the thermodynamic regime, the average vertex degree is a constant. The authors introduce a deterministic geometric graph (DGG) with nodes in a grid and provide an upper bound to the probability that the Hilbert–Schmidt norm of the difference between the normalized Laplacian matrices of the RGG and DGG is greater than a certain threshold in both the connectivity and thermodynamic regime. Using this result, they show that the RGG and DGG normalized Laplacian matrices are asymptotically equivalent with high probability (w.h.p.) in the full range of the connectivity regime. The authors use the regular structure of the DGG to show that the limiting eigenvalue distribution of the RGG normalized Laplacian matrix converges to a distribution with a Dirac atomic measure at zero. In the thermodynamic regime, the authors approximate the eigenvalues of the regularized normalized Laplacian matrix of the RGG by the eigenvalues of the DGG regularized normalized Laplacian and provide an error bound which is valid w.h.p. and depends upon the average vertex degree.

In 30, T. Si Salem and G. Neglia in collaboration with G. Iosifidis (TU Delft, Netherlands) study the fairness of dynamic resource allocation problem under the

In 14, K. Avrachenkov and V.S. Borkar (Indian Institute of Technology Bombay, IITB, India) present a novel reinforcement learning algorithm for multi-armed restless bandits with average reward, using the paradigms of Q-learning and Whittle index. Specifically, the authors leverage the structure of the Whittle index policy to reduce the search space of Q-learning, resulting in major computational gains. Rigorous convergence analysis is provided, supported by numerical experiments. The numerical experiments show excellent empirical performance of the proposed scheme.

Then, in 48, 64, K. Avrachenkov in collaboration with V.S. Borkar (Indian Institute of Technology Bombay, IITB, India), U. Ayesta (IRIT, CNRS) and F. Robledo (Univ. Basque Country, Spain) study reinforcement learning for restless bandits with discounted reward. Specifically, they present QWI and QWINN, two algorithms capable of learning the Whittle indices for the total discounted criterion. The key feature is the usage of two timescales, a faster one to update the state-action Q-values, and a relatively slower one to update the Whittle indices. In our main theoretical result the authors show that QWI, which is a tabular implementation, converges to the real Whittle indices. QWINN, an adaptation of QWI algorithm using neural networks to compute the Q-values on the faster timescale, is able to extrapolate information from one state to another and scales naturally to large state-space environments. Numerical computations show that QWI and QWINN converge much faster than the standard Q-learning algorithm, neural-network based approximate Q-learning and other state of the art algorithms.

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. Federated learning allows clients to collaboratively learn statistical models while keeping their data local.

Federated learning was originally used to train a unique global model to be served to all clients, but this approach might be sub-optimal when clients' local data distributions are heterogeneous. In order to tackle this limitation, recent personalized federated learning methods train a separate model for each client while still leveraging the knowledge available at other clients. In 38, O. Marfoq and G. Neglia in collaboration with L. Kameni and R. Vidal (Accenture Labs) exploit the ability of deep neural networks to extract high quality vectorial representations (embeddings) from non-tabular data, e.g., images and text, to propose a personalization mechanism based on local memorization. Personalization is obtained by interpolating a collectively trained global model with a local

The cross-silo FL setting corresponds to the case of few (2–50) reliable clients, each holding medium to large datasets, and is typically found in applications such as healthcare, finance, or industry. While previous works have proposed representative datasets for cross-device FL, few realistic healthcare cross-silo FL datasets exist, thereby slowing algorithmic research in this critical application. In 41, O. Marfoq and G. Neglia in collaboration with S. Ayed, S. Silva, M. Lorenzi (Inria team EPIONE), E. Cyffers, P. Mangold, A. Bellet, M. Tommasi (Inria team MAGNET) and other collaborators from Owkin, Inc., EPFL, FeML, Inc., University of Southern California, École Polytechnique, Institut Polytechnique de Paris, Univ. Hospital Bonn, Helmholtz Munich, and Univ. of California at Berkeley propose a novel cross-silo dataset suite focused on healthcare, FLamby (Federated Learning AMple Benchmark of Your cross-silo strategies), to bridge the gap between theory and practice of cross-silo FL. FLamby encompasses 7 healthcare datasets with natural splits, covering multiple tasks, modalities, and data volumes, each accompanied with baseline training code. As an illustration, they additionally benchmark standard FL algorithms on all datasets. The proposed flexible and modular suite allows researchers to easily download datasets, reproduce results and re-use the different components for their research. FLamby is available on GitHub.

