A robust on-line learning algorithm for intelligent control systems

This paper describes a novel error extraction approach for exploiting the strength of Levenberg-Marquardt (LM) optimization technique in intelligent control systems. Since the target value of the control signal is unknown, tuning of the controller parameters becomes a tedious task if the knowledge about the system and the environment is limited. The suggested methodology utilizes the sliding model control (SMC) technique. The error extraction scheme postulates the form of error on the applied control signal using the discrepancy from the prescribed reaching dynamics. The devised approach has been tested on the non-linear Duffing oscillator, which has been forced to follow a periodic orbit radically different from the natural one. The results obtained through a series of simulations have confirmed the high precision and robustness advantages without knowing the analytical details of the system under investigation. The issues of observation noise and the stability in the parametric space have approximately been addressed from the point of SMC perspective.