This paper revisits combinatorial Thompson sampling (CTS) in the semi-bandit setting with sleeping arms, where only a subset of arms is available each round under combinatorial constraints. Such settings frequently arise in networking applications and pose significant challenges due to the need for combinatorial optimization over a dynamically changing action space. A canonical example is wireless mesh routing, where fluctuating link availability and routing constraints lead to a time-varying set of feasible paths between source and destination nodes.The existing works have three key limitations: (1) the lack of worst-case regret guarantees for CTS in semi-bandit settings, even without accounting for sleeping arms; (2) the absence of theoretical guarantees for CTS under adversarially varying arm availability; and (3) the subpar empirical performance of CTS with Gaussian priors (CTS-G).To address these issues, we propose CL-SG, a simple yet effective CTS-G variant that samples a single shared Gaussian seed each round to coordinate exploration across arms. Theoretically, CL-SG achieves a worst-case regret bound of $\tilde O(\sqrt {mNT} )$ and a matching lower bound of $\Omega (\sqrt {mNT} )$. Our analysis can also easily recover a regret bound for CTS-G. With experiments driven by real-world datasets, CL-SG consistently outperforms diverse benchmark algorithms such as CTS-G and CTS-B. We open-source our implementation and experiments to support reproducibility and further research.

Distributed Mixture-of-Experts (MoE) model training requires all-to-all (A2A) communication for dynamic token routing. While it is empirically observed that A2A becomes a bottleneck, no quantitative framework exists to predict when this bottleneck will occur. This paper presents an analytical model of computation and communication time in distributed MoE training that accounts for load imbalance, hierarchical network topology, and hybrid parallelism. We derive a load-imbalance factor as a function of the expert parallelism degree and the number of local experts, and introduce a critical expert parallelism threshold that predicts the transition from compute-bound to communication-bound. Through validation on a real-world GPU cluster, the model achieves prediction errors below 5% for the compute-to-communication ratio. We also discuss how the proposed model guides practitioners in configuring parameters in practice.

Low earth orbit (LEO) satellite networks, while essential for data services in mobile and remote scenarios in the forthcoming sixth-generation (6 G), suffer from intrinsic dynamics like frequent handovers and channel variability that severely challenge transport-layer reliability. To overcome these challenges, this paper proposes a packet transmission scheme based on adaptive forward erasure correction (FEC) that integrates streaming codes with deep reinforcement learning (DRL). We employ a deep recurrent Q-network (DRQN) to dynamically orchestrate the transmission of repair packets based on real-time channel state information, thereby achieving an optimal balance between bandwidth efficiency and transmission reliability at the transport layer. The proposed system is validated using real-life Starlink traces via a trace-driven emulation platform built upon Linux's eBPF technology. Experimental results demonstrate improved latency-overhead trade-offs over the conventional fixedinterval FEC, providing adaptive transport-layer error control for the next-generation satellite networks.

Low Earth Orbit (LEO) satellite networks play an important role in today’s global communication system, which offer extensive coverage and fast data transmission to support real-time communications on a global scale. However, the stable operation of such satellites requires an adequate supply of energy, which poses a real challenge. When a satellite is in sunlight, the solar energy meets the satellite’s operation needs. On the other hand, when the satellite is during a solar eclipse, it can only rely on its battery. Although there is sufficient solar energy in space, irrational use of the batteries will rapidly reduce the limited cycle life of the satellites. In this paper, we propose an energy-efficient routing scheme for LEO satellite networks based on artificial potential fields and reinforcement learning (RL). First, we design a satellite sleep strategy to set certain satellites to sleep mode to reduce unnecessary energy consumption, considering the uneven distribution of traffic on the Earth. Next, we design an Artificial Potential Field-guided Reinforcement Learning (APFRL) routing algorithm, which utilizes potential fields as the guiding information for selecting the next hop and reinforcement learning to find the solutions under such guidance. Experimental results demonstrate that our scheme significantly reduces the energy and cycle life consumption of the satellites, while outperforming the state-of-the-art schemes in terms of data delivery ratio.

