The problem of adaptive equalization is classically solved using supervised and unsupervised methods that employ certain classes of moments associated with the signal of interest and the observed signal. Information theoretic learning (ITL) methodologies, however, was responsible for establishing new approaches both to the supervised and the unsupervised cases. These approaches have in common the use of criteria based on the (implicit or explicit) estimation of the probability density functions of the signals of interest, which, in practical scenarios characterized by nonlinearity / nongaussianity, can allow a wider exploration of the available statistical information in the process of optimization of the filter parameters. In this project, two aspects of the problem of ITL-based equalization will be addressed. The first is related to the deconvolution of signals that present statistical dependence between samples (i.e. "colored signals"). The second concerns the extension of the ITL-based equalization paradigm to encompass IIR filters, which will give rise to a general optimal filtering framework and to efficient adaptive methodologies.
News published in Agência FAPESP Newsletter about the scholarship: