Parallel Data Vault Methods for Larger Scale Stochastic Dynamic Programming

Abstract

The decomposition of the finite difference approximation to stochastic dynamic programming problems is described for the optimal control of nonlinear, continuous time dynamical systems. The stochastic components include both Gaussian and Poisson random white noise. A parallel data vault mass storage method is developed to take advantage of the decomposition, and therefore to help alleviate Bellman's curse of dimensionality in dynamic programming computations. It is shown that data vault memory on the data parallel Connection Machine type computational model can be enhance the efficiency of the decomposition performance. Extension of the data vault technique to a more general stochastic optimal control problem is discussed. Performance on the Connection Machine for larger scale stochastic dynamic programming problems, such as resource management problems with up to a projected six states, are illustrated and discussed.