Approximations and bounds in discrete stage markov decision processes

In this paper, an approximate version of the value iteration dynamic programming algorithm is proposed. Each iteration of that method consists roughly in computing an approximation Vn+ 1 of T(Vn) , where V n is the current cost-to-go function and T is the usual dynamic programming operator. The method of aDproxlmation to be used is not fixed; it can be choosen among many available methods (spline interpolation or aDproximation, finite elements methods, etc.) and may vary from iteration to iteration.