Model Quantization for Resource-Efficient DNN Implementation of MAC Protocols in Industrial IoT

Artificial Intelligence (AI) is becoming a key enabler in Industrial Internet of Things (IIoT) applications, supporting intelligent functionalities such as predictive maintenance, energy optimization, and adaptive communication. However, deploying Deep Neural Networks (DNNs) on resource-constrained IoT devices (IoTDs) remains a significant challenge due to limited computational resources, memory, and energy availability. In this paper, we investigate the effectiveness of model quantization techniques in reducing the computational and memory demands of DNNs, with a focus on the implementation of MAC protocols learned via Multi-Agent Deep Reinforcement Learning (MADRL). Specifically, we evaluate Post-Training Quantization (PTQ), static and dynamic, targeting inference on general-purpose CPUs embedded in Microcontroller Units (MCUs). We assess the performance of these techniques based on deployment metrics (disk usage, runtime memory, and inference time) and communication metrics (e.g., throughput). Experimental results demonstrate that quantization significantly improves deployability on constrained hardware while maintaining good levels of communication performance, enabling efficient integration of AI-based MAC protocols in HoT networks.