Active preference learning in product design decisions.

In the earliest product design stages, it can be challenging to decide which product concepts make the cut and are worth working out in more detail. Especially with the advent of technologies like concept generators that can generate a vast number of candidates, often a lot of assumptions are made when choosing the most promising candidates. While performance attributes can be simulated or otherwise estimated under different loads, it can be difficult to balance many desirable and undesirable performance attribute values. Here, active preference learning offers a solution methodology that leverages easy-to-give designer feedback to rank a large set of product concepts. Experimental evaluation of pertinent active preference learning algorithms demonstrates that accurate concept rankings can be learned given only minimal user effort, even when there is noise in the designer's feedback. Moreover, this paper proposes and evaluates the use of a graph kernel to learn geometric preferences affecting the ranking, in addition to simulated performance attributes. [ABSTRACT FROM AUTHOR]