The objective of this research project is to create techniques for capturing and reusing design procedures for parametric design. The proposed techniques will be implemented as a transparent software layer that will be placed over existing design tools, transforming them into trainable design automation tools. The designer will train these tools by solving a few sample design problems. Once trained, the tools will be capable of solving similar problems with a minimum of human intervention. The proposed approach will use decision tree learning algorithms to learn which attributes of the design best indicate which parameter the designer will change at any given point in the design procedure. Background knowledge will be used to improve the learning accuracy and to generalize the learned procedures. Neural networks will be used to learn representations for the implicit design constraints.
If successful, this work will help to preserve design knowledge in a reusable form. A tremendous amount of useful information is generated during the process of designing, but often only the final result is recorded, in the form of detailed drawings. When design information is lost, future efforts to maintain the design may result in unexpected failures. Similarly, future designers may have to reinvent what was once known, or they may repeat past mistakes. The proposed work directly addresses these issues by creating techniques for capturing the procedures by which design problems are solved. Solution procedures are the distillation of the designer's understanding of a problem, thus preserving them provides a valuable resource for others who must solve similar problems.
