Learning Symbolic Formulations in Design
Optimization
Somwrita Sarkar and Andy Dong
University of Sydney, Australia
John S. Gero
George Mason University, USA
This paper presents a learning and inference mechanism for unsupervised learning
of semantic concepts from purely syntactical examples of design optimization
formulation data. Symbolic design formulation is a tough problem from
computational and cognitive perspectives, requiring domain and mathematical
expertise. By conceptualizing the learning problem as a statistical pattern
extraction problem, the algorithm uses previous design experiences to learn design
concepts. It then extracts this learnt knowledge for use with new problems. The
algorithm is knowledge-lean, needing only the mathematical syntax of the
problem as input, and generalizes quickly over a very small training data set. We
demonstrate and evaluate the method on a class of hydraulic cylinder design
problems.
Motivation
Design formulation and reformulation significantly affect the results of any
automated optimization exercise, and is a difficult problem from both
computational and cognitive perspectives. It is a difficult problem for
many reasons – there is no known formal process that takes an abstract set
of design requirements as input and produces a symbolic mathematical
design model as output [1]; It is knowledge intensive and requires both
domain and mathematical expertise from designers; the numbers of
variables, parameters, objectives and constraints in a problem often exceed
© Springer Science + Business Media B.V. 2008
J.S. Gero and A.K. Goel (eds.), Design Computing and Cognition ’08, 533