A Novel Modeling Approach for Cumulative Belief Rule-Base With Joint Optimization and Rule Synthesis
针对累积信念规则库系统过度依赖专家经验、规则合成不合理的问题，提出一种联合优化模型和基于领域的合成因子计算方法，实现参数与结构自动优化及规则精简，在稻米口味评估和基准分类任务中平衡了模型复杂度与推理精度。
Cumulative belief rule-based system (CBRBS) is a recent representative of explainable artificial intelligence (XAI). However, the use of CBRBS as XAI still faces many challenges, e.g., over-reliance on expert experience and applying unreasonable rule synthesis in the existing modeling process. Hence, a novel modeling approach is proposed for constructing CBRBS in the aim of providing a better XAI, in which a joint optimization model is proposed first to describe the mathematical model of parameter and structure optimization, and the corresponding algorithm is further designed to automatically achieve the joint optimization of CBRBS. Afterward, a domain-based calculation method of synthesis factor is proposed to develop a new rule synthesis method for CBRBS, which not only achieves the reduction of inefficient and inconsistent rules but also takes into account interpretability and generalization ability. In experimental analysis, the proposed modeling approach is employed to construct CBRBS for handling rice taste assessment and benchmark classification problems. The comparison results show that the proposed approach makes it possible for CBRBS to achieve a good balance between model complexity and inference accuracy. More importantly, the resulting CBRBS has better accuracy and lower complexity than some existing rule-based systems and classical classifiers.