Application of orthogonal arrays and MARS to inventory forecasting stochastic dynamic programs

This paper describes the solution to inventory forecasting problems using a statistical perspective of the high-dimensional continuous-state stochastic dynamic programming (SDP) optimization model. In particular, the accuracy of the OA/MARS SDP solution method (Chen et al., 1999, Oper. Res., to appear), which employs orthogonal arrays and multivariate adaptive regression splines, is examined via simulations which vary certain user-specified parameters. For continuous-state SDP, the current definition of high-dimensional is more than five state variables. Most continuous-state problems require an approximate solution through discretization of the state space and estimation of the future value function. Under a statistical perspective, the discretization and the future value function are analogous to an experimental design and an unknown mean response. (C) 1999 Elsevier Science B.V. All rights reserved.