Evolutionary trends in the use of artificial intelligence in support of architectural design
From Firenze University Press Journal: TECHNE
Gian Luca Brunetti, Dipartimento di Architettura e Studi Urbani, Politecnico di Milano
The phase of expert systems
In the last forty years, a revolution has taken place in the field of artificial intelligence for architectural design. From the 1960s to the 1980s, the prevalent line of attack to the problem of assisting design by means of artificial intelligence techniques was characterised by the prevalence of so-called ex-pert systems, based on the application of instructions of the kind if-then-else, and on knowledge acquisition through the input of experts (Newell, 1982), in a perspective of problem-solving structured in design option spaces (Newell and Simon 1972)1. But after a period of euphoria about their possibilities in the 1980s, it became clear that those systems had critical weaknesses, mainly deriving from their difficulty in taking into account the complex, numerous and often implicit contextual conditions of problems that are at play in architectural design, and that are resolved by human beings mainly on the basis of common sense — the most challenging aspect to pursue through the application of inductive/deductive logic. It is no coincidence that the most interesting results of expert systems in the construction sector back then were to be found in applications characterised by substantial pattern-matching requirements — like building diagnoses, or conformity verifications — rather than generative tasks (Rosenman et al., 1986; Schwarz et al., 1994).What makes architectural design problems difficult to approach with knowledge-based expert systems is, on the one hand, the fact that in them, the objects of decisions are not only the ways to solve problems, but also the design objectives themselves, which are often implicit and nuanced (Dorst and Cross, 2001). But, these premises did not prevent the overall balance of expert systems from being fairly positive overall, as many of the techniques on which they are founded (such as object-oriented programming, constraint propagation, integration of relational databases) have been capillarily integrated into the CAD systems in use to day.
From stochastic methods to probabilistic approaches
The emergence of the limits of knowledge-based expert systems coincided, on the one hand, with the growth of metaheuristic approaches founded on stochastic processes (among the most used of which there are evolutionary algorithms — Goldberg 2002 –, support vector machines — Cortes and machines Vapnik, 1995 –, and the nearest neigh-bour methods — Brunetti, 2020), and on the other hand, with probabilistic, Bayesian methods (Pearl, 1988), which are indeed strong in that which rule-based systems are weak: precisely, the ability to deal with uncertainties and ambiguities that are typical of reality. Today metaheuristic methods are mainly used for optimisation, and would be potentially suited to cover a substantial part of the design process, but they are, above all, utilised for almost nothing more than fine-tuning already defined architectural solutions. This is due to the technical difficulty of applying them to problems characterised by great degrees of freedom, like the ones typical of preliminary architectural design (Brunetti, 2016). The most important among those methods have proven to be genetic algorithms, which today are often utilised both in the energy (Gan et al., 2019) and the structural fields (Boonstra et al. 2020), and are also present in the form-find-ing arena (Boonstra et al., 2021).The golden age of probabilistic approaches culminated around 2010, generating numerous design support systems in the field of architecture (Sokol et al., 2017), which have not, however, wiped out rule-based ap-proaches nor metaheuristic methods, but have coexisted and hybridised with them. The result of such hybridisation is the generation of composite design support systems, robust and adapted to the needs of modernity. They are ubiquitous today, from the field of optimisation (Wu and Wang, 2020) to design generation (Liu and Wu, 2015), sometimes through the use of shape grammars (Wang and Zhang, 2020).
DOI: https://doi.org/10.36253/techne-13739
Read Full Text: https://oaj.fupress.net/index.php/techne/article/view/13739
