7.1.1 marmote

• Name:
  MARkovian MOdeling: The Environment
• Keyword:
  Markov model
• Functional Description:

  marmote is a C++ library 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. The current release features the library marmoteMDP for modeling Markov Decision Processes and solving them.

  This software was started within the ANR MARMOTE project: ANR-12-MONU-00019 under the name marmoteCore. Within the marmote project, the code conforms the latest C++ standards and the library is available on multiple platforms via a conda distribution.

• Release Contributions:
  Version 1.2.0 was released in 2024. This version further improves the code portability and introduces structural changes improving code maintainability. From the functional point of view, it introduces a Python interface to the library: pyMarmote for the core, and pyMarmoteMDP for the library for manipulating Markov Decision Processes. A complete set of examples is available on the project's website, both for the C++ and the Python languages. For the latter, a notebook-based tutorial demonstrates the principal functionalities of the library.
• News of the Year:
  The software was awarded an Inria ADT (Action of Technological Development). Thanks to the software engineers of Inria’s SED (Service Experimentation Development), a Python interface was built for the core and for the application library devoted to Markov Decision Processes. Many improvements have been brought to internal self-testing and to programming examples, in particular via Python tutorial notebooks.
• URL:
  https://­marmote.­gitlabpages.­inria.­fr/­marmote/
• Publications:
  hal-03770430v1, hal-03781620v1, hal-02395100v1, hal-04176076v1
• Contact:
  Alain Jean-Marie
• Participants:
  Alain Jean-Marie, Patrick Brown, Emmanuel Hyon
• Partner:
  Université Paris Nanterre

Flamby.

• Contributors:
  Giovanni Neglia and Othmane Marfoq were part of a large international consortium, coordinated by the company Owkin and including EPFL, FeML Inc., University of Southern California, École Polytechnique, Institut Polytechnique de Paris, University Hospital Bonn, Helmholtz Munich, University of California at Berkeley and other Inria teams (Epione and Magnet).
• Description:
  Federated Learning (FL) enables several clients holding sensitive data to collaboratively train machine learning models, without centralizing data. 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. FLamby (Federated Learning AMple Benchmark of Your cross-silo strategies) is a novel cross-silo dataset suite focused on healthcare 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. It also includes benchmark standard FL algorithms for all datasets.
• Project link:
  https://­github.­com/­owkin/­flamby
• Publications:
  Ogier du Terrail et al, FLamby: Datasets and Benchmarks for Cross-Silo Federated Learning in Realistic Healthcare Settings, Advances in Neural Information Processing Systems (NeurIPS), 2022 https://­inria.­hal.­science/­hal-03900026.
• Contact:
  Jean Ogier Du Terrail jean.du-terrail@owkin.com
• Release contributions:
  Neo members contributed to defining the benchmark, developing the code, and writing the corresponding paper.

Similarity caching traces.

• Contributors:
  Abdelkarim Hafid (previously master intern in Neo), Caelin Kaplan , Giovanni Neglia , Anirudh Sabnis (previously visiting PhD student in Neo), Tareq Si Salem (previously PhD student in Neo).
• Description:
  A similarity cache can return an item that is similar, but not necessarily identical, to the query element. Similarity caching systems can be effectively employed in several application areas, like multimedia retrieval, recommender systems, genome study, and machine learning training/serving. We produced and released four distinct traces to evaluate similarity caching policies. Each trace includes a catalog of items and a time series of requests, with both items and requests represented as real vectors. In three datasets, these representations were derived from machine learning models, while in the fourth, the items correspond to tiles in a 360° video, and each request represents the tiles a user is watching.
• Project link:
  https://­sim-cache.­gitlabpages.­inria.­fr/­similarity-caching-traces/
• Publications:
  • A. Sabnis, T. Si Salem, G. Neglia, M. Garetto, E. Leonardi and R. K. Sitaraman, GRADES: Gradient Descent for Similarity Caching, IEEE INFOCOM 2021 https://­inria.­hal.­science/­hal-03484133
  • A. Sabnis, T. Si Salem, G. Neglia, M. Garetto, E. Leonardi and R. K. Sitaraman, GRADES: Gradient Descent for Similarity Caching, IEEE/ACM Transactions on Networking, Vol 31, n 1, 2023 https://­inria.­hal.­science/­hal-03906099
• Contact:
  Giovanni Neglia .
• Release contributions:
  The Neo researchers were at the origin of this project and carried it out in collaboration with close partners. Three of the four traces were derived from existing datasets, where a machine learning model implicitly defines item embeddings within a space equipped with a similarity metric. The fourth trace is synthetic, generated using a Markovian model to simulate user navigation in a 360° video.

8.1.1 Learning and Balancing Unknown Loads in Large-Scale Systems

In 20, Diego Goldsztajn in collaboration with Sem Borst (Technical University of Eindhoven, the Netherlands) and Johan van Leeuwaarden (Tilburg University, the Netherlands) considered a system of identical server pools where tasks with exponentially distributed service times arrive as a time-inhomogeneous Poisson process. An admission threshold is used in an inner control loop to assign incoming tasks to server pools, while in an outer control loop, a learning scheme adjusts this threshold over time to keep it aligned with the unknown offered load of the system. In a many-server regime, they prove that the learning scheme reaches an equilibrium along intervals of time when the normalized offered load per server pool is suitably bounded and that this results in a balanced distribution of the load. Furthermore, they established a similar result when tasks with Coxian distributed service times arrive at a constant rate and the threshold is adjusted using only the total number of tasks in the system. The novel proof technique developed in this paper, which differs from a traditional fluid limit analysis, allowed them to handle rapid variations of the first learning scheme, triggered by excursions of the occupancy process that have vanishing size. Moreover, our approach allowed them to characterize the asymptotic behavior of the system with Coxian distributed service times without relying on a fluid limit of a detailed state descriptor.

8.1.2 Server Saturation in Skewed Networks

In 40, 21, Diego Goldsztajn in collaboration with Sem Borst (Technical University of Eindhoven, the Netherlands) and Johan van Leeuwaarden (Tilburg University, the Netherlands) considered a model inspired by compatibility constraints that arise between tasks and servers in data centers, cloud computing systems and content delivery networks. The constraints are represented by a bipartite graph or network that interconnects dispatchers with compatible servers. Each dispatcher receives tasks over time and sends every task to a compatible server with the least number of tasks, or to a server with the least number of tasks among $d$ compatible servers selected uniformly at random. They focus on networks where the neighborhood of at least one server is skewed in a limiting regime. This means that a diverging number of dispatchers are in the neighborhood which are each compatible with a uniformly bounded number of servers; thus, the degree of the central server approaches infinity while the degrees of many neighboring dispatchers remain bounded. They proved that each server with a skewed neighborhood saturates, in the sense that the mean number of tasks queueing in front of it in steady state approaches infinity. Paradoxically, this pathological behavior can even arise in random networks where nearly all the servers have at most one task in the limit.

8.1.3 The Granger–Johansen Representation Theorem for Integrated Time Series on Banach Space

In 23, Konstantin Avrachenkov in collaboration with Phil Howlett (University of South Australia, Australia), Brendan Beare (University of Sydney, Australia), Massimo Franchi (Sapienza Universita di Roma, Italy) and John Boland (University of South Australia, Australia) proved an extended Granger-Johansen Representation Theorem (GJRT) for finite—or infinite—order integrated autoregressive time series on Banach space. They assumed only that the resolvent of the autoregressive polynomial for the series is analytic on and inside the unit circle except for an isolated singularity at unity. If the singularity is a pole of finite order, the time series is integrated of the same order. If the singularity is an essential singularity, the time series is integrated of order infinity. When there is no deterministic forcing, the value of the series at each time is the sum of an almost surely convergent stochastic trend, a deterministic term depending on the initial conditions and a finite sum of embedded white noise terms in the prior observations. This is the extended GJRT. In each case the original series is the sum of two separate autoregressive time series on complementary subspaces—a singular component which is integrated of the same order as the original series and a regular component which is not integrated. The extended GJRT applies to all integrated autoregressive processes irrespective of the spatial dimension, the number of stochastic trends and cointegrating relations in the system and the order of integration.

