Robust identification of dynamical systems by neurocomputing

If a dynamical system has a fine feature or a dynamics under-represented in the data used for its identification, the ordinary criterion for training neural networks such as the quadratic criterion often leads to very large identification errors sometimes during the operation of the identifier. To cope with this problem, general risk-averting criteria were proposed by Lo (1996) for training neural networks for robust system identification. This paper studies the numerical feasibility of this approach, and compares the performances of the neural identifiers trained with respect to the risk-averting error criterion and those trained with respect to the quadratic error criterion