Data-Based Generation of Residential Floorplans Using Neural Networks

Most generative design applications used in architectural design are developed with rule-based approaches, based on rules collected from expert knowledge and experience. In other domains, machine learning and, more in particular, neural networks have proven their usefulness and added value in replacing these hard-coded rules or improving applications when combining these two strategies. Since the space allocation problem still remains an open research question and common generative design techniques showed their limitations trying to solve this problem, new techniques need to be explored. In this paper, the application of neural networks to solve the space allocation problem for residential floor plans is tested. This research aims to expose the advantages as well as the difficulties of using neural networks by reviewing existing neural network architectures from different domains and by applying and testing them in this new context using a dataset of residential floor plans.

Automated summary

Most generative design applications in architecture rely on rule-based approaches, which collect rules from expert knowledge and experience. However, in other domains, the use of machine learning and neural networks has proven to be more effective in replacing these hard-coded rules or enhancing applications. This paper explores the application of neural networks to solve the space allocation problem in residential floor plans, reviewing existing architectures from different areas and testing them with a dataset of floor plans.