Neuro-optimal control

SUMMARY NEURO-OPTIMAL CONTROL Keywords: Intelligent control, neuro-control, optimal control, dynamic neural networks, second order nonlinear dynamic optimization, adjoint theory. In this dissertation, the improvement of intelligent optimal control algorithms and the application of dynamic neural networks (DNN) in support of optimal control calculations are presented. Intelligent control covers a wide range of technologies related to hard-sciences such as the optimal control theory and soft-sciences such as artificial intelligence (AI) and heuristics. Using the broad knowledge spectrum, hard and soft, in these technologies, new concepts emerge and computationally efficient software structures become feasible. The application of artificial neural networks (ANN) to dynamic system control has been constrained by the non-dynamic nature of popular network architectures. Many of the difficulties are large network sizes, long training times, etc. These problems can be overcome with DNN. In this dissertation, intelligent optimal control problem is considered as a nonlinear optimization with dynamic equality constraints, and DNN as a control trajectory priming system. Direct-descent-curvature or direct second order descent algorithm has been used for the optimal control computations. This algorithm is compared with the Hamiltonian methods in the literature. The algorithm has generated more robust solutions than the others with respect to conjugate points. The time varying optimal feedback gains are also generated along the trajectory as byproducts. In this study, intelligent optimal control algorithm has been developed. The resulting algorithm operates as an auto-trainer for DNN (a self-learning structure) and generates optimal feed-forward control trajectories in a significantly smaller number of iterations. In this way, optimal control trajectories are encapsulated and generalized by DNN. Speeding up trajectory calculations opens up avenues for real-time intelligent optimal control with virtual global feedback. The adjoint theory has been used in the training of DNN which is considered as a quasi-linear dynamic system. The updating of weights (identification of parameters) are based on steepest descent gradient (SDG), scaled conjugate gradient (SCG), Davidon-Fletcher-Powell (DFP) and Broyden-Fletcher-Goldfarb-Shanno (BFGS) methods with line search. All methods has been successfully applied to DNN. Simulation results are given for controlling VDP and CSTR which are nonlinear second order systems.