期刊
EUROPEAN JOURNAL OF OPERATIONAL RESEARCH
卷 293, 期 2, 页码 658-680出版社
ELSEVIER
DOI: 10.1016/j.ejor.2020.12.055
关键词
Multiple criteria analysis; Multiple criteria sorting; Active learning; Preference elicitation; Interactive methods
资金
- Polish National Science Center under SONATA BIS project [DEC-2019/34/E/HS4/00045]
- Foundation for Polish Science (FNP - START Scholarship)
This study explores the interactive elicitation of holistic preference information for multiple criteria sorting and introduces several active learning strategies for selecting alternatives. Experimental results demonstrate that heuristic strategies based on current classification analysis can achieve competitive results compared to strategies predicting future stages.
We consider an interactive elicitation of holistic preference information for multiple criteria sorting approached with a threshold-based value-driven procedure. We introduce several active learning strategies for selecting, in each stage of interaction, an alternative that the Decision Maker (DM) should assign to its desired class. To identify the best assignment-based question, we evaluate each candidate alternative in terms of either ambiguity in its possible assignments at the current stage of interaction or its potential contribution to reducing uncertainty in the assignments of all alternatives once the question is answered. The performance of the proposed heuristic strategies is experimentally verified in view of computational time as well as the average and maximal numbers of questions that need to be answered by the DM until the classification recommended by all compatible preference models is sufficiently robust. We demonstrate that competitive results can be obtained with the heuristics that select the next question based on the analysis of current classification results as compared to the strategies looking ahead the current stage, which takes significantly more time. We also show how the performance of the questioning strategies is affected when, e.g., considering various problem sizes, imposing different stopping criteria for the preference elicitation, or reducing the flexibility of an assumed preference model by fixing the class thresholds. (C) 2021 Elsevier B.V. All rights reserved.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据