Multiple recent works show that client's private information can still be disclosed to an adversary who just eavesdrops the messages exchanged between the targeted client and the server. In 46, G. Neglia, together with I. Driouich, C. Xu, and F. Giroire (Inria Coati team) and E. Thomas (Amadeus) propose a novel model-based attribute inference attack in federated learning which overcomes the limits of gradient-based ones. Furthermore, they provide an analytical lower-bound for the success of this attack. Empirical results using real world datasets confirm that this attribute inference attack works well for both regression and classification tasks. Moreover, they benchmark this novel attribute inference attack against the state-of-the-art attacks in federated learning. The attack results in higher reconstruction accuracy especially when the clients' datasets are heterogeneous (as it is common in federated learning). Most importantly, their model-based fashion of designing powerful and explainable attacks enables an effective quantification of the privacy risk in FL.

Small IoT devices, such as drones and lightweight battery-powered robots, are emerging as a major platform for the deployment of AI/ML capabilities. Autonomous and semiautonomous device operation relies on the systematic use of deep neural network models for solving complex tasks, such as image classification. The challenging restrictions of these devices in terms of computing capabilities, network connectivity, and power consumption are the main limits to the accuracy of latencysensitive inferences. In 34, G. Castellano and G. Neglia, together with F. Pianese and T. Zhang (Nokia Bell Labs) and D. Carra (Univ. of Verona, Italy) present ReBEL, a split computing architecture enabling the dynamic remote offload of partial computations or, in alternative, a local approximate labeling based on a jointly-trained classifier. Their approach combines elements of head network distillation, early exit classification, and bottleneck injection with the goal of reducing the average end-to-end latency of AI/ML inference on constrained IoT devices.

In the context of the exploratory action IDEM (Information and Decision Making), in 40, 55, 56 and 57, 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.

In 13, K. Avrachenkov together with V.S. Borkar, S. Moharir and S. M. Shah (all from Indian Institute of Technology Bombay, India) introduce a model of graph-constrained dynamic social choice with reinforcement modeled by positively

In 22, S. Dhamal (Chalmers Univ. of Technology, Sweden) in collaboration with W. Ben-Ameur and T. Chahed (both from Telecom SudParis, France), E. Altman, A. Sunny (from IIT Palakkad, India) and S. Poojary (Qualcomm India), study a stochastic game with a dynamic set of players, for modeling and analyzing their computational investment strategies in distributed computing. Players obtain a certain reward for solving a problem, while incurring a certain cost based on the invested time and computational power. The authors relate the model to blockchain mining, and show that the framework is applicable to certain other distributed computing settings as well. The authors consider a particular yet natural scenario where the rate of solving the problem is proportional to the total computational power invested by the players. It is shown that, in Markov perfect equilibrium, players with cost parameters exceeding a certain threshold, do not invest; while those with cost parameters less than this threshold, invest maximal power. With extensive simulations and insights through mean field approximation, the authors study the effects of players' arrival/departure rates and the system parameters on the players' utilities.

Asymmetry and/or incompleteness of information in the context of Game Theory typically leads to situations where one of the players can take an extra advantage over her opponent.

One such situation is analyzed in 58 by S.M. Perlaza, K. Sun and A. Jean-Marie.
In this paper, two-player, two-action, zero-sum games are studied under the following assumptions:
(1) One of the players (the leader) publicly and irrevocably commits to choose its actions by
sampling a given probability measure (strategy);
(2) The leader announces its action, which is observed by its opponent (the follower) through a binary channel; and (3) the follower chooses its
strategy based on the knowledge of the leader’s strategy and the noisy observation of the leader’s action. Under these conditions, the equilibrium is shown to always exist and be often different from the Nash and Stackelberg equilibria. Even subject to noise, observing the actions of the leader is either beneficial or immaterial to the follower for all possible commitments: acquiring information indeed leads to an improved payoff. When the commitment is observed subject to a distortion, the equilibrium does not necessarily exist. Nonetheless, the leader might still obtain some benefit in some specific cases subject to equilibrium refinements. For instance,

As a preliminary to this analysis, K. Sun has systematically reviewed in 59 the properties of Nash equilibria in two-player, two-action, zero-sum games.

This research is part of the IDEM project (Section 9.4).

Standard models in Economics assume that rational players will seek to optimize some objective utility or profit. Current studies in behavioral Economics, acknowledging that real economic agents depart from this ideal behavior, attempt at incorporating “subjective” components in these optimization programs.

