Modeling of User Design Preferences in Multiobjective Optimization of Roof Trusses

Many conceptual design programs determine optimality by relying on quantifiable design objectives such as minimizing weight and deflection. Qualitative design criteria related to architectural or constructability requirements, however, are not typically considered. This paper presents a preference-prediction model that evaluates user preference as an explicit design objective in the topology and geometry optimization of roof trusses. During optimization, each truss design is characterized by using a set of nine quantifiable features. A trained neural network inputs the characteristic truss feature values for each proposed design and outputs an integer value that indicates the user’s preference for that design. To reduce the number of user interactions required to collect the required neural network training data, a Kohonen self-organizing map clusters designs into groups on the basis of feature similarities. Clustering allows users to indicate preferences for groups of similar trusses instead of for individual trusses. A rough-set reduct technique removes outliers from the user-preferred truss groups before the feature information and user preference assignments are used to train the predictive back-propagation neural network. The preference-prediction model is implemented in an existing multiobjective optimization program to determine the benefits of including satisfaction of user preferences as a design objective. Trial results indicate that more Pareto-optimal designs are identified, which increases the extent of the front, when user preferences are stated. Also, the Pareto-optimal set includes designs that have features reflecting the imposed user preferences and features that are structurally optimal.