A learning theory approach to the construction of predictor models (Q1422498)

A new approach for the construction of predictor models for unknown dynamical systems is presented. Dynamical models that return a prediction interval for the output of an unknown remote system are studied. The study is organized as follows. After an introduction, in the second section of the paper two key elements of the approach are introduced: models that return an interval as output (Interval Predictor Model) and the notion of consistency with observed data. In the third section, the computational results for the construction of interval models, using linear regression structures are presented. In the fourth section the fundamental issue of assessing the reliability of a data-consistent model, with respect to its ability to predict the future behaviour of the unknown system is given. The results are extended in the fifth section to the case of weakly dependent observations. To illustrate the nature of the presented results, a simple numerical example is proposed.