A self-organizing controller for dynamic processes using neural networks

The use of multilayer feedforward neural networks for self-organizing control of poorly defined dynamic processes is proposed. The control policy at any stage can be expressed in terms of linguistic control rules. These rules are treated quantitatively using fuzzy logic theory and implemented through neural networks to obtain deterministic control actions. The performance of the controller is evaluated at each time step using a performance measure, and the rules which contributed to previous control actions are reinforced. The neural network is then trained using backpropagation to accommodate such reinforcement in controller output. The performance of these controllers was tested through simulation on a number of linear dynamical systems as well as on nonlinear systems such as the cart-pole balancing problem. The results indicate that the controller has strong adaptive properties, and it can be built with only a little knowledge about the system dynamics. The performance of the controller was also observed to be highly robust to system parameter changes