In the last few years, considerable research efforts have focused on measuring and improving Starlink network performance, especially for user terminals (UTs) in stationary scenarios. However, the performance of Starlink networks in mobility settings, particularly with frequent changes in the UT's orientation, and the impact of environmental factors, such as transient obstructions, has not been thoroughly studied, leaving gaps in understanding the causes of performance degradation. Recently, researchers have started identifying the communicating satellites to evaluate satellite selection strategies and the impact on network performance. However, existing Starlink satellite identification methods only work in stationary, obstruction-free scenarios, as they do not account for UT mobility, obstructions or detect dynamic beam switching events. In this paper, we reveal that the UT can perform multiple dynamic beam switching attempts to connect to different satellites when the UT-satellite link is degraded. This degradation can occur either due to the loss of line-of-sight (LoS) from changes in the FOV or obstructions, or due to poor signal quality, extending UT-satellite handovers beyond the well-known 15-second regular handover interval. We propose a mobility-aware Starlink satellite identification method that detects dynamic beam switching events, and plausibly explain network performance using UT's diagnostic data and connected satellite information. Our findings demystifies the mobile Starlink network performance degradations, which is crucial to enhance the end-to-end performance of transport layer protocols and in diverse application scenarios.

Low Earth Orbit (LEO) satellite networks are transforming global connectivity by enabling high-speed, low-latency Internet access. Particularly, they significantly facilitate areas where terrestrial networks are not deployed or destroyed. Meanwhile, LEO satellite technology is experiencing an unprecedented surge in development. This paper provides a comprehensive and up-to-date overview of LEO satellite networks. First, the evolution of LEO satellites is introduced, followed by an exploration of the components and communication architecture within LEO satellite networks using representative examples. Second, key technologies, including routing, handover management, and digital twins, are summarized, and some practical application scenarios are discussed. The performance of LEO satellite networks, illustrated by SpaceX’s Starlink, is then evaluated to understand its scheduling algorithm and network characteristics, which can inform future satellite-related algorithms and architecture design. Finally, as LEO constellations continue to expand, practical operations face significant challenges in management, technology, and security. Consequently, we highlight some open research issues to provide potential inspiration for academia and industry in satellite networking.

Low earth orbit (LEO) satellite networks are pivotal for sixth-generation (6G) wireless systems, yet their high-speed mobility induces frequent packet loss, causing severe head-of-line blocking delays under traditional retransmission mechanisms. While streaming forward erasure correction (FEC) can mitigate retransmissions, existing packet loss models fail to capture the unique dynamics of LEO networks, causing difficulties in the design and analysis of FEC schemes. This paper addresses this problem through the following contributions. First, based on real-world Starlink measurements, we reveal the inadequacy of conventional loss models such as those based on Markov chains. Second, we propose a Markovian arrival process (MAP) to model LEO packet loss. Using an expectation-maximization (EM) algorithm to fit Starlink traces, we demonstrate its superior accuracy over existing models. Third, based on MAP modeling, we show that the decoding delay of a typical streaming FEC scheme with fixed repair insertion intervals can be analyzed by approximating it as the busy period of a MAP/D/1 queue. Using matrix-analytic methods, we provide a numerical recipe to compute this delay. Simulations validate the precision of the model in predicting delay, offering practical guidelines for FEC design in LEO networks.

Wormhole attacks and related topology-fabrication attacks distort internet of things (IoT) routing by creating fake one-hop adjacencies that bypass path-selection logic and evade common defenses. These fabricated links leave an observable mismatch between the network’s logical topology and each node’s physical channel footprint (i.e., transmit, receive, and idle states). In this paper, we propose the number of channel footprint inconsistencies (NCFI), a lightweight metric that quantifies these mismatches to identify nodes impacted by wormholes, one of the most challenging attacks. NCFI does not require specialized hardware or tight clock synchronization, making it suitable for resource-constrained IoT devices. Our approach enables nodes to report channel-state traces to a central verifier, which models consistency checking as a constraint satisfaction problem (CSP) solved via graph coloring. Additionally, NCFI can be computed in a distributed manner by node pairs to validate links locally. Although this work primarily targets wormhole detection, the proposed methodology is readily extensible to other state-manipulation threats, including replay, man-in-the-middle, and route falsification. Beyond detection, we introduce wormhole endpoint weighted centroid localization (WE-WCL) to estimate attacker locations. Simulations demonstrate that NCFI improves detection accuracy by at least 30% in broadcast and 16% in unicast scenarios compared with state-of-the-art baselines, while WE-WCL achieves 1.15× to 4× lower localization error.

We revisit combinatorial Thompson sampling (CTS) for semi-bandits with sleeping arms, where arm availability varies over time and actions must satisfy combinatorial constraints, as in wireless mesh routing with fluctuating link availability. Despite its practical relevance, CTS has been hindered by several long-standing problems: (i) the absence of worst-case regret guarantees in the semi-bandit setting even without sleeping arms, (ii) the lack of theory under adversarially varying availability, and (iii) the consistently weak empirical performance of CTS with Gaussian priors (CTS-G). This paper resolves these long-standing issues by providing the first worst-case regret analysis of CTS-G, proving an upper bound of Õ(m√(NT)) and a matching lower bound of (m√(NT)). To bridge the gap between theory and practice, we further propose CL-SG, a simple CTS-G variant that samples a single shared Gaussian seed each round to coordinate exploration across arms. We show that CL-SG achieves an improved regret bound of Õ(√(mNT)), together with a matching lower bound Ω(√(mNT)). Experiments on real-world datasets demonstrate that CL-SG consistently outperforms strong baselines including CTS-G and CTS-B, and we open-source our implementation for reproducibility.

