Resource Allocation in Hybrid Radio-Optical IoT Networks using GNN with Multi-task Learning

TL;DR

This work addresses resource allocation in hybrid RF-OWC IoT networks by casting scheduling as a bi-objective MINLP (throughput maximization vs delay minimization) and proposing a scalable data-driven framework, DGET, that blends a transductive GNN with an inductive GNN and a Transformer classifier to predict link activations and technologies under partial observability. DGET trains with ground-truth optimal schedules and a consistency loss to align evolving embeddings with recorded network states, achieving near-optimal performance with over 90% classification accuracy and substantially reduced inference complexity compared to solving the MINLP. Experimental results show that hybrid RF-OWC networks outperform RF-only configurations at higher loads, delivering up to 20% AoI reduction while maintaining comparable energy consumption, and demonstrating robustness to outdated channel information. The framework offers practical implications for scalable dual-band IoT deployments and suggests future work on mobility, real-time feedback, hardware validation, and energy-harvesting-aware policies for sustainable operation.

Abstract

This paper addresses the problem of dual-technology scheduling in hybrid Internet of Things (IoT) networks that integrate Optical Wireless Communication (OWC) alongside Radio Frequency (RF). We begin by formulating a Mixed-Integer Nonlinear Programming (MINLP) model that jointly considers throughput maximization and delay minimization between access points and IoT nodes under energy and link availability constraints. However, given the intractability of solving such NP-hard problems at scale and the impractical assumption of full channel observability, we propose the Dual-Graph Embedding with Transformer (DGET) framework, a supervised multi-task learning architecture combining a two-stage Graph Neural Networks (GNNs) with a Transformer-based encoder. The first stage employs a transductive GNN that encodes the known graph topology and initial node and link states. The second stage introduces an inductive GNN for temporal refinement, which learns to generalize these embeddings to the evolved states of the same network, capturing changes in energy and queue dynamics over time, by aligning them with ground-truth scheduling decisions through a consistency loss. These enriched embeddings are then processed by a classifier for the communication links with a Transformer encoder that captures cross-link dependencies through multi-head self-attention via classification loss. Simulation results show that hybrid RF-OWC networks outperform standalone RF systems by handling higher traffic loads more efficiently and reducing the Age of Information (AoI) by up to 20%, all while maintaining comparable energy consumption. The proposed DGET framework, compared to traditional optimization-based methods, achieves near-optimal scheduling with over 90% classification accuracy, reduces computational complexity, and demonstrates higher robustness under partial channel observability.

Figures (17)

• Figure 1: Illustration of a hybrid RF and OWC deployment in a hospital environment. RF ensures broad coverage and mobility for staff and patient monitoring, while OWC provides secure, interference-free communication for critical areas such as operating rooms and diagnostic imagingKatz2024.
• Figure 2: A visualization of the RF-OWC model showing a hybrid RF-OWC network exchanging different types of applications' data.
• Figure 3: Illustration of the communication model and the lifespan of messages.
• Figure 4: An example of a graph modeling for a network with $N$ devices over a period of time $\mathcal{T}$. $E_{t_{i-j}}$ represents the allowed communication links from device $i$ to device $j$. $E_{c_{i-j}}$, on the other hand, represents the chosen communication technology for that link. For illustration, we have omitted some of the communication links that cannot be established from a device $i$ to a device $j$ to showcase the environment visibility (i.e., $E_{t_{i-j}}=0$ ). The edges in yellow/blue represent the chosen links and their value (e.g., $E_{c_{2-3}}=2$ refers to OWC) at each time $k \in \mathcal{T}$.
• Figure 5: Visualization of device-link transductive embeddings from the network graph, resulting in a numerical representation for each device in the graph. Each temporal input graph $\mathcal{G}^{\text{I}}_{k}(\mathcal{V},\mathcal{E})$ is transformed into embeddings with a size $|\mathcal{N}| \times Z$, with $Z$ the embedding dimension.