Quantifying the Distance of Residential Plans from the Spatial Logic of the Iranian House Using a Graph Neural Network

Quantifying the Distance of Residential Plans from the Spatial Logic of the Iranian House Using a Graph Neural Network

Negar Shahiddokht1

1) Ph.D. in Artificial Intelligence, Amirkabir University of Technology, Tehran, Iran

IOI Code : ABGG-ABKFCA
Publication : 11th.International Conference on Researches in Science & Engineering & 8th.International Congress on Civil, Architecture and Urbanism in Asia(11icrsie.ir)
Abstract :
    Architectural judgements about the relationship between contemporary housing and the spatial logic of the Iranian house have been expressed largely in qualitative terms — claims that, lacking any measurable criterion, can be neither refuted nor confirmed, and that play no operational role in design. This paper attempts to translate part of that judgement into quantity. Five configurational rules documented in the space syntax literature for the introverted Iranian house are first recast as computable descriptors on the access graph. A critic based on a graph neural network with attention-based aggregation in the readout layer is then introduced; it takes a plan as a graph and yields a continuous index termed typological fidelity. To make the claim testable, a synthetic corpus of five thousand plans was generated with a shape grammar in which a rupture parameter varies in a controlled manner from the courtyard configuration to the apartment configuration. Models without topological information recover only part of this parameter, and the graph network reduces their error by roughly twenty-three percent — a difference significant at below one in ten thousand. To test the objection of circularity, the model was applied without retraining to a corpus generated by a second, independent grammar with graphs forty percent larger; the ordering of the three typological bands was preserved, and pairwise ranking accuracy was about eighty-three percent, rising to ninety-three percent for pairs separated by more than two tenths. The attention mechanism, without significantly altering accuracy, enables node-level detection: in ninety-seven percent of cases the rule-violating space is ranked first. Finally, placing the critic in the selection loop of a simple generator improves the syntactic quality of its output on all five rules. The study is a proof of concept and does not rely on field survey data.

Keywords :