Letter to the Editor—A Critique of the Norman-White Dynamic Programming Approximation

Abstract

This note points out that the Norman-White procedure utilizes the full apparatus of Bellman's value iteration, but stops after one iteration. This would seem wasteful unless either one iteration could be shown to give uniformly good results, or better stopping rules could not be found. A simple counterexample shows that the possible percentage cost error of the Norman-White procedure is unbounded. The note then discusses the excellent results of the approximation for their two examples, and mentions a stopping criterion for nondiscounted value iteration that is due to D. J. White himself.