The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical Engineering University of Southern California

Technical Report USC-SIPI-361

“Rule Based Systems for Data Processing”

by Hongwei Wu

August 2004

This thesis is fucused on fuzzy set and fuzzy logic theories for uncertainty modeling and processing. We examine ten of the linguistic connector word models in Mamdani fuzzy logic rule-based systems (FLRBS), so that a FLRBS structure can be established to simultaneously model and process uncertainties in data, descriptive and connector words. Computational complexity problem of interval type-2 FLRBSs is addressed. Inner- and outer-bound sets are derived for the type-reduced set. A design method is then proposed to incorporate these bound sets, so that the resulting interval type-2 FLRBS can operate without type-reduction, and without performance degeneration. Theories of FLRBSs are applied to the acoustic feature based classification of ground vehicles. To evaluate the fuzzy logic rule-based classifiers in an appropriate way, a Bayesian classifier is also constructed, and experiments are conducted to compare their performances. It was observed during classification experiments that a simple majority voting adaptive operational mode can improve the classification performance. This leads us to analyze the majority voting based spatial and spatio-temporal decision fusion. For spatial decision fusion, lower and upper bounds for the accuracy rate of the fused decision are provided, and are applied to inverse problems. For spatio-temporal decision fusion, three strategies are proposed to implement the majority vote, and are compared in terms of implementation costs and accuracy rate of the fully-fused decision. The studies described herein let us conclude that:

1. FLRBSs provide a natural framework to incorporate both objective and subjective knowledge, and to process uncertainties in data, descriptive and connector words.

2. Interval type-2 FLRBSs have potentials to be used in real-time applications.

3. FLRBSs can be applied for pattern classification problems, especially when 1) there is direct correspondence from sub-categories to fuzzy logic rules, 2) there are feature uncertainties1q within each sub-category, and 3) both subjective knowledge and training samples are available for classifier optimization. 4. Majority vote is a simple and efficient way for decision fusion. It can be implemented in various strategies in the spatio-temporal fusion scenario, and both implementation cost and performance issues should be considered when making choices among these strategies.

To download the report in PDF format click here: USC-SIPI-361.pdf (10.1Mb)