PaperlayerView exact revision on arXiv

Source-linked AI summary

Mobility-aware Seamless Service Migration and Resource Allocation in Multi-edge IoV Systems

Zheyi Chen, Sijin Huang, Geyong Min, Zhaolong Ning, Jie Li, Yan Zhang

arXiv:2503.13494v1cs.NIcs.AI

TL;DR

Dynamic vehicle mobility and limited MEC resources make seamless service migration and resource allocation difficult, while prior approaches often rely on prior knowledge or omit efficient allocation. SR-CL decouples the MINLP, uses asynchronous actor-critic learning for migration, and derives convex-optimal resource allocation; experiments report better convergence and delay performance than benchmarks across scenarios.

  • Problem

    High vehicle mobility and finite BS coverage complicate seamless service, while existing solutions commonly rely on prior knowledge and rarely optimize resource allocation during migration.

  • Method

    SR-CL decouples service migration from resource allocation, applies asynchronous actor-critic DRL to migration, and derives optimal per-server allocation through convex optimization.

  • Results

    SR-CL achieves better convergence and delay performance than benchmark methods under different scenarios using real-world vehicle trajectories and a testbed.

  • Takeaways & Limitations

    The framework provides a mobility-aware integrated approach for seamless service migration and resource allocation in multi-edge IoV systems.

  • Takeaways & Limitations

    The method faces rapidly growing time complexity for huge action and state spaces, while continuous-space exploration and strict allocation constraints can cause invalid exploration.

Abstract

from arXiv · show

Mobile Edge Computing (MEC) offers low-latency and high-bandwidth support for Internet-of-Vehicles (IoV) applications. However, due to high vehicle mobility and finite communication coverage of base stations, it is hard to maintain uninterrupted and high-quality services without proper service migration among MEC servers. Existing solutions commonly rely on prior knowledge and rarely consider efficient resource allocation during the service migration process, making it hard to reach optimal performance in dynamic IoV environments. To address these important challenges, we propose SR-CL, a novel mobility-aware seamless Service migration and Resource allocation framework via Convex-optimization-enabled deep reinforcement Learning in multi-edge IoV systems. First, we decouple the Mixed Integer Nonlinear Programming (MINLP) problem of service migration and resource allocation into two sub-problems. Next, we design a new actor-critic-based asynchronous-update deep reinforcement learning method to handle service migration, where the delayed-update actor makes migration decisions and the one-step-update critic evaluates the decisions to guide the policy update. Notably, we theoretically derive the optimal resource allocation with convex optimization for each MEC server, thereby further improving system performance. Using the real-world datasets of vehicle trajectories and testbed, extensive experiments are conducted to verify the effectiveness of the proposed SR-CL. Compared to benchmark methods, the SR-CL achieves superior convergence and delay performance under various scenarios.

1 INTRODUCTION

The paper addresses seamless service migration and resource allocation for mobile vehicles in multi-edge IoV systems. SR-CL decouples the joint problem, uses asynchronous actor-critic learning for migration, and derives resource allocation analytically.

  • Motivation: High vehicle mobility and finite BS coverage make uninterrupted service difficult, while poor migration decisions can increase response delay and degrade QoS.Migration must account for vehicle mobility, task attributes, and available MEC resources.
  • Motivation: Existing migration studies often neglect resource allocation, hindering performance improvement on resource-constrained MEC servers.MEC servers face continuous task inflows with diverse resource demands.
  • Framework: SR-CL formulates long-term QoS optimization and decouples the original MINLP into service-migration and resource-allocation sub-problems.The objective includes migration, communication, and computation delays.
  • Framework: The migration sub-problem uses an asynchronous actor-critic method whose delayed-update actor chooses migrations and one-step-update critic evaluates them.The critic guides actor updates during policy exploration.
  • Framework: The resource-allocation sub-problem is solved by convex optimization, using Hessian analysis, Lagrangian functions, and KKT conditions to derive optimal allocations per MEC server.The method is designed for given migration decisions.
  • Evaluation: Experiments using Rome vehicle trajectories and a testbed report better convergence and delay performance than benchmark methods under different scenarios.The evaluation focuses on complex mobility patterns in the city center.

2 RELATED WORK

Prior work studies service migration, resource allocation, or computation offloading, but commonly separates mobility from resource decisions. The paper positions SR-CL as an integrated mobility-aware approach using DRL and convex optimization.

