A Unified Framework for Energy-Aware Edge AI Offloading Strategy

Thesis Abstract

The rapid growth of edge computing and AI workloads imposes a critical energy efficiency challenge, as constrained edge devices must balance local processing, offloading costs, and response latency while preserving user quality of experience. This study addresses the problem of designing a unified framework for energy-aware offloading decisions that harmonizes device-level, network, and cloud resources to minimize energy consumption without compromising AI performance. The aim is to develop a theoretical and empirical framework that integrates decision-theoretic models, learning-based offloading policies, and system-level energy metrics to optimize offloading under heterogeneous network conditions and workload characteristics. Specific objectives are to (i) formulate an energy-aware offloading model capturing local execution energy, transmission energy, latency, and task accuracy; (ii) derive a decision policy space grounded in reinforcement learning and optimization theory; (iii) implement a modular offloading framework that supports adaptive policy selection across CPU, GPU/accelerator, and cloud resources; (iv) evaluate the framework on a realistic edge-to-cloud testbed with representative AI workloads; and (v) assess robustness to dynamic network conditions and model drift through scenario-based experiments. The methodology adopts a mixed-methods design combining formal modeling, simulation, and empirical validation. The population comprises edge devices (Raspberry Pi 4 and NVIDIA Jetson Nano), wireless access points, and cloud servers hosting lightweight and heavy AI models. A stratified sample of 60 edge devices and 6 edge-to-cloud deployment scenarios is used to capture heterogeneity in hardware, network bandwidth, and latency. Data collection employs instrumentation of energy consumption (via shunt-based power measurement and device-native counters), kernel-level telemetry (CPU/GPU utilization, memory bandwidth, and cache misses), and network metrics (throughput, packet loss, RTT). Task workloads include a suite of 10 neural network inference tasks across computer vision and natural language processing, with varying input sizes to produce diverse offloading decisions. Instruments include calibrated power meters, software profilers, and a telemetry collector synchronized with a central data repository. Analytical techniques encompass (i) energy–latency–accuracy trade-off modeling using multi-objective optimization and Pareto-front analysis; (ii) supervised and reinforcement learning approaches (DQN, PPO) for policy learning, with transfer learning across devices; (iii) regression and time-series analysis to model energy consumption dynamics; (iv) Statistical hypothesis testing (ANOVA and post-hoc tests) to compare policy performance under different network regimes; and (v) ablation studies to quantify the contribution of each framework component. The analysis will leverage a simulation-enhanced empirical pipeline where a discrete-event network simulator informs offloading decisions, complemented by real-world measurements from the testbed. The model specification includes a dynamic offloading decision function f(t, s, a, b) mapping state features (device energy state, network conditions, workload characteristics) to action choices (local execution, edge offload, or cloud offload), constrained by latency budgets and energy caps. Expected findings include (i) a reproducible energy-aware offloading policy that achieves measurable energy savings (15–35%) with comparable or improved latency across heterogeneous networks; (ii) a validated energy model enabling rapid estimation of offloading energy under varying conditions; (iii) evidence that hybrid policies combining model-based optimization with learning-based adaptation outperform static heuristics; and (iv) robust performance under network variability due to adaptive policy selection and transfer learning across devices. The study contributes to knowledge by presenting a unified framework that synthesizes energy-aware optimization with reinforcement learning for edge AI offloading, extending existing theories of multi-objective decision making and computational offloading to incorporate energy dynamics and model heterogeneity. The theoretical contribution includes a formalization of an energy-aware offloading equilibrium and a practical blueprint for modular deployment in real-world systems. The main conclusion posits that integrated, adaptive offloading policies grounded in a unified framework can significantly reduce total energy consumption without sacrificing AI accuracy or response times, even under fluctuating network conditions. Recommendations emphasize standardization of energy-aware offloading interfaces, incorporation of these policies into edge-native AI frameworks, and further exploration of privacy-preserving offloading mechanisms and cross-layer optimization for emergent 5G/6G edge ecosystems.