An algorithm for identification of belief measures Wei Chen Computer Science, University of Nebraska at Omaha wchen@mail.unomaha.edu Abstract: Zhengyuan Wang Mathematics, University of Nebraska at Omaha zhenyuanwang@mail.unomaha.edu usually given by experts or obtained from other empirical ways. Incompleteness and inconsistence Identification of a belief measure is to test whether a containing in the estimation of belief measure are the measure by given meet the requirement of belief major problems for its application. As mentioned above, measure, if contraction exists in given measure, then belief measure can be computed from a basic belief measure is generated has small error to given one. probability assignment for every element in the set we An algorithm is developed to search for the best fitting concerned. If we can get all basic probability for every basic probability assignment based on a given “blank” element in the universal set, if we can find such a of belief measure. Through the algorithm, a belief probability assignment, that belief measure on them has measure can be identified if it exists. When the exact no or least error comparing with the estimation of it by identification fails, a genetic algorithm is activated to given , we could avoid or reduce incompleteness and search for a global optimal solution and obtain an inconsistence which come from subjective views and approximate optimal belief measure. other resources. Finding such basic probability assignments is identification of belief measure. By 1. Introduction identification we use a procedure of fitting an optimal consistent belief measure to estimation by given, this Concept of belief measure is introduced in Dempster-shafer (D-S) theory, which provides a method optimal belief measure has as low as possible error to estimation measure from ,for example , experts. for uncertainty inference and information fusion. The The belief measure is one of the most important degree of belief based on belief measure can be got types of fuzzy measures in common use. From from basic probability assignment. Different belief information obtained from the existing experience of measures can be combined to form a new belief experts, can be estimation measure of target attributes measure; this process can be used to combine sets. However, it does not satisfy, usually, the information or called information fusion. Through this requirement being a belief measure. Based on given set combination, the uncertain on the events can be reduced function, we may find an approximate optimal belief [3]. measure through a good search method. There is little In belief measure, a generalization of classical research in this process of information uncertainty. In probability theory that permits the assignment of our algorithm, we first find out whether there is a probability masses (beliefs) to all subsets of the probability assignment meets the requirement. if the universe and not just to the basic elements. A measure estimation of belief measure satisfy the requirement, or of belief for an assertion is then computed as a there is no contradiction within the estimation, the subinterval of [0, 1], where the length of the interval algorithm measures the uncertainty. corresponding probability assignment; if contradiction In real life, the estimation of belief measure is will give positive answer, also give exists, a genetic algorithm is activated to search an approximate solution. In the following section, basic In our case, let the universal set X be a finite set concepts and connection of belief measure and { x1 , x2 ,..., xn }. The belief measure on E j , j=1, 2,…, s, probability assignment are given ( E j ) are give by human experts subjectively. We 2. Background need to find 2n 1 values to fit m( Fk ) a basic probability assignment such that for a belief measure Belˆ( E j ) built on them contribute least sum of error 2.1 Basic Concept of Belief Measure compare with the ( E j ) by given. Ideally, if there exists a probability assignment contributes no error, this assignment is the solution without contradiction, the In this section, we introduce some basic definitions given measure is identified as a belief measure, if we and theorem about belief measure can not find such assignment, and an optimal solution is search to achieve least error . Definition 2.1. Let P (P(X)) be the power set of P(X). If p is a discrete probability measure on (P(X), P (P(X))) with p( ) 0 , then the set of function m: P(X) [0, 1] determined by I 3. Identification of belief measure General structure of identification algorithm is two any E P(X) (2-1) is called a basic probability assignment o P(X). part (see figure 1), one is finding contradiction, if Theorem 2.1. A set function m: P(X) [0, 1] is a basic contradiction does not exist, the solution without error is probability assignment if and only if generated in this part; else a genetic algorithm is m(E) P({E}) for (1) m( ) 0 ; (2) m( E ) 1 ; activated. E X Definition 2.2. If m is a basic probability 3.1 coding the sets assignment on