PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for Hybrid 5G Services in Mobile Edge Computing Systems: Learn from a Digital Twin

Rui Dong, Changyang She, Wibowo Hardjawana, Yonghui Li, Branka Vucetic

arXiv:1907.01523v1eess.SPcs.ITcs.LG

TL;DR

The paper tackles energy-efficient operation of hybrid URLLC and delay-tolerant services in multi-AP MEC under QoS constraints. It uses a digital twin to train a DL-based user-association model offline, while APs optimize resource allocation and offloading. Simulations report lower normalized energy consumption and lower complexity than an existing method, approaching the global optimum.

  • Problem

    Hybrid MEC must optimize user association, resource allocation, and offloading for URLLC and delay-tolerant services in a non-convex problem.

  • Method

    A digital twin trains a DNN for MME-managed user association, while each AP optimizes resource allocation and offloading for the selected association.

  • Results

    The low-complexity AP algorithm can save more than 87 % energy compared with the baselines, while DL-based association approaches the global optimum.

  • Takeaways & Limitations

    The framework combines digital-twin-based learning with optimization to support energy-efficient hybrid 5G services in MEC.

Abstract

from arXiv · show

In this work, we consider a mobile edge computing system with both ultra-reliable and low-latency communications services and delay tolerant services. We aim to minimize the normalized energy consumption, defined as the energy consumption per bit, by optimizing user association, resource allocation, and offloading probabilities subject to the quality-of-service requirements. The user association is managed by the mobility management entity (MME), while resource allocation and offloading probabilities are determined by each access point (AP). We propose a deep learning (DL) architecture, where a digital twin of the real network environment is used to train the DL algorithm off-line at a central server. From the pre-trained deep neural network (DNN), the MME can obtain user association scheme in a real-time manner. Considering that real networks are not static, the digital twin monitors the variation of real networks and updates the DNN accordingly. For a given user association scheme, we propose an optimization algorithm to find the optimal resource allocation and offloading probabilities at each AP. Simulation results show that our method can achieve lower normalized energy consumption with less computation complexity compared with an existing method and approach to the performance of the global optimal solution.

I. INTRODUCTION

The paper addresses energy-efficient hybrid 5G MEC by jointly optimizing user association, resource allocation, and offloading under URLLC and delay-tolerant service requirements. It combines a digital-twin-trained DL framework with per-AP optimization to obtain efficient, near-optimal decisions in changing networks.

  • Motivation: Hybrid MEC must satisfy URLLC delay and reliability constraints while serving delay-tolerant services and improving users’ energy efficiency.Short packets can be lost in deep fading, and finite blocklength coding causes nonzero decoding errors even at high SNR.
  • Problem: Jointly optimizing user association, resource allocation, and offloading probabilities is a non-convex, complicated problem in multi-AP hybrid-service MEC.The objective is normalized energy consumption, defined as energy consumption per bit, subject to quality-of-service requirements.
  • Method: The proposed digital twin combines real-network data and theoretical rules to evaluate energy, delay, and packet-loss outcomes for candidate decisions.It monitors network variations and updates the twin so the DL system can operate in non-stationary environments.
  • Method: A central server trains a DNN offline for user association, sends it to the MME, and updates it as the real network changes.For a given association scheme, each AP optimizes resource allocation and task offloading.
  • Results: The per-AP optimization algorithm converges to the global optimal solution with linear complexity despite the AP problem being non-convex.Exploration policies generate association schemes whose normalized energy consumption, delay, and reliability are evaluated in the digital twin for labeled training samples.
  • Results: Simulations show lower normalized energy consumption and less computing complexity than an existing solution while approaching the global optimum.The framework is designed for hybrid 5G services in multi-AP MEC systems.

II. RELATED WORK

Prior MEC studies addressed energy efficiency, latency, reliability, learning-based resource management, or user association, but generally did not resolve hybrid 5G service provision. The paper therefore develops a multi-AP MEC framework combining digital-twin modeling, DL-based association, and AP-level optimization.

  • Conventional MEC optimization: Earlier studies optimized energy efficiency or energy–latency tradeoffs mainly for single-service or single-AP settings.Examples include minimizing power under latency and reliability constraints and maximizing energy efficiency subject to delay constraints.
  • Research gap: Prior literature mainly focused on one service type and neglected service heterogeneity, although multi-access MEC work optimized resource management and user association.Hybrid 5G service provision in MEC systems remained unclear and required further study.
  • Learning-based MEC: Existing learning-based MEC work studied offloading and resource management, including energy-harvesting systems, renewable energy, and single-AP task offloading.These studies provided machine-learning insights but did not consider 5G services.
  • System and approach: The considered system contains multiple APs serving URLLC and delay-tolerant users, with AP-level resource allocation and task offloading.The MME manages user association, while APs solve the resulting single-AP problems.
  • Digital-twin framework: The digital twin represents topology, channels, queueing models, and theoretical rules, while monitored non-stationary parameters update DNN training.The previous trained DNN initializes the updated DNN rather than training from scratch.

