Abstract

This study presents a Large Language Model (LLM)-assisted closed-loop network optimization framework that integrates Network Simulator 3 (NS-3) v3.37 with a large language model accessed through the OpenAI Application Programming Interface (API) and a deterministic Python controller. The framework executes simulations, extracts performance metrics, validates LLM-generated candidate configurations, and applies controller-side safeguards before launching the next NS-3 run. The revised Wireless Fidelity (Wi-Fi) experiments use a reconstructed IEEE 802.11 g model and evaluate overloaded User Datagram Protocol (UDP) scenarios with 3 and 10 station nodes. In addition to the initial LLM-guided optimization, the study compares the LLM controller with Bayesian Optimization under the same search space, evaluation budget, simulator, metrics, and utility function. The results are scenario-dependent rather than universally favoring either optimizer. In 3 station nodes (3-STA) validation, Bayesian Optimization achieved the higher utility, 112.587 ± 0.011 versus 89.200 ± 0.012 for the LLM. In 10-STA validation, the LLM-guided configuration achieved 24.996 Mbps aggregate throughput with 0.004% packet loss and 0.424 ms latency, compared with 27.151 Mbps, 9.482% loss, and 180.562 ms for Bayesian Optimization under the same evaluation setting. Reliability analysis showed that controller-side safeguards were essential: 57 API calls produced 36 valid JavaScript Object Notation (JSON) responses, 21 malformed outputs, 29 duplicate proposals, and zero manual interventions.