The problem of pattern recognition in stochastic data is approached by relating the characteristics of signals to the properties of filters. The data are restricted to binary quantities so that information measures used in coding theory can be applied. A method is developed for transforming patterns in binary data into a set of coefficients representing crosscorrelations of all orders. These describe a filter or model of the process being investigated, and form the basics for the caloulation of conditional expectations. A laboratory computer has been constructed to analyse patterns in blocks of five binary digits received in parallel from a punched paper tape reader. This computer has been used to test the method experimentally in three rather different pattern recognition problems. The first was prediction on a single Markov time series. The second was adaptive error correction for bock codes. Finally, a process with interacting control variables was optimised experimentally by analysing correlations between the output and perturbations on the controls. In each case the filter or model includes nonlinear relationships. The results of these experimental applications demonstrate the effectiveness and generality of method. The extension of the technique to non-binary data is indicated.
Doctor of Philosophy (Ph.D.)
Electrical and Computer Engineering