PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning Based Joint Resource Scheduling Algorithms for Hybrid MEC Networks

Feibo Jiang, Kezhi Wang, Li Dong, Cunhua Pan, Wei Xu, Kun Yang

arXiv:1911.09598v1eess.SP

TL;DR

H-MEC offloading must minimize UE energy while jointly adapting MEC locations, associations, and resources in dynamic conditions. H2O combines LS-FCM placement, U-PSO sample generation, and scheduling-aware DNN inference for scalable online decisions. In evaluation, H2O reaches near-optimal performance similar to heuristic methods while producing actions in less than 0.03 second for 100 UEs, with PSO taking 80 times longer CPU time.

  • Problem

    The paper addresses real-time UE-energy minimization with joint placement, association, and resource-allocation decisions in dynamic H-MEC networks.

  • Method

    H2O combines LS-FCM for UAV and GV placement, U-PSO for MINLP sample generation, and a scheduling-layer DNN for per-UE online decisions.

  • Results

    Less than 0.03 second for 100 UEs; H2O achieves similar near-optimal performance to PSO while requiring 80 times less CPU time.

  • Takeaways & Limitations

    The framework makes real-time offloading and resource allocation viable for H-MEC networks in fast-changing environments within the reported evaluation scope.

Abstract

from arXiv · show

In this paper, we consider a hybrid mobile edge computing (H-MEC) platform, which includes ground stations (GSs), ground vehicles (GVs) and unmanned aerial vehicle (UAVs), all with mobile edge cloud installed to enable user equipments (UEs) or Internet of thing (IoT) devices with intensive computing tasks to offload. Our objective is to obtain an online offloading algorithm to minimize the energy consumption of all the UEs, by jointly optimizing the positions of GVs and UAVs, user association and resource allocation in real-time, while considering the dynamic environment. To this end, we propose a hybrid deep learning based online offloading (H2O) framework where a large-scale path-loss fuzzy c-means (LSFCM) algorithm is first proposed and used to predict the optimal positions of GVs and UAVs. Secondly, a fuzzy membership matrix U-based particle swarm optimization (U-PSO) algorithm is applied to solve the mixed integer nonlinear programming (MINLP) problems and generate the sample datasets for the deep neural network (DNN) where the fuzzy membership matrix can capture the small-scale fading effects and the information of mutual interference. Thirdly, a DNN with the scheduling layer is introduced to provide user association and computing resource allocation under the practical latency requirement of the tasks and limited available computing resource of H-MEC. In addition, different from traditional DNN predictor, we only input one UE information to the DNN at one time, which will be suitable for the scenarios where the number of UE is varying and avoid the curse of dimensionality in DNN.

I. INTRODUCTION

The paper targets real-time energy-efficient offloading in dynamic H-MEC networks, where conventional optimization can be too slow and learning methods require labeled data. It proposes H2O, combining LS-FCM, U-PSO, and a DNN with scheduling to support scalable online decisions.

  • Motivation: Conventional heuristics lack guaranteed solution quality, while search and convex-relaxation methods can require too many iterations for fast-changing environments.The optimization must be re-solved when UE numbers or positions change significantly.
  • Motivation: Heuristic algorithms offer global search and accuracy but require long computation times, whereas supervised learning requires substantial labeled training data.
  • Proposed framework: H2O jointly targets UE energy minimization by optimizing GV and UAV positions, user association, and resource allocation in real time.The platform includes GSs, GVs, and UAVs with edge-cloud capabilities.
  • Proposed framework: LS-FCM locates UAVs and GVs by fixing GS centers and replacing distance with a large-scale path-loss component in clustering.
  • Proposed framework: U-PSO solves task admission and resource allocation, using fuzzy membership information to capture small-scale fading and mutual interference while generating DNN training samples.A U-based roulette-wheel strategy provides initial solutions intended to accelerate PSO convergence.
  • Proposed framework: The DNN uses one UE's membership information at a time for real-time offloading decisions, and a scheduling layer checks constraints and admission control.The framework is designed for varying UE counts and dynamic scenarios, with new online samples fed back into the sample database.

