DESIGN OPTIMIZATION PROBLEM REFORMULATION USING SINGULAR VALUE DECOMPOSITION Somwrita Sarkar Graduate Student email: ssar3264@mail.usyd.edu.au University of Sydney, Australia Andy Dong Senior Lecturer email: andy.dong@usyd.edu.au University of Sydney, Australia John S. Gero Research Professor email: john@johngero.com George Mason University, USA ABSTRACT This paper presents a design optimization problem reformulation method based on Singular Value Decomposition (SVD), dimensionality reduction, and unsupervised clustering. The method calculates linear approximations of the associative patterns of symbol co-occurrences in a design problem representation to infer induced interaction/coupling strengths between variables and constraints. Unsupervised clustering of these approximations is used to identify useful reformulations. These two components of the method automate a range of reformulation tasks that have traditionally required different solution algorithms. We explain the method using an analytic model-based decomposition problem, and apply the method to an analytic hydraulic cylinder design problem as an example of heuristic design “case” identification, and to non-analytic problems expressed in FDT and DSM forms as examples of design decomposition. An Aircraft Concept Sizing (ACS) problem is used to empirically validate the method’s performance. The results show that the method can be used to infer multiple well-formed reformulations starting from a single problem representation in a knowledge-lean and training lean manner. Keywords: design methodology, design automation, unsupervised learning, pattern extraction, SVD INTRODUCTION Design problem reformulation significantly affects the final results of any design optimization process. Artificial intelligence methods have been proposed to automate some problem formulation tasks in design optimization. Cagan et al. [1] present a review of methods that use artificial intelligence techniques for optimization in engineering design. Schwabacher et al. [2] explore inductive machine learning techniques such as decision tree induction to automate various tasks in design optimization. Other notable research exploring similar issues include work by