Neural network-based adaptive control with a probing signal

We investigate properties of inputs to neural networks in indirect adaptive control systems. Motivated by the conventional adaptive control literature, we introduce the concept of “sufficient richness” that the input signal contains enough frequencies in order to deal with the identification and control of nonlinear systems using multilayered neural networks (MLNNs). The proposed control algorithm generates the control input by summation of the output of the neural network controller (NNC) and the probing signal obtained from signals in the system. In the algorithm the probing signal has attributes of random signals and may excite the weights of the MLNN because of many spectral lines. Some examples were used to explain the characteristics of the proposed algorithm. From results of the simulations, we see that the probing signal plays an important role in operating the control system online