| Intelligent control that involves both intelligence and control theory to solve complex, nonlinear, multivariate systems is a challenging problem. Among the various neurocontrol schemes, adaptive critic designs (ACDs), proposed by Werbos, is the most powerful and yet the most difficult to apply. In this research, a new form of neurocontrol, the Natural Adaptive Critic Design (NACD) is proposed. NACD effectively combines ACDs and Trajectory Pattern Method (TPM) proposed by Rastegar. Through TPM, NACD transforms the learning of complex desired control signal into the learning of a set of Fourier coefficients and thus allowing Heuristic Dynamic Programming (HDP), the simplest form of ACDs to be used.; In this dissertation, both theoretical analysis and simulations of NACDs are provided. For theoretical analysis, the advantage of progressive learning is illustrated through a simple nonlinear system. Both Jacobian analysis and effects of decoupling Fourier coefficients are addressed. For simulation, NACD is used to control a 3D robot manipulator. To compensate plant disturbances and noises, PID controllers are used in conjunction with NACDs. Through our theoretical analysis and simulation, we found that NACD is an effective neurocontrol technique. However, one disadvantage of NACD is its prolong training time. To speed up the training time and to exploit the inherent parallel nature of neural networks, parallel implementation of NACD is used for training the 3D robot controller. The training data is partitioned into several parts and trained simultaneously on different computers via Parallel Virtual Machine. The results show a significant saving in time, achieving an average speedup ratio of 4 for 6 computers running in parallel.; In summary, a new form of neurocontrol technique, Natural Adaptive Critic Design (NACD) is proposed. Theoretical analysis, computer simulation and parallel implementation of NACD are addressed. Although NACD is still in its early stage of development, NACD combines the simplicity of TPM and power of ACD and is, therefore, potentially a truly intelligent controller for complex, nonlinear, multivariate systems. (Abstract shortened by UMI.)... |
