Is Entropy Enough to Evaluate the Probability
Transformation Approach of Belief Function?
Deqiang HanInstitute of Integrated Automation,
Xiβan Jiaotong University,Xiβan,Shaanxi, China, 710049.
Jean DezertONERA
The French Aerospace Lab29 Av. Division Leclerc, 92320 Chatillon, France.
[email protected] Han, Yi Yang
Institute of Integrated Automation,Xiβan Jiaotong University,
Xiβan,Shaanxi, China, [email protected]; [email protected]
Abstract β In Dempster-Shafer Theory (DST) of ev-idencee and transferable belief model (TBM), the prob-ability transformation is necessary and crucial fordecision-making. The evaluation of the quality of theprobability transformation is usually based on the en-tropy or the probabilistic information content (PIC)measures, which are questioned in this paper. Anotheralternative of probability transformation approach isproposed based on the uncertainty minimization to ver-ify the rationality of the entropy or PIC as the evalua-tion criteria for the probability transformation. Accord-ing to the experimental results based on the comparisonsamong different probability transformation approaches,the rationality of using entropy or Probabilistic Infor-mation Content (PIC) measures to evaluate probabilitytransformation approaches is analyzed and discussed.
Keywords: TBM, uncertainty, pignistic probabilitytransformation, evidence theory, decision-making.
1 IntroductionEvidence theory, as known as Dempster-Shafer Theory(DST) [1, 2] can reason with imperfect information in-cluding imprecision, uncertainty, incompleteness, etc.It is widely used in many fields in information fusion.There are also some drawbacks and problems in evi-dence theory, i.e. the high computational complexity,the counter-intuitive behaviors of Dempsterβs combi-nation rule and the decision-making in evidence the-ory, etc. Several modified, refined or extended mod-els were proposed to resolve the problems aforemen-tioned, such as transferable belief model (TBM) [3] pro-posed by Philippe Smets and Dezert-Smarandache The-ory (DSmT) [4] proposed by Jean Dezert and FlorentinSmarandache, etc.
The goal of uncertainty reasoning is the decision-making. To take a decision, the belief assignment val-ues for a compound focal element should be at firstassigned to the singletons. So the probability transfor-mation from belief function is crucial for the decision-making in evidence theory. The research on probabilitytransformation has attracted more attention in recentyears.
The most famous probability transformation in evi-dence theory is the pignistic probability transformation(PPT) in TBM. TBM has two levels including credallevel and pignistic level. At the credal level, beliefsare entertained, combined and updated while at thepignistic level, the PPT maps the beliefs defined onsubsets to the probability defined on singletons, then aclassical probabilistic decision can be made. In PPT,belief assignment values for a compound focal elementare equally assigned to the singletons belonging to thefocal element. In fact, PPT is designed according toprinciple of minimal commitment, which is somehowrelated with uncertainty maximization. But the goalof information fusion at decision-level is to reduce theuncertainty degree. That is to say more uncertaintymight not be helpful for the decision. PPT uses equalweights when splitting masses of belief of partial un-certainties and redistributing them back to singletonsincluded in them. Other researchers also proposed somemodified probability transformation approaches [5β13]to assign the belief assignment values of compound fo-cal elements to the singletons according to some ra-tio constructed based on some available information.The typical approaches include the Sudanoβs probabili-ties [8] and the Cuzzolinβs intersection probability [13],etc. In the framework of DSmT, another probabilitytransformation approach was proposed, which is calledDSmP [9]. DSmP takes into account both the values of
the masses and the cardinality of focal elements in theproportional redistribution process. DSmP can also beused in Shaferβs model within DST framework.
In almost all the research works on probability trans-formations, the entropy or Probabilistic InformationContent (PIC) criteria are used to evaluate the prob-ability transformation approaches. Definitely for thepurpose of decision, less uncertainty should be betterto make a more clear and solid decision. But does theprobability distribution generated from belief functionswith less uncertainty always rational or always be ben-efit to the decision? We do not think so. In this pa-per, an alternative probability transformation approachbased on the uncertainty minimization is proposed.The objective function is established based on the Shan-non entropy and the constraints are established basedon the given belief and plausibility functions. The ex-perimental results based on some provided numericalexamples show that the probability distributions gen-erated based on the proposed alternative approach havethe least uncertainty degree when compared with otherapproaches. When using the entropy or PIC to evalu-ate the proposed probability transformation approach,the probability distribution with the least uncertaintyseemingly should be the optimal one. But some riskyand strange results can be derived in some cases, whichare illustrated in some numerical examples. It can beconcluded that the entropy or PIC, i.e. the uncertaintydegree might not be enough to evaluate the probabilitytransformation approach. In another word, the entropyor PIC might not be used as the only criterion to makethe evaluation.
2 Basics of evidence theory andprobability transformation
2.1 Basics of evidence theory
In Dempster-Shafer theory [2], the elements in theframe of discernment (FOD) Ξ are mutually exclusive.Define the function π : 2Ξ β [0, 1] as the basic prob-ability assignment (BPA, also called mass function),which satisfies:β
π΄βΞπ(π΄) = 1, π(β ) = 0 (1)
Belief function and plausibility function are definedrespectively in (2) and (3):
π΅ππ(π΄) =β
π΅βπ΄π(π΅) (2)
ππ(π΄) =β
π΄β©π΅ β=β π(π΅) (3)
and Dempsterβs rule of combination is defined as fol-lows: π1,π2, ...,ππ are πmass functions, the new com-
bined evidence can be derived based on (4)
π(π΄) =
β§β¨β©0, π΄ = β ββ©π΄π=π΄
β1β€πβ€π
ππ(π΄π)ββ©π΄π β=β
β1β€πβ€π
ππ(π΄π), π΄ β= β (4)
Dempsterβs rule of combination is used in DST toaccomplish the fusion of bodies of evidence. But the fi-nal goal of the information fusion at decision-level is tomake the decision. The belief function (or BPA, plausi-bility function) should be transformed to the probabil-ity, before the probability-based decision-making. Al-though there are also some research works on makingdecision directly based on belief function or BPA [14],probability-based decision methods are the develop-ment trends of uncertainty reasoning and theories [15].This is because the two-level reasoning and decisionstructure proposed by Smets in his TBM is appealing.
2.2 Pignistic transformation
As a type of probability transformation approach, theclassical pignistic probability in TBM framework wascoined by Philippe Smets. TBM is a subjective andnon probabilistic interpretation of evidence theory. Itextends the evidence theory to the openβworld propo-sitions and it has a range of tools for handling belieffunctions including discounting and conditioning, etc.At the credal level of TBM, beliefs are entertained, com-bined and updated while at the pignistic level, beliefsare used to make decisions by transforming beliefs toprobability distribution based on pignistic probabilitytransformation (PPT). The basic idea of the pignistictransformation consists in transferring the positive be-lief of each compound (or nonspecific) element onto thesingletons involved in that element split by the cardi-nality of the proposition when working with normalizedBPAs.Suppose that Ξ = {π1, π2, ..., ππ} is the FOD. The
PPT for the singletons is illustrated as follows [3]:
BetPπ(ππ) =β
ππβπ΅, π΅β2Ξ
π(π΅)
β£π΅β£ (5)
where 2Ξ is the power set of the FOD. Based on thepignistic probability derived, the corresponding deci-sion can be made.But in fact, PPT is designed according to the idea
being similar to uncertainty maximization. In general,the PPT is just a simple averaging operation. The massvalue is not assigned discriminately to the different sin-gletons involved. But for information fusion, the aim isto reduce the degree of uncertainty and to gain a moreconsolidated and reliable decision result. The high un-certainty in PPT might be not helpful for the decision.Several researchers aim to modify the traditional PPT.Some typical modified probability transformation ap-proaches are as follows.
