| The use of neural networks in control applications is gaining more and more acceptance because of its unique ability to learn from examples. However, the control schemes that have been developed require off-line training of the controller, or measuring some plant characteristic, and/or assuming a plant model. In this research, a novel method that allows on-line training is suggested. The method uses two neural networks, one for the controller, and one for the model. The training of both networks is done simultaneously using the relation between different outputs and their errors. This relation is extracted and used to quantify the error of the model as well as that of the controller so that both of them can be updated. The advantage of this method is that the training set is cut in half, and the method is general enough to handle different problems. The method is incorporated into a computer program using "C" language and applied to robot positioning control for a two degrees of freedom robot with revolute joints. In this example, the Cartesian coordinates are the input to the system. The system has to construct the correct corresponding angles which give an output similar to the input. The new method is also applied to the temperature control of a dual duct air conditioning system. In this case, given the desired space temperature, the controller changes the dampers positions so that the mixing of cold and hot air streams results in attaining the desired temperatures. |
