To design sustainable architecture, theory encourages architects to rely on automated exploration processes. In practice, the problems encountered are often multicriteria and under constraint. This paper compares different constraint handling strategies, approachable to designer, for processes involving evolutionary algorithms. Four methods are tested on a case study from professional practice. Two methods rely on parametric models: the penalty function method and the use of hyperparameters. The others involve the use of generative techniques: a rule-based method and a repair algorithm that takes the form of an agent-based model. This study highlights the significant impact of the choice of the constraint management method on exploration performance. Among other results, it appears that models involving the use of generative techniques are more efficient than those using parametric models. This calls for the development of dedicated tools.

Il existe différentes méthodes génératives qui permettent une conception performancielle. Les algorithmes génétiques apparaissent comme la solution la plus populaire dans la littérature scientifique. Néanmoins, en pratique, les problèmes rencontrés nécessitent souvent une intégration de contraintes strictes que d’autres techniques génératives fonctionnant avec des règles, comme les automates cellulaires, sont plus aptes à intégrer. Dans cette étude, nous comparons l’efficacité de ces deux techniques appliquées à un même cas d’étude issu de la pratique professionnelle. Il s’agit de la conception d’une enveloppe pour une tour de bureaux à Madagascar qui intègre à la fois le confort thermique et la qualité de vue. Il ressort de cette expérience que l’approche la plus efficace est une méthode hybride qui utilise à la fois un automate cellulaire pour intégrer les contraintes, et un algorithme génétique conduisant l’optimisation multicritère.

The use of genetic algorithms as generative and performance design techniques often involves, in practice, constraint handling, which can be a complex task. Moreover, environmental simulations are computationally expensive and managing constraints can avoid wasting time on infeasible solutions. Despite these two incentives, and the benefits of an immense literature, both applied and theorical, on constrained optimization, there are only few guidelines and tools directly applicable by architects to address this issue. This paper proposes to fill this gap by identifying, classifying, and implementing different constraint management techniques available to architects. Seven methods have been tested for a highly constrained envelope design problem, consisting in the optimization of a sun-shading system. Three of them are easily replicable to different types of projects while the four others need to find a problem-specific heuristic. It appears that the second category is more efficient but implies the use of generative techniques that are more difficult to implement than parametric models.