Inference Accuracy-Aware Service Deployment and Model Selection in Edge Environments
摘要
随着边缘计算与人工智能技术的快速发展,深度神经网络(Deep Neural Networks,DNNs)已被广泛应用于边缘环境中,为终端用户提供智能化服务.然而,面对有限的边缘资源与多样化用户需求之间的矛盾,如何实现高效的服务部署与模型选择已成为一个挑战性难题.现有研究大多未充分考虑推理精度对服务体验的重要影响,难以有效应对动态变化的系统状态与用户需求,进而导致资源浪费或服务降级.为了解决这一重要挑战,本文提出了一种新颖的边缘环境下推理精度感知的服务部署与模型选择(Service Deployment and Model Selection,SDMS)方法.首先,利用自注意力机制对未来时间隙中的用户请求流量分布进行预测,并结合剪枝搜索策略对服务部署方案进行优化.其次,设计了一种新型的基于深度模仿学习(Deep Imitation Learning,DIL)的引导式深度强化学习(Deep Reinforcement Learning,DRL)框架,通过引入专家经验以加速收敛并快速逼近最佳的模型选择与资源分配决策.基于真实用户通信流量数据和 DNN 推理数据集,大量实验验证了所提出 SDMS 方法的有效性.实验结果表明,相比于其他基准方法,SDMS 方法在不同场景下均展现出了更加优越的服务收益、延迟与精度满足率,有效实现了多目标之间的平衡.
With the rapid advancement of edge computing and artificial intelligence technologies,Deep Neural Networks(DNNs)have been extensively deployed in edge environments to provide intelligent services to end-users.However,a critical challenge arises due to the inherent contra-diction between the limited computational and storage resources of edge servers and the increas-ingly diverse and dynamic demands of users.Efficient service deployment and model selection in such resource-constrained and demand-heterogeneous scenarios have become a significant prob-lem.Existing studies often neglect the critical impact of inference accuracy on overall service quality,which leads to suboptimal system performance when handling dynamically changing sys-tem states and user requirements.This oversight can result in resource underutilization,service degradation,and an inability to satisfy multiple objectives simultaneously.Addressing this chal-lenge requires a holistic approach that jointly considers prediction of user demand,resource con-straints,model performance,and service quality requirements.To tackle this problem,this paper proposes a novel inference-accuracy-aware service deployment and model selection framework for edge environments,referred to as Service Deployment and Model Selection(SDMS).The core i-dea of SDMS is to optimize both the deployment of services across edge nodes and the selection of DNN models for inference tasks to maximize service utility while satisfying latency and accuracy constraints.Specifically,the proposed approach consists of two key components.First,a self-at-tention mechanism is employed to predict the distribution of user request traffic in future time slots,capturing temporal correlations and dynamic patterns in user demands.This prediction is then combined with a pruning-based search strategy to identify near-optimal service deployment schemes,thereby ensuring that limited edge resources are allocated efficiently to meet expected demand.Second,a novel guidance-based Deep Reinforcement Learning(DRL)framework is de-signed,which is enhanced by Deep Imitation Learning(DIL).By incorporating expert knowledge and historical experience into the learning process,the DIL-guided DRL agent can significantly accelerate convergence and rapidly approach near-optimal decisions for model selection and re-source allocation.This integration of imitation learning and reinforcement learning allows the framework to adaptively balance multiple objectives,including service latency,inference accura-cy,and overall system throughput,under dynamically changing conditions.The proposed SDMS framework is extensively evaluated using real-world user traffic traces and DNN inference data-sets.A comprehensive set of experiments is conducted to assess its effectiveness in various realis-tic scenarios,including varying traffic patterns,heterogeneous model characteristics,and differ-ent resource availability levels.Experimental results demonstrate that the SDMS method consist-ently outperforms existing baseline approaches in terms of service utility,latency,and accuracy satisfaction rates.In particular,the framework exhibits a robust ability to maintain high infer-ence accuracy while meeting strict latency requirements,effectively achieving a balance among multiple conflicting objectives.In summary,the SDMS method provides a systematic and effec-tive solution for inference-accuracy-aware service deployment and model selection in edge compu-ting environments.By leveraging predictive modeling,pruning-based optimization,and imita-tion-guided reinforcement learning,the framework can dynamically adapt to user demand fluctua-tions and resource constraints,thereby enhancing the quality of service delivered to end users.The proposed approach offers a valuable reference for the design of intelligent edge systems where both computational efficiency and inference accuracy are critical for user satisfaction and overall system performance.The contributions of this work demonstrate the potential of combining ad-vanced AI techniques with edge resource management to address real-world challenges in next-generation intelligent services.