consortiums

PROJECT 5

SELECTION AND CLASSIFICATION

Shing Chang, Stanley Lee, Steve Hanna
Kansas State University

Proposed Grant $122,272

Abstract:
The proposed research is in response to the research vision described in Sailor 21 of the Navy Personnel Research Studies and Technology (NPRST). Specifically, we propose to study and develop a set of decision-making tools for the Navy selection and classification processes. Due to the recent advances in fuzzy set theory, fuzzy and evidence reasoning can be used to support traditional decision support

techniques, such as, discriminate analysis, multiple attribute decision making (MADM) and multiple objective decision making (MODM). The proposed tools would allow decision makers to judge a sailor's composite competency for a future assignment/position based on the multiple uncertain levels of all sources of information related to a sailor's isolated component abilities, experiences, skills, and other qualifications. The uncertainty information on factors is modeled by fuzzy linguistic variables while the system issues of tracking a recruit's capability for a career in the Navy are handled by uncertainty reasoning tools. The novice research in this proposal highlights the following areas. First, the proposed model considers all input factors over time, including Armed Forces Vocational Aptitude Battery (ASVAB) scores. Second, the proposed solution approach by fuzzy MADM or MODM provides not just a few "crisp" solutions but rather a set of possible solutions ranked by possibility scores, which take into accounts a recruit's "whole person" profiles, such as, test scores, job specific skills, social understandings, motivations, leadership, etc.

The proposed project will be validated by the anslysis of different approaches for selection and classification modeling. We propose to compare the proposed uncertainty reasoning tools to different types of multivariate decision analysis approaches, such as, multivariate, multi-response regression analysis, statistical discriminate analysis using linear and quadratic functions, and neural networks. Almost all current selection and classification research is based on simple linear regression model where the independent variable is a composite measure of multiple responses and the predictor variables include intellectual and behavioral measures, which are usually ill defined. Treating these estimates as absolute values in regression analysis will lead to large errors in the model validation phase. In addition, the regression model is restricted by model assumptions, e.g., constant variance on the response for all factor combinations, which may not be true in practice. The single outcome measure also reduces the information that could have been gained from a multivariate analysis. By modeling uncertain variables, such as behavioral measures in fuzzy sets, the proposed fuzzy multivariate analysis will provide more accurate description of a real world event.