B. Computation Tasks and Communication Packets

The system models short URLLC packets and longer delay-tolerant packets, with tasks processed locally or offloaded to MEC servers over OFDMA links. Channel, packet, and computation assumptions support rate and offloading models.

  • Task and packet models: Tasks are conveyed in one packet, with packet size and required CPU cycles represented separately.
  • Task and packet models: URLLC packets have constant size and computational demand, and arrivals follow a Bernoulli process with one packet or silence per slot.
  • Task and packet models: Delay-tolerant packet sizes and inter-arrival times may follow general distributions, but their packets are assumed much longer than URLLC packets.
  • Offloading and association: Users can offload tasks to one MEC server, with β indicating AP association and β_m,k indicating whether user k is associated with AP m.
  • Wireless transmission: The wireless model uses OFDMA, where N_k subcarriers are allocated per user; URLLC packets experience flat-fading quasi-static channels.
  • Wireless transmission: Only 1 bit CSI is available at each transmitter, determining whether delay-tolerant packets are offloaded with probability one or zero based on a channel-gain threshold.

2) Offloading Policy of Delay Tolerant Services:

The delay-tolerant offloading policy is channel-independent because long packets may span multiple coherence times. MEC APs therefore use processor sharing to prevent short packets from being blocked behind long packets.

  • Offloading policy: For long packets, the offloading policy does not depend on the current small-scale channel gain, and each packet is offloaded with probability x_b.
  • Queueing models: Local servers serve packets FCFS, while each AP adopts a processor-sharing server that equally divides service among packets.
  • Queueing models: URLLC packets have no queue before uplink transmission because their peak arrival rate equals the one-slot wireless transmission rate.
  • Queueing models: Delay-tolerant arrivals can exceed the transmission rate, so packets may wait in a communication queue before uplink transmission.
  • Queueing models: With both short and long packets, processor sharing is reported to outperform FCFS.
  • Optimization objective: The formulation analyzes QoS constraints and minimizes the maximum energy consumption per bit across users by optimizing association, resources, and offloading probabilities.

B. Stability of Delay Tolerant Services

Delay-tolerant service stability requires processing and transmission rates to meet arrival rates, while the optimization minimizes normalized energy consumption under these and QoS constraints. The problem is decomposed across APs and the MME, with deep learning used for association.

  • Stability conditions: Delay-tolerant queueing systems are stable when average service rates are equal to or higher than average arrival rates.
  • Stability conditions: Local-server processing rates must exceed average data arrivals without exceeding the server's maximal computing capacity.
  • Stability conditions: Communication-queue stability requires the wireless link's average transmission rate to equal or exceed the average data arrival rate.
  • Energy optimization: The objective is normalized energy consumption, with energy per bit evaluated for local processing and MEC offloading.
  • Energy optimization: Fairness is incorporated by minimizing the maximal normalized energy consumption among users.
  • Decomposition and learning: User association is managed by the MME, whereas APs optimize resource allocation and offloading probabilities for a given association scheme.
  • Decomposition and learning: The MME uses a deep learning algorithm for user association while accounting for each AP's optimal resource allocation and task-offloading behavior.

E. Structure of Deep Learning

The framework uses a digital twin to train a DNN for user association, then solves resource allocation and offloading for each association scheme while enforcing QoS feasibility.

  • Both P2 and P3 are non-convex, motivating separate optimization procedures for resource allocation and deep-learning-based user association.
  • The DNN takes large-scale channel gains and average task arrival rates as inputs and outputs a user association scheme.
  • The digital twin evaluates explored association schemes, and the scheme with minimum normalized energy consumption is retained for DNN training.
  • Given a user association scheme, P2 decomposes into multiple single-AP problems whose solutions determine resource allocation and offloading probabilities.
  • The algorithm uses binary searches over normalized energy consumption and subcarrier allocation, then checks the QoS-dependent constraints for feasibility.
  • If the constraints cannot be satisfied, P2 is infeasible, the AP cannot guarantee all associated users' QoS requirements, and the association scheme is excluded from DNN training.

B. Optimal Offloading Probability

The paper derives optimal offloading probabilities for URLLC services by sequentially enforcing delay, error, power, and energy requirements.

  • The URLLC procedure first finds the minimum service rate satisfying end-to-end and queueing-delay violation constraints.
  • It then finds the minimum channel-gain threshold and transmit power satisfying decoding-error and maximal-transmit-power constraints.
  • The normalized energy objective first decreases and then increases with the channel-gain threshold, yielding a unique optimal threshold solution.
  • Substituting the optimal threshold into the service model gives the optimal offloading probability in closed form.