8.1.4 An Information Theoretic Metric for Measurement Vulnerability to Data Integrity Attacks on Smart Grids

In 28, Samir Medina Perlaza , Xiuzhen Ye, Iñaki Esnaola, and Robert F. Harrison (University of Sheffield, UK) studied a novel metric that describes the vulnerability of the measurements in power systems to data integrity attacks. The new metric, coined vulnerability index (VuIx), leverages information theoretic measures to assess the attack effect in terms of the fundamental limits of the disruption and detection tradeoff. The result of computing the VuIx of the measurements in the system yields an ordering of their vulnerability based on the degree of exposure to data integrity attacks. This new framework is used to assess the measurement vulnerability of IEEE 9-bus and 30-bus test systems and it is observed that power injection measurements are significantly more vulnerable to data integrity attacks than power flow measurements. A detailed numerical evaluation of the VuIx values for IEEE test systems is provided.

8.1.5 Development of pyMarmote and pyMarmoteMDP

The development of the Marmote library has continued, led by Alain Jean-Marie and Emmanuel Hyon of Univ. Paris-Nanterre. The software was awarded an Inria ADT (Action of Technological Development), during which several software engineers of Inria's SED (Service Experimentation Development) helped improve the robustness and portability of the library. Most of all, they endowed the library with a Python interface, named pyMarmote for the core, and pyMarmoteMDP for the application library devoted to Markov Decision Processes. A tutorial providing an introduction to the capabilities of Marmote via its Python interfaces was presented by Alain Jean-Marie and Emmanuel Hyon in 65. It is now part of the online documentation of the package.

8.2.1 Recovering Static and Time-Varying Communities Using Persistent Edges

In 13, Konstantin Avrachenkov in collaboration with Maximilien Dreveton (EPFL, Switzerland) and Lasse Leskelä (Aalto University, Finland) developed spectral methods for recovering communities in temporal networks. In the case of fixed communities, spectral clustering on the simple time-aggregated graph (i.e., the weighted graph formed by the sum of the interactions over all temporal snapshots) does not always produce satisfying results. To utilize information carried by temporal correlations, they proposed to employ different weights on freshly appearing and persistent edges. They showed that spectral clustering on such weighted graphs can be explained as a relaxation of the maximum likelihood estimator of an extension of the degree-corrected stochastic block model with Markov interactions. They also studied the setting of evolving communities, for which they used the prediction at time $t-1$ as an oracle for inferring the community labels at time $t$. They demonstrated the accuracy of the proposed methods on synthetic and real data sets.

8.2.2 Almost Exact Recovery in Noisy Semi-supervised Learning

Graph-based semi-supervised learning methods combine the graph structure and labeled data to classify unlabeled data. In 12, Konstantin Avrachenkov and Maximilien Dreveton (EPFL, Switzerland) studied the effect of a noisy oracle on classification. In particular, they derived the maximum a posteriori (MAP) estimator for clustering a degree corrected stochastic block model when a noisy oracle reveals a fraction of the labels. They then proposed an algorithm derived from a continuous relaxation of the MAP, and established its consistency. Numerical experiments show that the proposed approach achieves promising performance on synthetic and real data sets, even in the case of very noisy labeled data.

8.2.3 The Generalization Error of Machine Learning Algorithms

The generalization error is one of the standard metrics used to evaluate the generalization capabilities of machine learning algorithms. Selecting a good algorithm in terms of generalization can be framed as a problem of Empirical Risk Minimization (ERM). When this minimization problem is regularized by incorporating a penalty term based on relative entropy, the resulting unique solution is a Gibbs probability measure (a Gibbs algorithm), which possesses several notable properties. Interestingly, the worst-case data-generating probability measure under this framework is also a Gibbs probability measure. Exploiting this dual characterization of Gibbs measures has led to several significant results, as outlined below.

In 62, Samir Medina Perlaza and Xinying Zou introduced the method of gaps, a novel technique for deriving closed-form expressions in terms of information measures for the generalization error of machine learning algorithms. The method relies on two central observations: 1) the generalization error is an average of the variation of the expected empirical risk with respect to changes on the probability measure (used for expectation); and 2) these variations, also referred to as gaps, exhibit closed-form expressions in terms of information measures. The expectation of the empirical risk can be either with respect to a measure on the models (with a fixed dataset) or with respect to a measure on the datasets (with a fixed model), which results in two variants of the method of gaps. The first variant, which focuses on the gaps of the expected empirical risk with respect to a measure on the models, appears to be the most general, as no assumptions are made on the distribution of the datasets. The second variant develops under the assumption that datasets are made of independent and identically distributed data points. All existing exact expressions for the generalization error of machine learning algorithms can be obtained with the proposed method. Also, this method allows obtaining numerous new exact expressions, which improves the understanding of the generalization error; establishes connections with other areas in statistics, e.g., hypothesis testing; and potentially, might guide algorithm designs.

In 25, Samir Medina Perlaza , Alain Jean-Marie , and Gaetan Bisson (Univ. Polynésie française), Iñaki Esnaola (University of Sheffield, UK), and Stefano Rini (National Chiao Tung University, Taiwan) studied the ERM problem with relative entropy regularization (ERM-RER) under the assumption that the reference measure is a $\sigma$-finite measure, and not necessarily a probability measure. Under this assumption, which leads to a generalization of the ERM-RER problem allowing a larger degree of flexibility for incorporating prior knowledge, numerous relevant properties are stated. Among these properties, the solution to this problem, if it exists, is shown to be a unique probability measure, mutually absolutely continuous with the reference measure. Such a solution, which is a Gibbs probability measure, exhibits a probably-approximately-correct guarantee for the ERM problem independently of whether the latter possesses a solution. For a fixed dataset and under a specific condition, the empirical risk is shown to be a sub-Gaussian random variable when the models are sampled from the solution to the ERM-RER problem. The generalization capabilities of the solution to the ERM-RER problem (the Gibbs algorithm) are studied via the sensitivity of the expected empirical risk to deviations from such a solution towards alternative probability measures. Finally, an interesting connection between sensitivity, generalization error, and lautum information is established.

In 57, José Francisco Daunas Torres and Samir Medina Perlaza , along with Iñaki Esnaola (University of Sheffield, UK) and H. Vincent Poor (Princeton University, USA), analyzed the effect of relative entropy asymmetry in the context of ERM-RER. Two regularizations are considered:

1. the relative entropy of the measure to be optimized with respect to a reference measure (Type-I ERM-RER); and
2. the relative entropy of the reference measure with respect to the measure to be optimized (Type-II ERM-RER).

The main result is the characterization of the solution to the Type-II ERM-RER problem and its key properties. By comparing the well-understood Type-I ERM-RER with Type-II ERM-RER, the effects of entropy asymmetry are highlighted. The analysis shows that in both cases, regularization by relative entropy forces the solution's support to collapse into the support of the reference measure, introducing a strong inductive bias that can overshadow the evidence provided by the training data. Finally, it is shown that Type-II regularization is equivalent to Type-I regularization with an appropriate transformation of the empirical risk function.

In 36, the same authors studied the solution to ERM with $f$-divergence regularization (ERM-fDR) under mild conditions on $f$. Under such conditions, the optimal measure is shown to be unique. Examples of the solution for particular choices of the function $f$ are presented. Previously known solutions to common regularization choices are obtained by leveraging the flexibility of the family of $f$-divergences. These include the unique solutions to empirical risk minimization with relative entropy regularization (Type-I and Type-II). Furthermore, the analysis of the solution unveils the following properties of $f$-divergences when used in the ERM-fDR problem:

• $f$-divergence regularization forces the support of the solution to coincide with the support of the reference measure, introducing a strong inductive bias that dominates the evidence provided by the training data; and
• any $f$-divergence regularization is equivalent to a different $f$-divergence regularization with an appropriate transformation of the empirical risk function.

