Machine Learning of Design Rules: Methodology and Case Study

Abstract

This paper describes a methodology for applying machine learning to problems of conceptual design, and presents a case study of learning design rules for wind bracings in tall buildings. Design rules are generated by induction from examples of minimum weight designs. This study investigates the applicability of machine learning methods that are capable of constructive induction, that is of automatically searching for and generating problem‐relevant attributes beyond those originally provided. The decision rules generated by machine learning programs specify design configurations that are recommended, typical, infeasible, or those that are to be avoided. The learned rules captured some of the essential expert's understanding of the design characteristics involved in selecting wind bracings for tall buildings. These results are promising and demonstrate a potential practical usefulness of the proposed methodology for automated generating of design rules.