Learning-theoretic approach to probabilistic model-set identification

Abstract

A model-set identification algorithm is proposed in a probabilistic framework based on the leave-one-out technique. It provides a nominal model and a bound of its uncertainty for a provided plant assuming that the effect of the past inputs decays with a known bound. Since it does not require further assumptions on the true plant dynamics or on the noise, a risk to make inappropriate assumptions is small. The number of assumptions is shown to be minimum in the sense that identification is impossible after removing the assumption made here. An algorithm similar to the proposed one is constructed based on a mixing property. A simple plant is identified by means of the proposed algorithm for illustration.