Local Linear Regression and Low-Discrepancy Sampling for Approximate Dynamic Programming

Abstract: Approximate Dynamic Programming is the standard method for the numerical solution of the well-known Bellman's equations. With such technique, two main issues arise: (i) the choice of a class of models to approximate the value functions; (ii) the definition of an efficient sampling of the domain where estimates of the value functions are computed. In this work the use of local linear regression based models is investigated when low-discrepancy sampling methods are used to sample the state space.