A learning approach for incorporation of imperfect knowledge in decision support system design

Abstract

This paper discusses the decision support system design problem with a finite set of multiattributed alternatives and imperfect knowledge about: the decision situation structural model, about the impacts of alternative courses of action, and about the value perspectives of the decisionmakers. "Imperfect" knowledge refers to available information that may be imprecise relative to the degree of refinement with which the assessment is made, ambiguous in the sense of giving rise to inconsistencies with principles and laws of an assumed decision situation model, and incomplete in that otherwise needed elements are missing. Principles for analyzing decision problems with imprecise and incomplete information are presented for the case of outcome certainty. Extensions of this, for cases involving risk, are proposed. A procedure based on these results is developed to rank the set of alternatives. It can be formulated as a set of linear programming problems. Knowledge of the functional form of the utility function such as whether it is multilinear or additive, or equivalent independence conditions on the set of attributes such as preferential independence or mutual preferential independence, are not necessary for implementation of the resulting algorithms. The amount of information needed is reduced when some knowledge of these conditions is available. These results are analogous to well known results in multiattribute utility theory in which precise and complete information is required. A decision support systems design methodology, that has implications for the knowledge base management subsystem and the model base management subsystem, results. The decision support design problem is formulated as a learning process in which the decisionmaker is encouraged to successively improve understanding of the decision situation, and adopt value judgements accordingly. Inconsistencies in the knowledge base are, for the most, the result of biases and inadequate heuristics that are used in the acquisition and representation of information. Extensions of existing methods and new approaches to identify, avoid, and resolve inconsistencies are discussed.

Special attention is paid to the complicated roles of the analyst in using the interactive screening procedure proposed. Due to the flexibility, and hence lack of structure, in assessing the required information, facilitation skills are required that go beyond those of more conventional and structured approaches. Identification of a minimal set of information required to induce a linear order on the set of alternatives is dicussed. A conceptual design of a proposed system to aid in the information processing activities of identification, acquisition, aggregation, evaluation, and interpretation of the required information for decision support is presented.