1) Sudanoβs probabilities: Sudano [8] proposedsome interesting alternatives to PPT denoted by PrPl,PrNPl, PraPl, PrBel and PrHyb, respectively. Sudanouses different kinds of mappings either proportional tothe plausibility, to the normalized plausibility, to allplausibilities and to the belief, respectively or a hybridmapping.
2) Cuzzolinβs intersection probability: In theframework of DST, Fabio Cuzzolin [13] proposed an-other type of transformation. From a geometric inter-pretation of Dempsterβs combination rule, an intersec-tion probability measure was proposed from the propor-tional repartition of the total non specific mass (TNSM)by each contribution of the non-specific masses involvedin it.
3) DSmP: Dezert and Smarandache proposed theDSmP as follows: Suppose that the FOD is Ξ ={π1, ..., ππ}, the DSmPπ(ππ)can be directly obtained by:
DSmPπ(ππ) = π({ππ}) + (π({ππ}) + π)β
(β
πβ2Ξ
ππβπβ£πβ£β₯2
π(π)βπ β2Ξ
πβπβ£π β£=1
π(π )+πβ β£πβ£ ) (6)
In DSmP, both the values of the mass assignment andthe cardinality of focal elements are used in the pro-portional redistribution process. DSmP does an im-provement of all Sudano, Cuzzolin, and BetP formulas,in the sense that DSmP mathematically makes a moreaccurate redistribution of the ignorance masses to thesingletons involved in ignorance. DSmP works in boththeories: DST and DSmT as well.
There are still some other definitions on modifiedPPT such as the iterative and self-consistent approachPrScP proposed by Sudano in [5] and a modified PrScPin [12]. Although the approaches aforementioned aredifferent, all the probability transformation approachesare evaluated based on the degree of uncertainty. Lessuncertainty means that the corresponding probabilitytransformation result is better. According to such aidea, the probability transformation approach shouldattempt to enlarge the belief differences among all thepropositions and thus to derive a more reliable decisionresult. Is this definitely rational? Is the uncertaintydegree always proper or enough to evaluate the prob-ability transformation? In the following section, someuncertainty measures are analyzed and an alternativeprobability transformation approach based on uncer-tainty minimization is proposed to verify the rationalityof the uncertainty degree as the criteria for evaluatingthe probability transformation.
3 An alternative probabilitytransformation based on un-certainty minimization
3.1 Evaluation criteria for probabilitytransformation
The metrics depicting the strength of a critical decisionby a specific probability distribution are introduced asfollows:1) Normalized Shannon entropySuppose that ππ is a probability distribution, where
π β Ξ, β£Ξβ£ = π and the β£Ξβ£ represents the cardinalityof the FOD Ξ. The evaluation criterion for the proba-bility distribution derived based on different probabilitytransformation is as follows [12].
EH =
β βπβΞ
ππ log2(ππ)
log2 π(7)
The dividend in (7) is the Shannon entropy and the di-visor in (7) is maximum value of the Shannon entropyfor {ππβ£π β Ξ},β£Ξβ£ = π . Obviously πΈπ» is normalized.The larger the πΈπ» is, the larger the degree of uncer-tainty is. The less the πΈπ» is, the less the degree of un-certainty is. When πΈπ»= 0, there is only one hypothesishas a probability value of 1 and the rest has 0, the agentor system can make decision correctly. When πΈπ»= 1,it is impossible to make a correct decision, because allthe ππ,βπ β Ξ are equal.2) Probabilistic Information ContentProbabilistic Information Content (PIC) criterion is
an essential measure in any threshold-driven automateddecision system. A PIC value of one indicates the totalknowledge to make a correct decision.
PIC(π ) = 1 +1
log2 πβ βπβΞ
ππ log2(ππ) (8)
Obviously, PIC = 1 β EH. The PIC is the dual ofthe normalized Shannon entropy. A PIC value of zeroindicates that the knowledge to make a correct deci-sion does not exist (all the hypotheses have an equalprobability value), i.e. one has the maximal entropy.As referred above, for information fusion at decision-
level, the uncertainty seemingly should be reduced asmuch as possible. The less the uncertainty in prob-ability measure is, the more consolidated and reliabledecision can be made. Suppose such a viewpoint isalways right and according to such an idea, an alter-native probability transformation of belief function isproposed.
