Hybrid Neural Network/Knowledge Base Automated Classification System

We propose to design a hybrid neural network and knowledge base automatic classification system which dynamically constructs an optimal decision tree. Our proposed hybrid system offers several advantages. Expert system knowledge bases will be used to construct the decision hierarchy with neural network modules acting as classifiers at the nodes of the tree. Neural networks such as Fuzzy ARTMAP can be trained with ease and little computational effort. The proposed research extends the concept of self-organization from within the neural network outward to the overall structure of the decision hierarchy to support dynamic construction of an optimal decision tree. We will analyze the feasibility of the approach via demonstration and comparison with a previously developed ATR system. The military significance of our Phase I work includes adaptive and robust classification of potential targets given large quantities of features from preprocessed sensor data. The commercial application will lead to a general supervised classification system capable of rapid decision making and dynamic reorganization of its internal structure. In addition to the planned use in our own intelligent agent software, applications also exist for the technology to speech processing and medical imaging systems. We also see commercial software applications via the development of a neural network/expert systems shell targeted to aerospace, electrical, mechanical and civil engineering.