Is training of adaptive equalizers still useful?

We present a novel unsupervised adaptive equalizer. It has the same computational complexity, convergence speed and steady-state MSE as a trained LMS adaptive DFE, but it is not subject to error propagation. Therefore it can equalize even severe and/or quickly varying channels. This follows from the very structure of the equalizer, which allows a completely reversible transition between (i) a linear structure in the starting mode: the decoupled cascade of a recursive adaptive predictor and a transversal phase equalizer and (ii) a classical DFE in the tracking mode. The equalizer behaviour is fully satisfactory during hours of real underwater communications. It reaches the standard of trained equalizers. Hence the question in the title