Learning-Based Collaborative Service Placement for AI-oriented Mobile Edge Computing

Authors:

Abstract
Next generation mobile communication technology (6G) intelligent endogenous network deeply integrates network connectivity with artificial intelligence (AI) elements. It is urgent to push AI services to the edge of the network. How to deploy computing intensive AI models in edge devices with limited resources, and consider the collaborative training and inference between edge-edge or edge-cloud is a challenge. The partition of complex edge AI models into small subtasks and the collaborative executions among the edge nodes can efficiently solve this problem. Therefore, in this paper, exploiting the dispersive resource across edge nodes and devices, we propose a collaborative service placement strategy efficiently deploying all the sub-services among edge nodes and considering the service migration due to the user’s mobility. Moreover, we will explore the trade-offs between different metrics, i.e., accuracy, latency and migration cost, helping to improve the efficiency of deploying AI application at edge. To overcome the unknown system dynamics, we formulate the problem as a combinatorial contextual Multi-armed Bandit (MAB) problem. Then based on MAB theory and greedy method, we give a online learning algorithm to cope with the system dynamics. Theoretical analysis and evaluations demonstrate the proposed algorithm can effectively learn and adapt to the dynamic network environment and user demands, reduce the latency and migration cost while satisfying the minimal service requirement.

This publication has 11 references indexed in Scilit: