冲突证据融合的优化方法

多数研究者认为, 用修改数据模型(证据体)的方法来解决冲突证据组合问题较为合理. 然而, 已有的基于修改数据模型的方法仅考虑如何提高冲突证据组合结果的聚焦程度. 实际上, 它们并没有考虑如何通过修正来消减证据之间的冲突. 显然, 若融合结果由冲突证据组合得到, 那么其可信性必然较低且会给随后的融合过程带来较大的风险. 针对此问题, 沿用折扣系数法, 该文基于证据距离准则提出了一种折扣系数(可靠度)优化学习模型, 优化过程同时考虑提高聚焦程度和消减冲突, 通过使折扣修正后组合结果的基本概率赋值(Basic probability assignment, BPA)与直言BPA (Categorical BPA, CBPA)之间的距离最小来寻优, 其中证据可靠度大小的序关系作为约束条件, 它依据证据的虚假度确定. 典型算例验证了所提方法比现有的一些组合方法, 在聚焦能力和冲突消减两方面都更合理.

An Optimal Method for Combining Conflicting Evidences

Most researchers hold that revising mass function based methods are reasonable to deal with the problem of conflicting evidence combination. However, the existing methods of revising mass function only consider improving focusing degree of combination results. Actually, they did not effectively reduce conflict among evidences by revision. Obviously, the fusion result of conflicting evidences has low credibility and will certainly bring risks to subsequent fusion process. To solve this problem, by adopting the idea of discounting, this paper proposes an optimal model to learn discounting factors (reliability) based on evidence distance criterion which considers improving focusing degree and reducing conflict simultaneously. The procedures of optimization are achieved through minimizing the distance between combined basic probability assignment (BPA) of revised mass function and categorical BPA (CBPA). The permutation of reliabilities associated with evidences, which is regarded as constraint condition, is determined according to their falsity. Typical examples illustrate that the presented method is more reasonable than some existing methods both in reducing conflict and improving focusing degree.