From General Space Planner to Learning-Based Generation: A Systematic Mapping Study of Representation, Objectives, and Control

Abstract: Recent advances in learning-based artificial intelligence (AI) show that shared representations (e.g., pixelized -images and tokenized corpora) enable training multi-task foundation models, whereas digital building layout representations have diverged. Grounded in Eastman’s General Space Planner, this study systematically maps 52 studies over a half-century (1971–2025) to trace how building layout representations evolved with expanding objectives and emerging control mechanisms. Publications surged in the final half-decade (46%), with representations shifting from simplified toward free-formed representations as controlling mechanisms and objectives diversify. The trends suggest a bottleneck, as consistent transfer of learning-based building layout generation is difficult without a shared representation.