Learning and Recovering Additive and Multiplicative Value Functions: A Criterion Validation of Multiattribute Utility Techniques

摘要

This paper describes two experimental studies in which subjects were taught additive and multiplicative value functions for the evaluation of diamonds. After learning, subjects were sent to a decision analyst who used standard multiattribute utility elicitation techniques to recover these value functions. Comparison of the taught and recovered functions allowed us to evaluate techniques. In the real-world experiment, internal band auditors served as subjects in a criterion validation study. Subjects provided both holistic and SMART models of commercial loan classification. Both types of models resulted, overall, in about the same level of accuracy. This level of accuracy was slightly better than a least squares solution using the same variables. Taken together, we found the studies suggestive of two strategies for coping with complex structures in MAUM. The first is to attempt to reduce the complexity by searching for simple and independent sets of attributes that lend themselves to additive modeling. The second is to increase the model complexity, if you believe the underlying preferences are non-additive and the deviations from additivity are not too extreme.