Robust interactive decision-analysis (RID)

Po-Young Chu, Purdue University

Abstract

In Chapter One of the thesis we describe the stringent information requirements and the resultant measurement problems associated with traditional decision tree analysis in practice. We propose an interactive procedure for performing decision analysis, called Robust Interactive Decision Analysis (RID), which permits a Decision Maker (DM) to interactively express strong binary preferences and only imprecise state probability and utility function assessments. In Chapter Two we present an example to illustrate the RID method, discuss its operational perspectives, and implement the methodology as a decision support system. Conceptually, the operation of the RID method can be viewed as a state-space pruning system which has incorporated a decomposition concept and several pruning operators to search for optimal strategy. In implementing the RID system, we have followed several important system design principles based upon psychological theories. In Chapter Three of the thesis, we experimentally examine the behavioral premises upon which the RID methodology is based. Behavioral results indicate that the RID methodology is viable as subjects can provide strong preference inputs and imprecise (ordinal or interval) state probability assessments consistently. A normative DM's choice model is also presented for discussions and comparisons with observed DM's choice behavior. In Chapter Four of the thesis, we pursue the computational efficiency issues related to the RID method. The three pruning operators, namely the vector dominance, the preference dominance, and the statistical dominance operators, are shown very effective in pruning dominated (inefficient) strategies. With a limited number of strong preferences from a DM, the RID method can solve complex decision tree problems efficiently. The decomposition concept and other pruning operators provide the RID method with the reduction in computational effort which is approximately proportional to the number of states of nature. It is noted that Moskowitz's original recursion algorithm generally requires computational effort proportional to the squared term of states of nature. Chapter Five contains our concluding remarks and topics for future research. The RID method obviates the need of the precise measurement of the underlying state probabilities and DM's utility functions. Based upon our experiment results, we believe the RID methodology is viable, efficient, and robust.

Degree

Ph.D.

Advisors

Wong, Purdue University.

Subject Area

Management|Artificial intelligence

COinS