In 29, 49, Xinying Zou , Samir Medina Perlaza , Eitan Altman , along with Iñaki Esnaola (University of Sheffield, UK), introduced the worst-case probability measure over the data as a tool for characterizing the generalization capabilities of machine learning algorithms. More specifically, the worst-case probability measure is a Gibbs probability measure and the unique solution to the maximization of the expected loss under a relative entropy constraint with respect to a reference probability measure. Fundamental generalization metrics, such as the sensitivity of the expected loss, the sensitivity of the empirical risk, and the generalization gap, are shown to have closed-form expressions involving the worst-case data-generating probability measure. Existing results for the Gibbs algorithm, such as characterizing the generalization gap as a sum of mutual information and lautum information, up to a constant factor, are recovered. A novel parallel is established between the worst-case data-generating probability measure and the Gibbs algorithm. Specifically, the Gibbs probability measure is identified as a fundamental commonality of the model space and the data space for machine learning algorithms.

In 64, the same authors and H. Vincent Poor (Princeton University, USA), studied the worst-case data-generating (WCDG) probability measure as a tool for characterizing the generalization capabilities of machine learning algorithms. Such a WCDG probability measure is shown to be the unique solution to two different optimization problems:

1. the maximization of the expected loss over the set of probability measures on the datasets whose relative entropy with respect to a reference measure is not larger than a given threshold; and
2. the maximization of the expected loss with regularization by relative entropy with respect to the reference measure.

Such a reference measure can be interpreted as a prior on the datasets. The WCDG cumulants are finite and bounded in terms of the cumulants of the reference measure. To analyze the concentration of the expected empirical risk induced by the WCDG probability measure, the notion of $(ϵ,\delta )$-robustness of models is introduced. Closed-form expressions are presented for the sensitivity of the expected loss for a fixed model. These results lead to novel expressions for the generalization error of arbitrary machine learning algorithms. These expressions can be broadly divided into two classes. The first one involves a WCDG probability measure, while the second one involves a Gibbs algorithm. This finding reveals that an exploration into the generalization error of the Gibbs algorithm facilitates the derivation of overarching insights applicable to any machine learning algorithm.

8.2.4 Density-based Clustering

Many of clustering algorithms are based on density estimates in $R^d$. Building geometric graphs on the dataset $X$ is an elegant way of doing this. In fact, the connected components of a geometric graph match exactly with the high-density clusters of the 1-Nearest Neighbor density estimator. In 41, Louis Hauseux , Konstantin Avrachenkov and Josiane Zerubia (Ayana) showed that the natural way to generalize geometric graphs is to use hypergraphs with a more restrictive notion of connected component called K-Polyhedron. Herein, they proved that K-Polyhedra correspond to high-density clusters of K-Nearest Neighbors density estimator. Furthermore, the percolation phenomenon is omnipresent behind the family of clustering algorithms they studied in this paper.

Many of clustering algorithms for a point cloud $X_n\subset R^d$ in the Euclidean space are based on density estimates. In fact, the density function $f$ of point generation contains the relevant information. It is quite natural to try to extract what Hartigan called `high-density clusters'. One elegant solution to do this task consists in constructing a graph whose nodes are the points of the cloud and whose edges connect nearby points. One wants the connected components of this graph to reflect the high-density clusters. Some very classical algorithms such as (Robust) Single-Linkage or (H)DBSCAN work in this way. It is particularly helpful because its connected components correspond exactly to the high-density clusters of the density estimator of the 1-Nearest Neighbor. An example of the Single-Linkage will show us the mathematical phenomenon at the heart of these algorithms: the percolation. In 43, Louis Hauseux defined and measured the percolation rate to evaluate the performance of such algorithms. By way of example, the Robust Single-Linkage algorithm is analyzed and its percolation rate is calculated. This showed theoretically why it is actually preferable to use $K$-Nearest Neighbours rather than 1-NN. However, convergence in $K$ towards a perfect estimator is very slow, so this analysis explains why in practice $K=10$ often represents a good trade-off.

In 42, Louis Hauseux , Konstantin Avrachenkov and Josiane Zerubia (Ayana) firstly studied how to measure the theoretical performance of hierarchical clustering algorithms which depend on a scale parameter; and secondly, improved the State-of-the-Art of the family of hierarchical clustering with respect to the introduced new measure. Single-Linkage is perhaps the simplest and the most famous algorithm belonging to this family of hierarchical clustering. Nowadays, the State-of-the-Art clustering algorithm, HDBSCAN, works in a similar way to Single-Linkage (with some refinements analyzed in this work). Specifically, they explained why the percolation phenomenon is omnipresent behind this family of clustering algorithms. Previously, they defined an index named “percolation rate” to measure the theoretical capability of algorithms to identify different high-density levels. In 42, they showed that using hypergraphs is a natural way to generalize Single-Linkage with higher-order interactions. New high-order connected components on hypergraphs, `$K$-polyhedra', have much better percolation rates than the classic `robustification' of Single-Linkage (used e.g. by HDBSCAN), the $K$-Robust Single-Linkage components. They investigate in detail the important cases of $R^2$ and $R^3$ for $K\in \{1,2,3\}$.

8.2.5 Reinforcement Learning for Networks and Weakly Coupled MDPs

The Whittle index policy is a heuristic that has shown remarkably good performance (with guaranteed asymptotic optimality) when applied to the class of problems known as Restless Multi-Armed Bandit Problems (RMABPs). In 27, Konstantin Avrachenkov in collaboration with Francisco Robledo and Urtzi Ayesta (IRIT) and Vivek Borkar (IITB, India) presented QWI and QWINN, two reinforcement learning algorithms, respectively tabular and deep, to learn the Whittle index for the total discounted criterion. The key feature is the use of two time-scales, a faster one to update the state-action Q-values, and a relatively slower one to update the Whittle indices. As the main theoretical result they showed that QWI, which is a tabular implementation, converges to the real Whittle indices. They then presented QWINN, an adaptation of QWI algorithm using neural networks to compute the Q-values on the faster time-scale, which is able to extrapolate information from one state to another and scales naturally to large state-space environments. For QWINN, they showed that all local minima of the Bellman error are locally stable equilibria, which is the first result of its kind for DQN-based schemes. Numerical computations showed that QWI and QWINN converge faster than the standard Q-learning algorithm, neural-network based approximate Q-learning and other state of the art algorithms.

In 47, Konstantin Avrachenkov in collaboration with Francisco Robledo and Urtzi Ayesta (IRIT) introduced the Lagrange Policy for Continuous Actions (LPCA), a reinforcement learning algorithm specifically designed for weakly coupled Markov Decision Process (MDP) problems with continuous action spaces. LPCA addresses the challenge of resource constraints dependent on continuous actions by introducing a Lagrange relaxation of the weakly coupled MDP problem within a neural network framework for Q-value computation. This approach effectively decouples the MDP, enabling efficient policy learning in resource-constrained environments. They presented two variations of LPCA: LPCA-DE, which utilizes differential evolution for global optimization, and LPCA-Greedy, a method that incrementally and greadily selects actions based on Q-value gradients. Comparative analysis against other state-of-the-art techniques across various settings highlight LPCA's robustness and efficiency in managing resource allocation while maximizing rewards.

Influence maximization (IM) has been widely studied in recent decades, aiming to maximize the spread of influence over networks. Despite many works for static networks, fewer research studies have been dedicated to the IM problem for dynamic networks, which creates many challenges. An IM method for such an environment, should consider its dynamics and perform well under different network structures. To fulfill this objective, more computations are required. Hence, an IM approach should be efficient enough to be applicable for the ever-changing structure of a network. In 17, Konstantin Avrachenkov and Kishor Patil in collaboration with Haleh Dizaji (University of Klagenfurt, Austria) proposed an IM method for dynamic networks which uses a deep Q-learning (DQL) approach. To learn dynamic features from the network and retain previously learned information, incremental and transfer learning methods have been applied. Experiments substantiate the good performance of the DQL methods and their superiority over compared methods on larger sizes of tested synthetic and real-world networks. These experiments illustrate better performance for incremental and transfer learning methods on real-world networks.

