Site-Specific Architectural Design of Multi-Story Residential Buildings with Hierarchical Expert Iteration

Abstract

Purpose: Residential floor-plan design is a combinatorial problem without a closed-form solution. Deep generative models mostly produce images or coarse room boxes of single-story apartments without reference to the plot, and general-purpose large language models lack robust geometric reasoning. We investigate how preliminary designs for detached houses of one or more stories can be generated as editable building models on real plots.

Methods: We present MinoSketch, a deployed system for plots in Germany. A large language model elicits requirements in natural language. A Bayesian network learned from 168 captured architectural designs drives a masked sequential sampler that yields consistent room programs. Open geodata on parcels, buildings, terrain, and development plans constrain and inform the design. A hierarchical design engine, structured into options, is trained by Expert Iteration in synthetic self-play against a cost function developed with architects. Metropolis sampling, mixed-integer optimization, and deep reinforcement learning were evaluated and abandoned.

Results: On 60 fixed design tasks, the engine produces valid designs in 97–98% of cases. The share of self-play episodes reaching a valid design rose from about 63% to 95%. Masked sampling eliminated over-filled stories in the room programs.

Conclusions: A search-based, hierarchical, and trainable approach can generate site-specific designs without public training data. The deployed system exports building models in the Industry Foundation Classes format (IFC 4x3) for professional workflows.