SUBSPACE IDENTIFICATION AS MULTI-STEP PREDICTIONS OPTIMIZATION

摘要

We will show that apparently complicated and not easy to understand expressions with geometrical projections appearing in the algorithms of subspace identification for linear state space models, can be quite simply derived from optimizing these models for multi-step predictions on measured data samples using least squares. Furthermore we will show the advantages which brings the use of multi-step predictions instead of single-step predictions arising from regression methods.