II. SYSTEM MODEL AND PROBLEM FORMULATION

The system models a hybrid MEC network with fixed GSs and optimizable GVs and UAVs serving UE computation tasks. Its energy-minimization formulation jointly represents execution or offloading decisions, resource allocation, and mobile MEC locations as an MINLP.

  • Network model: The H-MEC network contains edge-enhanced GSs, GVs, and UAVs; GS locations are fixed, while GV and UAV locations can be optimized.
  • Task and association model: Each UE has one computation task and may execute it locally or offload it to a UAV, GV, or GS, with at most one execution location.
  • Task and association model: Task parameters include required CPU cycles F_i, transmitted data size D_i, and latency or QoS requirement T req.The paper assumes the same T req for all tasks.
  • Resource and energy model: The formulation accounts for local execution, offloading rates, computation capacities, transmission and execution power, latency constraints, and limited UE and MEC resources.UAV, GV, and GS links use distinct channel and interference assumptions, including shared-channel interference for UEs offloading to the same GV.
  • Optimization problem: The optimization minimizes UE energy over binary admission or association variables, continuous resource allocations, and UAV and GV locations.Problem P1 is difficult because it combines large-scale mixed-integer optimization, real-time decisions, and a dynamic environment.
  • Optimization problem: Problem P1 is an MINLP because admission decisions are binary while resource allocations and UAV and GV locations are continuous.The proposed framework uses clustering, PSO, and DNN inference to address these difficulties.

A. Algorithm Overview

H2O separates offline sample-based learning from online optimization. It trains a DNN to map system inputs to resource-allocation solutions, then uses lightweight per-UE inference for online offloading decisions.

  • Algorithm structure: H2O has four components: GV and UAV location optimization, sample collection, DNN offline learning, and DNN online decision-making.The procedure is divided into offline training and online optimization phases.
  • Offline training: The offline phase treats P1 as a mapping from system parameters to resource-allocation solutions and trains a DNN using generated samples.The remote cloud server performs the training because it has substantial computational and storage capabilities.
  • Online decision: During online execution, each UE supplies membership values and the DNN outputs its offloading choice and computing-resource allocation.This replaces repeated high-complexity heuristic optimization with simple algebraic calculations.

B. Location optimization based on LS-FCM

LS-FCM locates UAVs and GVs by adapting fuzzy c-means to fixed GS positions and large-scale path loss. Its iterative fuzzy memberships and cluster centers minimize aggregate path loss, supporting energy-aware placement.

  • LS-FCM modifications: LS-FCM fixes GS locations as cluster centers while iteratively optimizing the remaining UAV and GV centers.Fixed GS centers do not participate in the iteration process.
  • Fuzzy clustering formulation: FCM assigns each UE a fuzzy membership grade for each cluster, with UE positions represented as the clustering data points.The i-th UE is represented by its position (x_i, y_i).
  • LS-FCM modifications: The algorithm replaces UE–MEC distance with large-scale path loss in its clustering objective.This incorporates propagation conditions into location optimization rather than relying only on geometric distance.
  • Fuzzy clustering formulation: LS-FCM updates cluster centers and memberships iteratively until convergence, using the weighting exponent τ to control cluster fuzziness.The stopping procedure uses either a maximum iteration count or a sufficiently small objective-function variation.
  • Location assignment: Minimizing total large-scale path loss provides the basis for selecting UAV and GV positions and implicitly reduces the system energy objective.After hard classification, centers are ordered by the computing resources required by their assigned UEs before assignment to mobile edge platforms.

C. Computing offloading selection and computing resource allocation based on U-PSO

U-PSO combines particle-swarm optimization with dynamic fuzzy membership information to solve offloading and resource-allocation decisions. It uses high-quality optimization samples for DNN training while incorporating channel fading and interference into initialization and prediction inputs.

