Approximate Dynamic Programming with Postdecision States as a Solution Method For Dynamic Economic

Abstract

I introduce and evaluate a new stochastic simulation method for dynamic economic models. It is based on recent work in the operations research and engineering literatures (Van Roy et. al, 1997; Powell, 2007; Bertsekas, 2011). The baseline method involves rewriting the household’s dynamic program in terms of post-decision states. This makes it possible to choose controls optimally without computing an expectation. I add a subroutine to the original algorithm that updates the values of states not visited frequently on the simulation path; and adopt a stochastic stepsize that efficiently weights information. Finally, I modify the algorithm to exploit GPU computing.

Keywords: Numerical Solutions, Approximations, Heterogeneous Agents, Nonlinear Numerical Solutions, Dynamic Programming

JEL Classification: C60, C61, C63, D52