Belief-evidence fusion through successive rule refinement in a hybrid intelligent system

A hybrid intelligent system that is able to sucessively refine knowledge stored in its rulebase is developed. The existing knowledge (referred to as belief rules), which may initially be defined by experts in a particular domain, is stored in the form of rules in the rulebase and is refined by comparing it with new knowledge (referred to as evidence rules) extracted from data sets trained under a neural network. Based on measurement, assessment and interpretation of rule similarity, belief rules existing in the rulebase may be found to be confirmed, contradicted or left unsupported by the new training data. New evidence rules may also be discovered from a training data set. This rule comparison is unique in the sense that rules are viewed and compared in geometric manner. As rules evolve in existence in the rulebase during the belief-evidence fusion process, their bounds, strengths and certainties are also revised. The hybrid intelligent system is tested with different data sets, including hypothetical data sets and actual data sets. the system outperforms conventional backpropagation neural learning systems in terms of accuracy and predictability especially when the data is sparse or arrives in bursts, or when the initial knowledge is incorrect. Ordering effects inherent in incremental systems, however, is difficult to address as neural network learning can be unpredictable. The performance of the system increases if predictive classification of unclassified test data is performed