Automatically Learning Preliminary Design Knowledge from Design Examples

Abstract: Acquisition of structural design knowledge for knowledge‐based systems is difficult because humans acquire their design knowledge through years of experience, resulting in a diverse, unstructured and implicit body of knowledge. Research and development of knowledge bases for structural design have identified distinguishing characteristics of design knowledge. One approach to acquiring design knowledge using machine learning techniques is to transform design examples into a generalized representation of design knowledge. In this paper, we suggest the use of a dependency network as the generalized representation. A four‐step learning process is presented: grouping design variables into skeletal concepts, learning default values and values ranges for each variable, deriving relationships among numerical valued variables, and learning patterns among nominal valued variables. The resulting representation provides a basis for producing an initial design given a set of design requirements.