P(X), then the set function Bel: P(X) How the sets present in the algorithm matter a lot. A [0, 1] determined by Bel(E)= m( F ) for any E P(X) F E (2-2) good coding can make algorithm efficient. A binary is called a belief measure on (P(X), P (P(X))), or more string is used in the coding, every bit presents whether exactly, a belief measure induced from m. corresponding element in the universe set appear in Theorem 2.2. If Bel is a belief measure on ( P(X), P subset, the ith binary bit is 1 in its codes; otherwise the bit is set as 0. We use bnbn 1...b1 to present the code (P(X))), then (BM1) (BM2) Bel( )=0 string. After coding, we transform bit string of the set n into a real number as indices of the sets, that is Bel(X)=1; n (BM3) Bel( Ei ) i 1 current set. If the ith element appears in the specific I {1, 2,..., n}, I (1)| I | 1 Bel ( Ei ) i I where { E1,..., En } is any set finite subclass of P(X); Theorem 2.3. The basic probability assignment and n i bk 2 n 1 . For example, if the universe set is k 1 {x1 , x2 , x3 , x4 , x5 , x6 } , a set {x3 , x5 } can be coded as 010100, so it is called set E20 . its corresponding belief measure satisfy condition: m( E ) (1)| E F | Bel ( F ) for any E (X) F E The method of coding helps to find out whether (2-3) Ei E j . The operation is easy to implemented through Where |E-F| denotes the cardinality number of the set (E-F).[2], [3] logic “or”. For example, sets {x1, x2 , x3} and {x2 , x3} are coded as 111 and 110, [(111) or (110)] 2.2 Problem descriptions =(111), so we {x2 , x3} {x1, x2 , x3} . Proposition 3.1 can get conclusion as ( E j ) , j=1, 2,s, are given, Given a universe set X , every element in the power set P(X) Ei X ,if using the coding method Goal function: s f min ( ( E j ) Belˆ( E j )) 2 above and Ei E j then i j where E j P(X), j=1, 2,s Proof: if Ei E j , the bit strings for Ei and E j are si and s j . So there exist at least one element xs Belˆ( E j ) can be computed from Belˆ( E j ) m( F ) F E j contained in E j but not Ei ; accordingly, at least one bit, let say n (3-1) j 1 bs on which is 1 in s j but 0 in si . So n i bk(i ) 2n 1 j bk( j ) 2n 1 . k 1 k 1 This statement says the index of any set is greater than all of its subset. This information can help to locate the subset of any given set. A solution is a set of vector with cardinality number of 2n 1 . We simplify the basic probability assignment m( Ei ) to mi , i=1, …, 2n 1 . We decode the index i to binary vector with dimension n. For example m5 m(101) , the binary vector 101 denotes Ei x1 and x3 .Similarly, m6 m(110) denotes x2 and x3 . Then we will design a genetic n n algorithm for find 2 2 of m( Ei ) , i=1,…, 2 1 , to is consist of 3.2 Finding contradiction As mentioned above, if we can find a set of probability assignment m( Fk ) , k 1... 2n 1 , such that for a belief measure Belˆ( E j ) built on them contribute get optimal (3-1). Step1: randomly generate big population of potential no sum of error compare with the ( E j ) by given, vectors ( m1,..,m2 n 1 ). For every vector, we call it a j 1...2n 1 , we say the there are no contradict between chromosome. Every piece of binary number stored in belief measure. In this situation, an algorithm of the chromosome denote to corresponding value of identification could: 2 n 1 , j=1,…P. P is the number of ( j) pupation. We call vector g i the ith gene on jth mi 1) initialize all basic probability assignment blank or m( Fk ) 0 k 1... 2n 1 2n 1 , update all basic probability assignment with according belief measure. If set Fi 2) From 1 to ( j) i=1,… chromosome, mi ( j ) g i ( j ) / g k ( j ) . All the genes generated must satisfies the restriction, 2 n 1 g i [0,1] is singleton, contains only one element (this is easy ( j) and mi 1 . The length of every chromosome to know when we decode the index of the set), such depends on the accuracy of every gene on it. In our case, as {xs } , else m( Fi ) ( Fi ) m( Fi ) ( Fi ) then F j Fi , j i m( F j ) , and , every gene has same characteristics, it share the same all structure and bits. Then using the position of the gene, m( F j ) , j i , are available, 3) Check n 2 1 whether every m( Fi ) [0,1] j 1 we and m( Fi ) 1 , if it is true, there is no contradiction i 1 among belief measure, and the current assignment is the solution, else there exists contradiction and a genetic algorithm is activate to search an optimal solution. 3.3 Search optimal assignment using genetic algorithm can decode every substring chromosome into a real value in [0,1] for corresponding basic probability assignment. Step 2: According to the goal of optimization, we choose a criteria for computing the goodness of chromosomes. For every chromosome, when generated, can get an evaluation of its goodness using its error to belief measure given. s r 2 ( ( E j r ) Belˆ( E j r )) 2 (3-2) Where Belˆ( E j r ) m ( r ) ( F ) (3-3) j 1 A genetic algorithm is involved in search optimal solution. We use model this problem in the following: of binary on F E j goodness function: Gr step 5. m( ) 2 2 (3-4) r Step 5: control the termination of the search. Although the GA is good for global searching, there is a ( 1 r p ). All situation that the error of best chromosome can not be chromosomes can get Gr [01] and r 2 . We sort all chromosome behaves. If it changes smaller and smaller, chromosomes on their goodness in descending order. So In the end, we decode the chromosome to the vector m( 2 ) Where is min r2 the best result of this generation among the population convergent to 0, we can observe how the error of best then we can exit the search and report the chromosome. we need. This vector is the optimal belief measure. are list on the top, the worst cases are on the button. 