Resource allocation for efficient AI inference in wireless sensing edge networks

Integrating AI inference into wireless sensing edge networks presents notable challenges due to limited resources, changing environments, and diverse devices. In this study, we proposed a novel resource allocation framework that enhances energy efficiency, reduces latency, and ensures fairness across distributed edge nodes for AI inference. The framework models a multi-objective optimization problem that reflects the interdependence of computation, communication, and energy at each device. We also develop a decentralized algorithm based on dual decomposition and projected gradient ascent, by using local data. The extensive simulations demonstrate that our proposed method reduces the average inference latency by 31.4% and energy consumption by 27.8% compared to the greedy and round-robin techniques. The system utility is improved by up to 59.2%, and fairness, measured using Jain’s index, remains within 8% of the ideal. Additionally, throughput analysis further confirms that our approach gains up to 49 tasks/sec, outperforming existing strategies by more than 40%. These findings show that the resource-aware AI inference approach is scalable, energy-efficient, and appropriate for real-time use in multi-user wireless edge networks. • Proposes a decentralized resource allocation framework for AI inference in wireless sensing edge networks. • Jointly optimizes latency, energy consumption, and fairness under dynamic workloads. • Employs a dual decomposition-based algorithm with minimal control signaling for scalability. • Achieves up to 31.4% lower latency and 27.8% lower energy use compared to greedy and round-robin schemes. • Maintains Jain’s fairness index above 0.95 while improving system utility by up to 59.2.