Inter-satellite links enable low Earth orbit satellite networks (LSNs) to communicate and relay messages over vast distances, including remote regions lacking reliable terrestrial infrastructure. Leveraging this capability, constellations such as Starlink and Kuiper aim to deliver global internet coverage. However, the reliability of message propagation, i.e., the probability that a message successfully traverses the network, remains poorly understood. In this paper, we develop a topological model for LSNs and apply statistical analyses to evaluate network reliability. We derive expressions for the probability of successful message propagation between arbitrary source and destination satellites, both within a single orbital plane and across two orbital planes. The expressions are validated through simulations, demonstrating high accuracy. These analytical results provide valuable insights for determining the optimal number of satellites in LSN design.

Environmental awareness is critical for maintaining robust connectivity in low Earth orbit (LEO) satellite networks. Starlink terminals construct a 123×123 obstruction map by passively recording signal quality from connected satellites. However, this process is slow: even under continuous operation, meaningful map coverage typically requires 6 to 12 hours, impractically slow for dynamic environments or short-term deployments. Building on Starlink’s existing capability, we investigate how intelligent satellite selection can further accelerate obstruction map construction, without requiring architectural changes. We formulate the problem as a multi-armed bandit with sleeping arms and introduce SkyBandit, a lightweight online learning algorithm that balances link quality with exploration of uncharted sky regions. SkyBandit achieves a regret bound of $O\left( {\sqrt {NT{\text{ln}}T} } \right)$, where N is the number of observed orbits over T rounds. Experiments using a refined version of xeoverse LEO emulator along with real Starlink measurement data demonstrate that SkyBandit reduces obstruction map convergence time to just around 3 hours, while preserving high-quality connections. To our knowledge, this is the first work to formalize and optimize obstruction map construction in LEO networks via principled sequential learning.

Low Earth Orbit (LEO) satellite ISPs promise universal Internet connectivity, yet their interaction with content delivery remains poorly understood. We present the first comprehensive measurement study decomposing Starlink's web content delivery performance decomposed across Point of Presence (PoP), DNS, and CDN layers. Through two years of measurements combining 225K Cloudflare AIM tests, M-Lab data, and active probing from 99 RIPE Atlas and controlled Starlink probes, we collect 6.1M traceroutes and 10.8M DNS queries to quantify how satellite architecture disrupts terrestrial CDN assumptions. We identify three distinct performance regimes based on infrastructure density. Regions with local content-rich PoPs achieve near-terrestrial latencies with the satellite segment dominating 80-90

Topological attacks, including Sybil and wormhole attacks, modify a network topology by adding fake nodes and links to bypass typical routing paths, significantly degrading the Internet of Things (IoT) network performance. The presence of fake nodes and links introduces inconsistencies between the modified topology and the actual channel states of the nodes. Therefore, a fast and precise method for detecting topological attacks is crucial to identify and count these inconsistencies, which reveal the presence of an attack and affected nodes. In this paper, we propose a straightforward numerical metric called the Number of Channel Footprint Inconsistencies (NCFI) to count inconsistencies between nodes' transmission, reception, and idle states on the channel and the network topology. In the proposed method, each node reports its channel states to a central node, which uses a Constraint Satisfaction Problem (CSP)-based approach in graph coloring to calculate the NCFI. Additionally, each pair of IoT nodes can use the NCFI in a distributed manner to detect the presence of fake links between them. Our simulation results indicate that the proposed method can accurately detect the presence of attackers in the network compared to the state-of-the-art methods.

In this paper, we study unknown-game bandits, where multiple agents play a general-sum game repeated over T rounds. In each round, each agent independently selects an action and observes the reward for that action. The game is unknown to every agent, meaning each agent has no knowledge about the underlying game structure, the number of other agents, or their actions and rewards. Such unknown-game bandits have wide applications in computer and communication networks, including congestion control and network selection. The goal of each agent is to minimize swap regret, which measures the performance gap from a broader class of competitors than the traditional external regret that only compares against competitors always playing a fixed action. Our main contribution is to bridge the gap in the literature by proving the first swap-regret bound with a time-dependence of (O) over tilde (T-1/4) if the proposed learning algorithm based on optimistic follow-the-regularized-leader (OFTRL) is played by all agents involved in the game, where (O) over tilde(center dot) hides log-arithmic factors. This regret bound demonstrates a faster convergence rate with respect to the number of rounds T compared to the state-of-the-art swap regret bound of O(T-1/2). Furthermore, we demonstrate the efficacy of the proposed algorithm through an application in heterogeneous network selection with both numerical and simulation-based experiments.