4. Simulations and conclusions Step 3: generate new chromosomes on randomly chosen parents. A random number generator is used in this step. For every chromosome with Gr , we can assign 4.1 Some simulations a probability value to it as In real world, the belief measure is given subjectively Pr Gr by experts or other empirical ways. We give sets of , r=1, 2,p (3-5) Gi We usei random number generator to get a value 1 known estimation of belief measures and then use our from 0 to 1. This number will fall into an interval [ Pj , Pj , ], j is the parent chosen. (Let P0 0 and j=1, listed in the Table 1 and 2. Setting situation as: universe set X {x1, x2 , x3} ; in 2,…,p). The generations of new chromosome have 3 case 1 and 2, belief measures are given as in table 1 and ways, crossover, mutation and realignment. All of them 2, the results of identification shows in the tables. j 1 j i 0 i 0 p algorithm to get optimal solutions. The comparisons are are two points’ operators. The crossover means we The results shows the proposed identification randomly choose two positions on the target pair of algorithm can not only find optimal solution, but also chromosomes, the string between two points get can find whether there is contradiction among in exchanged into each other. (See figure 2). The mutation estimation of belief measures by given. is choosing two positions randomly on the target pair of chromosomes and restore the opposite values in 4.2 Conclusions corresponding position, such as change 1 to 0 and 0 to 1. (See figure 3). The realignment is choosing two The identification algorithm proposed in the paper is positions randomly on the target pair of chromosomes, used to get a set of probability assignments by the chromosomes become 3 substrings for each, and we subjective belief measures given. The coding method is relocate these 3 substrings to get a new chromosome. used to present the power set, which is embedded in the (See figure 4). Step 4: examine the new chromosomes and update algorithm and helps to deal with relationship between the population. As mentioned above, every time a pair part is checking existence of contradiction, the second of new chromosomes generate. For each chromosome, part is genetic algorithm activated to find optimal corresponding error and goodness function is computed. solution if contradiction exists. Without contradiction, We use these values to resort the population. If new the algorithm is very fast, but genetic algorithm search chromosome is better than the worst one on the button takes time. The reason to use it is that genetic algorithm of the list, it will insert into the list, otherwise the new is global search, if time is enough, the solution found is chromosome is removed. If we can get an error of global optimal. the sets. The algorithm is composed of two parts; first chromosome equal to 0, then we can direct report this Due to the rareness of previous work having done in chromosome and exit the searching. Otherwise go to the this field, there are further improvement can be done, a lot of search methods and soft computing methods can Set be used. When the problem of identification of belief measure is solved, the belief measure get more abroad application in information fusion, decision making and other fields related to uncertainty management. {x1} {x2} {x1, x2} {x3} {x1, x3} {x2, x3} {x1, x2, x3} given by Probability by cmp Bel cmp 0.1 0.2 0.2 0.3 0.5 0.9 1 0.2 0.1 0.2 0.3 0.05 0.1 0.05 0.1 0.1 0.2 0.3 0.05 0.1 1.0 by contradictio n No Error=0 Table 1. In case 1, the belief measures are given without contradiction, the result of identification algorithm shows error is 0. Set Figure 1. Structure of identification algorithm Probability by cmp Bel by cmp contradiction given {x1} 0.2 0.069745 0.0697 yes {x2} 0.1 0.16503 0.1650 {x1, x2} 0.5 0 0.2348 {x3} 0.3 0.33399 0.3340 {x1, x3} 0.55 0.061886 0.4656 {x2, x3} 0.5 0.36935 0.8684 {x1, x3} 1 0 1.0000 x2, by Error=0.0067 Table 2. In case 2, the belief measures are given with contradiction, the result of identification algorithm shows error is small Reference [1] Z. Wang and G. J. Klir, Fuzzy Measure Theory, Plenum Press, New York, 1992. Figure 2. Crossover [2] S. T. Wierzchon, “An algorithm for identification of fuzzy measure”, Fuzzy Sets and Systems 9 (1983) 69-78. [3] G. Shafer, A Mathematical Theory of Evidence, Princeton University Press, Princeton (1976) Figure 3. Mutation Figure 4. Realignment