Applications of unsupervised learning to some Problems of digital communications

The Bayes optimal m-ary digital communication receiver structure is derived using a recursive formula from the theory of unsupervised learning pattern classification. The receiver structure is optimal for a model which includes intersymbol interference, Markov symbol and noise sequences, and unknown parameters. The optimum receiver is found as a function of the noise density. In the particular case of Gauss-Markov noise, the receiver is shown to consist of (1) discrete-time pre-whitening, (2) correlation, (3) energy correction, (4) expontiation, and (5) delay-feedback "filtering" followed by zero-memory linear operations and minimum selection.