An Improved Method Of Convex Evidence Theory

There are a lot of uncertain and inconsistent information fusion problems in real world todetermine. DS evidence theory is the most common method to deal with such problems, hasbeen widely applied in many fields such as Agriculture, Education, Health Care, Military, etc.But with the increase in application and research, DS evidence theory gradually exposes someproblems. When dealing with high-conflict evidence, DS evidence theory has paradoxes, andthe fusion results contrary to people’s intuitive sense. Also the fusion results in dealing withordered proposition fusion are often not recognized by domain experts. Obviously thelimitations of evidence theory fusion rules emerge in this kind of application. The convexevidence theory is proposed to solve this problem. This paper is an in-depth analysis of theshortcomings of the original convex evidence theory, and proposes appropreate improvementprogram. Specifically, this paper includes the following:This paper analyzes the original convex evidence theory’s characteristics, combined withpractical examples analyzes the cause of fusion results do not meet people’s intuitive sense.Also it analyzed the advantages and disadvantages of Gaussian convex evidence theory, alsogive some examples. Then this paper clearly proposes that it’s better to divide the process intotwo steps: g-value’s calculation and redistribution of mass, and then it proposes improvementprograms for the two steps. For g-value’s calculation, this paper first proposed thecharacteristic-value method. This method abandoned centroid-method used in the originalconvex function evidence theory. First of all, it calculates the basic trust function’scharacteristic-value by mass sum and convexity. Then it finds the position’s index of max value.Last it calculates g-value with the weighted sum of two indexes, and the weight here for each isjust the basic trust function’s characteristic-value. This method takes three factors into account:the max value’s index and the sum of basic trust functions as well as its convexity. Thusg-value we got is more convincing in a practical sense. For distribution of mass, the paper putsforward the concept of initial distribution. This will actually turn the process into initialdistribution and redistribution. The initial distribution ensures that the results of fusion get fitraw data as possible. Redistribution ensures the total amount of two basic trust functions.Under the combined effect of these two improved steps, the fusion results have also beensignificantly improved.This article also gives some other improvement ideas. For g-value’s calculation, the paperalso presents large-numbers-centroid method. The method combines features of the originalconvex evidence theory’s method and characteristics-value method. For the redistribution ofmass, the paper also presents a multiplication-increases method, a natural-number-prescribing method and an arithmetic-sequence method, these methods can not compare which one isbetter, because they are based on different application requirements, each proposed has certaincharacteristics.This paper also presents some assisted algorithms commonly used in convex evidencetheory, such as a test method of whether basic trust function having convex nature, a generatingmethod of basic trust functions and a test framework of whether all fusion results calculated bya model having convex nature. These methods actually play a key role in researching theconvex evidence theory. Such as the generating method can give all the basic trust functions intheory, when combined it with the method of testing convex nature, we get a framework cantest for the fusion results.