Model-based pattern analysis of signals using frames
Ho Soo Lee · 1985
A model-based system for analyzing signals is described, where signals are hierarchically decomposed to produce hierarchical AND/OR graphs, and frames are employed as vehicles for conveying knowledge of signals. Children nodes of an AND node in hierarchical graphs are not completely independent as can be seen in 'nearly decomposable problems'. The basic principle for the search mechanism resembles the procedure for locating a specific spot on a map by a human in that global constraints are exploited first for exploring primary features which play important roles in perception, and local tests are conducted from these features. Both data-driven and model-driven hypothesis formation strategies are used in a mixed fashion: the former is invoked in the initial stage or when a process for instantiating expectations generated by prototypes halts, and the latter is used for generating expectations. The task of analyzing signals is achieved by maintaining global compatibility based on features which have been confirmed or already hypothesized. The compatibility with previously instantiated features is examined by evaluating the scoring function which utilizes a set of constraints among features. While searching, two kinds of pruning activities are in effect, front-end filtering and cutoff pruning. Ad hoc hypothesis modification networks are devised to facilitate dynamic hypothesis modifications when a hypothesized feature would not be instantiated, which corresponds to one of plausible prototypes. In this research, two supporting techniques for the analysis process have been developed: an algorithm for extracting a set of significant points from curves, and a pattern classification technique, called 'relation-based correlation', where global relations among features are used to check the global compatibility. The proposed methodology for analyzing signals has been implemented in Franz Lisp running under Eunice.