ABSTRACT
Hybrid Intelligent Architectures synergistically combine the strengths of diverse computational intelligence paradigms and avail of both domain knowledge and training data to solve difficult learning tasks. In particular, several researchers have studied some aspects of combining symbolic and neural/connectionist approaches, such as initializing a network based on existing rules, or extracting rules from trained neural networks. In this chapter, we present a complete system that embeds initial domain knowledge and/or statistical information into a custom neural network, refines this network using training data, and finally extracts back refined knowledge in the form of a refined rule base with an associated inference engine. Two successful applications of this hybrid architecture are described.