  • Service Migration: Related migration studies address QoS, migration overhead, network status, user mobility, or proactive migration through methods including particle swarm optimization, Lyapunov optimization, and DQN.These studies represent varied migration objectives and decision mechanisms.
  • Service Migration: Earlier migration methods use optimization theories, heuristics, or value-based DRL, with limitations involving iterations, rule-setting overhead, local optima, or action-space size.Value-based DRL performs well when the action space is small.
  • Resource Allocation: MEC resource-allocation studies combine mathematical optimization and DRL to address computation rate, energy consumption, task execution delay, and related constraints.Examples include weighted MSE, quadratic transformation, difference-of-convex methods, and model-free DRL.
  • Research Gap: Existing resource-allocation work generally does not adequately model user or vehicle mobility during decision making.Changing geographic locations complicate seamless service provision.
  • Research Gap: The paper claims to be the first to integrate DRL with convex optimization for joint seamless service migration and resource allocation in multi-edge IoV systems.The stated scope is mobility-aware operation in multi-edge IoV environments.

3 SYSTEM MODEL AND PROBLEM FORMULATION

The system models mobile vehicles, MEC servers, migration, communication, computation, and resource constraints in discrete time. The resulting long-term MINLP is NP-hard and is decoupled into migration and resource-allocation sub-problems.

  • System Model: The system contains a MEC controller, multiple BS-associated edge nodes, and intelligent vehicles that offload tasks through 5G to service instances.Tasks are processed at MEC servers and returned to vehicles.
  • Migration Model: Vehicle movement can disconnect it from the edge node hosting its service, creating a trade-off between backhaul communication delay and service-migration delay.Migration also requires suitable computational resources at the destination node.
  • Migration Model: Migration decisions select the edge node hosting each vehicle’s service instance, with hop distance determining whether migration occurs and contributing to migration delay.Migration delay increases with service-data amount and hop distance.
  • Communication Model: Communication delay includes vehicle-to-BS transmission and, when source and destination differ, backhaul transmission.Backhaul delay depends on task data amount and hop distance, while download delay is treated as negligible.
  • Computation Model: Computation delay depends on task CPU-cycle requirements and the computational resource proportion allocated to each service instance.Each MEC server has a maximum CPU frequency and allocation decisions are collected over vehicles.
  • Problem Formulation: The objective minimizes long-term migration, communication, and computation delays while enforcing discrete migration and bounded resource-allocation constraints.Each service instance runs on one edge node per time slot, and allocations on each node obey a proportion-sum constraint.
  • Problem Complexity: The joint problem is NP-hard, as shown by reducing a specific instance to the Knapsack Problem and extending it across multiple time slots.The construction uses limited total MEC resources and delay-based objective values.
  • Problem Decoupling: Fixing migration decisions decouples the problem into migration optimization for long-term delays and resource allocation for computation delay.The resource-allocation sub-problem is optimized after migration decisions determine the relevant service-instance placement.

4 THE PROPOSED SR-CL

SR-CL models service migration as sequential decision-making and combines an improved actor-critic DRL method with convex-optimization-based resource allocation. Its asynchronous actor-critic design addresses large action spaces while resource allocation is derived within the migration process.

  • Resource allocation: SR-CL derives optimal resource allocation through convex optimization and embeds it into DRL, reducing the action-space dimension.The training procedure obtains resource allocation after executing migration decisions and uses it when calculating rewards and updating the networks.
  • Service migration model: Service migration is modeled as an MDP in which the agent observes states, selects migration actions, receives rewards, and transitions between states.The state includes vehicle and task information together with prior migration decisions; rewards reflect long-term migration, communication, and computation delays.
  • Policy learning: DRL evaluates trajectories using discounted cumulative rewards and state-action value functions to guide policy learning toward an optimal policy.The expected reward evaluates actions, while the optimal state-action value identifies the best action at each state.
  • Improved DRL: The framework addresses value-based DRL limitations from huge action spaces and biased target-network estimates, while policy-based DRL can suffer high-variance training.These limitations motivate the improved actor-critic approach for dynamic multi-edge IoV service migration.
  • Improved DRL: The improved DRL method uses an actor-critic architecture in which the actor generates migration actions and the critic evaluates their Q-values.The critic uses one-step updates to guide actor updates and reduce action-value estimation errors.
  • Training procedure: The training procedure initializes actor, critic, target networks, and replay memory, then iteratively samples experiences, updates the critic and actor, and softly updates target networks.The algorithm also updates vehicle locations, executes migration decisions, derives resource allocation, and calculates rewards within each episode.

