Initializing multilayer perceptrons with interconnected neurons

Multilayer perceptrons with interconnected neurons (MLPWINs) are known to be universal approximators of dynamic systems. A common way to initialize MLPWIN for identifying a plant in the parallel formulation is simply to set the initial dynamic state of the MLPWIN equal to zero in both its training and operation. This causes the MLPWIN to have a poor transient performance. The paper proposes two methods of initializing an MLPWIN to improve or eliminate such poor transient performance: an initial dynamic state of the plant is converted into that of the MLPWIN by a look-up table, if only a finite number of initial dynamic states of the plant are of interest, or by a feedforward network, otherwise