The accurate prediction of Quality of Experience (QoE) in video streaming services is essential for optimizing user satisfaction and network performance. While traditional Quality of Service (QoS) metrics provide objective measurements of network behavior, they often fail to reflect the subjective nature of user experience. This paper investigates the use of Machine Learning models to estimate QoE based on QoS indicators. Building upon the recently published SNESet dataset, we evaluate a range of modern regression techniques, including randomization-based neural networks, symbolic regression and Kolmogorov-Arnold Networks, alongside other traditional and ensemble-based models. A central focus of this study is the explainability of such new models, which enables the extraction of domain-relevant insights from the learned relationships. Using model-agnostic techniques for explainable Artificial Intelligence and uncertainty quantification, we assess the confidence of such models in their predictions and analyze the contribution of individual features to the estimated QoE. Our results underscore the need for explainable QoE prediction systems, closing the gap between data-driven modeling and domain expertise.
A Matching model describes the waiting times suffered by items before they match with other items and disappear immediately without service. The matching relation is described by a compatibility graph. It is an easy representation of multiple types synchronizations between items. When we add an edge in the compatibility graph, one expects that the expected total number of items decreases. Unexpectedly this is not always true and there is a performance paradox. Extending previous results, we show new matching models with performance paradoxes which seems to be more frequent than one can expect.
In cybersecurity, moving-target defense (MTD) mitigates reconnaissance attacks from botnets by periodically resetting server identities, disrupting adversarial intelligence gathering. We model this defense mechanism as a dynamic coupon collection process, where servers correspond to coupon types and botnet attacks to coupon acquisitions. The defender’s intervention (removing collected coupons) represents server resets, restoring their uncompromised state. In this paper, we derive various results regarding the number of compromised servers under various stochastic attack-defense dynamics.
Modern cyberattacks increasingly involve coordinated teams of adversaries, posing new challenges for detection and defense. This paper introduces a probabilistic framework for modeling such multi-agent attacks, combining Markov Decision Processes with an augmented Markov game that captures dynamic team composition under detection events. The model integrates state-dependent detection probabilities, implicit coordination via reward structures, and efficient value iteration for policy computation. Through a stochastic grid-world case study, we analyze the performance trade-offs between isolated and coordinated strategies. Our results reveal that while isolated agents perform better in simple, single-target scenarios, coordinated strategies significantly improve success rates in complex, multi-target operations by systematically managing exposure. This work provides a formal foundation and practical methodology for analyzing emergent attack behaviors in adversarial environments.
This paper addresses decision-making problems where the objective is to maximize the best outcome along a trajectory in a Markov Decision Process (MDP), rather than its cumulative reward. Such extremal objectives naturally arise in risk-sensitive applications, including cybersecurity and resilience planning, but fall outside the scope of classical MDP theory due to their non-Markovian nature. We propose a principled transformation that augments the MDP state space with a deterministic memory variable tracking the maximal reward, yielding an equivalent total-reward MDP solvable by dynamic programming. This construction enables the extraction of Markovian policies through a provision function tailored to each initial state. We evaluate our framework on a malware containment problem, showing that the memory-based baseline policy significantly outperforms a greedy myopic strategy across all configurations. Our results demonstrate improved worst-case cost and containment time, highlighting the utility of memory-augmented planning for extremal performance control in stochastic systems.
This paper addresses the problem of coordinating a large population of heterogeneous electrical loads, such as electric vehicles (EVs) and water heaters (WHs), under global operational constraints. We extend the Moment Constrained Optimal Transport for Control (MCOT-C) framework to accommodate multiple classes of agents with distinct dynamics and cost structures. Our formulation relies on a mean-field limit that captures agent heterogeneity through class-specific distributions. We propose a scalable gradient descent algorithm and a Model Predictive Control (MPC) scheme that enables online adaptation of this algorithm to uncertain or progressively revealed agent information. The proposed approach is validated through numerical experiments on real datasets [8, 9] for EVs and WHs, demonstrating the effectiveness of this method in enforcing global constraints while preserving agent-level dynamics.
Research suggests that large-scale human cooperation is driven by shared narratives that encode common beliefs and values. This study explores whether such narratives can similarly nudge LLM agents toward collaboration. Therefore, we let LLM agents play a (networked) finitely repeated public goods game after being primed with different stories. Our experiments address four questions: (1) How do narratives influence negotiation behavior? (2) What differs when agents share the same story versus different ones? (3) What happens when the agent numbers grow? (4) Are agents resilient against self-serving participants? We find that story-based priming significantly affects collaboration. Common stories improve collaboration and benefit all participants, while different story priming reverses this effect, favoring self-interested agents. These patterns persist across network sizes and structures. These findings have implications for multi-agent coordination and AI alignment. Code is available at github.com/storyagents25/story-agents.
