Dynamic rule adjustment approach for optimizing belief rule-base expert system

The belief rule-base (BRB) inference methodology, which uses the evidential reasoning (RIMER) approach, has been widely popular in recent years. As an expert-system methodology using the RIMER approach, BRB is used for storing various types of uncertain knowledge in the form of belief structure. Several structure-learning approaches have been proposed in recent years. However, these approaches are deficient in various aspects, do not have repeatability, hold incomplete data, and are constrained by the associated scale-utility value. Moreover, considering the influence of the number of rules for a BRB system, two scenarios are designed to reveal the relationship between structure feature and fewer/excessive rules. Excessive rules may lead to a BRB that is equipped with an over-complete structure, whereas significantly fewer rules may result in a BRB with an incomplete structure. To solve these problems, we initially proposed to develop an adjusted structure that is leading to the establishment of a complete structure instead of incomplete and over-complete structures. By scenario analysis and experimental verification through parameter learning of BRBs, we summarize several features of two scenarios, which can be used to reveal certain number of key BRB properties. Finally, density and error analyses are introduced to dynamically prune or add rules to construct the complete structure, particularly that of the BRB comprising multiple-antecedent attributes. We verify the effectiveness of the proposed approach by testing its use in a practical case study on oil pipeline-leak detection and demonstrate how the approach can be implemented.