Comparison of experimental designs in continuous-state stochastic dynamic programming

An approximation algorithm for high-dimensional, continuous-state stochastic dynamic programming was first presented based on an orthogonal array (OA) state space discretization and a Multivariate Adaptive Regression Splines (MARS) value function approximation. Given the popularity of Number Theoretic Methods (NTM), this paper compares OA-based experimental designs and NTMs for state space discretization using a ninedimensional inventory forecasting problem. The statistical model employed for future value function approximation is Artificial Neural Networks (ANN).