LLM-enhanced-EA-driven automated optimization framework for joint indoor deployment of multiple BSs and RISs

Abstract

The joint deployment of base stations (BSs) and reconfigurable intelligent surfaces (RISs) is considered a promising but underexplored solution to simultaneously improve indoor coverage performance and reduce network costs. Classical evolutionary algorithms (EAs) can solve this bi-objective non-convex mixed integer programming problem. However, EAs often suffer from efficiency issues due to convergence on local optima, and require redesigning the mechanism to concentrate on the fruitful areas of the search space, especially when the space is large. In this work, we use the large language model (LLM)-enhanced non-dominated sorting genetic algorithm II (NSGA-II) to directly solve the bi-objective optimization problem, assisted by an LLM to enhance its concentrated exploration capability by extending the exploration in language space. To minimize manual engineering, we designed an automated LLM-based module that generates tailored optimization prompts from templates, user descriptions, and reference materials. Simulations show that LLM-NSGA-II, guided by these automated prompts, matches the performance of manual engineering. When applied to joint BS and RIS deployment, LEAO achieves a cost efficiency of 55.7% ± 7.9% across five independent runs, exceeding the strongest baseline by 11.2 percentage points and indicating a more favorable balance between deployment cost and indoor coverage.