The Kelly mechanism is a proportional allocation auction widely adopted in decentralized resource allocation systems to share an infinitely divisible resource among competing agents. We analyze the sequential game it induces when agents have α -fair utilities and behave strategically. Our main result proves that synchronous best-response updates drive bids to the unique Nash equilibrium at a linear rate for α∈{0,1,2} . Extensive simulations reveal that best-response dynamics reach equilibrium significantly faster than previously proposed no-regret learning algorithms.
In many games, the payoff of individual users is a function of a collective outcome that is common to all agents. Often, the users may be interested in jointly steering this outcome to a desired value. This work presents such a scenario, where the collective outcome is strictly monotone in the joint strategy of the agents. Further, the agents are irrational in their perception of a random cost component in their payoff. This irrationality is modelled using prospect theory. A coordinator steers the game to a desired collective outcome, by designing incentives. These incentives modify the responses of the users. Owing to the potential structure of the game, the system converges to a Nash equilibrium at which the desired collective outcome is obtained.
The significant computational demands of modern machine learning (ML) models raise growing concerns about their environmental impact and economic cost during training. To address this, data summarization techniques, which involve selecting a smaller, representative subset of the training data, offer a promising avenue for optimizing energy consumption without compromising model performance. This article investigates the influence of various data summarization algorithms—specifically random sampling, Facility Location (FL), and CRAIG—on three critical aspects of supervised learning (SL) training: model accuracy, energy consumption, and overall efficiency, which we quantify as the ratio of accuracy to energy consumption. Our extensive experimental findings reveal that while sophisticated methods like FL and CRAIG aim to select highly representative subsets, random sampling consistently achieved a robust balance of accuracy and efficiency. This is particularly evident when accounting for the often significant pre-processing energy overheads incurred by more complex selection strategies. Furthermore, we emphasize the pivotal role of early stopping criteria in optimizing the overall energy efficiency of the training process. Our analysis demonstrates that strategic adjustments to these criteria can substantially reduce the number of training epochs required for convergence, thereby mitigating considerable energy waste.
The wireless communication between a transmitter and a group of nodes, such as a control center interacting with several drones engaged in a specific mission, may be vulnerable to malicious eavesdropping attacks. Cons of traditional communication protocol aimed to maximize total secrecy rate is that might lead to neglecting communication with some nodes, and, to failure of nodes’ mission. To deal with such cons we suggest a maximizing minimal secrecy rate protocol, or, in other words, a guaranteed (expected) secrecy rate protocol. The problem is modeled and solved in a (Bayesian) game-theoretical framework between a transmitter and an adversary. We prove that such protocol supports uninterrupted communication and maintains its stability even if the transmitter has incomplete information on what type of antenna (directional or omni-directional) the adversary has. Finally, an approach to solve an adversary dilemma of which of the antennas to engage is suggested and illustrated.
Over the years, cybercriminals have refined their tools by developing increasingly stealthy malware. One effective approach involves temporarily putting malware into a dormant state to evade detection, then reactivating it at the right moment to relaunch the infection. Notable examples, such as Emotet and Marap, well illustrate this strategy: the former is capable of remaining inactive for several weeks before launching new attacks, while the latter passively gathers minimal system data while awaiting further instructions. These behaviors suggest the presence of two distinct categories of infected nodes: active and dormant. Motivated by these practical illustrations, we propose an SI ^2 R model in which an infected node can transition into a dormant, undetectable state and be reactivated at a speed controlled by the attacker. The goal is to identify optimal transition strategies (both for dormancy and reactivation) to optimize the peak number of dormant nodes reached during the process, prepared to strike at the most effective moment. This model provides a mathematical framework for analyzing delayed cyber threats and the attacker’s timing strategies while also informing proactive defense mechanisms against such latent threats (DISTRIBUTION A. Approved for public release: distribution unlimited).
Many queueing-based service system, such as call centers, telecommunication networks, and ride booking platforms, frequently face user abandonment due to limited service resources. Admission control thus becomes an important strategy for service providers to reject users likely to abandon after entering the system. These policies aim to minimize the long-run average system cost by balancing the costs of rejection, abandonment, and server idling. The quality of service (QoS) experienced by the user-base directly depends on the admission control policy of the service provider. In turn, users respond to QoS experienced by offering an abandonment rate at stationarity. Since the admission control policy also depends on this abandonment rate, a strategic interaction emerges between the service provider and its user-base, which can be modeled as a one-period, two-player non-cooperative game. This paper investigates the existence of pure-strategy Nash equilibrium (PSNE) in such interactions. Under a mild condition, we prove the monotonicity of PSNE with respect to key system parameters, such as arrival and service rates. In scenarios where a PSNE does not exist, we explore an equilibrium set, which captures cyclical fluctuations of the system between low and high abandonment rate regimes.
