Decoupled Data-Based Approach for Learning to Control Nonlinear Dynamical Systems

This article addresses the problem of learning optimal control policy for a nonlinear stochastic dynamical. is subject to “curse dimensionality” associated with dynamic programming method. proposes novel decoupled data-based (D2C) algorithm that this using decoupled, “open-loop–closed-loop,” approach. First, an open-loop deterministic trajectory optimization solved black-box simulation model dynamical system. Then, closed-loop developed around by linearization dynamics about nominal trajectory. By virtue linearization, linear quadratic regulator based can be used control. We show performance D2C approximately optimal. Moreover, suggests significant reduction in training time compared other state-of-the-art algorithms.