3.2 Probability transformation of belieffunction based on uncertainty min-imization
To accomplish the probability transformation, the be-lief function (or the BPA, the plausibility function)
should be available. The relationship between the prob-ability and the belief function are analyzed as follows.Based on the viewpoint of Dempster and Shafer, the
belief function can be considered as a lower probabil-ity and the plausibility can be considered as an upperprobability. Suppose that ππ β [0, 1] is a probabilitydistribution, where π β Ξ. For a belief function definedon FOD Ξ, suppose that π΅ β 2Ξ, the inequality (9) issatisfied:
π΅ππ(π΅) β€β
πβπ΅ππ β€ ππ(π΅) (9)
This inequality can be proved according to the proper-ties of the upper and lower probability.Probability distributions β¨ππβ£π β Ξβ© also must meet
the usual requirements for probability distributions, i.e.{0 β€ ππ β€ 1,βπ β Ξβ
πβΞ ππ = 1(10)
It can be taken for granted that there are several proba-bility distributions {ππβ£π β Ξ} consistent with the givenbelief function according to the relationships defined in(9) and (10). This is a multi-answer problem or one-to-many mapping relation. As referred above, the proba-bility is used for decision, so the uncertainty seeminglyshould be as little as possible. We can select one prob-ability distribution from all the consistent alternativesaccording to the uncertainty minimization criterion anduse the corresponding probability distribution as the re-sult of the probability transformation.The Shannon entropy is used here to establish the
objective function. The equations and inequalities in(9) and (10) are used to establish the constraints. Theproblem of probability transform of belief function hereis converted to an optimization problem under con-straints as follows:
Min{ππβ£πβΞ}
{β β
πβΞ
ππ log2(ππ)
}
π .π‘.
β§β¨β© π΅ππ(π΅) β€ βπβπ΅ ππ β€ ππ(π΅)
0 β€ ππ β€ 1,βπ β ΞβπβΞ ππ = 1
(11)
Given belief function (or the BPA, the plausibility), bysolving (11), a probability distribution can be derived,which has least uncertainty measured by Shannon en-tropy and thus is seemingly more proper to be used indecision procedure.It is clear that the problem of finding a minimum en-
tropy probability distribution does not admit a uniquesolution in general. The optimization algorithm usedis the Quasi-Newton followed by a global optimizationalgorithm [16] to alleviate the effect of the local ex-tremum problem. Other intelligent optimization algo-rithms [17, 18] can also be used,such as Genetic Algo-rithm (GA), Particle Swarm Optimization (PSO), etc.
4 Analysis based on examplesAt first, two numerical examples are first provided toillustrate some probability transformation approaches.To make the different approaches reviewed and pro-posed in this paper more comparable, the examplesin [6, 12] are directly used here. The PIC is used toevaluate the probability transformation.
4.1 Example 1
For FOD Ξ = {π1, π2, π3, π4}, the corresponding BPAis as follows:
π({π1}) = 0.16, π({π2}) = 0.14, π({π3}) = 0.01,π({π4}) = 0.02,π({π1, π2}) = 0.20, π({π1, π3}) = 0.09,π({π1, π4}) = 0.04, π({π2, π3}) = 0.04,π({π2, π4}) = 0.02, π({π3, π4}) = 0.01,π({π1, π2, π3}) = 0.10, π({π1, π2, π4}) = 0.03,π({π1, π3, π4}) = 0.03, π({π2, π3, π4}) = 0.03,π(Ξ) = 0.08.
