Graph-Based Autoregressive Machine Learning for High Performance Residential Layout Generation

Abstract

The design of residential layouts extends beyond the distribution of space; it also requires the carefulcoordination of spatial organization and environmental performance. However, most machine-learning-based floor-plan generation methods treat performance as a post-generation evaluation criterion, rather than as an input condition that guides the generation process itself. This thesis develops a prediction-based generative framework that conditions residential layout generation on environmental-performance targets during the generation process. The proposed framework operates on structured geometric data rather than images. It first predicts room-area shares and then assigns rooms to spatial zones informed by environmental performance metrics using a conditional graph-based autoregressive model. A separate connectivity model predicts room adja- cencies, while a schematic synthesis stage assembles candidate layouts from the predicted areas, zones, and connections. The generated layouts are evaluated as a population, and non-dominated alternatives are se- lected along a Pareto front, preserving competing performance objectives as explicit design trade-offs rather than collapsing them into a single optimized outcome. This study uses two physically coupled environmental contexts, Quality View and acoustic performance, as representative performance conditions. Because both are affected by spatial arrangement and the window- to-wall ratio, two study-specific environmental performance indicators—the Quality View Coverage Index (QVCI) and the Acoustic Quality Index (AQI)—were developed and incorporated into the machine-learning training process. By demonstrating that these coupled performance conditions can guide layout generation, the framework establishes a basis for extending the same approach to additional environmental contexts in future work. The performance-conditioned spatial zoning model achieved approximately (86%) accuracy in predicting room-zone assignments, with a cross-entropy loss of (0.20). More importantly, it demonstrated strong target controllability, with target prediction accuracy reaching approximately (94%) and a root-mean-square error of (0.06) between requested and achieved performance values. Sweeping the conditioning targets produced large and correctly directed changes in the achieved metrics, including directional gaps of (+0.71) and (+0.54) for the two environmental performance axes and (+0.66) for the zone-spread target. The slider-responsiveness results further show that the model can steer both environmental performance and the spatial distribution of rooms across the plan. These results demonstrate that the proposed pipeline can generate diverse and plausible residential lay- outs that respond to user-defined performance targets. Rather than producing a layout first and evaluating it afterward, the framework generates candidate layouts already shaped by performance trade-offs. Envi- ronmental performance is therefore treated not as a post-generation check, but as a condition that guides generation itself.