Browse Prior Art Database

Sequential Pattern Recognition Machine

IP.com Disclosure Number: IPCOM000079287D
Original Publication Date: 1973-Jun-01
Included in the Prior Art Database: 2005-Feb-26
Document File: 3 page(s) / 49K

Publishing Venue

IBM

Related People

Hopkins, WD: AUTHOR [+3]

Abstract

Detection and classification of patterns which are characterized by unknown parameters and state variables, must be accomplished in conjunction with a method for estimating those quantities. Further, it is often desirable to perform this combined task in a sequential fashion. In general, the estimation and recognition must be accomplished in the presence of nonstationary observation noise.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 53% of the total text.

Page 1 of 3

Sequential Pattern Recognition Machine

Detection and classification of patterns which are characterized by unknown parameters and state variables, must be accomplished in conjunction with a method for estimating those quantities. Further, it is often desirable to perform this combined task in a sequential fashion. In general, the estimation and recognition must be accomplished in the presence of nonstationary observation noise.

The present technique makes use of a feature of sequential filtering algorithms or devices, which heretofore has been considered as detrimental. Specifically, the technique uses the property of divergence of the innovations produced by sequential filters, in order to solve pattern recognition problems. This divergence occurs When the model assumed in the construction of a filter differs from that model which generates the observations. Based on this idea, a pattern recognition problem can be solved by passing a sequence of observations, made on a pattern, through an optimum (Kalman-Bucy) linear filter, which is constructed by assuming that the observations were generated by an arbitrary one-of-N possible dynamic systems. The divergence of the resulting innovations sequence is monitored, in order to determine if the observations are consistent with the model assumed in the construction of the filter. The decision- making process is based on the behavior of the innovations relative to a prescribed set of thresholds.

Assume a sequence {X(i)} of vectors generates a sequence {Y(i)} of vector observations of the form Y(i) = g(i) (X(i), t(i), alpha) + n(i) (1).

Where the noise, eta(i), is zero-mean and white, with covariance matrix R(i), and t(i) is the observation time for the i/th/ measurement. The sequence, {X(i)}, of state vectors is assumed to be generated by one-of-N possible processes of the form (j)

X(i) = f(i) (t(i), X(i-1), alpha) ; j = 1, 2, ---, N (2) where alpha is an unknown constant parameter vector. In order to solve the pattern recognition problem, it is required to determine the value of
j.

The technique for solving this problem is to repeatedly assume different values of j and test the alternative hypotheses H(o) : X(i) = f(i)/(j)/ ( t(i), X(i-1), alpha )(3) H(i) : X(i) not = f(i) ( t(i), X(i-1) alpha ).

The procedure for testing the hypotheses (3) is to assume H(o) true, for the chosen value of j. Under this assumption, and with any appropriate linearizations of (1) and (2), each vector of the observation sequence {Y(i)} is processed through a sequential linear filter, in order to estimate X(m) and alpha. For any specified set of observations, the innovations (equation (9)) are collected, and a vector of these innovations is formed. The norm of...