The corresponding belief functions are calculated andlisted as follows:
π΅ππ({π1}) = 0.16, π΅ππ({π2}) = 0.14,π΅ππ({π3}) = 0.01, π΅ππ({π4}) = 0.02,π΅ππ({π1, π2}) = 0.50, π΅ππ({π1, π3}) = 0.26,π΅ππ({π1, π4}) = 0.22, π΅ππ({π2, π3}) = 0.19,π΅ππ({π2, π4}) = 0.18, π΅ππ({π3, π4}) = 0.04,π΅ππ({π1, π2, π3}) = 0.74, π΅ππ({π1, π2, π4}) = 0.61,π΅ππ({π1, π3, π4}) = 0.36, π΅ππ({π2, π3, π4}) = 0.27,π΅ππ(Ξ) = 1.00.
The corresponding plausibility functions are calcu-lated and listed as follows:
ππ({π1}) = 0.73, π π({π2}) = 0.64, π π({π3}) = 0.39,π π({π4}) = 0.26,π π({π1, π2}) = 0.96, π π({π1, π3}) = 0.82,π π({π1, π4}) = 0.81, π π({π2, π3}) = 0.78,π π({π2, π4}) = 0.74, π π({π3, π4}) = 0.50,π π({π1, π2, π3}) = 0.98, π π({π1, π2, π4}) = 0.99,π π({π1, π3, π4}) = 0.86, π π({π2, π3, π4}) = 0.84,π π(Ξ) = 1.00.
Suppose the probability distribution as the unknownvariables. Based on the plausibility functions and thebelief functions, the constraints and the objective func-tion can be established according to (11). The probabil-ity distribution can be derived based on the minimiza-tion. The results of some other probability transforma-tion approaches are also calculated. All the results arelisted in Table 1 (on the next page) to make the com-parison between the approach proposed in this paper(denoted by Un min) and other available approaches.
4.2 Example 2
For FOD Ξ = {π1, π2, π3, π4}, the corresponding BBAis as follows:
π({π1}) = 0.05, π({π2}) = 0.00, π({π3}) = 0.00,π({π4}) = 0.00,
Table 1 Probability Transformation Results ofExample 1 based on Different Approaches
π1 π2 π3 π4 PICBetP [3] 0.3983 0.3433 0.1533 0.1050 0.0926PraPl [8] 0.4021 0.3523 0.1394 0.1062 0.1007PrPl [8] 0.4544 0.3609 0.1176 0.0671 0.1638PrHyb [8] 0.4749 0.3749 0.0904 0.0598 0.2014PrBel [8] 0.5176 0.4051 0.0303 0.0470 0.3100FPT[11] 0.5176 0.4051 0.0303 0.0470 0.3100DSmP 0[9] 0.5176 0.4051 0.0303 0.0470 0.3100PrScP [10] 0.5403 0.3883 0.0316 0.0393 0.3247PrBP1 [12] 0.5419 0.3998 0.0243 0.0340 0.3480PrBP2 [12] 0.5578 0.3842 0.0226 0.0353 0.3529PrBP3 [12] 0.0605 0.3391 0.0255 0.0309 0.3710Un min 0.7300 0.2300 0.0100 0.0300 0.4813
π({π1, π2}) = 0.39, π({π1, π3}) = 0.19,π({π1, π4}) = 0.18,π({π2, π3}) = 0.04,π({π2, π4}) = 0.02, π({π3, π4}) = 0.01,π({π1, π2, π3}) = 0.04, π({π1, π2, π4}) = 0.02,π({π1, π3, π4}) = 0.03, π({π2, π3, π4}) = 0.03,π(Ξ) = 0.00.
The corresponding belief functions are calculated andlisted as follows:
π΅ππ({π1}) = 0.05, π΅ππ({π2}) = 0.00,π΅ππ({π3}) = 0.00, π΅ππ({π4}) = 0.00,π΅ππ({π1, π2}) = 0.44, π΅ππ({π1, π3}) = 0.24,π΅ππ({π1, π4}) = 0.23, π΅ππ({π2, π3}) = 0.04,π΅ππ({π2, π4}) = 0.02, π΅ππ({π3, π4}) = 0.01,π΅ππ({π1, π2, π3}) = 0.71, π΅ππ({π1, π2, π4}) = 0.66,π΅ππ({π1, π3, π4}) = 0.46, π΅ππ({π2, π3, π4}) = 0.10,π΅ππ(Ξ) = 1.00.
