[4M5-GS-2f] Machine learning

4:15 PM - 4:30 PM JST(7:15 AM - 7:30 AM UTC)

[4M5-GS-2f-04] LCaMO: A Role-Separated GA–LLM Framework for Weapon Balance Optimization in FPS Games

〇Manato Ishihara¹, Ryota Nakamura¹ (1. Musashino University)

Keywords:

Multi-objective Optimization,Large Language Model,Explainability,Genetic Algorithm,Hypothesis-Driven Search

Weapon balance adjustment in competitive FPS games remains challenging because evolutionary search processes are hard to explain. This study proposes LCaMO, a hypothesis-driven hybrid optimization method combining a genetic algorithm (GA) and a large language model (LLM). The method has three phases: (1) multi-objective search with NSGA-II, (2) LLM-based causal hypothesis generation, and (3) statistical revalidation. To reduce hallucinations, the LLM is limited to hypothesis generation; numeric parameters are chosen from a predefined intervention catalog. Experiments on five assault rifles in an Apex Legends Season 27-like simulation show that LCaMO reaches full convergence (5/5 weapons within 45–55%, range 48.4–52.7%), reduces win-rate MSE by 96.8% from the default, and is 2.1× faster than a GA-only baseline (50 generations; 1/5 convergence).