Techniques in computational stochastic dynamic programming

Abstract

Differential dynamic programming (DDP) is a variant of dynamic programming in which a quadratic approximation of the cost about a nominal state and control plays an essential role. The method uses successive approximations and expansions in differentials or increments to obtain a solution of optimal control problems. The DDP method is due to Mayne [11, 8]. DDP is primarily used in deterministic problems in discrete time, although there are many variations. Mayne [11] in his original paper did give a straight-forward extension to continuous time problems, while Jacobson and Mayne [8] present several stochastic variations. The mathematical basis for DDP is given by Mayne in [12], along the relations between dynamic programming and the Hamiltonian formulation of the maximum principle. A concise, computationally oriented survey of DDP developments is given by Yakowitz [16] in an earlier volume of this series and the outline for deterministic control problems in discrete time here is roughly based on that chapter. Earlier, Yakowitz [15] surveys the use of dynamic programming in water resources applications, nicely placing DDP in the larger perspective of other dynamic programming variants. Also, Jones, Willis and Yeh [9], and Yakowitz and Rutherford [17] present brief helpful summaries with particular emphasis on the computational aspects of DDP.