The corresponding plausibility functions are calcu-lated and listed as follows:
ππ({π1}) = 0.90, π π({π2}) = 0.54, π π({π3}) = 0.34,π π({π4}) = 0.29,π π({π1, π2}) = 0.99, π π({π1, π3}) = 0.98,π π({π1, π4}) = 0.96, π π({π2, π3}) = 0.77,π π({π2, π4}) = 0.76, π π({π3, π4}) = 0.56,π π({π1, π2, π3}) = 1.00, π π({π1, π2, π4}) = 1.00,π π({π1, π3, π4}) = 1.00, π π({π2, π3, π4}) = 0.95,π π(Ξ) = 1.00.
Suppose the probability distribution as the unknownvariables. Based on the plausibility functions and thebelief functions, the constraints and the objective func-tion can be established according to (11). The probabil-ity distribution can be derived based on the minimiza-tion. The results of some other probability transfor-mation approaches are also calculated. All the resultsare listed in Table 2 to make the comparison betweenthe approach proposed in this paper and other availableapproaches.
N/A in Table 2 means βNot availableβ. DSmP 0means the parameter π in DSmP is 0.
Table 2 Probability Transformation Results ofExample 2 based on Different Approaches
π1 π2 π3 π4 PICPrBel [8] N/A due to 0 value of singletonsFPT[11] N/A due to 0 value of singletonsPrScP [10] N/A due to 0 value of singletonsPrBP1 [12] N/A due to 0 value of singletonsPraPl [8] 0.4630 0.2478 0.1561 0.1331 0.0907BetP [3] 0.4600 0.2550 0.1533 0.1317 0.0910PrPl [8] 0.6161 0.2160 0.0960 0.0719 0.2471PrBP2 [12] 0.6255 0.2109 0.0936 0.0700 0.2572PrHyb [8] 0.6368 0.2047 0.0909 0.0677 0.2698DSmP 0[9] 0.5162 0.4043 0.0319 0.0477 0.3058PrBP3 [12] 0.8823 0.0830 0.0233 0.0114 0.5449Un min 0.9000 0.0900 0.0000 0.0100 0.7420
Based on the experimental results listed in Table 1and Table 2, it can be concluded that the probabil-ity derived based on the proposed approach (denotedby Un min) has significantly lower uncertainty whencompared with the other probability transformation ap-proaches. The difference among all the propositions canbe further enlarged, which is seemingly helpful for themore consolidated and reliable decision.
Important remark: In fact, there exist fatal deficien-cies in the probability transformation based uncertaintyminimization, which are illustrated in following exam-ples.
4.3 Example 3
The FOD and BPA are as follows [4]:
Ξ = {π1, π2}, π({π1}) = 0.3,π({π2}) = 0.1, ,π({π1, π2}) = 0.6
Based on different approaches, the experimental re-sults are derived as listed in Table 3
Table 3 Probability Transformation Results ofExample 3 based on Different Approaches
π1 π2 PICBetP 0.6000 0.4000 0.0291PrPl 0.6375 0.3625 0.0553PraPl 0.6375 0.3625 0.0553PrHyb 0.6825 0.3175 0.0984DSmP 0.001 0.7492 0.2508 0.1875PrBel 0.7500 0.2500 0.1887DSmP 0 0.7500 0.2500 0.1887Un min 0.9000 0.1000 0.5310
DSmP 0 means the parameter π in DSmP is 0 andDSmP 0.001 means the parameter π in DSmP is 0.001.Is the probability transformation based on PIC max-
imization (i.e. entropy minimization) rational ?It can be observed, in our very simple example
3, that all the mass of belief 0.6 committed {π1, π2}is actually redistributed only to the singleton {π1}
using the Un min transformation in order to get themaximum of PIC.
