Sparse probabilistic classification models for nonlinear channel equalization

We propose a sparse probabilistic learning approach for nonlinear channel equalization in wireless communication systems, by using the relevant vector machine (RVM) technique. In particular, we propose two versions of the RVM based equalizer: 1) maximum a posterior RVM (MAP-RVM), 2) marginalized RVM (MRVM). Compared to the standard support vector machine (SVM) method, the proposed RVM approach not only provides a better performance by introducing probabilistic criteria to obtain the equalizer´s parameters, but also requires a much lower detection complexity by obtaining much sparser models. The RVM based equalizers also achieve nearly the same performance as the optimal Bayesian detector. Furthermore, the RVM based equalizers are shown to be robust to time variations of channels.