8.4.1 Games among selfish and team stations in polling systems

In 30, Khushboo Agarwal and Eitan Altman focus on a particular polling system known as the cyclic Bernoulli polling (CBP) system, where a server moves cyclically between the stations and serves the queue at a station with a certain probability when polled. Each station follows either a gated or partially exhaustive service discipline. In the steady state of such a system, they study a new game-theoretic aspect, where, the stations strategically choose the probability of accepting or rejecting the service from the server when polled. They examine three variants of non-cooperative games among stations: (i) each station selfishly minimizes its expected waiting time, (ii) a team game where each station minimizes the expected workload of the system, and (iii) stations act with partial cooperation, incurring an additional linear cost. They begin by presenting a new result for the CBP system regarding the continuity of expected waiting times in relation to the probabilities selected by the stations. For each game, they then investigate the existence and uniqueness of the Nash equilibrium (NE). In some cases, the NE is explicitly derived, while in others, characterizing the NE remains challenging due to the complex dependence of waiting times on the non-trivial buffer occupancy equations. Nonetheless, they analyze the NE and its properties through numerical experiments. Notably, in many instances, stations opt to accept service with a probability less than 1—a trend observed even among selfish stations.

8.4.2 Leveraging Noisy Observations in Zero-Sum Games

In 55, 32, Emmanouil Marios Athanasakos and Samir Medina Perlaza studied an instance of zero-sum games in which one player (the leader) commits to its opponent (the follower) to choose its actions by sampling a given probability measure (strategy). The actions of the leader are observed by the follower as the output of an arbitrary channel. In response to that, the follower chooses its action based on its current information, that is, the leader's commitment and the corresponding noisy observation of its action. Within this context, the equilibrium of the game with noisy action observability is shown to always exist and the necessary conditions for its uniqueness are identified. Interestingly, the noisy observations have important impact on the cardinality of the follower's set of best responses. Under particular conditions, such a set of best responses is proved to be a singleton almost surely. The proposed model captures any channel noise with a density with respect to the Lebesgue measure. As an example, the case in which the channel is described by a Gaussian probability measure is investigated.

8.4.3 Constrained Correlated Equilibria

In 33, Eitan Altman , in collaboration with Omar Boufous, Mikaël Touati, and Mustapha Bouhtou (Orange Labs), and Rachid El-Azouzi (Avignon Univ.), introduces constrained correlated equilibrium, a solution concept combining correlation and coupled constraints in finite non-cooperative games. They study the conditions for equilibrium in the general case of an arbitrary correlation device and coupled constraints in the extended game. In the particular case of constraints induced by a feasible set of probability distributions over action profiles, they first show that canonical correlation devices are sufficient to characterize the set of constrained correlated equilibrium distributions and provide conditions for their existence. Second, they demonstrate that constrained correlated equilibria of the mixed extension of the game do not lead to additional equilibrium distributions. Third, they show that the constrained correlated equilibrium distributions may not belong to the polytope of correlated equilibrium distributions. Finally, they illustrate these results through numerical examples.

8.4.4 Positional and conformist effects in voluntary public good provision

The literature featuring game–theoretical models aimed at explaining the effect of the status concerns on the voluntary provision of a public good is generally focused on snob agents, driven by a desire for exclusiveness. However, the social context literature highlights that status concerns can give rise to a desire, in some individuals to be different from the “common herd”, and in some others to conform with other people. In 15, Francisco Cabo (Univ. Valladolid), Mabel Tidball (INRAe) and Alain Jean-Marie analyze a two-player public good game under two different settings: The standard case with two positional players, versus the case in which the positional player faces a conformist player. Giving entrance to conformism has two main implications. Strong status concerns by both players can lead to a (socially) virtuous cycle in which the conformist player wishes to imitate the contributing behavior of the positional player, and the latter wishes to increase contribution to distinguish herself from the former. Then, the contribution to the public good can be higher than in the case with only snob agents. This higher contribution can increase social welfare, but only if endowments are not too large and the status concern of the positional player is not excessively high.

8.4.5 Equilibrium bids for reverse auctions with budget constraints

Alain Jean-Marie and Mabel Tidball (INRAe) consider in 60 a relatively new class of auctions: the budget-constrained reverse auction. In this auction, offers are accepted until some predefined budget is exhausted. Payments are discriminatory: bidders receive their own bid as payment when they win the bid. Bidders have a private value for their item which is uniformly distributed between 0 and 1; they are risk-neutral and items are indivisible. For small budget and an arbitrary number of bidders, as well as in the case of two bidders, an equilibrium bidding function is identified. When possible, this equilibrium bidding function is compared with that of target-constrained reverse auctions.

8.5.1 Local Approximation of Secrecy Capacity

In 31, Emmanouil Marios Athanasakos , Hariprasad Manjunath Hegde , and Nicholas Kalouptsidis studied the wiretap channel using Euclidean Information Theory (EIT). The focus is on a scenario in which a small amount of information shall be efficiently transmitted subject to compression and secrecy constraints. The information-theoretic problem is transformed into a linear algebra problem and obtain the perturbed probability distributions such that secrecy is achievable. An approximate estimate of the secrecy capacity is obtained by solving a generalized eigenvalue problem.

8.5.2 Caching

Probabilistic analysis of caches has been the topic of Younes Ben Mazziane 's PhD Thesis 51.

In 18, Giovanni Neglia , in collaboration with Francescomaria Faticanti (École Normale Supérieure de Lyon, previously postdoc in Neo), investigates "optimistic" online caching policies, distinguished by their use of future request predictions derived, for example, from machine learning models. Traditional online optimistic policies, grounded in the Follow-The-Regularized-Leader (FTRL) algorithm, incur a higher computational cost compared to classic policies like Least Frequently Used (LFU) and Least Recently Used (LRU). This is due to each cache state update necessitating the resolution of a constrained optimization problem. To address this problem, they introduce and analyze the "batched" version of two distinct FTRL-based optimistic policies. In this approach, the cache updates occur less frequently, thereby amortizing the update cost over time or over multiple requests. Rather than updating the cache with each new request, the system accumulates a batch of requests before modifying the cache content. First, they present a batched version of the Optimistic Bipartite Caching (OBC) algorithm, which works for single requests, then introduce a new optimistic batched caching policy, the Per-Coordinate Optimistic Caching (PCOC) algorithm, derived from the per-coordinate-based FTRL. They demonstrate that these online algorithms maintain "vanishing regret" in the batched case, meaning their average performance approaches over time that of an optimal static file allocation, regardless of the sequence of file requests. They then compare the performance of these two strategies with each other and against optimistic versions of LFU and LRU. Their experimental results indicate that this batched optimistic approach outperforms traditional caching policies on both stationary and real-world file request traces.

Commonly used caching policies, such as LRU or LFU, exhibit optimal performance only under specific traffic patterns. Even advanced machine learning-based methods, which detect patterns in historical request data, struggle when future requests deviate from past trends. Recently, a new class of policies has emerged that is robust to varying traffic patterns. These algorithms address an online optimization problem, enabling continuous adaptation to the context. They offer theoretical guarantees on the regret metric. However, the high computational complexity of these solutions hinders their practical adoption. In an ongoing work 56, Giovanni Neglia in collaboration with Damiano Carra (Univ. of Verona, Italy) introduce a new variant of the gradient-based online caching policy that achieves groundbreaking logarithmic computational complexity relative to catalog size, while also providing regret guarantees. This advancement allows them to test the policy on large-scale, real-world traces featuring millions of requests and items—a significant achievement, as such scales have been beyond the reach of existing policies with regret guarantees. Their experimental results demonstrate for the first time that the regret guarantees of gradient-based caching policies offer substantial benefits in practical scenarios.

In 14, Younes Ben Mazziane , Sara Alouf , and Giovanni Neglia , in collaboration with Daniel S. Menasche (Federal Univ. of Rio de Janeiro, Brazil), address the challenge of similarity caching, which allows requests for an item to be served by a similar item—a concept applicable in recommendation systems, multimedia retrieval, and machine learning. They note that while many similarity caching policies, such as SIM-LRU and its generalization RND-LRU, have been proposed, their performance analysis, particularly regarding hit ratio, remains limited. To address this, they extend the popular time-to-live (TTL) approximation from classic caching to similarity caching. Specifically, they propose the RND-TTL approximation method to estimate the hit ratio of the RND-LRU similarity caching policy. This involves introducing the RND-TTL cache model and tuning its parameters to emulate RND-LRU's behavior. The parameter tuning requires solving a fixed-point system of equations, for which they provide both an algorithm for numerical resolution and sufficient conditions for convergence. Their approach is evaluated using both synthetic and real-world traces, demonstrating its effectiveness in approximating the hit ratio of RND-LRU.

