Stochastic complexity-based model selection with false alarm rate control in optical spectroscopy

Abstract

Stochastic complexity-based penalization criteria can prove efficient and robust in spectroscopy applications for unsupervised identification and concentration estimation of spectrally interfering chemical components. It is shown here how the so-called Normalized Maximized Likelihood (nMDL) introduced in [17] can be tailored to provide control of the detection performances in terms of probability of false alarm. Numerical experiments conducted on realistic simulated optical spectroscopy signals evidence that the nMDL approach outperforms standard information criteria in terms of model selection performances. Moreover, the ability to control false alarm rates with the proposed modified nMDL criterion is demonstrated on simulations.