XSXinghua Sun

Xinghua Sun

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Xinghua Sun has a total of 115 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Offloading Optimization in Intent-Driven Edge Computing: An LLM-Based Distributed Framework

    by Mingqi Han, Xinghua Sun, Chenyuan Feng, Wen Zhan, Xijun Wang, Tony Q. S. Quek

    Meeting the demand for personalized, low-latency services in 5G networks requires the convergence of intent-driven edge computing. However, existing frameworks are limited: centralized approaches incur high overhead and privacy risks, while distributed solutions overlook the flexibility of natural-language intents and their impacts on offloading optimization. To solve this, we propose a distributed LLM-based joint intent comprehension and offloading optimization framework, in which each edge server utilizes an LLM agent to locally interpret natural-language intents and optimize offloading decisions in a distributed manner. To enable distributed coordination, we introduce a self-reasoning prompt to guide the LLM in predicting neighbors’ decisions, thus capturing correlations among UEs. Moreover, we propose the Weighted Reward-aligned DPO (WR-DPO) algorithm to enhance decision capability. High-quality samples are utilized to mitigate reward noise and incorporate reward-gap weighting and regularization for stable convergence. Simulation results demonstrate that WR-DPO significantly outperforms other baselines with a 30% average performance gain, and demonstrates strong generalization capability in inexperienced scenarios.

  • Minimizing Task-Oriented Age of Information for Remote Monitoring With Pre-Identification

    by Shuying Gan, Chenyuan Feng, Chao Xu, Xinghua Sun, Xiang Chen, Xijun Wang

    The emergence of new intelligent applications has fostered the development of a task-oriented communication paradigm, where a comprehensive, universal, and practical metric is crucial for unleashing the potential of this paradigm. To this end, we introduce an innovative metric, the Task-oriented Age of Information (TAoI), to measure whether the content of information is relevant to the system task, thereby assisting the system in efficiently completing designated tasks. We apply TAoI to a wireless monitoring system tasked with identifying targets and transmitting their images for subsequent analysis. To minimize TAoI and determine the optimal transmission policy, we formulate the dynamic transmission problem as a Semi-Markov Decision Process (SMDP) and transform it into an equivalent Markov Decision Process (MDP). Our analysis demonstrates that the optimal policy is threshold-based with respect to TAoI. Building on this, we propose a low-complexity relative value iteration algorithm tailored to this threshold structure to derive the optimal transmission policy. Additionally, we introduce a simpler single-threshold policy, which, despite a slight performance degradation, offers faster convergence. Comprehensive experiments and simulations validate the superior performance of our optimal transmission policy compared to two established baseline approaches.

  • Joint Codebook and UE Selection for mmWave Multi-gNB Coordination Based on Dual-Encoder

    by Mingqi Han, Xinghua Sun, Huadong Li, Xijun Wang, Qiaofeng Xue, Tony Q. S. Quek

    To meet the growing demand for network throughput, millimeter-wave (mmWave) communication has emerged as a promising solution. However, the inherent directional transmission of mmWave introduces severe interference in multi-gNB networks. While Deep Reinforcement Learning (DRL)-based joint selection of User Equipments (UEs) and codebooks is widely used for network optimization, conventional Multi-layer Perceptron (MLP)-based models suffer from performance degradation when the number of gNBs and UEs varies. In this paper, we propose a Dual-Encoder Q-network (DEQ) model to capture the correlations among gNBs and UEs for joint UE and codebook selection. Furthermore, we implement a sequential selection process for each gNB to enhance inter-cell coordination. Simulation results demonstrate that the proposed DEQ significantly improves network throughput while ensuring UE fairness across both seen and unseen dense scenarios.

  • To Sense or Not To Sense: A Comparative Study of Energy Efficiency in Energy-Aware Random Access Networks

    by Anshan Yuan, Xinghua Sun, Jiyun Qiu, Wen Zhan, Yaru Fu

    With the rapid proliferation of low-power machine-type devices (MTDs) in the Energy-Aware Internet of Things (EAIoT), a central challenge for IEEE 802.15.4-based wireless sensor networks is to maximize device lifetime while ensuring reliable data delivery. This paper addresses this challenge by proposing a general and practical energy model that facilitates the analysis of energy efficiency in both Carrier Sense Multiple Access (CSMA) and Aloha networks, and solves a constrained optimization problem: maximizing network lifetime under a lifetime throughput constraint. In particular, by jointly considering finite energy budgets, carrier sensing time, data traffic, and various power consumption levels, we derive closed-form expressions for the lifetime and lifetime throughput of MTDs operating under both CSMA and Aloha. Our analysis reveals a nuanced relationship: while a fundamental trade-off exists in saturated conditions, a win-win situation that enhances both lifetime and lifetime throughput emerges in the unsaturated regime. Building on this, we derive the optimal lifetime for both protocols under the lifetime throughput constraint, which facilitates a systematic comparative study. This comparison identifies when, and to what extent, carrier sensing can significantly improve MTDs’ lifetime while ensuring reliable packet delivery. The proposed analyses are validated through extensive simulations and a case study on IEEE 802.15.4-based networks, providing pivotal insights for the practical design of energy-efficient access protocols.