  • U-PSO role: U-PSO uses PSO to generate high-quality samples for offline DNN training, shifting task admission and resource allocation decisions to online inference.This addresses the long computation time of heuristic optimization in fast-changing environments.
  • U-PSO design: Traditional PSO is adapted because admission decisions are binary, constraints require checking, and random initialization ignores channel-quality information.The modified particle representation includes admission variables and allocated computing resources.
  • U-PSO design: The particle encodes each UE’s admission or offloading decision together with its allocated computing resource.The encoding concatenates the decision variables a_i and resource variables f_i.
  • U-based initialization: U-based roulette-wheel initialization uses fuzzy membership information to select candidate H-MECs and accelerate convergence from higher-quality starting points.The selection probabilities are derived from membership values and are used to initialize the particle solution.
  • Dynamic membership information: The dynamic fuzzy membership matrix incorporates small-scale fading and mutual interference, reflecting changing channel and interference conditions in real time.It both guides U-PSO initialization and provides a concise UE–H-MEC representation for the DNN.

D. Offline DNN training and online implementation

The framework trains a DNN on U-PSO solutions and uses a scheduling layer to produce constraint-aware task admission and resource allocation from each UE's membership values. This reduces online computation while accommodating changing UE populations.

  • Offline DNN training: The optimization problem is modeled as a mapping from fuzzy membership matrix U to optimal task admission and resource allocation.
  • Offline DNN training: DNN parameters are trained by minimizing average supervised loss with gradient descent and back-propagation over generated samples.The training set uses desired and actual outputs to define the loss.
  • DNN input-output design: The DNN maps each UE's membership vector to task admission and allocated computing resources.The input dimension depends on the number of H-MECs rather than the number of UEs.
  • Scheduling layer: The scheduling layer checks output constraints and returns the DNN decision only when all constraints are satisfied; otherwise, the UE executes locally.Its constraint and decision layers enforce feasibility after DNN inference.
  • Online implementation: During online operation, each UE inputs its membership values and receives a resource-allocation solution through simple algebraic calculations.New UE arrivals trigger updated H-MEC positioning and membership calculation before DNN inference.
  • Online implementation: A differential feedback check sends sufficiently different new inputs back to U-PSO for resolution as new training samples.The difference is evaluated using Euclidean distance against database samples.

IV. SIMULATION RESULTS

Simulations evaluate the H2O framework against PSO, Greedy, Random, and Local baselines. The proposed method closely matches PSO's energy performance while substantially reducing computation time.

  • Simulation setup: The simulations use three UAVs, one GV, and one GS, with specified LS-FCM, U-PSO, and DNN parameters.
  • DNN configuration: The DNN uses an 80%-training and 20%-testing split, and six hidden layers are selected after testing layer and node configurations.Thirty hidden nodes per layer achieve the minimum testing loss above six hidden layers.
  • Energy consumption: PSO achieves the lowest average energy consumption, while the proposed method has almost the same energy consumption as PSO as devices increase from 10 to 100.Greedy consumes less energy than Random, and Local has the highest consumption because it admits no devices.
  • Task admission: The proposed method admits the most devices for offloading across device counts and task-latency requirements.Greedy admits more devices than Random, whereas Local admits none.
  • Computation time: For 100 devices, the proposed method generates an offloading action in less than 0.03 second, while PSO requires 80 times longer CPU time.The authors report similar performance to PSO with substantially lower computation time.

V. CONCLUSION

H2O combines LS-FCM, U-PSO, and a scheduling-layer DNN for real-time offloading in H-MEC networks. It achieves near-optimal performance while reducing CPU time by more than several orders of magnitude, supporting real-time optimization in fast-changing environments.

  • H2O uses LS-FCM for GV and UAV positioning, U-PSO for MINLP-based sample generation, and a scheduling-layer DNN for real-time task admission and resource allocation.
  • H2O achieves similar near-optimal performance to heuristic methods while reducing CPU time by more than several orders of magnitude.
  • The evaluation compares average admitted devices as the number of devices varies from 10 to 100.
  • The evaluation also compares average admitted devices as T_req varies from 1s to 3s.
Loading 1911.09598v1…