8.5.3 Design and cross-layer optimization of low cost RIS-assisted communication systems

In 16, Eitan Altman , in collaboration with Antoine Dejonghe and Zwi Altman (Orange Labs) and Francesco De Pellegrini (Avignon Univ.), introduces a new cross-layer low-complexity scheme for the online optimization of RIS-assisted communication systems in Orthogonal Frequency-Division Multiple Access (OFDMA) systems. This scheme jointly combines base station (BS) and reconfigurable intelligent surface (RIS) configuration with fair user equipment (UE) scheduling, which is critical for high-performance deployments. A RIS beam synthesis method is proposed for RIS configuration. The proposed solution embeds two nested control loops: i) a fast control loop operating at the OFDMA slot scale, consisting of a standard UE proportional fair scheduler; ii) a slow control loop operating at the OFDMA frame scale, which adapts the RIS configuration to the UEs' spatial distribution and maximizes the UEs' aggregated performance. The slow control loop is based on an online stochastic approximation algorithm, with convergence to the optimal rest point proved. In a reference scenario, the proposed scheduler achieves a gain of 47% in mean spectral efficiency for NLOS UEs over a baseline scheme.

8.5.4 Optimal Flow Admission Control in Edge Computing via Safe Reinforcement Learning

With the uptake of intelligent data-driven applications, edge computing infrastructures necessitate a new generation of admission control algorithms to maximize system performance under limited and highly heterogeneous resources. In 38, Eitan Altman , in collaboration with Andrea Fox, Francesco de Pellegrini (Avignon Univ.), and Francesco Bronzino (ENS Lyon), studies how to optimally select information flows which belong to different classes and dispatch them to multiple edge servers where applications perform flow analytic tasks. The optimal policy is obtained via the theory of constrained Markov decision processes to take into account the demand of each edge application for specific classes of flows, the constraints on computing capacity of edge servers and the constraints on access network capacity. They develop DRCPO, a specialized primal-dual Safe Reinforcement Learning (SRL) method which solves the resulting optimal admission control problem by reward decomposition. DRCPO operates optimal decentralized control and mitigates effectively state-space explosion while preserving optimality. Compared to existing Deep Reinforcement Learning (DRL) solutions, extensive results show that it achieves 15% higher reward on a wide variety of environments, while requiring on average only 50% learning episodes to converge. Finally, they further improve the system performance by matching DRCPO with load-balancing in order to dispatch optimally information flows to the available edge servers.

8.5.5 Learning optimal edge processing with offloading and energy harvesting

In 19, Eitan Altman , in collaboration with Andrea Fox and Francesco De Pellegrini (Avignon Univ.), introduces a Markovian model for systems where multiple battery-operated devices perform data processing and energy harvesting in parallel. The study focuses on modern portable devices capable of executing sophisticated AI models on sensed data, with data-dependent complexity and significant energy costs. The model addresses scenarios where part of the computational burden is offloaded to an edge server that polls devices at a given rate. The authors derive structural properties of an optimal policy for a single device-server system, leading to the development of a new model-free reinforcement learning method specialized for monotone policies, termed Ordered Q-Learning. This method provides a fast procedure to learn the optimal policy, independent of devices' battery capacities, the cost and value of data batch processing, and the dynamics of the energy harvesting process. Additionally, the server's polling strategy is optimized by combining this policy improvement technique with stochastic approximation methods. Extensive numerical results offer insights into system properties and demonstrate that the proposed learning algorithms outperform existing baselines.

8.5.6 Covert Communications

In 46, Philippe Nain , in collaboration with Amir Reza Ramtin and Don Towsley (Univ. Massachusetts), James Z. Hare and Lance Kaplan (DEVCOM Army Research Laboratory), and Venugopal Veeravalli (Univ. of Illinois), investigates the damage that an adversary can effect while remaining covert in the presence of the Cumulative Sum (CuSum) procedure. An adversary is covert if the time to detection is on the same order as the time to false alarm. Damage is given as an increasing function of the Kullback–Leibler divergence of the adversarial actions and the normal distribution prior to the adversarial attack. By analyzing the problem with a focus on the growth function g(n), which measures the cumulative expected log-likelihood ratio after n time slots following the change, conditions are established under which the adversary remains covert. An analysis of the impact of different adversarial strategies on damage is also provided.

9.1.1 Accenture “Plan de Relance” (PLR) contract on the topic “Energy-Aware Federated Learning” (Oct 2022 – September 2024)

Participants: Giovanni Neglia, Charlotte Rodriguez.

• Contractor: Accenture Labs
• Collaborators: Laura Degioanni, Laetitia Kameni, Richard Vidal

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. Publications in 2024: 63.

9.1.2 Contracts with Nokia

9.1.3 Cifre contract with NSP-SmartProfile “Automated and responsible recommendation systems for digital marketing” (August 2022 – July 2025)

Participants: Konstantin Avrachenkov, Ibtihal El Mimouni.

• Contractor: NSP-SmartProfile
• Collaborators: Hervé Baile, Julien Musso

SmartProfile is a marketing platform that allows to collect, to enhance and to analyze marketing data. Digital marketing campaigns continue to expand across all digital channels and media. The 'mass marketing' strategies implemented by most companies show limits in terms of performance and acceptance by clients, as well as in terms of their impact on the environment. In opposite to these practices, we believe that current technologies, particularly in terms of Artificial Intelligence (AI), should make marketing interactions more efficient and virtuous. Through this research project, we want to create an alternative solution to mass marketing by switching to an intelligent, automated and eco-responsible system, which will support the heterogeneity of data and the diversity of sectors, and whose purpose is to recommend the best content by determining the most relevant target and taking into account the communication constraints. This contract complements the Cifre thesis of Ibtihal El Mimouni.

9.1.4 Cifre contract with Orange Labs “Analytical modeling of large-scale wireless networks integrating RIS” (March 2023 – March 2026)

Participants: Eitan Altman, Julian Alfonso Santos Bustos.

• Contractor: Orange Labs
• Collaborators: Jean-Marc Kelif

A Reconfigurable Intelligent Surface (RIS) is a programmable surface structure that allows one to control the reflection of electromagnetic (EM) waves by changing the electric and magnetic properties of the surface. In the absence of RIS, short wavelentghs signals as in 5G, are subject to a huge attenuation when there is no direct line of sight channel. Within our collaboration, we shall evaluate and optimize the position of RIS.

This contract complements the Cifre thesis of J. Santos.

9.1.5 QITI 3IA Start-It-Up contract on the topic “Reinforcement Learning for Conversational Recommender Systems (RLCRS)” (April 2023 – March 2024)

Participants: Konstantin Avrachenkov, Hariprasad Manjunath Hegde.

• Contractor: QITI
• Collaborators: Christophe Bremard, Guillaume Dion

Qiti is a start-up created in Nice in 2021 which, among other things, develops a Conversational Recommender Systems (CRS) for insurance holders and insurers. The CRS should reduce the load on the workers of the insurers and to simplify the process of insurance establishment and modification. The goal of the present cooperation is to test and to improve various Reinforcement Learning schemes for CRS. The post-doc of H. Manjunath is funded by this contract.

9.1.6 Cifre contract with SAP “Privacy and fairness for ML” (December 2021 – December 2024)

Participants: Caelin Kaplan, Giovanni Neglia.

• Contractor: SAP Labs France
• Collaborators: Anderson Santana de Oliveira

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. This contract complements the Cifre thesis of C. Kaplan. Publications in 2024: 44, 53, 61.

10.1.1 Inria associate team not involved in an IIL or an international program

10.1.2 STIC/MATH/CLIMAT AmSud projects

10.1.3 Participation in other International Programs

10.2.1 Visits of international scientists

10.2.2 Visits to international teams

10.3.1 Horizon Europe

10.3.2 H2020 projects

NF-FOUNDS PC9 PEPR 5G

Participants: Khushboo Agarwal, Eitan Altman, Samir Medina Perlaza.

