Generative representations for artificial architecture and passive solar performance

This paper explores how the use of generative representations influences the quality of solutions in evolutionary design problems. A genetic programming system is developed with individuals encoded as generative representations. Two research goals motivate this work. One goal is to examine Hornby´s features and measures of modularity, reuse and hierarchy in new and more complex evolutionary design problems. In particular, we consider a more difficult problem domain where the generated 3D models are no longer constrained by voxels. Experiments are carried out to generate 3D models which grow towards a set of target points. The results show that the generative representations with the three features of modularity, regularity and hierarchy performed best overall. Although the measures of these features were largely consistent with those of Hornby, a few differences were found. Our second research goal is to use the best performing encoding on some 3D modeling problems that involve passive solar performance criteria. Here, the system is challenged with generating forms that optimize exposure to the Sun. This is complicated by the fact that a model´s structure can interfere with solar exposure to itself; for example, protrusions can block Sun exposure to other model elements. Furthermore, external environmental factors (geographic location, time of the day, time of the year, other buildings in the proximity) may also be relevant. Experimental results were successful, and the system was shown to scale well to the architectural problems studied.