DIRECT CONTROL AND COORDINATION USING NEURAL NETWORKS

The performance of an industrial process control system equipped with a conventional controller may be degraded severely by a long system-time delay, dead zone and/or saturation of actuator mechanisms, model and/or parameter uncertainties, and process noises. The coordinated control of multiple robots is another challenging problem. In a multiple-robot system, each robot is a stand-alone device equipped with commercially designed servo controllers. When such robots hold a solid object, failure of their effective coordination may damage the object and/or the robots. To overcome these problems, we propose to design a direct adaptive controller and a coordinator using neural networks. One of the key problems in designing such a controller/coordinator is to develop an efficient training algorithm. A neural network is usually trained using the output errors of the network, not controlled plant. However, when a neural network is used to directly control a plant, the output errors of the network are unknown, because the desired control actions are unknown. A simple training algorithm is proposed that enables the neural network to be trained with the output errors of the controlled plant. The only a priori knowledge of the controlled plant is the direction of its output response. A detailed analysis of the algorithm is presented and the associated theorems are proved. Due to its simple structure, algorithm and good performance, the proposed scheme has high potential for handling the difficult problems arising from industrial process control and multiple system coordination.