Associative spatial networks in architectural design: Artificial cognition of space using neural networks with spectral graph theory

Abstract

This paper looks at a new way of incorporating unsupervised neural networks in the design of an architectural system. The approach involves looking the whole lifecycle of a building and its coupling with its environment. It is argued that techniques such as dimensionality reduction are well suited to architectural design problems whereby complex problems are commonplace. An example project is explored, that of a reconfigurable exhibition space where multiple ephemeral exhibitions are housed at any given time. A modified growing neural gas algorithm is employed in order cognize similarities of dynamic spatial arrangements whose nature are not known a priori. By utilising the machine in combination with user feedback, a coupling between the building system and the users of the space is achieved throughout the whole system life cycle.