Searcharxiv⌕ Search

arXiv subjects

Yuelin Liu

Publications and source records attributed to Yuelin Liu.

6 recordsLinked to original sources
arXiv2026

Multi-Turn Distributed Inference with Mixture of Experts for 6G Edge--Cloud Networks

Mixture-of-Experts (MoE) architectures are increasingly deployed across 6G edge--cloud networks, where sparse activation reduces the computational footprint of each inference to only a fraction of the full expert set. However, MoE inference in edge-cloud networks creates a tension between KV state locality and elastic expert dispatch. KV state relocation incurs substantial transfer overhead, while expert computation benefits from spreading across the network to exploit available capacity. This tension is amplified in multi-turn inference, where each turn extends the KV state that must persist across the dialogue. To this end, we present StateFlow, a distributed inference policy that decouples persistent KV state from transient sparse computation. StateFlow pins KV state at a sticky serving site for cross-turn reuse and jointly optimizes expert dispatch and aggregation placement across the network. We further implement a real-world testbed with kernel-level network emulation and experimental results show that StateFlow sustains more than $2\times$ higher stable dialogue concurrency than the distributed baseline solutions and reduces turn-level p95 latency under multi-turn inference by 53.0\%.

cs.DC↗
arXiv2025

Incremental DRL-Based Resource Management for Dynamic Network Slicing in an Urban-Wide Testbed

Multi-access edge computing provides localized resources within mobile networks to address the requirements of emerging latency-sensitive and computing-intensive applications. At the edge, dynamic requests necessitate sophisticated resource management for adaptive network slicing. This involves optimizing resource allocations, scaling functions, and load balancing to utilize only essential resources under constrained network scenarios. However, existing solutions largely assume static slice counts, ignoring the re-optimization overhead associated with management algorithms when slices fluctuate. Moreover, many approaches rely on simplified energy models that overlook intertemporal resource scheduling and are predominantly evaluated through simulations, neglecting critical practical considerations. This paper presents an incremental cooperative Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm for resource management in dynamic edge slicing. The proposed approach optimizes long-term slicing benefits by reducing delay and energy consumption while minimizing retraining overhead in response to slice variations. Furthermore, we implement an urban-wide edge computing testbed based on OpenStack and Kubernetes to validate the algorithm's performance. Experimental results demonstrate that our incremental MADDPG method outperforms benchmark strategies in aggregated slicing utility and reduces training energy consumption by up to 50% compared to the re-optimization approach.

eess.SY↗
arXiv2025

Towards Practical Operation of Deep Reinforcement Learning Agents in Real-World Network Management at Open RAN Edges

Deep Reinforcement Learning (DRL) has emerged as a powerful solution for meeting the growing demands for connectivity, reliability, low latency and operational efficiency in advanced networks. However, most research has focused on theoretical analysis and simulations, with limited investigation into real-world deployment. To bridge the gap and support practical DRL deployment for network management, we first present an orchestration framework that integrates ETSI Multi-access Edge Computing (MEC) with Open RAN, enabling seamless adoption of DRL-based strategies across different time scales while enhancing agent lifecycle management. We then identify three critical challenges hindering DRL's real-world deployment, including (1) asynchronous requests from unpredictable or bursty traffic, (2) adaptability and generalization across heterogeneous topologies and evolving service demands, and (3) prolonged convergence and service interruptions due to exploration in live operational environments. To address these challenges, we propose a three-fold solution strategy: (a) advanced time-series integration for handling asynchronized traffic, (b) flexible architecture design such as multi-agent DRL and incremental learning to support heterogeneous scenarios, and (c) simulation-driven deployment with transfer learning to reduce convergence time and service disruptions. Lastly, the feasibility of the MEC-O-RAN architecture is validated on an urban-wide testing infrastructure, and two real-world use cases are presented, showcasing the three identified challenges and demonstrating the effectiveness of the proposed solutions.

cs.NI↗
arXiv2025

Cooperative Task Offloading through Asynchronous Deep Reinforcement Learning in Mobile Edge Computing for Future Networks

Future networks (including 6G) are poised to accelerate the realisation of Internet of Everything. However, it will result in a high demand for computing resources to support new services. Mobile Edge Computing (MEC) is a promising solution, enabling to offload computation-intensive tasks to nearby edge servers from the end-user devices, thereby reducing latency and energy consumption. However, relying solely on a single MEC server for task offloading can lead to uneven resource utilisation and suboptimal performance in complex scenarios. Additionally, traditional task offloading strategies specialise in centralised policy decisions, which unavoidably entail extreme transmission latency and reach computational bottleneck. To fill the gaps, we propose a latency and energy efficient Cooperative Task Offloading framework with Transformer-driven Prediction (CTO-TP), leveraging asynchronous multi-agent deep reinforcement learning to address these challenges. This approach fosters edge-edge cooperation and decreases the synchronous waiting time by performing asynchronous training, optimising task offloading, and resource allocation across distributed networks. The performance evaluation demonstrates that the proposed CTO-TP algorithm reduces up to 80% overall system latency and 87% energy consumption compared to the baseline schemes.

cs.LG↗
arXiv2020

Phase transition of disordered random networks on quasi-transitive graphs

Given a quasi-transitive infinite graph $G$ with volume growth rate ${\rm gr}(G),$ a transient biased electric network $(G,\, c_1)$ with bias $λ_1\in (0,\,{\rm gr}(G))$ and a recurrent biased one $(G,\, c_2)$ with bias $λ_2\in ({\rm gr}(G),\infty).$ Write $G(p)$ for the Bernoulli-$p$ bond percolation on $G$ defined by the grand coupling. Let $(G,\, c_1,\, c_2,\, p)$ be the following biased disordered random network: Open edges $e$ in $G(p)$ take the conductance $c_1(e)$, and closed edges $g$ in $G(p)$ take the conductance $c_2(g)$. Our main results are as follows: (i) On connected quasi-transitive infinite graph $G$ with percolation threshold $p_c\in (0,\, 1),$ $(G,\, c_1,\, c_2,\, p)$ has a non-trivial recurrence/transience phase transition such that the threshold $p_{c}^{*}\in (0,\, 1)$ is deterministic, and almost surely $(G,\, c_1,\, c_2,\, p)$ is recurrent for $p p_c^*.$ There is a non-trivial recurrence/transience phase transition for $(G,\, c_1,\, c_2,\, p)$ with $G$ being a Cayley graph if and only if the corresponding group is not virtually $\mathbb{Z}$. (ii) On $\mathbb{Z}^d$ for any $d\geq 1,$ $p_c^{*}= p_c$. And on $d$-regular trees $\mathbb{T}^d$ with $d\geq 3$, $p_c^{*}=(λ_1\vee 1) p_c$, and thus $p_c^{*}>p_c$ for any $λ_1\in (1,\,{\rm gr}(\mathbb{T}^d)).$ As a contrast, we also consider phase transition of having unique currents or not for $(\mathbb{Z}^d,\, c_1,\, c_2,\, p)$ with $d\geq 2$ and prove that almost surely $(\mathbb{Z}^2,\, c_1,\, c_2,\, p)$ with $λ_1<1\leqλ_2$ has unique currents for any $p\in [0,1]$.

math.PR↗