Floor Plan Design Collaborator: A Data-Driven Approach to Assist Human Architects in Design Exploration

Abstract

After a long AI winter since the 1980s, artificial intelligence is now experiencing a renaissance due to enhanced computing power and access to vast amounts of data. Today, machines can talk, sing, and draw like human experts. Despite this progress, we are still far from the vision where human designers and AI collaboratively discuss and develop designs. This study argues that a data-driven approach holds great potential in the design process by quickly learning from existing examples and generating new alternatives for exploration. To support this claim, the study presents a generative framework that learns from existing examples and generates new designs. Specifically, the proposed framework employs Bayesian networks to encode site layout data and floor plan examples, generating new design examples through a Markov Chain Monte Carlo (MCMC) sampling procedure. Experiments on real-world examples demonstrate that the framework effectively summarizes the statistical information of given design examples and generates unseen examples based on the learned knowledge. The transparency of the data representation and the inner workings of the proposed framework facilitate an active feedback loop in the iterative learning and generation process between human designers and machines. Observations throughout the study reveal intrinsic limitations and potential improvements of contemporary optimization-based approaches from the perspective of both lateral and vertical design development.