A. Jean-Marie, together with F. Cabo (Univ. Valladolid) and M. Tiball (INRAe), have considered this question in the context of private contribution to a public good. The standard situation is that selfish players will not find optimal to contribute. However, players may exhibit a positional behavior (they get joy from contributing more than their opponent, and displeasure from contributing less), or a conformist behavior (the closer to the average they are, the happier they are). In that case, their private optimum may lead to a postitive contribution to the public good. The paper 45, 54 considers the case of two positional players. Nash equilibria of the static game are computed and their dependence on the parameters of the problem is thoroughly studied. A dynamic model of myopic players featuring inertia is also built. It is shown that inertia can lead to an overshooting behavior: players may contribute more than what they would at the equilibrium, for some periods of time. Preliminary results for the case where one player is positional and the other one conformist, are presented in 44

M. Datar and E. Altman in collaboration with H. Le Cadre (Inria team Inocs), consider in
21 a marketplace in the context of 5G network slicing, where Application Service Providers (ASP), i.e., slice tenants, providing heterogeneous services, are in competition for the access to the virtualized network resource owned by a Network Slice Provider, who relies on network slicing. The authors model the interactions between the end users (followers) and the ASPs (leaders) as a Stackelberg game. The authors show that the competition between the ASPs results in a multi-resource Tullock rent-seeking game. To determine resource pricing and allocation, the authors devise two market mechanisms. First, it is assumed that the ASPs are pre-assigned with fixed shares (budgets) of infrastructure, and rely on a trading post mechanism to allocate the resource. Under this mechanism, the ASPs can redistribute their budgets in bids and customize their allocations to maximize their profits. In case a single resource is considered, it is proved that the ASPs' coupled decision problems give rise to a unique Nash equilibrium. Second, when ASPs have no bound on their budget, the problem is formulated as a pricing game with coupling constraints capturing the shared resource finite capacities, and the market prices are derived as the duals of the coupling constraints. In addition, it is proved that the pricing game admits a unique variational equilibrium. The authors implement two online learning algorithms to compute solutions of the market mechanisms. A third fully distributed algorithm based on a proximal method is proposed to compute the Variational equilibrium solution of the pricing game. Finally, the authors run numerical simulations to analyse the market mechanism's economic properties and the convergence rates of the algorithms.

In 33, O. Boufous (Orange Labs and Avignon Univ.) R. El-Azouzi (LIA, Avignon Univ.), M. Touati, M. Bouhtou (both from Orange labs) together with E. Altman, consider the problem of learning a correlated equilibrium of a finite non-cooperative game and show a new adaptive heuristic, called Correlated Perturbed Regret Minimization (CPRM) for this purpose. CPRM combines regret minimization to approach the set of correlated equilibria and a simple device suggesting actions to the players to further stabilize the dynamics. Numerical experiments support the hypothesis of the pointwise convergence of the empirical distribution over action profiles to an approximate correlated equilibrium with all players following the devices' suggestions. Additional simulation results suggest that CPRM is adaptive to changes in the game such as departures or arrivals of players.

In 35, O. Chuchuk and G. Neglia together with M. Schulz and D. Duellmann (CERN, Switzerland) explore the benefits of caching for existing scientific workloads, taking the Worldwide LHC (Large Hadron Collider) Computing Grid as an example. It is a globally distributed system that stores and processes multiple hundred petabytes of data and serves the needs of thousands of scientists around the globe. Scientific computation differs from other applications like video streaming as file sizes vary from a few bytes to terabytes and logical links between the files affect user access patterns. These factors profoundly influence caches’ performance and, therefore, should be carefully analyzed to select which caching policy to deploy or to design new ones. In this work, they study how the hierarchical organization of the LHC physics data into files and groups of files called datasets affects the request patterns. They then propose new caching policies that exploit dataset-specific knowledge and compare them with file-based ones. Moreover, they show that limited connectivity between the computing and storage sites leads to the delayed hits phenomenon and estimate the consequent reduction in the potential benefits of caching.

A similarity cache can reply to a query for an object with similar objects stored locally. In some applications of similarity caches, queries and objects are naturally represented as points in a continuous space. This is for example the case of

Although many similarity caching policies have been proposed, we still do not know how to compute the hit rate even for simple policies, like SIM-LRU and RND-LRU that are straightforward modifications of classic caching algorithms. In 32, Y. Ben Mazziane, S. Alouf, and G. Neglia, together with D. S. Menasche (Federal Univ. of Rio de Janeiro, Brazil) propose the first algorithm to compute the hit rate of similarity caching policies under the independent reference model for the request process. In particular, they show how to extend the popular time-to-live approximation in classic caching to similarity caching. The algorithm is evaluated on both synthetic and real world traces.

