Belief rule based expert system for classification problems with new rule activation and weight calculation procedures

Classification problems are significant because they constitute a meta-model for multiple theoretical and practical applications from a wide range of fields. The belief rule based (BRB) expert system has shown potentials in dealing with both quantitative and qualitative information under uncertainty. In this study, a BRB classifier is proposed to solve the classification problem. However, two challenges must be addressed. First, the size of the BRB classifier must be controlled within a feasible range for better expert involvement. Second, the initial parameters of the BRB classifier must be optimized by learning from the experts’ knowledge and/or historic data. Therefore, new rule activation and weight calculation procedures are proposed to downsize the BRB classifier while maintaining the matching degree calculation procedure. Moreover, the optimal algorithm using the evidential reasoning (ER) algorithm as the inference engine and the differential evolution (DE) algorithm as the optimization engine is proposed to identify the fittest parameters, including the referenced values of the antecedent attributes, the weights of the rules and the beliefs of the degrees in the conclusion. Five benchmarks, namely, iris, wine, glass, cancer and pima, are studied to validate the efficiency of the proposed BRB classifier. The result shows that all five benchmarks could be precisely modeled with a limited number of rules. The proposed BRB classifier has also shown superior performance in comparing it with the results in the literature.