The "universal" approximating/learning feature of neural network (NN), widely and extensively used for control design, is contingent upon some critical conditions, either of which, if not satisfied, would render such feature vanished. In this paper, we show that these conditions are literally linked with several fundamental issues that have been overlooked in most existing NN-based control designs, either unconsciously or deliberately. We further propose a collective approach to explicitly address these issues, establishing a strategy enabling the NN unit to be fully functional in the control loop during the entire process of system operation and ensuring the more reliable and more effective NN-associated control performance. This is achieved by incorporating the control with a new structural NN unit, consisting of a group of diversified neurons with self-adjusting subneurons, each being driven/stimulated by input signals confined within a compact set. Meanwhile, the continuity of the control signal and the boundedness of all the closed-loop signals are ensured. Both the theoretical analysis and numerical simulation validate the effectiveness of the proposed method.The "universal" approximating/learning feature of neural network (NN), widely and extensively used for control design, is contingent upon some critical conditions, either of which, if not satisfied, would render such feature vanished. In this paper, we show that these conditions are literally linked with several fundamental issues that have been overlooked in most existing NN-based control designs, either unconsciously or deliberately. We further propose a collective approach to explicitly address these issues, establishing a strategy enabling the NN unit to be fully functional in the control loop during the entire process of system operation and ensuring the more reliable and more effective NN-associated control performance. This is achieved by incorporating the control with a new structural NN unit, consisting of a group of diversified neurons with self-adjusting subneurons, each being driven/stimulated by input signals confined within a compact set. Meanwhile, the continuity of the control signal and the boundedness of all the closed-loop signals are ensured. Both the theoretical analysis and numerical simulation validate the effectiveness of the proposed method.
Keywords: Artificial neural networks; Control design; Neurons; Periodic structures; Reliability; Systems operation.