An analysis on parameter optimisation for a belief rules-based classification system

Given the non-linear property of the belief rule-base (BRB) approach, attribute weights (AW), rule weights (RW), and belief degrees of rule consequences (BDRC) should be carefully optimised in advance for a better performance. However, the impact of optimisation on these three types of parameters has not been carefully analysed. In this paper, we analyse the effect of optimisation on three types of parameters, beginning with those calculated by the belief-based method. Specifically, we adopt an archive-based differential evolution algorithm with three objective functions (namely, arithmetic mean accuracy (AMA), cross-entropy (CE), and mean absolute difference (MAD)) for optimisation. Then, we leverage the Wilcoxon signed-rank tests to analyse the benefits obtained from the optimisation. Experiments indicate that optimising these parameters is necessary, and the fitness function should be carefully designed to achieve better performance. Furthermore, CE is the most effective objective function for optimising RW. And, MAD outperforms others for BDRC.