Multicriterion decision merging: Competitive development of an aboriginal whaling management procedure

• 成果类型: Article
• 署名作者: Givens, GH
• 署名单位: Colorado State University System; Colorado State University Fort Collins
• 刊物名称: JOURNAL OF THE AMERICAN STATISTICAL ASSOCIATION
• ISSN/ISSBN: 0162-1459
• 发表日期: 1999
• 页码: 1003-1014
• 关键词: sensitivity analysis techniques; computer-models

摘要

International Whaling Commission management of aboriginal subsistence whaling will eventually use an aboriginal whaling management procedure (AWMP) chosen from a collection of candidate procedures after grueling simulation testing. An AWMP is a fully automatic algorithm designed to operate on the results of an assessment (i.e., a statistical estimation problem relying on sparse series of whale abundance data) to produce a catch limit in each year of real or simulated management. An AWMP should, as much as possible, meet the conflicting objectives of low population risk, high satisfaction of needed catch, and high rate of population recovery. The choice of the best procedure falls naturally in the multicriterion decision making framework, because one of several candidates must be chosen on the basis of high-dimensional simulated performance summaries over a wide range of assumptions about whales and whaling. However, standard multicriterion decision making methods are impractical and unsatisfying for this problem. A method is developed to merge competing procedures into a new procedure that is an admissible Bayes rule. The approach is constructive rather than selective, meaning:that it is not intended to produce an automatic winner, but rather a promising new candidate. This merging approach allows the best performance aspects of competing procedures to be combined. Ideally, and in examples shown, the newly constructed procedure outperforms all previous candidates. The approach also permits tuning of a single procedure to enhance performance or to more closely reflect design goals, without a simulation-intensive search-over the tuning parameter space. These methods are generalizable to a larger class of decision problems.