Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DNN Partitioning, Task Offloading, and Resource Allocation in Dynamic Vehicular Networks: A Lyapunov-Guided Diffusion-Based Reinforcement Learning Approach

Zhang Liu, Hongyang Du, Junzhe Lin, Zhibin Gao, Lianfen Huang, Seyyedali Hosseinalipour, Dusit Niyato

TL;DR

The paper addresses dynamic, long-horizon DNN task execution in Vehicular Edge Computing by jointly partitioning DNNs, offloading intermediate data, and allocating edge resources while guaranteeing system stability. It introduces a Lyapunov-based decoupling that turns a challenging MINLP into per-slot deterministic problems, and proposes MAD2RL, a diffusion-model–driven multi-agent DRL framework built on the QMIX architecture with a convex optimization subroutine for resource allocation. The approach leverages diffusion models to efficiently generate partitioning and offloading decisions across many mobile vehicles and edge nodes, yielding improved convergence, queue stability, and task completion time over several benchmarks on real-world traces. The work demonstrates practical impact by enabling low-latency, stable DNN inference in dynamic VEC scenarios, with potential extensions to multi-RSU coordination and offline expert data integration.

Abstract

The rapid advancement of Artificial Intelligence (AI) has introduced Deep Neural Network (DNN)-based tasks to the ecosystem of vehicular networks. These tasks are often computation-intensive, requiring substantial computation resources, which are beyond the capability of a single vehicle. To address this challenge, Vehicular Edge Computing (VEC) has emerged as a solution, offering computing services for DNN-based tasks through resource pooling via Vehicle-to-Vehicle/Infrastructure (V2V/V2I) communications. In this paper, we formulate the problem of joint DNN partitioning, task offloading, and resource allocation in VEC as a dynamic long-term optimization. Our objective is to minimize the DNN-based task completion time while guaranteeing the system stability over time. To this end, we first leverage a Lyapunov optimization technique to decouple the original long-term optimization with stability constraints into a per-slot deterministic problem. Afterwards, we propose a Multi-Agent Diffusion-based Deep Reinforcement Learning (MAD2RL) algorithm, incorporating the innovative use of diffusion models to determine the optimal DNN partitioning and task offloading decisions. Furthermore, we integrate convex optimization techniques into MAD2RL as a subroutine to allocate computation resources, enhancing the learning efficiency. Through simulations under real-world movement traces of vehicles, we demonstrate the superior performance of our proposed algorithm compared to existing benchmark solutions.

DNN Partitioning, Task Offloading, and Resource Allocation in Dynamic Vehicular Networks: A Lyapunov-Guided Diffusion-Based Reinforcement Learning Approach

TL;DR

The paper addresses dynamic, long-horizon DNN task execution in Vehicular Edge Computing by jointly partitioning DNNs, offloading intermediate data, and allocating edge resources while guaranteeing system stability. It introduces a Lyapunov-based decoupling that turns a challenging MINLP into per-slot deterministic problems, and proposes MAD2RL, a diffusion-model–driven multi-agent DRL framework built on the QMIX architecture with a convex optimization subroutine for resource allocation. The approach leverages diffusion models to efficiently generate partitioning and offloading decisions across many mobile vehicles and edge nodes, yielding improved convergence, queue stability, and task completion time over several benchmarks on real-world traces. The work demonstrates practical impact by enabling low-latency, stable DNN inference in dynamic VEC scenarios, with potential extensions to multi-RSU coordination and offline expert data integration.

Abstract

The rapid advancement of Artificial Intelligence (AI) has introduced Deep Neural Network (DNN)-based tasks to the ecosystem of vehicular networks. These tasks are often computation-intensive, requiring substantial computation resources, which are beyond the capability of a single vehicle. To address this challenge, Vehicular Edge Computing (VEC) has emerged as a solution, offering computing services for DNN-based tasks through resource pooling via Vehicle-to-Vehicle/Infrastructure (V2V/V2I) communications. In this paper, we formulate the problem of joint DNN partitioning, task offloading, and resource allocation in VEC as a dynamic long-term optimization. Our objective is to minimize the DNN-based task completion time while guaranteeing the system stability over time. To this end, we first leverage a Lyapunov optimization technique to decouple the original long-term optimization with stability constraints into a per-slot deterministic problem. Afterwards, we propose a Multi-Agent Diffusion-based Deep Reinforcement Learning (MAD2RL) algorithm, incorporating the innovative use of diffusion models to determine the optimal DNN partitioning and task offloading decisions. Furthermore, we integrate convex optimization techniques into MAD2RL as a subroutine to allocate computation resources, enhancing the learning efficiency. Through simulations under real-world movement traces of vehicles, we demonstrate the superior performance of our proposed algorithm compared to existing benchmark solutions.
Paper Structure (55 sections, 1 theorem, 54 equations, 9 figures, 1 table, 1 algorithm)

This paper contains 55 sections, 1 theorem, 54 equations, 9 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background Research
  3. Motivation and Main Challenges
  4. Summary of Contributions
  5. Paper Organization
  6. Related Works
  7. DNN Inference and Edge Computing
  8. DRL for Resource Management
  9. Lyapunov Optimization for System Stability
  10. Footprints of Diffusion Models in Optimization
  11. System Model
  12. AI-Powered Vehicular Network Model
  13. DNN Partitioning Model
  14. Communication Model
  15. V2I Communications
  16. ...and 40 more sections

Key Result

Lemma 1

For any feasible set of $\{\bm{\varphi}(t),\bm{\xi}(t),\bm{F}(t)\}_{t\in\mathcal{T}}$, which satisfies constraints eq:Qloc_constraint to eq:maxfk_constraint, the Lyapunov drift-plus-penalty function $\Lambda(\bm{Q}(t))$ can be upper bounded as follows: where $d_i(t)$ is given in eq:total_delay, and $\chi$ is a constant given in eq:constant_value, which is shown at the top of the previous page.

Figures (9)

  • Figure 1: VGG16 layer-wise runtime and output data size.
  • Figure 2: A schematic of VEC-assisted DNN-based task partitioning and offloading.
  • Figure 3: Illustration of the diffusion model tailored for generating the optimal decisions of DNN partitioning and task offloading for CV $i$ at time slot $t$.
  • Figure 4: The overall architecture of the MAD2RL algorithm.
  • Figure 5: Vehicular network visualization.
  • ...and 4 more figures

Theorems & Definitions (3)

  • Remark 1
  • Lemma 1
  • proof