A new method for applying choosing by advantages (CBA) multicriteria decision to a large number of design alternatives

Abstract

The design of buildings, civil infrastructure and other complex systems in our built environment involves considering many, often conflicting, design criteria. Architecture, engineering and construction (AEC) project teams often use multi-criteria decision making (MCDM) methods to help them arrive at a preferred design solution. An emergent MCDM method in practice today is Choosing by Advantages (CBA) which has been successfully applied to many AEC projects. This method has several benefits over traditional MCDM methods (such as the weighted sum): CBA does not allow to hide a compensation of money for value, CBA helps differentiate between alternatives based on the decision context, CBA reduces time to reach consensus, and it manages better subjective trade-offs by basing decisions on importance of agreed advantages. CBA is usually applied between two to ten alternatives, and it has never been used for more than one hundred alternatives. Hence, this study contributes to knowledge by developing and testing a new method for applying CBA to hundreds or thousands of alternatives. The new method involves clustering alternatives into a few representative design alternatives based on feature similarity using the K-means method. Preferences between these representative design alternatives are then generalized using linear regression. An experiment involving student subjects was conducted to measure the level of accuracy in which preferences can be generalized by the proposed method. The experiment considered 1,000 different building design alternatives. CBA was applied on representative alternatives using three, six, eight, and ten clusters. The study measured errors, correlations, and consistency of the predictions for each cluster configuration. When eight clusters were used for creating representative alternatives, decisions were always consistent to those made with random alternatives, and correlation with the predicted preference was higher with lower error compared to other cluster configurations tested.