Combining Generative Adversarial Networks and Reinforcement Learning for Floor Plan Layout Generation

Abstract

Generative Adversarial Networks (GANs) are among artificial intelligence (AI) methods for generating architectural floor plan layouts to approximate spatial distribution with a reasonable degree of accuracy. However, when used exclusively, GAN-based tools may fail to capture architectural patterns and often produce unrealistic layouts. To address this limitation, researchers have proposed integrating Reinforcement Learning (RL) into GANs. While RL has been combined with generative algorithms in other fields, a systematic multi-scenario integration of GANs and RL remains underexplored in architecture. This paper introduces a new solution by combining RL and GANs to generate floor plan layouts tailored to user requirements. The research design involves three different integration strategies: (1a) mere generation, where RL refines GAN outputs by eliminating inconsistencies and errors; (1b) objective optimization, where RL targets measurable attributes such as spatial size and morphological legibility; and (1c) refinement of non-quantifiable attributes, where RL incorporates user feedback to improve flexibility and perceived comfort. Additionally, the study employs House-GAN++ as the GAN model and the PPO algorithm as the RL framework. Three case studies are presented to evaluate performance. Results demonstrate that integrating RL with GANs yields floor plan layouts more responsive to user needs than those produced by GANs alone. Each scenario illustrates how RL optimizes GAN-generated outputs according to functional, measurable, and perceptual goals. The methodology acknowledges user expectations and translates them into realistic, adaptable plans. Key outcomes include more realistic layouts, designs with distinctive characteristics, and user-customized floor plans created through interaction. The proposed framework enables automatic floor plan generation that combines design, optimization, and user input at the conceptual stage. This integration enhances architectural design processes by balancing computational efficiency with user-oriented adaptability, thus broadening the potential of AI-assisted design. © 2025, Education and research in Computer Aided Architectural Design in Europe. All rights reserved.