Artificial intelligence–based dependent task offloading in next-generation edge networks

Abstract

The exponential growth of the Internet of Things (IoT) and the stringent latency requirements of Fifth Generation (5G) and future Sixth-Generation (6G) networks are driving the rapid evolution of Multi-Access Edge Computing (MEC). While MEC brings computational resources closer to end users to alleviate the burden on core networks, modern mobile applications such as Augmented Reality (AR), real-time video analytics, and autonomous driving have evolved from simple, independent requests into complex workflows modeled as Directed Acyclic Graphs (DAGs) with strict inter task dependencies. Traditional optimization techniques and heuristic methods often fail to address the high dimensionality, topological complexity, and dynamic nature of these heterogeneous environments. Furthermore, existing learning based solutions frequently neglect the structural dependencies of tasks or fail to account for the statistical heterogeneity inherent in distributed edge networks. To overcome these limitations, this thesis leverages advanced Artificial Intelligence (AI) techniques, including Graph Neural Networks (GNNs), Deep Reinforcement Learning (DRL), and Federated Learning (FL), to enable intelligent and robust resource orchestration. In this context, the present thesis investigates the joint optimization of dependent task offloading, resource allocation, and mobility management to enhance system efficiency, reliability, and privacy. Specifically, we propose three AI-driven frameworks to address the complexities of DAG based applications under realistic network constraints: First we present an Energy-Aware Multi-user Dependent Task Offloading and Resource Allocation (EMDTORA) framework, which integrates Graph Attention Networks (GAT) with the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. This way, the system explicitly learns the structural embeddings of application DAGs. This allows the agent to identify efficient execution the sequence of dependent tasks that strictly bounds the application’s completion time and prioritize their execution. Simulation results demonstrate that EMDTORA significantly outperforms standard baselines in reducing the weighted Energy Time Cost (ETC). Following, we address the critical challenges of data privacy and statistical heterogeneity in distributed networks. Specifically, we introduce a Federated ensemble reinforcement learning (FEDORA) framework for directed acyclic graph (DAG)-based task Offloading and resource allocation in MEC. Since, conventional centralized training poses severe privacy risks and incurs high communication overheads, FEDORA employs a decentralized training architecture where edge nodes train local models and share only gradient updates. To mitigate client drift caused by diverse user behaviors, we utilize the FedProx aggregation algorithm combined with a novel Ensemble DRL architecture that decouples discrete offloading decisions from continuous resource allocation. This approach ensures robust convergence and generalization across heterogeneous environments without requiring raw data exchange. Finally, we tackle the stochastic nature of high speed mobility in Vehicular Edge Computing (VEC) by proposing a Dependent Task Offloading in Vehicular Edge Computing Using Trajectory-Aware Deep Reinforcement Learning and incorporate a mobility-aware scheme which traditionally often leads to high task failure rates for delay sensitive applications. By leveraging Transformer-based sequence modeling to accurately predict Roadside Unit (RSU) connectivity windows, this framework enables the DRL agent to proactively manage dependent task execution. This predictive capability ensures that tasks are only offloaded when reliable completion is guaranteed, thereby outperforms traditional methods in terms of delay, and energy.

Date	30 Apr 2026
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Aris Leivadeas (Supervisor) & Julien Gascon-Samson (Co-supervisor)