In 25 P. Nain in collaboration with G. Vardoyan (QuTech, TU Delft, The Netherlands), S. Guha (Univ. Arizona, Tucson) and D. Towsley (Univ. Massachusetts, Amherst), studies the performance of a quantum switch that distributes tripartite entangled states to sets of users. The entanglement switching process requires two steps: first, each user attempts to generate bipartite entanglement between itself and the switch; and second, the switch performs local operations and a measurement to create multipartite entanglement for a set of three users. A simple variant of this system is studied, wherein the switch has infinite memory and the links that connect the users to the switch are identical. This problem formulation is of interest to several distributed quantum applications, while the technical aspects of this work result in new contributions within queueing theory. The state of the system is modeled as continuous time Markov chain and performance metrics of interest (probability of an empty system, switch capacity, expectation and variance of the number of qubit-pairs stored) are computed via the solution of a two-dimensional functional equation obtained by reducing it to a boundary value problem on a closed curve.

In 37, S. Perlaza in collaboration with M. Haddad (Univ. d’Avignon), P. Wiecek (Wrocław Univ. of Science and Technology), O. Habachi (Univ. Clermont Auvergne), and S.M. Shah (National Institute of Technology, Srinagar, India) consider a perfect coordinated water-filling game, in which each user transmits solely on a given carrier. The main goal of the proposed algorithm (FEAT) is to achieve close to optimal performance, while keeping a given level of fairness. The key idea within FEAT is to minimize the ratio between the best and the worst utilities of the users. This is done by ensuring that, at each iteration (channel assignment), a user is satisfied with this assignment as long as it does not lose much more than other users in the system. It has been shown that FEAT outperforms most of the related algorithms in many aspects, especially in interference-limited systems. Indeed, with FEAT we can ensure a near-optimal, fair and energy efficient solution with low computational complexity. In terms of robustness, it turns out that the balance between being nearly globally optimal and good from an individual point of view seems hard to sustain with a significant number of users. Also notice that, in this regard, global optimality gets less affected than the individual one, which offers hope that such an accurate water-filling algorithm can be designed around competition in interference-limited systems.

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 43 and 60, S. Perlaza, E. Altman and S. ul-Zuhra in collaboration with H. V. Poor (Princeton Univ.) and M. Skoglund (KTH) characterized an achievable information-energy region of simultaneous information and energy transmission over an additive white Gaussian noise channel. This analysis is performed in the finite block-length regime with finite constellations. More specifically, a method for constructing a family of codes is proposed and the set of achievable tuples of information rate, energy rate, decoding error probability (DEP) and energy outage probability (EOP) is characterized. Using existing converse results, it is shown that the construction is information rate, energy rate, and EOP optimal. The achieved DEP is, however, sub-optimal.

In 36, A. Dejonghe (Orange Labs and Avignon Univ.), Z. Altman, (Orange Labs), F. de Pellegrini (Avignon Univ.) and E. Altman, consider the problem of the integration of RIS elements in upcoming 6G networks. This paper introduces a new link-layer scheme for RIS-enabled communications building on existing models for the physical layer of RIS technologies. The scheme is able to integrate the selection of precoders/beams in a codebook and scheduling of UEs at once. Furthermore, elements for the integration of the proposed scheme in current 3GPP-5G specifications are addressed. The scheduler combines a slow mechanism operating at the downlink OFDMA frame scale with a standard proportional fair scheduler operating at the OFDMA slot scale and accommodating both LOS and nonLOS UEs. The authors introduce an optimization framework for the proposed scheduler whose performance is hence simulated in a reference scenario. The spectral efficiency figures in their tests confirm a large gain of the scheme against a baseline direct path scheme. Finally, the scheduler optimization permits to achieve a further improvement of 15-20% for non line-of-sight users.

