Recognition of brain states and subject's intention from electroencephalogram (EEG) is a challenging problem for brain-computer interaction. Signals recorded from each of EEG electrodes represent noisy spatio-temporal overlapping of activity arising from very diverse brain regions. However, un-mixing methods such as Cortical Current Density (CCD) can be used for estimating activity of different brain regions. These methods not only improve spatial resolution but also signal to noise ratio, hence the classifiers computed using this activity may ameliorate recognition performances. However, these methods lead to a multiplied number of channels, leading to the question -- ``How to choose relevant and discriminant channels from a large number of channels?''. In the current paper we present a channel selection method and discuss its application to the recognition of anticipation related potentials from surface EEG channels and CCD estimated cortical potentials. We compare the classification accuracies with previously reported performances obtained using Cz electrode potentials of 9 subjects (3 experienced + 6 naive). As hypothesised, we observed improvements for most subjects with channel selection method applied to CCD activity as compared to surface-EEG channels and baseline performances. This improvement is particularly significant for subjects who are naive and did not show a clear pattern on ERP grand averages.