User-Guided Design Optimization and Search
What algorithmic optimization methods can designers use to efficiently address the dynamic, multi-objective, and explorative nature of architectural design? How should computational optimization tools be adapted to address the specificities of the architectural design problem domain? How can they be made to interact with the designer in ways that educate, engage, and inspire creative imagination?
Dynamic multi-objective evolutionary algorithms (DMOEAs) are an emerging optimization approach for DMOPs, but they have not been explored in the field of architecture. This work addresses this gap in previous research through the development and testing of the first interactive DMOEA-based design tool for the conceptual design phase of architectural design called Design Breeder. Design Breeder’s core optimization algorithm is tested against leading static and dynamic multi-objective evolutionary algorithms. The results demonstrate that Design Breeder provides an advantage over its competitors in discovering diverse solutions while providing a competitive convergence capability.
Design Breeder integrates user preferences in a progressive fashion through the application of a hybrid approach. This is done through a reference-direction approach to guide the search at the global scale, and a solution ranking approach to guide the search at the local scale. In Figure 1 a diagram showing several reference-directions being combined is shown. This hybrid approach works especially well in high dimensional objective spaces where tradeoffs may change drastically at a global scale, but only moderately at a local scale. This interaction engages the decision maker in a manner that requires a level of comparative thought that other approaches involving a reference point, direction, or weighting may not provide. This progressive approach narrows the scope of the search in high-dimensional objective spaces and makes the optimization tractable, while engaging the user’s abilities to help search a complex multi-dimensional space.
Design Breeder also allows users to guide the search process by adjusting the objective and decision spaces dynamically during the search as shown in Figure 2. It includes procedures to help maintain diversity when simultaneous changes occur in the objectives and decision variables, something which no previous research has dealt with. The algorithm addresses the dual problems of convergence and diversity of solutions when changes occur through a combination of memory-based and prediction approaches. When the number of objectives changes, Design Breeder uses a memory-based approach and samples from novelty and Pareto archives to aid diversity and convergence respectively. To deal with changes in the number of decision variables, or time-dependent objective functions, the type and severity of the change is first computed.If the change is small, a prediction-based approach is applied in which already calculated objective values in the Pareto archive are scaled by a computed scale factor, sampled, and used to replace half of the current population.This aids convergence by using values that were known to have performed well previous to the change.The other half of the population is then replaced with random samples from the novelty archive to aid diversity.If the change is big, the Pareto archive is completely emptied and the current population of solutions is repopulated with samples from the novelty archive.When simultaneous changes occur (e.g., when the number of decision variables and objective functions change) these methods are used in combination.
Figure 1. Image showing multiple reference directions being defined to guide an optimization search.
Figure 2. An image of a user changing decision variables and objectives to guide a search process.
Figure 3. (Left)Diagram of a standard Pareto selection-based MOEA. (Right) Diagram of Design Breeder.
Figure 4. Sample of the hydronic facade system used for the dynamic multi-objective optimization tests.
Figure 5. Samples of facade designs from one of the optimization tests.