Leveraging LLMs to Close Simulation-to-Reality Gap in Wireless Mesh Network Configuration
Conference
Rodriguez, JT, Sadekeen, D, Sha, M et al. (2026). Leveraging LLMs to Close Simulation-to-Reality Gap in Wireless Mesh Network Configuration
. 129-136. 10.1109/DCOSS-IoT69657.2026.00023
Rodriguez, JT, Sadekeen, D, Sha, M et al. (2026). Leveraging LLMs to Close Simulation-to-Reality Gap in Wireless Mesh Network Configuration
. 129-136. 10.1109/DCOSS-IoT69657.2026.00023
Recent years have witnessed the rapid adoption of wireless mesh networks (WMNs) across various fields, including industrial automation, smart energy, and smart cities. Although WMNs usually perform well thanks to years of research, configuring these networks remains a challenging task. Simulations provide distinct advantages over physical deployment for network configuration, as they can be run more quickly, with much less overhead, and allow testing different configurations under controlled, identical conditions. However, recent studies have shown that the network configuration chosen in simulations may not perform well in real-world deployments due to the simulation-to-reality gap. Unfortunately, all existing solutions require collecting data from the physical deployment to effectively narrow the simulation-to-reality gap, a costly and labor-intensive process. In this paper, we present LLM-OPT, the first solution that produces high-quality network configuration models for WMNs without requiring data from physical deployments. LLM-OPT employs feature masking and textual serialization techniques to preprocess the simulation data and leverages LLM fine-tuning to generate network configuration models. Experimental results show that the network configuration model generated by LLM-OPT achieves 82% prediction accuracy and significantly outperforms all existing solutions (up to 34.2%).
