Adaptive multi-rate systematic models for data classification

During the past decade, there has been a tremendously increasing need for the development of intelligent systems. Generally, the more intelligence that is required of a system, the more sophisticated and complex the system needs to be. In addition, other non-linearity and uncertainty may be introduced into the systems. In order to predict the behavior of these systems prior to their development, approximate models. The concept of fuzzy logic can be effectively used in modeling are beneficial, nonlinear processes. The fuzzy logic modeling methodology is based on the idea that a nonlinear process can be represented in a small region by a simple mathematical input-output relationship that holds primarily in that region.; This thesis considers the architecture of learning systems that can explain their decisions through a rule-based knowledge representation algorithm. Two problems in learning are addressed: data classification and function approximation. A multi-rate data classifier for discrete-valued problems is developed. The system utilizes an information-theoretic algorithm for constructing informative rules from the data. These rules are then used to construct a computational network to perform parallel inference and posterior probability estimation. The network can be incrementally extended—new data can be incorporated without repeating the training on previous data.; In this dissertation, flexible adaptive multi-rate systematic models for identification and classification are developed, and application specific modules are constructed. The use of these modules is explained and demonstrations are presented. In addition, the learning rules are generated and adapted to train the fuzzy neural network modules using the generalized delta rule method. The results developed here are applied to the problem of speaker independent speech recognition. In this application, the system recognizes and classifies speech segments. The classification problem is considered difficult due to the variations in voice frequency and pronunciation from one speaker to another. Incorporating the fuzzy logic theory into the neural networks for this application shows a reasonable improvement in performance over conventional modeling identification systems that also perform well. These systematic data classification algorithms are considered equivalent in performance for this application.