Black-Box Optimisation Methods for Architectural Design

Abstract

Black-box optimization methods play an important role in automated design space exploration, but to-date have not been systematically compared on problems from architectural design optimization. This paper presents a quantitative comparison of the three major types of black-box optimization: metaheuristics, direct search, and model-based methods. We compare the performance of one representative algorithm of each type (including a genetic algorithm) on four performance based design problems of varying complexity and characteristics. Our results show that metaheuristics are greatly outperformed whenever evaluating tens of thousands of design candidates is not an option, and suggest direct search and model-based methods as viable and more efficient alternatives.