Universal Detection for Finite-State Channels and a Corrupted Message Set

We consider universal detection procedures for finite-state channels, when the message set is known to us only through a set of "training sequences," each being an output vector of some finite-state channel which is fed with the corresponding message. The message set is a collection of vectors, selected by Nature from a finite state probabilistic source. Although the probability measures of the source and the conditional probability measures of the channels are not known, a universal-decoding algorithm can achieve an error probability with an error exponent that for long enough messages, is equal to the error-exponent which is associated with the case where the probabilistic characterization of the source and both channels are fully known.