Classical theories of system identification are grounded on probabilistic assumptions under which various methods are guaranteed to converge, to be asymptotically efficient, etc. In this talk, we shall contend that theoretical guarantees can be obtained under way less assumptions than traditional theories do and shall make a case for the need to spend more research effort in this direction. This suggests a paradigm shift where prior knowledge only impacts on visible characteristics of the model, such as the extension of the identified region or the width of an interval used for prediction, while the model reliability is guaranteed under minimal a priori assumptions.
