Robust Selection of the Best Through Computer Simulation Experiments

Project: Research

View graph of relations

Description

Decision-making processes typically involve selecting the best decision among a set of competing alternatives and the best is often defined as the one with the largest or smallest mean performance. For instance, in risk management, investors call for the best portfolio of financial instruments (e.g., stocks, bonds, derivatives) to maximize the expected return. In inventory management, managers resort to the best decision rule for inventory control (e.g., the amounts to produce, pricing, etc.) to maximize the expected net profit of a firm. Where there is a finite (often small) number of alternatives, ranking and selection (R&S) serves as an important vehicle to select the best decision.Most R&S procedures in the literature assume that a simulation model is provided and the input distributions of the simulation model can be specified accurately a priori. However, when constructing a distributional model for a stochastic system in practice, the decision maker often faces uncertainty in the specification of the distribution family and/or estimation of the pertinent parameters. For instance, it may be difficult to specify covariance matrix of stock returns in risk management and to characterize the demand distribution in inventory management. In the proposed research, we use the term “ambiguity” to describe the above uncertainty issue. Ignoring the ambiguity may result in a misleading or false associated decision, especially when the ambiguity is deep and profound.This proposal is devoted to addressing the problem of how to select the best decision in the presence of ambiguity. In this proposal, we propose a robust framework in which the ambiguity is introduced by assuming that the input distribution of the simulation model belongs to a so-called “ambiguity set”. Under this framework, we compare decisions based on their worst-case performances over a pre-specified ambiguity set and select the decision with the best (smallest) worst-case performance. To achieve this target, we will design two-layer R&S procedures, either two-stage or fully sequential, under the indifference-zone formulation. The procedures identify the worst-case distribution in the first stage and the best decision in the second.

Detail(s)

Project number9042155
Grant typeGRF
StatusFinished
Effective start/end date1/09/142/10/18