Video streaming performance may degrade substantially in a mobile environment due to fast-changing wireless links. On the other hand, to provide ubiquitous services, heterogeneous static and mobile access and backbone networks will be integrated in the sixth-generation (6G) systems, so mobile users can take advantage of multiple access options for better services. Multi-path transport-layer protocols like Multi-Path QUIC (MPQUIC) show promise in utilizing multiple access links to address the impact of mobility. However, the optimal link selection that aims to provide statistical QoS guarantee for video streaming in a mobile environment with both user mobility and network mobility remains an open issue. In this paper, based on a lightweight Multi-Armed Bandit (MAB) technique, we develop a Q oS-driven C ontextual MAB (QC-MAB) framework for MPQUIC, which makes an intelligent access network selection and adaptively enables FEC coding to trade off delay, reliability and goodput. Extensive simulation results with ns-3 show that the proposed QC-MAB framework can outperform the state-of-the-art solutions. It achieves up to ten times lower video interruption ratio and three times higher goodput in highly dynamic mobile environments.

Two-path transmission with the involvement of LEO satellites is an increasingly common scenario. LEO networks may offer higher bandwidth compared with some terrestrial networks, such as cellular, but often come with increased variability in latency. Effective traffic allocation to maximize bandwidth utilization in such dynamic environments is challenging. This paper addresses two-path scheduling problem under a high dynamic environment by proposing the FTRL-WRR algorithm, which combines a lightweight, learning-based optimization algorithm with a weighted round-robin scheduler. We model traffic allocation as a 1-dimensional optimization problem and demonstrate the algorithm's effectiveness through evaluations in emulated Starlink-cellular scenarios. Results show that FTRL-WRR improves bandwidth utilization and reduces median completion time by up to 27.71%.

Integrating terrestrial networks together with non-terrestrial networks, such as Low Earth Orbit (LEO) satellites, can significantly enhance network reliability and reduce latency in unicast and multicast protocols. To use such a capability and reduce end-to-end latency, many transport-layer unicast protocols, such as Multipath TCP (MPTCP) and Multipath QUIC (MPQUIC), have adopted multi-connectivity (MC) by transmitting over multiple interfaces at the sender host. However, traditional and modern multicast protocols suppose that the sender host accesses the core network using a single path, leading to high end-to-end latency at destinations. In this regard, we first demonstrate in a real-world testbed that MC significantly improves network latency in multicasting compared to single-connectivity approaches. We then introduce a Bit Indexed Explicit Replication (BIER) Multi-Connectivity (BIER-MC) method designed to reduce end-to-end latency within the BIER protocol, as a modern multicast protocol. Our comparison indicates that BIER-MC outperforms traditional BIER implementations using multicast trees in terms of latency up to 5×, bandwidth usage up to 2.5×, and edge betweenness centrality up to 15×.

This paper studies a novel multi-agent combinatorial bandit problem called moving semi-bandits involving K agents and N arms, extending the problem of semi-bandits with adversarial rewards and stochastic arm availabilities (sleeping semi-bandits). The arms move across agents, making each arm available to at most one agent at a time, and the set of available arms for each agent changes over time. In each round, each agent plays up to m arms from their own available arm set simultaneously and observes the random loss for each played arm (i.e., semi-bandit feedback). The loss of each arm has no stochastic assumptions, and different agents may generate different random losses for each arm. The primary goal is to minimize the cumulative loss for all agents through collaboration. This bandit problem is motivated by real-world applications, such as traffic scheduling in wireless networks with multiple access points and task assignment for multiple crowdsourcing platforms. To address this challenge, we propose an efficient framework called Moving-FTPL, which guarantees a regret bound of O(N root NTK ln T) over T rounds. Moving-FTPL can reduce the total regret of all K agents by a factor of root K compared to scenarios where agents do not collaborate. Additionally, Moving-FTPL takes a step forward for the long-standing problems of a tighter regret bound for sleeping semi-bandits by significantly improving the state-of-the-art regret bound by a factor of m root N and improving the bound for sleeping adversarial bandits by a factor of root N. Furthermore, we showcase a crowdsourcing application to demonstrate the effectiveness of our proposed algorithm when compared with others.