Model quality: the roles of prior knowledge and data information

The authors discuss the basic issues involved in the problem of estimating a model's reliability. In particular, the role of prior information is scrutinized. The modeling errors can be divided into two categories, namely, systematic/bias errors and variability/random errors. All serious system identification experiments contain a model validation step. For an unfalsified model, the bias error has not been found to be significantly larger than the random errors. Hence, the traditional, statistical way to provide estimated standard deviations for the model is relevant for unfalsified models. In the case of many unfalsified models, a sound scientific approach is to choose the most powerful unfalsified one. The definition of most powerful depends on the intended application. For example, in robust control design an unfalsified model can be said to be most powerful if the H/sub infinity / error bound is minimized. How this concept relates to minimizing the mean square errors is discussed. A number of questions for further research are identified.<
>