• Project Acronym:
  NF-FOUNDS
• Project Title:
  Networks of the Future - Foundations of Future Communications Networks
• Program:
  ANR-22-PEFT-0010
• Coordinator:
  CEA (Dmitri Kténas), CNRS (Serge Verdeyme), IMT (Daniel Koffman)
• Duration:
  2023 - 2030
• Other Partners:
  EURECOM
• Abstract:
  The 5G network and the networks of the future represent a key issue for French and European industry, society and digital sovereignty. This is why the French government has decided to launch a dedicated national strategy. One of this strategy's priority ambitions is to produce significant public research efforts so the national scientific community contributes fully to making progress that clearly responds to the challenges of 5G and the networks of the future. In this context, the CNRS, the CEA and the Institut Mines-Télécom (IMT) are co-leading the '5G' acceleration PEPR to support upstream research into the development of advanced technologies for 5G and the networks of the future. Neo  is involved in the theme "Networks and Telecommunications" and more specifically in the targeted projet 9 (PC9) Foundations of Future Communications Networks (FOUNDS).
• Publications in 2024:
  30.

ANR PARFAIT

Participants: Eitan Altman, Ashok Krishnan Komalan Sindhu, Samir Medina Perlaza, Xinying Zou.

• Project Acronym:
  PARFAIT
• Project Title:
  Planning And leaRning For AI-Edge compuTing
• Coordinator:
  Avignon Univ.
• Duration:
  October 2021 - September 2025
• Other Partners:
  Conservatoire National des Arts et Métiers (CNAM), Univ. Savoie Mont Blanc (USMB)
• Abstract:
  The PARFAIT project develops theoretical foundations for distributed and scalable resource allocation schemes on edge computing infrastructures tailored for AI-based processing tasks. Algorithmic solutions will be developed based on the theory of constrained, delayed, and distributed Markov decision processes to account for edge service orchestration actions and quantify the effect of orchestration policies. Furthermore, using both game and team formulations, the project will pave the way for a theory of decentralized orchestration, a missing building block necessary to match the application quest for data proximity and the synchronization problems that arise when multiple edge orchestrators cooperate under local or partial system view. Finally, to achieve efficient online edge service orchestration, such solutions will be empowered with reinforcement learning techniques to define a suit of orchestration algorithms able to at once adapt over time to the applications' load and cope with the uncertain information available from AI-based applications' footprints.
• Publications in 2024:
  29, 38, 49, 62, 64.

Inria Challenge FedMalin

Participants: Othmane Marfoq, Giovanni Neglia, Angelo Rodio.

• Project Acronym:
  FedMalin
• Project Title:
  FEDerated MAchine Learning over the INternet
• Coordinator:
  Giovanni Neglia and Aurélien Bellet (Inria team Premedical)
• Duration:
  November 2022 - November 2026
• Abstract:

  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.

• Publications in 2024:
  34, 35, 45, 48, 58.

DIAMOND

Participants: Yaiza Bermudez, Samir Medina Perlaza.

• Project Acronym:
  DIAMOND
• Project Title:
  Data-Injection Attacks in Supervised Machine Learning Systems
• Funding Agency:
  Agence de l'Innovation de Défense (AID)
• Principal Investigators:
  Samir Medina Perlaza and Iñaki Esnaola (University of Sheffield)
• Duration:
  November 2024 - October 2027
• Note:
  The project funds the PhD scholarship of Yaiza Bermudez .

11.1.1 Scientific events: organisation

11.1.2 Scientific events: selection

11.1.3 Journal

11.1.4 Invited talks

• Sara Alouf delivered
  • a lecture "Performance modeling of similarity caching" at the MAKI Scientific Workshop 2024, Darmstadt;
  • a seminar "TTL Model for an LRU-Based Similarity Caching Policy" at IISc CNI Seminar Series, January 2024, Online.
• Konstantin Avrachenkov delivered a lecture "A Journey Through the History of Markov Chains" at 13ème Atelier en Évaluation des Performances, December 2-4, 2024, Toulouse.
• Diego Goldsztajn delivered a seminar "Asymptotically Optimal Policies for Weakly Coupled Markov Decision Processes" at LAAS Solace Seminar Series, October 2024, Toulouse.
• Louis Hauseux delivered a seminar "Density-based clustering: (Hyper-)Graphs & Percolation" at Università di Genova, April 2024, Genova, Italy.
• Alain Jean-Marie delivered a plenary talk "Challenges in Dynamic Stackelberg Games" at the 2024 IWOBIP, Rancagua, Chile.
• Samir Medina Perlaza delivered
  • a talk "From Statistical Empirical Risk Minimization with $f$-Divergence Regularization to Generalization Theory", at the Institut Henri Poincaré, GDR-IASIS, 5 July 2024, Paris, France;
  • an invited talk "An Exact Characterization of the Generalization Error of Machine Learning Algorithms" at the Polytech Annecy-Chambéry / LISTIC, Université Savoie Mont Blanc, Annecy Le Vieux, 21 February 2024. Host: Prof. Kave Salamantian.
• Giovanni Neglia delivered
  • an invited talk on "Personalized Federated Learning" at EPFL Control Seminar Series, February 23, 2024, Lausanne, Switzerland;
  • an invited talk on "Collaborative Inference Systems" at 3rd Workshop on Principles of Distributed Learning 2024, June 21, 2024, Nantes;
  • an invited talk on "Scalable Decentralized Algorithms for Online Personalized Mean Estimation" at the Workshop on distributed computing, optimization & learning 2024, May 22-23, 2024, Paris;
  • an invited talk "Scalable Decentralized Algorithms for Online Personalized Mean Estimation" at 13ème Atelier en Évaluation des Performances, December 2-4, 2024, Toulouse.

11.1.5 Tutorials

• Alain Jean-Marie delivered a tutorial on "An introduction to pyMarmote and pyMarmoteMDP for Markovian modeling" at the SIGMETRICS/PERFORMANCE 2024 conference 65.
• Samir Medina Perlaza delivered a tutorial on "Characterizing the Generalization Error of Machine Learning Algorithms Via Information Measures" in the IEEE Information Theory Workshop 2024.

11.1.6 Leadership within the scientific community

• Eitan Altman
  • is Fellow Member of IEEE;
  • is Member of WG 7.3 of IFIP on Computer System Modeling;
  • is the elected Vice Chairman of WG 6.3 of IFIP on Performance of Communications Systems.
• Konstantin Avrachenkov is a member of Conseil Scientifique & Pédagogique EUR DS4H Univ. Côte d'Azur.
• Samir Medina Perlaza
  • is a member of the Digital Presence Committee of the IEEE Information Theory Society;
  • is a workpackage leader of the PEPR - Réseaux du Futur – A project funded by the French National Agency for Research (ANR) via the project n°ANR-22-PEFT-0010 of the France 2030 program;
  • is the organizer of the PC9 Seminar on Wireless Communications, a national online seminar part of the PEPR – Réseaux du Futur.
• Giovanni Neglia
  • holds a Chair by the Interdisciplinary Institute for Artificial Intelligence 3IA Côte d'Azur, in the theme "Core Elements of AI".
  • participated in the panel on “Challenges and opportunities of Distributed Learning" organized by the NetworkingChannel, December 11, 2024, online.

11.1.7 Research administration

• Sara Alouf
  • is a member of the Colloquium Jacques Morgenstern Committee of Inria center at Université Côte d'Azur, since March 2023;
  • is a member of NICE, the Invited Researchers Committee of Inria center at Université Côte d'Azur, since June 2020;
  • is member of CLF, the training committee of Inria center at Université Côte d'Azur, November 2014 - 2024;
  • is vice-head of project-team Neo since January 2017;
  • is a member of the Selection Committee of the dAIEGDE European project.
• Konstantin Avrachenkov is a representant (suppléant, collège A) before the Center Committee of Inria center at Université Côte d'Azur.
• Louis Hauseux is
  • a representant (titulaire, collège C) before the Center Committee of Inria center at Université Côte d'Azur;
  • a representant member of the catering commission of Inria center at Université Côte d'Azur.
• Alain Jean-Marie
  • was the scientific coordinator of Inria activities in Montpellier (2008-2023); as part of this duty, he represented Inria at: the Scientific Council of the Doctoral School “Sciences and Agrosciences” of the Univ. of Avignon; at the Regional Conference of Research Organisms (CODOR); at the board of the Labex NUMEV;
  • is Head of project-team NEO since January 2017.
• Philippe Nain chaired the Inria ad-hoc commission in charge of evaluating applications for the 2025 bonus campaign (RIPEC-C3).
• Samir Medina Perlaza
  • is a representant (suppléant, collège A) before the Center Committee of Inria center at Université Côte d'Azur;
  • is a Member of the Bureau of the Réseaux, Information et Société Numérique Excellence Academy of Université Côte d'Azur;
  • represents Inria at the Conseil du Département Disciplinaire Informatique of Université Côte d'Azur.
• Giovanni Neglia
  • was a member of the competitive exam jury for a researcher position (CRCN/ISFP) at Inria Paris Center in 2024;
  • is an elected member of Inria evaluation committee, since September 2023;
  • is a member of the steering commitee of UniCA Graduate School of Digital Systems for Humans (DS4H) since September 2022.