The growing demand for wireless services necessitates efficient spectrum use, especially in 5G, NextG, and Internet of Things (IoT) network scenarios. Cognitive Radio (CR) systems can be employed to detect underutilized bands, allowing Secondary Users (SUs) to opportunistically access these bands, thus improving spectrum efficiency. Cooperative Spectrum Sensing (CSS) is essential for reliable detection of Primary Users (PUs), but full SU participation is energy-intensive. This article presents an energy-efficient sensing optimization algorithm that dynamically selects an optimal SU subset by iteratively assessing marginal gains in detection probability. SUs are ranked by Signal-to-Noise Ratio (SNR), and an iterative loop with a convergence threshold determines the final sensing set. Simulations show the algorithm significantly reduces the number of active SUs, still obtaining excellent detection performance.
To stimulate innovation and competition in the digital sector, the European Commission has adopted the Digital Markets Act (DMA), that aims at limiting the gatekeeping power of the biggest actors. For instance, such prevailing platforms may be forced to share with their competitors some, or all, of the data they collected on users that can help improve service quality. This paper proposes a model to analyze the impacts of such a regulation, by comparing (i) a laisser-faire situation where prevailing platforms are free to sell data to newcomer service providers, (ii) the case when platforms can decide to share data but charging newcomers is prohibited, and (iii) a scenario where data sharing is imposed and enforced by the regulator. We investigate the impacts of those scenarios on key metrics like User Welfare, provider revenues, and user demand repartition.
Online social networks exert a powerful influence on public opinion. Adversaries weaponize these networks to manipulate discourse, underscoring the need for more resilient social networksquery. To this end, we investigate the impact of network connectivity on Stackelberg equilibria in a two-player game to shape public opinion. We model opinion evolution as a repeated competitive influence-propagation process. Players iteratively inject messages that diffuse until reaching a steady state, modeling the dispersion of two competing messages. Opinions then update according to the discounted sum of exposure to the messages. This bi-level model captures viral-media correlation effects omitted by standard opinion-dynamics models. To solve the resulting high-dimensional game, we propose a scalable, iterative algorithm based on linear-quadratic regulators that approximates local feedback Stackelberg strategies for players with limited cognition. We analyze how the network topology shapes equilibrium outcomes through experiments on synthetic networks and real Facebook data. Our results identify structural characteristics that improve a network’s resilience to adversarial influence, guiding the design of more resilient social networks.
This paper investigates the coupling between a peer-to-peer (P2P) electricity market and a forecast market to alleviate the uncertainty faced by prosumers regarding their renewable energy sources (RES) generation. The work generalizes the analysis from Gaussian-distributed RES production to arbitrary distributions. The P2P trading is modeled as a generalized Nash equilibrium problem, where prosumers trade energy in a decentralized manner. Each agent has the option to purchase a forecast on the forecast market before trading on the electricity market. We establish conditions on arbitrary probability density functions (pdfs) under which the prosumers have incentives to purchase forecasts on the forecast market. Connected with the previous results, this allows us to prove the economic efficiency of the P2P electricity market, i.e., that a social optimum can be reached among the prosumers.
We study a model of auction representative of the 5G auction in France. We determine the optimal strategy of a bidder, assuming that the valuations of competitors are unknown to this bidder and that competitors adopt the straightforward bidding strategy. Our model is based on a Partially Observable Markov Decision Process (POMDP). This POMDP admits a concise statistics, avoiding the solution of a dynamic programming equation in the space of beliefs. In addition, under this optimal strategy, the expected gain of the bidder does not decrease if competitors deviate from straightforward bidding. We illustrate our results by numerical experiments, comparing the value of the bidder with the value of a perfectly informed one.
This paper provides an analysis of the impact of attacks on the Age of Information (AoI) in queuing systems. We consider single and tandem queue models, with and without preemption in service. We have shown that when preemption in service is allowed, attacks do not result in AoI reduction. For non preemptive servers, we show that attacks can reduce the AoI for both tandem and single server cases. For the single server case, we establish necessary and sufficient conditions for the existence of arrival and attacks rates that minimize the AoI and provide an explicit expression of the optimal attack rate when the arrival rate of updates is very large. Numerical results corroborate the analytical claims and show the accuracy of the obtained results when the arrival rate of update is low.
We analyze a system in which in each time slot one customer and one server arrive at the system according to a random process. Compatibilities between customers and servers are determined by a bipartite graph. An incoming customer (resp. server), if it finds a compatible server (resp. customer), they are matched and both leave the system. Otherwise, they are stored in a queue. We investigate the impact on the expected value of the unmatched customers and servers when we remove an edge from the compatibility graph. For a quasicomplete graph and a large family of matching policies, we provide necessary and sufficient conditions on the probability distribution of the arrivals such that a performance paradox occurs, i.e., such that removing an edge of the compatibility graph improves the performance of the system. This phenomenon can be seen as an analog of the Braess paradox in bipartite matching models.
