Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Energy Optimization of Multi-task DNN Inference in MEC-assisted XR Devices: A Lyapunov-Guided Reinforcement Learning Approach

Yanzan Sun, Jiacheng Qiu, Guangjin Pan, Shugong Xu, Shunqing Zhang, Xiaoyun Wang, Shuangfeng Han

TL;DR

This work tackles energy minimization for multi-task DNN inference in MEC-assisted XR devices by formulating a dual time-scale, bi-level optimization that jointly optimizes DNN partitioning and resource allocation. It introduces LyaPPO, a Lyapunov-guided reinforcement learning framework that solves a lower-level drift-plus-penalty resource-allocation problem and an upper-level partitioning problem via PPO, operating over distributed local, transmission, and edge queues. The method yields substantial energy savings across varying local resources, transmit power, and MEC capacity (up to around 46% in some scenarios) while ensuring queue stability. This approach advances XR edge intelligence by enabling dynamic, energy-efficient collaboration between XR devices and MEC servers with provable stability guarantees and scalable optimization.

Abstract

Extended reality (XR), blending virtual and real worlds, is a key application of future networks. While AI advancements enhance XR capabilities, they also impose significant computational and energy challenges on lightweight XR devices. In this paper, we developed a distributed queue model for multi-task DNN inference, addressing issues of resource competition and queue coupling. In response to the challenges posed by the high energy consumption and limited resources of XR devices, we designed a dual time-scale joint optimization strategy for model partitioning and resource allocation, formulated as a bi-level optimization problem. This strategy aims to minimize the total energy consumption of XR devices while ensuring queue stability and adhering to computational and communication resource constraints. To tackle this problem, we devised a Lyapunov-guided Proximal Policy Optimization algorithm, named LyaPPO. Numerical results demonstrate that the LyaPPO algorithm outperforms the baselines, achieving energy conservation of 24.79% to 46.14% under varying resource capacities. Specifically, the proposed algorithm reduces the energy consumption of XR devices by 24.29% to 56.62% compared to baseline algorithms.

Energy Optimization of Multi-task DNN Inference in MEC-assisted XR Devices: A Lyapunov-Guided Reinforcement Learning Approach

TL;DR

This work tackles energy minimization for multi-task DNN inference in MEC-assisted XR devices by formulating a dual time-scale, bi-level optimization that jointly optimizes DNN partitioning and resource allocation. It introduces LyaPPO, a Lyapunov-guided reinforcement learning framework that solves a lower-level drift-plus-penalty resource-allocation problem and an upper-level partitioning problem via PPO, operating over distributed local, transmission, and edge queues. The method yields substantial energy savings across varying local resources, transmit power, and MEC capacity (up to around 46% in some scenarios) while ensuring queue stability. This approach advances XR edge intelligence by enabling dynamic, energy-efficient collaboration between XR devices and MEC servers with provable stability guarantees and scalable optimization.

Abstract

Extended reality (XR), blending virtual and real worlds, is a key application of future networks. While AI advancements enhance XR capabilities, they also impose significant computational and energy challenges on lightweight XR devices. In this paper, we developed a distributed queue model for multi-task DNN inference, addressing issues of resource competition and queue coupling. In response to the challenges posed by the high energy consumption and limited resources of XR devices, we designed a dual time-scale joint optimization strategy for model partitioning and resource allocation, formulated as a bi-level optimization problem. This strategy aims to minimize the total energy consumption of XR devices while ensuring queue stability and adhering to computational and communication resource constraints. To tackle this problem, we devised a Lyapunov-guided Proximal Policy Optimization algorithm, named LyaPPO. Numerical results demonstrate that the LyaPPO algorithm outperforms the baselines, achieving energy conservation of 24.79% to 46.14% under varying resource capacities. Specifically, the proposed algorithm reduces the energy consumption of XR devices by 24.29% to 56.62% compared to baseline algorithms.
Paper Structure (19 sections, 1 theorem, 26 equations, 6 figures, 2 tables, 1 algorithm)

This paper contains 19 sections, 1 theorem, 26 equations, 6 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. System Model and problem formulation
  4. DNN Inference Task Model
  5. Computation and Communication Models
  6. Queue model
  7. Problem Formulation
  8. Lyapunov-guide Proximal Policy Optimization Algorithm
  9. Solution to the lower-level problem
  10. Solution to the upper-level problem
  11. Experiment Evaluations
  12. Experiment Setting
  13. Performance Comparison
  14. Performance Comparison under Different Resources
  15. Impact of the local computational resource
  16. ...and 4 more sections

Key Result

Theorem 1

Consider $\{\tau Q^e_{m,n}\}$, where each element corresponds to a tuple of indices $(m, n)$. Upon sorting this sequence, an ordered sequence $f_1 \ge f_2 \ge \cdots \ge f_{(\sum_{m \in \mathcal{M}}N_m)}$ can be obtained. Given a threshold $K$, if the condition $\sum_{k=1}^{K}f_k < F^e \le \sum_{k=1

Figures (6)

  • Figure 1: MEC-assisted Collaborative Inference Architecture for AI Applications on XR Devices.
  • Figure 3: Comparison of convergence in the training process
  • Figure 4: Comparison of algorithm performance with different numbers of deployed models $N_m$ on the XR device.
  • Figure 5: Comparison of algorithm performance under varying maximum local computational capacities $F_m^l$ of XR device.
  • Figure 6: Comparison of algorithm performance under varying maximum transmit power $p_m^{max}$ of XR device.
  • ...and 1 more figures

Theorems & Definitions (2)

  • Theorem 1
  • proof