A deeper analysis shows that with Un min transfor-mation, the mass of belief π{π1, π2} > 0 is always fullydistributed back to {π1} as soon as π({π1}) > π({π2})in order to obtain the maximum of PIC (i.e. the min-imum of entropy). Even in very particular situationswhere the difference between masses of singletons isvery small like in the following example:
Ξ = {π1, π2}, π({π1}) = 0.1000001,π({π2}) = 0.1, π({π1, π2}) = 0.7999999.
This previous modified example shows that the prob-ability obtained from the minimum entropy principleyields a counter-intuitive result, because π({π1}) isalmost the same as π({π2}) and so there is no solid rea-son to obtain a very high probability for π1 and a smallprobability for π2. Therefore, the decision based on theresult derived from Un min transformation is too risky.Sometimes uncertainty can be useful, and sometimes itis better to not take a decision than to take the wrongdecision. So the criterion of uncertainty minimizationis not sufficient for evaluating the quality/efficiency of aprobability transformation. There are also other prob-lems in the probability transformation based on uncer-tainty minimization principle, which are illustrated inour next example.
4.4 Example 4
The FOD and BPA are as follows: Ξ = {π1, π2, π3},with ,
π({π1, π2}) = π({π2, π3}) = π({π1, π3}) = 1/3.
Using the probability transformation based on uncer-tainty minimization, we can derive six different proba-bility distributions yielding the same minimal entropy,which are listed as follows:
π ({π1}) = 1/3, π ({π2}) = 2/3, π ({π3}) = 0;
π ({π1}) = 1/3, π ({π2}) = 0, π ({π3}) = 2/3;
π ({π1}) = 0, π ({π2}) = 1/3, π ({π3}) = 2/3;
π ({π1}) = 0, π ({π2}) = 2/3, π ({π3}) = 1/3;
π ({π1}) = 2/3, π ({π2}) = 1/3, π ({π3}) = 0;
π ({π1}) = 2/3, π ({π2}) = 0, π ({π3}) = 1/3.
It is clear that the problem of finding a probabil-ity distribution with minimal entropy does not admit aunique solution in general. So if we use the probabil-ity transformation based on uncertainty minimization,there might exist several probability distributions de-rived as illustrated in this Example 4. How to choosea unique one? In Example 4, depending on the choiceof the admissible probability distribution, the decisionresults derived are totally different which is a seriousproblem for decision-making support.
From our analysis, it can be concluded that the max-imization of PIC criteria (or equivalently the minimiza-tion of Shannon entropy) is not sufficient for evaluatingthe quality of a probability transformation and othercriteria have to be found to give more acceptable prob-ability distribution from belief functions. The searchfor new criteria for developing new transformations isa very open and challenging problem. Until findingnew better probability transformation, we suggest touse DSmP as one of the most useful probability trans-formation. Based on the experimental results shown inExamples 1β3, we see that the DSmP can always becomputed and generate a probability distribution withless uncertainty and it is also not too risky, i.e. DSmPcan achieve a better tradeoff between a high PIC value(i.e. low uncertainty) and the risk in decision-making.
5 Conclusion
Probability transformation of belief function can beconsidered as a probabilistic approximation of beliefassignment, which aims to gain more reliable decisionresults. In this paper, we focus on the evaluation crite-ria of the probability transformation function. Experi-mental results based on numerical examples show thatthe maximization of PIC criteria proposed by Sudanois insufficient for evaluating the quality of a probabilitytransformation. More rational criteria have to be foundand to better justify the use of a probability transfor-mation with respect to another one.
All the current probability transformations devel-oped so far redistribute the mass of partial ignorancesto the belief of singletons included in it. The redistri-bution is based either only on the cardinality of partialignorances, or eventually also on a proportionalizationusing the masses of singletons involved in partial igno-rances. However when the mass of a singleton involvedin a partial ignorance is zero, some probability trans-formations, like Cuzzolinβs transformation by example,do not work at all and thatβs why the π parameter hasbeen introduced in DSmP transformation to make itworking in all cases. In future, we plan to developa more comprehensive and rational criterion, whichcan take both the risk and the uncertainty degree intoconsideration, to evaluate the quality of a probabilitytransformation and to find an optimal probabilitydistribution from any basic belief assignment.
