LLM 다중 에이전트와 혼합정수계획법을 결합한 주거 평면도 자동 생성

초록

Automatic generation of residential floor plans from natural language requires both the semantic interpretation of user requirements and the creation of a structurally valid spatial layout. This paper proposes a system that translates natural language into structured spatial requirements using seven sequential LLM agents. It then mathematically guarantees structural integrity through Mixed-Integer Programming (MIP) optimization and automatically recovers from infeasible states using Minimal Infeasible Subset-based backtracking. In 30 scenarios, based on family types with 2-4 rooms and valid floor areas ranging from 65 to 137m², the proposed method achieved 100% open space connectivity (OC), 1.000 room accessibility (RA), 0.967 adjacency (Adj), and 0.933 prompt compliance (PC). When compared against Co-Layout and Pure LLM approaches, these results represent a 14.4%p improvement in compliance over Co-Layout and 17.0%p over Pure LLM, while also being the most efficient in terms of token usage. Notably, Pure LLM achieved only 43.3% in connectivity, empirically demonstrating the necessity of mathematical optimization for reliably ensuring spatial connectivity.

키워드