Universal neuroapproximation of dynamic systems for robust identification

Risk-sensitive criteria for system identification emphasize or de-emphasize large individual errors in an exponential manner and thereby induce risk-averting or risk-seeking identification performances. Identification by neural networks usually is not perfect and some errors are bound to be present. The availability of the risk-sensitive criteria allows us to obtain neural networks as system identifiers to better suit a large variety of applications. This paper shows that under mild regularity conditions, general risk-sensitive identification of dynamic systems by neural networks can be done to any desired degree of accuracy in both the series-parallel and parallel formulations