Design and configuration of a machine learning component for user profiling in a declarative design environment

MultiCAD is a design environment that generates geometric models of buildings based on abstract declarative descriptions. The increased number of solutions produced has called for an intelligent module selecting those closer to user's preferences. We have proposed and implemented such a module, featuring two components: a Decision Support Component, capturing user preferences based on attribute weight assignment techniques (used in SMART, AHP and via RR), and a Machine Learning Component, learning preferences by incrementally training a neural network committee based on user evaluated solutions. Alternative configurations must be compared before actual use of the ML Component takes place. Due to the practical limitation on the number of solutions that can be inspected and evaluated by human users, an automated mechanism plays the role of a group of virtual users. The best performing configuration, regarding virtual users' preferences, will be integrated to the system and evaluated against actual human evaluation results.