Incorporating user design preferences into multi-objective roof truss optimization

• Main Author: Bailey, Breanna Michelle Weir
• Other Authors: Raich, Anne M.
• Format: Others
• Language: en_US
• Published: Texas A&M University 2007
• Subjects: user preference; truss optimization
• Online Access: http://hdl.handle.net/1969.1/5932

Automated systems for large-span roof truss optimization provide engineers with the flexibility to consider multiple alternatives during conceptual design. This investigation extends previous work on multi-objective roof truss optimization to include the design preferences of a human user. The incorporation of user preferences into the optimization process required creation of a mechanism to identify and model preferences as well as discovery of an appropriate location within the algorithm for preference application. The first stage of this investigation developed a characteristic feature vector to describe the physical appearance of an individual truss. The feature vector translates visual elements of a truss into quantifiable properties transparent to the computer algorithm. The nine elements in the feature vector were selected from an assortment of geometrical and behavioral factors and describe truss simplicity, general shape, and chord shape. Using individual feature vectors, a truss population may be divided into groups of similar design. Partitioning the population simplifies the feedback process by allowing users to identify groups that best suit their design preferences. Several unsupervised clustering mechanisms were evaluated for their ability to generate truss classifications that matched human judgment and minimized intra-group deviation. A one-dimensional Kohonen self-organizing map was selected. The characteristic feature vectors of truss designs within user-selected groups provided a basis for determining whether or not a user would like a new design. After analyzing user inputs, prediction algorithm trials sought to reproduce these inputs and apply them to the prediction of acceptable designs. This investigation developed a hybrid method combining rough set reduct techniques and a back-propagation neural network. This hybrid prediction mechanism was embedded into the operations of an Implicit Redundant Representation Genetic Algorithm. Locations within the ranking and selection processes of this algorithm formed the basis of a study to investigate the effect of user preference on truss optimization. Final results for this investigation prove that incorporating a user's aesthetic design preferences into the optimization project generates more design alternatives for the user to examine; that these alternatives are more in line with a user's conceptual perception of the project; and that these alternatives remain structurally optimal.