The learned control of complex dynamic systems

This paper demonstrates how artificial intelligence techniques based on learning can be applied in the control of complex dynamic systems. The development of adaptive rule-based automatic controllers is first considered. The heuristic `IF condition THEN action´ control rules applied by humans were derived through observation of individual skills. The controller was tested and is compared with a rule-based controller in which the heuristic rule was derived by considering the systems dynamic equations only. Data gathered from experimental trials using these two controllers, on the physical system, was post-processed by an induction rule generator to see what form automatically generated rules would take. This is called passive learning. Finally, an artificial neural-net controller was constructed and trained to operate in a machine-learned controller domain. Comparative controller studies, in simulation, showed that the neural-net controller adapted to changing system parameters better than other controllers