11.2.1 Teaching

11.2.2 Supervision

11.2.3 Juries

International journals

• 12 articleK.Konstantin Avrachenkov and M.Maximilien Dreveton. Almost exact recovery in noisy semi-supervised learning.Probability in the Engineering and Informational SciencesNovember 2024, 1 - 22HALDOIback to text
• 13 articleK.Konstantin Avrachenkov, M.Maximilien Dreveton and L.Lasse Leskelä. Recovering Static and Time-Varying Communities Using Persistent Edges.IEEE Transactions on Network Science and Engineering112February 2024, 2087 - 2099HALDOIback to text
• 14 articleY.Younes Ben Mazziane, S.Sara Alouf, G.Giovanni Neglia and D. S.Daniel S. Menasche. TTL model for an LRU-based similarity caching policy.Computer Networks241March 2024, 110206HALDOIback to text
• 15 articleF.Francisco Cabo, A.Alain Jean-Marie and M.Mabel Tidball. Positional and conformist effects in voluntary public good provision.Journal of Public Economic Theory262March 2024HALDOIback to text
• 16 articleA.Antoine Dejonghe, Z.Zwi Altman, F.Francesco De Pellegrini and E.Eitan Altman. Design and cross-layer optimization of low cost RIS-assisted communication systems.IEEE Transactions on Wireless Communications2310May 2024HALDOIback to text
• 17 articleS. H.S. Haleh S. Dizaji, K.Kishor Patil and K.Konstantin Avrachenkov. Influence Maximization in Dynamic Networks Using Reinforcement Learning.SN Computer Science51January 2024, 169HALDOIback to text
• 18 articleF.Francescomaria Faticanti and G.Giovanni Neglia. Optimistic online caching for batched requests.Computer Networks244May 2024, 110341HALDOIback to textback to text
• 19 articleA.Andrea Fox, F. D.Francesco De Pellegrini and E.Eitan Altman. Learning optimal edge processing with offloading and energy harvesting.Computer Communications225September 2024, 324-338HALDOIback to text
• 20 articleD.Diego Goldsztajn, S.Sem Borst and J.Johan van Leeuwaarden. Learning and Balancing Unknown Loads in Large-Scale Systems.Mathematics of Operations ResearchMay 2024HALDOIback to text
• 21 articleD.Diego Goldsztajn, S.Sem Borst and J.Johan van Leeuwaarden. Server Saturation in Skewed Networks.Proceedings of the ACM on Measurement and Analysis of Computing Systems 82May 2024, 1-37HALDOIback to text
• 22 articleD.Diego Goldsztajn. Fluid Limits and Optimal Task Assignment Policies for Locally Pooled Service Systems.ACM SIGMETRICS Performance Evaluation Review523December 2024, 7-10HALDOI
• 23 articleP.Phil Howlett, B. K.Brendan K Beare, M.Massimo Franchi, J.John Boland and K.Konstantin Avrachenkov. The Granger–Johansen representation theorem for integrated time series on Banach space.Journal of Time Series AnalysisAugust 2024HALDOIback to text
• 24 articleA.Alain Jean-Marie and M.M. Tidball. Multiple Exploitation of a Renewable Resource — Optimal Solution of a Deterministic Singular Control Model.International Game Theory Review2602March 2024HALDOIback to text
• 25 articleS. M.Samir M Perlaza, G.Gaetan Bisson, I.Iñaki Esnaola, A.Alain Jean-Marie and S.Stefano Rini. Empirical Risk Minimization with Relative Entropy Regularization.IEEE Transactions on Information Theory7072024, 5122-5161HALDOIback to text
• 26 articleC.Chetansing Rajput and H.Hariprasad Manjunath. Metallic means and Pythagorean triples.Notes on Number Theory and Discrete Mathematics301March 2024, 184-194HALDOI
• 27 articleF.Francisco Robledo Relaño, V.Vivek Borkar, U.Urtzi Ayesta and K.Konstantin Avrachenkov. Tabular and Deep Learning for the Whittle Index.ACM Transactions on Modeling and Performance Evaluation of Computing SystemsJune 2024HALDOIback to textback to text
• 28 articleX.Xiuzhen Ye, I.Iñaki Esnaola, S. M.Samir M Perlaza and R. F.Robert F Harrison. An Information Theoretic Metric for Measurement Vulnerability to Data Integrity Attacks on Smart Grids.IET Smart Grid75October 2024, 583--592HALDOIback to textback to text
• 29 articleX.Xinying Zou, S. M.Samir M Perlaza, I.Iñaki Esnaola, E.Eitan Altman and H. V.H. Vincent Poor. The Worst-Case Data-Generating Probability Measure in Statistical Learning.IEEE Journal on Selected Areas in Information Theory52024, 175-189HALDOIback to textback to textback to text