AcknowledgmentThis work is supported by Grant for State Key Programfor Basic Research of China (973) No.2007CB311006and the Research Fund of Shaanxi Key Labora-tory of Electronic Information System Integration(No.200910A).
References[1] A. P. Dempster, βUpper and lower probabilities in-
duced by a multi-valued mappingβ, J. Annual MathStatist, Vol 38, No. 4, pp. 325β339, 1967.
[2] G. Shafer, A Mathematical Theory of Evidence,Princeton University Press, Princeton, 1976.
[3] P. Smets, R. Kennes, βThe transferable beliefmodelβ, Artificial Intelligence, Vol 66, No. 2, pp.191-234, 1994.
[4] F. Smarandache, J. Dezert (Editors), Applica-tions and Advances of DSmT for InformationFusion, Vol 3, American Research Press, 2009.http://www.gallup.unm.edu/Λsmarandache/DSmT-book3.pdf.
[5] J. Sudano, βPignistic probability transforms formixes of low- and high-probability eventsβ, Proc.of International Conference on Information Fusion2001, Montreal, Canada, August 2001, TUB3, pp.23-27.
[6] J. Sudano,βThe system probability information con-tent (PIC) relationship to contributing components,combining independent multisource beliefs, hybridand pedigree pignistic probabilitiesβ. Proc. of Inter-national Conference on Information Fusion 2002,Annapolis, July 2002. Vol 2, pp. 1277-1283.
[7] J. Sudano, βEquivalence between belief theories andnaive Bayesian fusion for systems with independentevidential data - Part I, The theoryβ, Proc. of Inter-national Conference on Information Fusion 2003,Cairns, Australia, July 2003, Vol 2, pp. 1239-1243.
[8] J. Sudano, βYet another paradigm illustrating ev-idence fusion (YAPIEF)β, Proc. of InternationalConference on Information Fusion 2006, Florence,July 2006, pp. 1-7.
[9] J. Dezert, F. Smarandache, βA new probabilistictransformation of belief mass assignmentβ, Inter-national Conference on Information Fusion 2008,Germany, Cologne, June 30π‘β β July 3rd 2008, pp.1-8.
[10] J. Sudano, βBelief fusion, pignistic probabilities,and information content in fusing tracking at-tributesβ, Radar Conference 2004, April 26th-29th2004, pp. 218 β 224.
[11] Y. Deng, W. Jiang, Q. Li, βProbability transfor-mation in transferable belief model based on fractaltheory (in Chinese)β, Proc. of Chinese Conferenceon Information Fusion 2009, Yantai, China, Nov.2009, pp. 10-13, 79.
[12] W. Pan, H. J. Yang, βNew methods of transform-ing belief functions to pignistic probability func-tions in evidence theoryβ, Proc. of InternationalWorkshop on Intelligent Systems and Applications,Wuhan, China, May 2009, pp. 1-5.
[13] F. Cuzzolin, βOn the properties of the Intersectionprobabilityβ, submitted to the Annals of Mathemat-ics and AI, Feb. 2007.
[14] W. H. Lv, βDecision-making rules based on beliefinterval with D-S evidence theory, (book section)β,Fuzzy Information and Engineering, pp. 619-627,2007.
[15] P. Smets, βDecision making in the TBM: the neces-sity of the pignistic transformationβ, InternationalJournal of Approximate Reasoning, Vol 38, pp. 133β147, 2005.
[16] R. Horst, P. M. Pardalos, N. V. Thoai, Introduc-tion to Global Optimization (2nd Edition), KluwerAcademic Publishers, New York, 2000.
[17] D. T. Pham, Intelligent Optimisation Techniques(1st Edition), Springer-Verlag, London, 2000.
[18] J. Kennedy, R. C. Eberhart, Y. H. Shi, Swarm In-telligence, Morgan Kaufmann Publisher, AcademicPress, USA, 2001