Achieving architectural logic in automated floor plan design: A multi-stage double DQN framework from topological graph synthesis to rectangular layout generation

Automatic floor plan generation has gained increasing attention in generative architectural design. However, existing studies often focus on geometric aesthetics or data-driven similarity, lacking explicit representation of spatial logic. This research proposes a multi-stage automated framework based on graph theory and reinforcement learning to model architectural reasoning. The method consists of two stages: spatial topology generation and geometric layout generation. First, floor plans are represented as graphs, where nodes denote functional spaces and edges represent adjacency. A Double Deep Q-Network (Double DQN) is employed to sequentially construct spatial topology, guided by a multi-objective reward function incorporating functional integrity, spatial connectivity, and topological coherence. A hierarchical curriculum learning strategy improves training stability and scalability. In the second stage, the generated topology is translated into rectangular layouts based on spatial constraints. Results demonstrate that the framework produces coherent and logically consistent floor plans without reliance on large datasets. Ablation studies confirm the effectiveness of curriculum learning, reward design, and network structure. This study highlights reinforcement learning as a tool for design reasoning, offering a new paradigm for explainable and intelligent architectural design.