Controllable and flexible residential floor plan layout design based on multi-agent deep reinforcement learning with layout prior size and similar experience abandon

Abstract

Automated floor plan generation can significantly enhance the efficiency of designers and reduce associated design costs. Nonetheless, ensuring the model’s controllability and flexibility is crucial for its practical applications, presenting distinct challenges for current methods. This paper presents an automated residential layout design method utilizing multi-agent deep reinforcement learning (MADRL) to address the challenges of spatial variability and customization requirements in architectural layouts. By simulating the collaborative design process through multiple agents, the proposed method effectively accommodates diverse layout environments while ensuring valid and personalized designs. A refined reward system was developed to guide agents in generating rational room arrangements and meeting different custom constraints. Additionally, layout prior size (LPS) was proposed to address the size selection challenge, effectively reducing the action space and enhancing layout quality. To further improve diversity, a similar experience abandon (SEA) mechanism was proposed, allowing efficient experience interaction among agents and eliminating redundant exploration of similar layouts. Experimental results demonstrate the proposed method’s capability to generate valid floor plans and provide diversified layout options under various design inputs and custom tasks. Meanwhile, the agent achieves a design consistent with the real layout in 1187 episodes, demonstrating the method’s compatibility. This paper highlights the potential of MADRL in advancing the automation and efficiency of architectural layout design, offering a novel solution for the integration of flexibility and controllability in residential planning.