23 end

SR-CL combines delayed actor-critic learning for service migration with convex optimization for resource allocation. Its implementation separates migration decisions in the MEC controller from parallel allocation decisions on edge nodes, while reducing exploration difficulty and handling resource constraints.

  • Service migration learning: The actor generates migration actions from system states, while the critic evaluates them and guides policy updates.The critic updates every step, whereas the actor is updated only after the critic has been updated λ times.
  • Service migration learning: Replay memory breaks training-sample correlation, and gradient clipping plus soft target updates improve network-update stability.Gaussian noise is also introduced during initial training to enhance policy exploration.
  • Convex resource allocation: Resource allocation is separated by MEC server because the allocation processes of different servers are independent and parallel.The resource-allocation proportion is constrained by lower and upper bounds.
  • Convex resource allocation: P4 is proved convex using its constraints and Hessian matrix, after which KKT conditions yield the optimal allocation for each MEC server.The resulting allocation is obtained under given migration decisions.
  • Integration and complexity: For every time slot, each MEC server obtains optimal service-instance allocation, supporting the DRL agent in learning better migration policies.The allocation decision is performed according to the derived solution, whose stated time complexity is O(1).
  • Integration and complexity: SR-CL reduces exploration complexity by combining DRL with convex optimization, while joint migration and allocation would otherwise enlarge hybrid action spaces and produce invalid exploration.The implementation places allocation decision makers on edge nodes and the remaining components in the MEC controller.

5 PERFORMANCE EVALUATION

SR-CL is evaluated with real-world vehicle trajectories, simulations, large-scale topology variations, and a Simu5G-based testbed. Across dynamic traffic, computational capacity, topology, and vehicle-speed scenarios, it generally achieves lower delay, stronger convergence, and improved response time than benchmark methods.

  • Convergence Comparison: SR-CL converges more stably than benchmark methods because convex optimization avoids directly exploring the full continuous resource-allocation action space.JSR may prematurely converge because its joint continuous action space is difficult to explore, whereas SR-CL combines improved DRL with convex optimization.
  • Delay Comparison: As computational capability increases, total system delay decreases, and SR-CL improves performance over DDPG, JSR, IDQN, GA, NM, and AM by around 3.63%, 9.90%, 30.96%, 15.95%, 20.29%, and 26.49%, respectively.Greater MEC capacity reduces computation delay by providing IVs with more available computational resources.
  • Delay Comparison: SR-CL adapts to dynamic traffic changes and improves total-delay performance over DDPG, JSR, IDQN, GA, NM, and AM by approximately 2.77%, 14.09%, 20.11%, 13.86%, 21.98%, and 36.15%, respectively.The reported results attribute this behavior to decoupled resource allocation and real-time DRL-based migration decisions.
  • Delay Comparison: SR-CL outperforms DDPG, JSR, IDQN, GA, NM, and AM in total delay, with improvements of approximately 3.62%, 11.42%, 54.31%, 29.29%, 16.42%, and 28.59%, respectively.This comparison is reported when the unit migration-delay coefficient is 4.5.
  • Delay Comparison: As network connectivity decreases, all methods experience higher delay, while SR-CL retains an advantage through adaptive interaction and optimal resource allocation.For the large-scale topology experiment, SR-CL improves performance over DDPG, JSR, IDQN, GA, NM, and AM by approximately 5.53%, 9.29%, 39.35%, 21.03%, 25.50%, and 25.46%, respectively.

6 CONCLUSION

SR-CL addresses seamless service migration and resource allocation in multi-edge IoV systems through decoupled optimization, improved deep reinforcement learning, and convex optimization. Experiments show better performance and convergence than benchmark methods across diverse scenarios.

  • SR-CL formulates long-term QoS optimization and decouples service migration from resource allocation into two sub-problems.
  • For service migration, SR-CL uses delayed and one-step update mechanisms in an improved deep reinforcement learning method.
  • For resource allocation, SR-CL derives the optimal allocation theoretically using convex optimization.
  • Across varied computational capabilities, migration delays, topologies, bandwidths, and vehicle counts, SR-CL outperforms benchmark methods with faster, more stable convergence.Real-world vehicle-trajectory datasets and testbed experiments validate the framework’s effectiveness and practicality.
Loading 2503.13494v1…