Work

Model-based decision support for subjective preference blending

Year: 2019
Type: dissertation
Abstract: ENGLISH SUMMARY: In a subjective preference blending problem (SPBP), the objective is to select a most pleasing alternative from a set of alternatives, based on the relative performances of the alternatives in respect of a number of (possibly conflicting) selection criteria. What distinguishes this problem from other types of selection problems is that there is no universally accepted objective function for scoring the alternatives according to the selection criteria. Instead, every decision maker has his or her own relative preferences of the alternatives with respect to the various selection criteria — hence the prefix subjective. Although various multi-criteria decision support frameworks or methodologies exist for estimating a particular decision maker’s value function in respect of combined selection criteria in an SPBP instance, these processes are only effective if the decision space of the problem instance is not too large. This is because value function estimation procedures typically involve pairwise comparisons of the criteria, which has quadratic-order time complexity. Large decision spaces, in fact, often induce the phenomenon of satisficing, where a decision maker settles for an acceptable alternative rather than attempting to discover the best alternative — either because of time constraints or because of a lack of a full characterisation of the decision space. The provision of automated decision support to a decision maker in respect of the exploitation and exploration of the decision space of an SPBP instance resides within the realm of interactive preference learning. The quest in preference learning is to predict the subjective preferences of a decision maker based on past observed preference information elicited from that decision maker or similarly minded decision makers. Preference learning becomes interactive if the decision maker is actively included in the (typically iterative) learning process, in which case such (subjective) learning is usually tailored to the specific decision maker involved. An interactive preference learning decision support system is proposed in this dissertation for facilitating solution of SPBP instances. The working of the system is based on an iterative involvement of the DM in the form of the elicitation of pairwise comparisons of alternatives strategically selected by the system so as to achieve a trade-off between efficiently exploring the entire decision space and exploiting well-performing areas of the decision space according to the current value function estimation of the decision maker. The design objectives of the system specifically include workload reduction of the decision maker and the accommodation of low to moderate levels of decision maker inconsistency during pairwise comparisons of alternatives. The effectiveness of the decision support system is assessed in view of its intrinsic ability to learn a decision maker’s value function by applying methods from inferential statistics to the output recommendations of the system. The system is finally also validated by applying it to two SPBP instances involving real human decision makers.
Language: English
Open Access status: green
OpenAlex ID: W2946174876
Cites: 136
Cited by:
Related to: 10