2) Delay Tolerant Services:

For delay-tolerant services, the algorithm searches feasible offloading probabilities and subcarrier allocations, using monotonicity and equivalence properties to establish feasibility and convergence.

  • For feasible cases, binary search finds the optimal offloading probability because the objective is convex in the offloading variable.
  • The method bounds each delay-tolerant user's offloading probability using average-rate, transmit-power, and local-server service-rate constraints.
  • If the lower bound exceeds the upper bound, the delay-tolerant single-user problem is infeasible.
  • The overall P2 feasibility test checks whether total subcarriers and total offloading probability satisfy the AP constraints, while binary search converges to the minimum normalized energy consumption η*.
  • Minimizing a user's offloading probability is equivalent to minimizing its allocated subcarriers, so minimizing total subcarriers also minimizes total offloading probability and MEC workload.

D. Complexity Analysis

The optimization complexity grows with the numbers of users and search precisions, while DNN training repeatedly updates parameters from stored digital-twin-generated samples.

  • The required search steps for normalized energy consumption increase linearly with Ku + Kb.
  • Subcarrier allocation uses precision-controlled searches, and delay-tolerant offloading probabilities are optimized over [0, 1].
  • URLLC offloading probabilities use a closed-form expression, which contributes to the stated complexity expression.
  • The DNN uses large-scale channel gains and task arrival rates to calculate user association outputs, while explored schemes provide training targets.
  • Training samples are stored in memory, sampled for experience replay, and used with Adam updates until the loss falls below σL.
  • After training, the MME uses the DNN to calculate a user association scheme for any α and λ.

VII. SIMULATION RESULTS

The simulations evaluate the proposed DL framework across network, learning, user-association, and non-stationary settings. The method achieves lower normalized energy consumption than baselines while reducing computation complexity and remaining close to optimal performance.

  • Simulation setup: The simulation varies user distribution, channel conditions, service arrivals, and network size to evaluate normalized energy consumption.The setup uses a path-loss model, lognormal shadowing, Rayleigh fading, and packet-arrival rates for delay tolerant and URLLC users.
  • DL implementation: The DNN uses one input layer, four hidden layers, and one output layer, with input features derived from α and λ.The implementation uses a learning rate of 0.001, 128 training samples per epoch, and memory for 1024 samples.
  • Bandwidth allocation and offloading: Around 89% and 87% normalized energy consumption are saved versus the ‘MEC’ and ‘Local’ schemes, respectively.The ‘MEC’ scheme offloads all packets, whereas the ‘Local’ scheme processes all packets at local servers.
  • User association: The proposed method achieves much smaller normalized energy consumption than three baselines and performs close to the optimal scheme.For M = 3, exhaustive search cannot obtain the optimal scheme because its complexity is too high.
  • Exploration policies: The DNN output improves exploration efficiency: direct use saves around 30 % versus ‘Highest α’, while ‘DL{10, 0}’ outperforms random exploration with fewer explorations.The results indicate that DNN outputs help improve the efficiency of the exploration policy.
  • Non-stationary environments: Updating the DNN is necessary because a fixed DNN can become worse than ‘Highest α’ when user density varies.The updated DNN achieves satisfactory performance after the network user distribution ratio changes.

VIII. CONCLUSION

The paper proposes a digital-twin-trained DL architecture for user association and a low-complexity AP-level optimizer for resource allocation and offloading. Simulations report substantial energy savings, lower complexity, and performance approaching the global optimum.

  • Objective: The framework minimizes normalized energy consumption for URLLC and delay tolerant services in a multi-AP MEC system.Normalized energy consumption is energy consumption per bit.
  • DL architecture: A digital twin trains the user-association algorithm offline at a central server before the DNN is sent to the MME.The MME manages user association in real time.
  • AP optimization: For a given user association scheme, a low-complexity algorithm optimizes resource allocation and offloading probabilities at each AP.The network can be decomposed into single-AP problems once user association is fixed.
  • Results: More than 87 % energy is saved compared with the baselines, while the DL association scheme achieves lower normalized energy consumption with less computing complexity and approaches the global optimal solution.These outcomes are reported for the proposed optimization and DL framework.

APPENDIX A

The appendix establishes convexity and monotonicity properties used to analyze the resource-allocation and offloading optimization. The proofs separately examine URLLC and delay tolerant service terms.

  • Convexity proofs: The normalized energy expression is shown to be convex in the delay tolerant offloading probability x_b,k.Both terms in the relevant expression are treated as convex in x_b,k.
  • URLLC analysis: For URLLC services, the required transmit power and offloading-related quantities are analyzed through their dependence on service rates and user counts.The proof uses decreasing and increasing relationships involving N_u,k and the threshold gain.
  • Delay tolerant analysis: For delay tolerant services, the proof derives monotonicity and convexity relationships for transmission power and offloading probability under the wireless-link rate constraint.The analysis considers the required average data rate and the inverse-function characteristic.
Loading 1907.01523v1…