  • Multi-LLM Cooperation-Based Joint Intent-Driven Network Slicing and Resource Allocation

    by Mingqi Han, Xinghua Sun, Huadong Li, Chenyuan Feng, Xijun Wang, Qiaofeng Xue, Tony Q. S. Quek

    The exponential growth of heterogeneous mobile applications has raised challenges in current Radio Access Networks (RAN). While the network slicing technique offers virtualized networks with independent resources to satisfy diverse Quality of Service (QoS) requirements, persistent challenges remain in dynamic resource allocation and intent-driven slice configuration. For resource allocation, conventional Deep Reinforcement Learning (DRL)-based approaches encounter challenges in addressing the high-dimensional state-action space in large-scale networks with massive Base Stations (BS) and users. For slice configuration, current static intent-driven slice management frameworks suffer from inflexible architectures that fail to adaptively optimize network slicing configurations. In this letter, we propose a multi-Large Language Model (LLM) cooperation-based Joint Network Slicing and Resource Allocation (JNSRA) algorithm to jointly optimize the configurations of intent-driven network slices and resource allocation in large-scale multi-BS networks. In JNSRA, we propose to train Agent LLM through LLM alignment, which can enhance cooperation among slices and BS by regarding joint resource allocation actions as sequences. Moreover, we propose a multi-LLM joint optimization framework to jointly optimize Agent, leading to enhanced network slicing and resource allocation. Simulation results illustrate that JNSRA outperforms other DRL and heuristic approaches, and the proposed multi-LLM collaboration can further enhance the reward, satisfaction ratio and throughput.


  • Euclidean Distance Matrix Completion via Asymmetric Projected Gradient Descent

    by Yicheng Li, Xinghua Sun

    This paper proposes and analyzes a gradient-type algorithm based on Burer-Monteiro factorization, called the Asymmetric Projected Gradient Descent (APGD), for reconstructing the point set configuration from partial Euclidean distance measurements, known as the Euclidean Distance Matrix Completion (EDMC) problem. By paralleling the incoherence matrix completion framework, we show for the first time that global convergence guarantee with exact recovery of this routine can be established given $\mathscr{O}(\mu ^2 \ r^3 \ ?^2 \ n \ log \ n)$ Bernoulli random observations without any sample splitting. Unlike leveraging the tangent space Restricted Isometry Property (RIP) and local curvature of the low-rank embedding manifold in some very recent works, our proof provides extra upper bounds that act as analogies of the random graph lemma under EDMC setting. The APGD works surprisingly well and numerical experiments demonstrate exact linear convergence behavior in rich-sample regions yet deteriorates rapidly when compared with the performance obtained by optimizing the s-stress function, i.e., the standard but unexplained non-convex approach for EDMC, if the sample size is limited. While virtually matching our theoretical prediction, this unusual phenomenon might indicate that: (i) the power of implicit regularization is weakened when specified in the APGD case; (ii) the stabilization of such new gradient direction requires substantially more samples than the information-theoretic limit would suggest.


  • Open Access

    CoCo-Fed: A Unified Framework for Memory- and Communication-Efficient Federated Learning at the Wireless Edge

    by Zhiheng Guo, Zhaoyang Liu, Zihan Cen, Chenyuan Feng, Xinghua Sun, Xiang Chen, Tony Q. S. Quek, Xijun Wang

    The deployment of large-scale neural networks within the Open Radio Access Network (O-RAN) architecture is pivotal for enabling native edge intelligence. However, this paradigm faces two critical bottlenecks: the prohibitive memory footprint required for local training on resource-constrained gNBs, and the saturation of bandwidth-limited backhaul links during the global aggregation of high-dimensional model updates. To address these challenges, we propose CoCo-Fed, a novel Compression and Combination-based Federated learning framework that unifies local memory efficiency and global communication reduction. Locally, CoCo-Fed breaks the memory wall by performing a double-dimension down-projection of gradients, adapting the optimizer to operate on low-rank structures without introducing additional inference parameters/latency. Globally, we introduce a transmission protocol based on orthogonal subspace superposition, where layer-wise updates are projected and superimposed into a single consolidated matrix per gNB, drastically reducing the backhaul traffic. Beyond empirical designs, we establish a rigorous theoretical foundation, proving the convergence of CoCo-Fed even under unsupervised learning conditions suitable for wireless sensing tasks. Extensive simulations on an angle-of-arrival estimation task demonstrate that CoCo-Fed significantly outperforms state-of-the-art baselines in both memory and communication efficiency while maintaining robust convergence under non-IID settings.