MS Thesis Presentation by Steven J. Rekuc
Thursday, July 7, 2005

(Dr.Chris Paredis, Chair)

"Eliminating Design Alternatives under Epistemic Uncertainty"


Typically, design is approached as a series of decisions in which designers select what they believe to be the best alternative. While this approach can be used to arrive at a final solution quickly, it is unlikely to result in the most-preferred solution. The reason for this is that all the decisions in the design process are coupled. To determine the most preferred value for a design variable considered in the first decision, one would really need to know the outcomes of all future decisions. By solving coupled decisions as a sequence of independent point decisions, one introduces uncertainty, leading to a solution different from the most preferred solution. This uncertainty can be characterized as epistemic (due to lack of knowledge) to differentiate it from aleatory uncertainty (due to inherent randomness).

Since point decisions often do not lead to the most-preferred design, I suggest a set-based approach motivated by the engineering practices at Toyota and based on the structure of the Branch and Bound Algorithm. This approach takes into account epistemic uncertainty with the understanding that due to this uncertainty, designers cannot always decide on a single value or alternative in a decision. Thus, decisions do not have to result in a single value or alternative, instead the designer decides on a set of possible solutions, eliminating the inferior designs from the set. For this approach to be feasible, an efficient method of eliminating inferior designs is necessary.

In this thesis, I derive a criterion for efficient elimination under epistemic uncertainty. To be efficient, the criterion takes into account common uncertainty uncertainty shared between design alternatives while eliminating all the design alternatives possible without the risk of eliminating the most preferred design. I show how in taking this uncertainty into account one is able to eliminate significantly more designs than without it. The effectiveness of this elimination with and without common uncertainty is demonstrated in both a beam design and a gearbox design.