Anzeige
Erschienen in:
2019 | OriginalPaper | Buchkapitel
Interactive Design Support for Architecture Projects During Early Phases Based on Recurrent Neural Networks
verfasst von : Johannes Bayer, Syed Saqib Bukhari, Andreas Dengel
Erschienen in: Pattern Recognition Applications and Methods
Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by
Abstract
In the beginning of an architectural project, abstract design decisions have to be made according to the purpose of the later building. Based on these decisions, a rough floor plan layout is drafted (and subsequently redrafted in successively more refined versions). This entire process can be considered an iterative design algorithm, in which high-level ideas and requirements are transformed into a specific building description.
Nowadays, this process is usually carried out in a manual and labor-intensive manner. More precisely, concepts are usually drafted on semi-transparent paper with pencils so that a when a new sheet of paper is put on an existing one, the old concept may serve as a template for the next step in the design iteration.
In this paper, we present a semi-automatic approach to assist the developer by proposing suggestions for solving individual design steps automatically. These suggested designs can be modified between two successive automatic design steps, hence the developer remains in control of the overall design process. In the presented approach, floor plans are represented by graph structures and the developer’s behavior is modeled as a sequence of graph modifications. Based on these sequences we trained a recurrent neural network-based predictor that is used to generate the design suggestions. We assess the performance of our system in order to show its general applicability.
The paper at hand is a extended version of our ICPRAM 2018 conference paper [1], in which we address the different aspects of our proposed algorithm, challenges we faced during our research as well as intended work flow in greater detail.