Explainable AI-Based Feature Selection Method for Energy Harvesting-Assisted IoT Network
Year: 2026
Type: conference-paper
Abstract: Energy Harvesting (EH)-assisted Backscatter IoT networks offer a promising solution for sustainable connectivity but face complex optimization challenges. While Deep Reinforcement Learning (DRL) provides a viable model-free solution, its black-box nature often obscures the underlying decision mechanisms, leading to redundant state representations and inefficient resource usage. To address these challenges, we propose XAI-LyDRL, a novel framework that integrates Lyapunov optimization with explainable artificial intelligence (XAI)-based feature selection. We derive three key mechanistic insights: (1) we empirically verify that the agent implicitly internalizes physical path-loss models, rendering explicit geometric distance features redundant; (2) we unveil a regime-dependent attention mechanism where the policy dynamically shifts focus from queue stability to channel quality based on resource constraints, confirming theoretical consistency with Lyapunov optimization; and (3) we identify effective state sparsity where optimal actions are primarily driven by bottleneck users. Guided by these insights, we construct a streamlined policy that prunes 40% of the state space. Simulation results demonstrate that this lightweight design achieves comparable stability and throughput to full-scale models while providing an empirical foundation for scalable deployment in resource-constrained edge devices.
Language: English
Topic: Internet of Things and AI
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204084634