Distributed Hierarchical Machine Learning for Joint Resource Allocation and Slice Selection in In-Network Edge Systems

TL;DR

This paper introduces a distributed hierarchical deep learning model for efficient resource allocation and slice selection in in-network edge systems, enabling near-optimal decisions with low latency for Metaverse applications.

Contribution

It proposes a novel DeepSets-based distributed hierarchical model trained offline to solve complex resource management problems in edge networks, ensuring permutation equivariance and fast inference.

Findings

Achieves over 95% accuracy in intra- and inter-slice allocation tasks.

Reduces execution time by 86.1% compared to exact solvers.

Maintains system costs within 6.1% of optimal solutions.

Abstract

The Metaverse promises immersive, real-time experiences; however, meeting its stringent latency and resource demands remains a major challenge. Conventional optimization techniques struggle to respond effectively under dynamic edge conditions and high user loads. In this study, we explore a slice-enabled in-network edge architecture that combines computing-in-the-network (COIN) with multi-access edge computing (MEC). In addition, we formulate the joint problem of wireless and computing resource management with optimal slice selection as a mixed-integer nonlinear program (MINLP). Because solving this model online is computationally intensive, we decompose it into three sub-problems (SP1) intra-slice allocation, (SP2) inter-slice allocation, and (SP3) offloading decision and train a distributed hierarchical DeepSets-based model (DeepSets-S) on optimal solutions obtained offline. In the proposed model, we design a slack-aware normalization mechanism for a shared encoder and task-specific decoders, ensuring permutation equivariance over variable-size wireless device (WD) sets. The learned system produces near-optimal allocations with low inference time and maintains permutation equivariance over variable-size device sets. Our experimental results show that DeepSets-S attains high tolerance-based accuracies on SP1/SP2 (Acc1 = 95.26% and 95.67%) and improves multiclass offloading accuracy on SP3 (Acc = 0.7486; binary local/offload Acc = 0.8824). Compared to exact solvers, the proposed approach reduces the execution time by 86.1%, while closely tracking the optimal system cost (within 6.1% in representative regimes). Compared with baseline models, DeepSets-S consistently achieves higher cost ratios and better utilization across COIN/MEC resources.