There are well-known issues in conjunction with eliciting probabilities, utilities, and criteria weights in real-life decision analysis. This article explores various computationally efficient methods for generating weights in multi-criteria decision support systems. Therefore, it constitutes an aid for MCDA modellers and tool designers in selecting surrogate methods for criteria weights. Given the challenges in eliciting precise criteria weights from decision-makers, this study evaluates a range of techniques for automatically generating surrogate weights, focusing on both ordinal and cardinal ranking approaches. With a thorough inquiry methodology never before used, we examine automatic multi-criteria weight-generating algorithms in this article. The methods tested include traditional rank-based models such as rank sum (RS), rank reciprocal (RR), and rank order centroid (ROC), alongside newer approaches like the sum reciprocal (SR) and cardinal sum reciprocal (CSR). The results show that the SR approach for the ordinal case and the CSR method for the cardinal case perform better in terms of robustness than other methods, even including the promising new geometric class of methods. It is also shown that linear programming (LP) performs poorly when compared to surrogate weight models. Additionally, as expected, the cardinal models perform better than the ordinal models. Unexpectedly, though, the well-established LP model’s performance is worse than previously thought.
There are well-known issues in eliciting probabilities, utilities, and criteria weights in real-life decision analysis. In this paper, we examine automatic multi-criteria weight-generating algorithms which are seen as one remedy to some of the elicitation issues. The results show that the newer Sum Rank approaches perform better in terms of both performance and robustness than older (classical) methods, also when compared to the new and promising geometric class of methods. Additionally, as expected the cardinal surrogate models perform better than their ordinal counterparts (with one exception) due to their ability to take more information into account. Unexpectedly, though, the well-established linear programming model’s performance is worse in this respect than previously thought, despite a promising mapping between linear optimisation and surrogate weight generation which is explored in the paper.