AI Explorations
1. Single Objective Optimization (SOO): To compare the performance of optimization algorithms three different optimization tools were used to maximize lux levels: metaheuristics, a DIRECT, and a model-based algorithm. Each was run 3 times. An environmental simulation in Climate Studio was run to achieve maximized mean illuminance results. It was observed that this would result in a reduced building volume.
2. Multi-Objective Optimization (MOO): To establish a financially viable massing, mean illuminance was maximized along with a profit calculation directly related to the way floor area, terrace area and ceiling heights affect market value and cost of construction. RBFMOpt within Opposum as well as SPEA2 & HypE within Octopus were selected as multi-objective optimization tools. Each of these tools was ran 3 times at 700 iterations contributing to a joint pareto front.
3. Un-Supervised Machine-Learning (UML): To gain further understanding on the various features resulting from the MOO, the resulting objectives and parameters were respectively clustered using Un-Supervised Machine Learning (UML). It was observed that the clustering based on objectives would lead to specific and obvious purposes, but based on parameters it would lead to types of features unrelated to the objectives. With the aim to understand the clustering results it was decided to compare parameters with each other with an example, in this case balcony width to the floor-to-floor heights. Based on the features correlation matrix these two parameters had a relatively contrasting relationship to one another, but when analyzing the resultant massing within each of the clusters it is possible to find features that lead to groups that share design characteristics unrelated to the massing.
4. Supervised Machine-Learning (SML): A surrogate model using Supervised Machine Learning was generated with the aim to access non-dominated cases on the pareto front of the MOO study in an expedited manner. This process cut down processing time significantly (around 4000%), but it was observed that the surrogate model could lead to misleading results and has a high inaccuracy in relation to those directly resulting from simulations. It was determined that this technique is not ideal to fulfill the objectives of this project unless there is access to an extensive and expedited initial sampling similar to Hypercube sampling.
5. Generative Deep-Learning (GDL): In the second phase of the project Generative Deep Learning (GDL) was used to produce speculative advertising images and a volumetric distillation of the building mass from the development resulting from the previous phase. 2D images of the building were generated using a style transfer algorithm by applying street art images as style. Additionally, a 3D voxelated model was produced from 2D section images generated with the previous process, which lead to a speculation of the sectional qualities of the interior space.