A hybrid intelligent architecture for revising domain knowledge

This dissertation introduces a novel Hybrid Intelligent Architecture (HIA) that augments a knowledge-based system with a connectionist model and a statistical model to help the former to refine its domain knowledge. Key innovations of the introduced HIA include its capability to learn fast, refine prior domain knowledge, enhance its generalization capability, and support its output decisions with the required explanation power. These abilities are made possible by efficiently incorporating the features of seven main modules in the structure of HIA.;The first module is a knowledge-based module that represents initial domain knowledge acquired from domain experts in a well defined format. The second is a statistical module that analyses available data sets and extracts correlation rules to supplement the extracted prior domain knowledge. The third is a mapping module where all extracted prior domain knowledge are mapped into a uniform initial connectionist architecture. This mapping module is also able to bound the search space for the optimal parameters of the connectionist architecture before training. Therefore, the learning time required to train the mapped initial architecture is reduced. The fourth module of HIA is a discretization module used as a front-end to the connectionist module, the fifth module of HIA, to discretize its inputs into multi-interval inputs. In the training phase of the connectionist module, an augmented backpropagation algorithm is developed to provide the system with the ability to refine its input characterization parameters while updating and learning new domain concepts. The sixth module of HIA is a rule extraction module. In this module, the topology of the trained connectionist architecture, that represents the refined domain knowledge and new learned concepts, is mapped into an updated rule-based system. This dissertation introduces three new rule extraction techniques. The suitability of each technique depends on the network architecture, nature of inputs, and the transparency level of the required explanation power. The last module of HIA is an integration module that integrates the outputs of the extracted rules and the trained connectionist architecture and provides the final explained and revised outputs of the system.;The introduced HIA concepts were applied to a real control problem and three classification problems. Experimental results show that a connectionist network generated by HIA generalizes better and faster than other connectionist architectures. Moreover, HIA shows its capability to refine prior domain knowledge and extract updated and new domain concepts. Furthermore, it supports its output decisions with a powerful rule-based explanation utility even in cases where prior domain knowledge is not available.