To satisfy demanding quality of service requirements for slice tenants and their end-users, Infrastructure Providers (InPs) need to perform efficient resource allocation. In the context of network slicing, the key challenge is the presence of multiple resource types and competing service providers (SPs) with heterogeneous characteristics and preferences. The main goal in this context is devising efficient mechanisms to create slices while ensuring fairness both among slice tenants as well as across their end-users. To address the aforementioned challenges, M. Datar, N. Modina, R. El-Azouzi and F. de Pellegrini (Avignon Univ.) formulate in 39 the multi-resource allocation for network slicing in the form of a Fisher Market model, where SPs act as buyers and a set of resources (divisible goods) is made available by the InP at different locations. Within the Fisher market framework, a generalized alpha-fairness resource allocation for SPs able to adapt the degree of fairness as a function of a nonnegative parameter alpha, striking the trade-off between fairness and efficiency. Given the resource prices, each SP with a certain budget – namely the market power of the SP – buys the optimal set of multi-resources to maximize its utility under the budget constraints. The Market Equilibrium is computed as a price vector for each resource type that ensures market clearance, i.e., the demand of a resource equals its supply. In this paper, it is shown that it is possible to let such market equilibrium correspond to the allocations maximizing alpha-fair utility, which is obtained under non-linear pricing. Furthermore, the authors obtain a closed-form of the pricing as a function of alpha and resources purchased by SP.

Neo members are involved in the
Inria-Nokia Bell Labs joint laboratory:
the joint laboratory consists of five ADRs (Action de
Recherche/ Research Action) in its third phase (starting October
2017). Neo members participated in the former ADR
“Distributed Learning and Control for Network Analysis” and participate in ADR
“Rethinking the network: virtualizing network functions, from middleboxes to application” (see §8.1.1);

Neo has contracts with
Accenture (see §8.1.2 and §8.1.3),
SAP (see §8.1.4),
and MyDataModels (see §8.1.5).

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 postdoc of G. Castellano was funded by this project.

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. O. Marfoq is funded by this project. A first 12-month pre-PhD contract has been followed by a PhD grant.

Deep neural networks have enabled impressive accuracy improvements across many machine learning tasks. Often the highest scores are obtained by the most computationally-hungry models. As a result, training a state-of-the-art model now requires substantial computational resources which demand considerable energy, along with the associated economic and environmental costs. Research and development of new models multiply these costs by thousands of times due to the need to try different model architectures and different hyper-parameters. In this project, we investigate a more algorithmic/system-level approach to reduce energy consumption for distributed ML training over the Internet. The postdoc of C. Rodriguez is funded by this project.

There are increasing concerns among scholars and the public about bias, discrimination, and fairness in AI and machine learning. Decision support systems may present biases, leading to unfair treatment of some categories of individuals, for instance, systematically assigning high risk of recidivism in a criminal offense analysis system. Essentially, the analysis of whether an algorithm’s output is fair (e.g. does not disadvantages a group with respect to others) depends on substantial contextual information that often requires human intervention. There are though several metrics for fairness that have been developed, which may rely on collecting additional sensitive attributes (e.g., ethnicity) before imposing strong privacy guarantees to be used in any situation. It is known that differential privacy has the effect of hiding outliers from the data analysis, perhaps compounding existing bias in certain situations. This project encompasses the search for a mitigating strategy. The PhD thesis of C. Kaplan is funded by this project.

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. This contract financed the PhD candidate M. Kamalov. His thesis was defended in December 2022.

In many use-cases of Machine Learning (ML), data is naturally decentralized: medical data is collected and stored by different hospitals, crowdsensed data is generated by personal devices, etc. Federated Learning (FL) has recently emerged as a novel paradigm where a set of entities with local datasets collaboratively train ML models while keeping their data decentralized.

FedMalin is a research project that spans 10 Inria research teams and aims to push FL research and concrete use-cases through a multidisciplinary consortium involving expertise in ML, distributed systems, privacy and security, networks, and medicine. We propose to address a number of challenges that arise when FL is deployed over the Internet, including privacy and fairness, energy consumption, personalization, and location/time dependencies. FedMalin will also contribute to the development of open-source tools for FL experimentation and real-world deployments, and use them for concrete applications in medicine and crowdsensing. The FedMalin Inria Challenge is supported by Groupe La Poste, sponsor of the Inria Foundation.

NEO members regularly perform reviews for journals such as IEEE/ACM Transactions on Networking, IEEE Transactions on Mobile Computing, IEEE Transactions on Network Science and Engineering, IEEE Transactions on Information Theory, IEEE Transactions on Wireless Communications, IEEE on Communications, IEEE Journal on Selected Areas in Communications, IEEE Transactions on Network and Service Management.

Note: UCA is the Univ Côte d'Azur.

S. Alouf met and discussed with two classes at Jules Ferry High School in Cannes on 9 May 2022, in the context of the French national program « 1 Scientifique – 1 Classe, Chiche ! ».