International peer-reviewed conferences

• 30 inproceedingsK.Khushboo Agarwal and E.Eitan Altman. Games among selfish and team stations in polling systems.VALUETOOLS 2024 - 17th EAI International Conference on Performance Evaluation Methodologies and ToolsMilan, ItalyDecember 2024HALback to textback to textback to text
• 31 inproceedingsE.-M.Emmanouil-Marios Athanasakos, N.Nicholas Kalouptsidis and H.Hariprasad Manjunath. Local Approximation of Secrecy Capacity.2024 32nd European Signal Processing Conference (EUSIPCO)Lyon, FranceAugust 2024, 745-749HALDOIback to text
• 32 inproceedingsE. M.Emmanouil M Athanasakos and S. M.Samir M. Perlaza. Leveraging Noisy Observations in Zero-Sum Games.IEEE Information Theory Workshop (ITW)Shenzhen, ChinaNovember 2024HALback to text
• 33 inproceedingsO.Omar Boufous, R.Rachid El-Azouzi, M.Mikaël Touati, E.Eitan Altman and M.Mustapha Bouhtou. Constrained Correlated Equilibria.60th Annual Allerton Conference on Communication, Control, and ComputingUrbana, United StatesSeptember 2024, 1-8HALDOIback to text
• 34 inproceedingsF.Fabio Busacca, S.Stefano Mangione, G.Giovanni Neglia, I.Ilenia Tinnirello, S.Sergio Palazzo and F.Francesco Restuccia. FedLoRa: IoT Spectrum Sensing Through Fast and Energy-Efficient Federated Learning in LoRa Networks.MASS 2024 - IEEE 21st International Conference on Mobile Ad-Hoc and Smart SystemsSeoul, South KoreaSeptember 2024, 295-303HALDOIback to textback to textback to text
• 35 inproceedingsM.Marina Costantini, G.Giovanni Neglia and T.Thrasyvoulos Spyropoulos. FedDec: Peer-to-peer Aided Federated Learning.SPAWC 2024 - IEEE 25th International Workshop on Signal Processing Advances in Wireless CommunicationsLucca, ItalySeptember 2024, 426-430HALDOIback to textback to textback to text
• 36 inproceedingsF.Francisco Daunas, I.Iñaki Esnaola, S. M.Samir M Perlaza and H. V.H. Vincent Poor. Equivalence of the Empirical Risk Minimization to Regularization on the Family of f-Divergences.Proceedings of the IEEE International Symposium on Information TheoryAthens, GreeceJuly 2024, 759-764HALDOIback to textback to text
• 37 inproceedingsF.Francesco Diana, O.Othmane Marfoq, C.Chuan Xu, G.Giovanni Neglia, F.Frédéric Giroire and E.Eoin Thomas. Attribute Inference Attacks for Federated Regression Tasks.39th Annual AAAI Conference on Artificial Intelligence - AAAI 2025Philadelphia (Pennsylvania), United StatesFebruary 2025HALback to text
• 38 inproceedingsA.Andrea Fox, F.Francesco de Pellegrini, F.Francescomaria Faticanti, E.Eitan Altman and F.Francesco Bronzino. Optimal Flow Admission Control in Edge Computing via Safe Reinforcement Learning.WiOpt 2024 - 22nd International Symposium on Modeling and Optimization in Mobile, Ad hoc, and Wireless NetworksSeoul, South KoreaOctober 2024HALback to textback to text
• 39 inproceedingsF.Franco Galante, G.Giovanni Neglia and E.Emilio Leonardi. Scalable Decentralized Algorithms for Online Personalized Mean Estimation.39th Annual AAAI Conference on Artificial Intelligence - AAAI 2025Philadelphia (Pennsylvania), United StatesFebruary 2025HALback to textback to text
• 40 inproceedingsD.Diego Goldsztajn, S.Sem Borst and J.Johan van Leeuwaarden. Server Saturation in Skewed Networks.SIGMETRICS/PERFORMANCE 2024 - Joint International Conference on Measurement and Modeling of Computer SystemsVenice, ItalyJune 2024, 97-98HALDOIback to text
• 41 inproceedingsL.Louis Hauseux, K.Konstantin Avrachenkov and J.Josiane Zerubia. Benefits of hypergraphs for density-based clustering.EUSIPCO 2024 - 32nd IEEE European Signal Processing ConferenceLyon, FranceAugust 2024, 2302-2306HALDOIback to text
• 42 inproceedingsL.Louis Hauseux, K.Konstantin Avrachenkov and J.Josiane Zerubia. Hypergraphs, percolation, and hierarchical clustering.Complex Networks 2024 - 13th International Conference on Complex Networks and their ApplicationsIstanbul, TurkeyDecember 2024HALback to textback to text
• 43 inproceedings L.Louis Hauseux. How can we theoretically measure the performance of density-based clustering algorithms? ACM SIGMETRICS 2024 Student Research Competition Venice, Italy June 2024 HAL back to text back to text
• 44 inproceedingsC. G.Caelin G Kaplan, C.Chuan Xu, O.Othmane Marfoq, G.Giovanni Neglia and A.Anderson Santana de Oliveira. A Cautionary Tale: On the Role of Reference Data in Empirical Privacy Defenses.PETS 2024 - 24th Privacy Enhancing Technologies Symposium20241Bristol, United KingdomJuly 2024, 525–548HALDOIback to textback to textback to text
• 45 inproceedingsB.Batiste Le Bars, A.Aurélien Bellet, M.Marc Tommasi, K.Kevin Scaman and G.Giovanni Neglia. Improved Stability and Generalization Guarantees of the Decentralized SGD Algorithm.ICML 2024 - The Forty-first International Conference on Machine LearningVienne, AustriaJuly 2024HALback to textback to textback to text
• 46 inproceedingsA. R.Amir Reza Ramtin, J. Z.James Z Hare, L.Lance Kaplan, P.Philippe Nain, V.Venugopal Veeravalli and D.Don Towsley. Quickest Change Detection in the Presence of Covert Adversaries.4th International Workshop on the Internet of Things for Adversarial Environments,Washington DC, United StatesOctober 2024HALback to text
• 47 inproceedingsF.Francisco Robledo, U.Urtzi Ayesta and K.Konstantin Avrachenkov. Deep reinforcement learning for weakly coupled MDP's with continuous actions.ASMTA 2024 - Analytical and Stochastic Modelling Techniques and Applications14826Lecture Notes in Computer ScienceVenise, ItalySeptember 2025, 67-80HALDOIback to text
• 48 inproceedingsA.Angelo Rodio and G.Giovanni Neglia. FedStale: leveraging Stale Updates in Federated Learning.ECAI 2024 - 27TH European Conference on Artificial Intelligence392Santiago De Compostela, SpainOctober 2024, 3071-3078HALDOIback to textback to textback to text
• 49 inproceedingsX.Xinying Zou, S. M.Samir M. Perlaza, I.Iñaki Esnaola and E.Eitan Altman. Generalization Analysis of Machine Learning Algorithms via the Worst-Case Data-Generating Probability Measure.AAAI 2024 - Conference on Artificial Intelligence3815Vancouver, CanadaMarch 2024, 17271-17279HALDOIback to textback to text

Edition (books, proceedings, special issue of a journal)

• 50 periodicalS.Sara Alouf, O.Oliver Hohlfeld and Z.Zhiyuan Jiang, eds. Preface: Special issue on ITC 2023.Performance Evaluation167March 2025, 102462HALDOIback to text

Doctoral dissertations and habilitation theses

• 51 thesisY.Younes Ben Mazziane. Probabilistic analysis for caching.Université Côte d'AzurMay 2024HALback to textback to text
• 52 thesisO.Olga Chuchuk. Data access optimisation at CERN and in the Worldwide LHC Computing Grid (WLCG).Université Côte d'AzurFebruary 2024HALback to text
• 53 thesisC.Caelin Kaplan. Inherent trade-offs in privacy-preserving machine learning.Université Côte d'AzurNovember 2024HALback to textback to text
• 54 thesisA.Angelo Rodio. Client heterogeneity in federated learning systems.Université Côte d'AzurJuly 2024HALback to textback to text

Reports & preprints

• 55 reportE.-M.Emmanouil-Marios Athanasakos and S. M.Samir M. Perlaza. Leveraging Noisy Observations in Zero-Sum Games.RR-9538Inria Sophia Antipolis - Méditerranée, Université Côte d'AzurJanuary 2024, 24HALback to text
• 56 miscD.Damiano Carra and G.Giovanni Neglia. An Online Gradient-Based Caching Policy with Logarithmic Complexity and Regret Guarantees.2024HALback to text
• 57 miscF.Francisco Daunas, I.Iñaki Esnaola, S. M.Samir M. Perlaza and H. V.H. Vincent Poor. Asymmetry of the Relative Entropy in the Regularization of Empirical Risk Minimization.October 2024HALback to textback to text
• 58 miscF.Francesco Diana, O.Othmane Marfoq, C.Chuan Xu, G.Giovanni Neglia, F.Frédéric Giroire and E.Eoin Thomas. Attribute Inference Attacks for Federated Regression Tasks.2024HALDOIback to textback to text
• 59 miscF.Franco Galante, G.Giovanni Neglia and E.Emilio Leonardi. Scalable Decentralized Algorithms for Online Personalized Mean Estimation.2024HAL
• 60 reportA.Alain Jean-Marie and M.Mabel Tidball. Equilibrium bids for reverse auctions when the budget is announced - Some preliminary results.RR-9545Inria; CEE-M. Centre d'Economie de l'Environnement - MontpellierMarch 2024, 19HALback to text
• 61 miscC.Caelin Kaplan, A.Angelo Rodio, T. S.Tareq Si Salem, C.Chuan Xu and G.Giovanni Neglia. Federated Learning for Collaborative Inference Systems: The Case of Early Exit Networks.2024HALback to textback to textback to text
• 62 miscS. M.Samir M. Perlaza and X.Xinying Zou. The Generalization Error of Machine Learning Algorithms.November 2024HALback to textback to textback to text
• 63 miscC.Charlotte Rodriguez, L.Laura Degioanni, L.Laetitia Kameni, R.Richard Vidal and G.Giovanni Neglia. Evaluating the Energy Consumption of Machine Learning: Systematic Literature Review and Experiments.2024HALback to textback to textback to text
• 64 reportX.Xinying Zou, S. M.Samir M. Perlaza, I.Iñaki Esnaola, E.Eitan Altman and H. V.H. Vincent Poor. An Exact Characterization of the Generalization Error of Machine Learning Algorithms.RR-9539INRIAMarch 2024, 80HALback to textback to textback to text

Educational activities

• 65 unpublishedA.Alain Jean-Marie and E.Emmanuel Hyon. An introduction to pyMarmote and pyMarmoteMDP for Markovian modeling - A tutorial.June 2024, DoctoralItalyHALback to textback to text