상세 보기
A deep reinforcement learning-driven method for generating alternative architectural spatial layouts
- Seo, Min Chae;
- Park, Bo Rang;
- Choi, Young Jae;
- Park, Jin Chul;
- Moon, Jin Woo
WEB OF SCIENCE
0SCOPUS
0초록
Recent studies have explored generating spatial layouts to enhance the efficiency and automation of architectural design, with deep reinforcement learning (DRL) emerging as a promising approach to addressing complex, dynamic requirements. This study proposes a DRL-based spatial layout generation method that models each room as an independent agent associated with a specific design objective, aiming to generate spatial layouts that reflect program-specific characteristics. In a residential layout scenario, spatial adjacency, spatial orientation, and energy efficiency are defined as representative design objectives. Each design objective is learned separately and then jointly within an integrated multi-objective setting in a multi-agent framework. After training, the model generates spatial layouts satisfying the given design conditions within 74 s. Compared to layouts generated under single-objective settings, the integrated multi-objective approach produces higher-quality layouts by more effectively complementing individual design objectives and enhancing overall spatial coherence. Furthermore, to evaluate the performance and scalability of the proposed model, generalization tests are conducted under altered environmental conditions. The results demonstrate that the model can adapt to new spatial configurations through additional learning. These findings highlight the potential of the proposed DRL-based framework as a systematic approach for analyzing programmatic design objectives and supporting automated architectural spatial layout generation.
키워드
- 제목
- A deep reinforcement learning-driven method for generating alternative architectural spatial layouts
- 저자
- Seo, Min Chae; Park, Bo Rang; Choi, Young Jae; Park, Jin Chul; Moon, Jin Woo
- 발행일
- 2026-07
- 유형
- Article
- 권
- 130
- 언어
- ENG
- 출판사
- Elsevier Ltd
- 발행국가
- 네덜란드
- ISSN
- E 2352-7102