Professional Documents
Culture Documents
Éloi Bossé Abstract. The problem of fusing identity declarations emanating from
Jean Roy, MEMBER SPIE different sources is explored and decision makers are offered a quanti-
Defence Research Establishment Valcartier tative analysis based on statistical methodology that can enhance their
Decision Support Section decision making processes regarding the identity of detected objects.
2459 Pie Xi North Boulevard The context is naval warfare where commanders and their staff require
P.O. 8800 access to a wide range of information to carry out their duties. This
Val-Bélair, G2J 1X5, Canada information provides them with the knowledge necessary to determine
E-mail: eloi.bosse@drev.dnd.ca the position, identity and behavior of the enemy. Statistical analysis
rooted in the Dempster-Shafer theory of evidence is used to further the
fusion of identity declarations. © 1997 Society of Photo-Optical Instrumentation
Engineers.
648 Opt. Eng. 36(3) 648–657 (March 1997) 0091-3286/97/$10.00 © 1997 Society of Photo-Optical Instrumentation Engineers
Downloaded From: http://opticalengineering.spiedigitallibrary.org/ on 05/12/2015 Terms of Use: http://spiedl.org/terms
Bossé and Roy: Fusion of identity declarations . . .
tially displayed tactical data using symbology and color for must be combined to obtain an estimate of an object’s iden-
NATO maritime units. tity. The italicized elements symbolize identity declaration
Figures 3 and 4 offer an adapted version of a tactical domains; they do not belong to the STANAG 4420 charts
information hierarchy for surface and air objects, respec- but were added to establish a relationship between the do-
tively; certain levels and entries have been omitted for sim- mains and their specific entities. Some entities can be fur-
plicity. The names in the various boxes, which are indica- ther divided into class and unit. These categories typify the
tive of the taxonomy used in STANAG 4420, represent identity of objects at two levels of specificity. The ‘‘unit’’
identity declaration entities. These entities illustrate the sort declaration domain characterizes uniquely a detected ob-
of identity declarations provided by various sources that ject. Also included in the hierarchy of tactical in-
formation is the threat designation; its subdivisions are of Q. The elements of 2Q, or equivalently the subsets of Q,
given in Fig. 5. The hierarchy of tactical information delin- are the class of general propositions and include the empty
eated by Figs. 3 to 5 should encompass many of the identity set B, which corresponds to a proposition that is known to
declarations provided by the organic and nonorganic be false, and the whole set Q, which corresponds to a
sources of Fig. 2. Furthermore, because this hierarchy is a proposition that is known to be true. We assume throughout
NATO standard, it provides a better base for achieving in- our discussion that the frame of discernment is finite.
teroperability in information exchange between nations Let A be a subset of Q. The evidential theory differen-
than uncontrolled alternatives. tiates between the measure of belief committed exactly to A
and the total belief committed to A. The first is character-
3 Fusion of Uncertain Information ized by the basic probability assignment and the latter by
the belief function. A function m is a basic probability
If a priori probability distributions and conditional prob-
assignment if it assigns a number m(A) to each subset
ability matrices are available and if probabilities can be
distributed to single elements, then the Bayesian approach AP2Q, in such a way that m~B!50, and ( A#Q m(A) 5 1.
should be used to fuse uncertain information. However, the The quantity m(A) is called A’s basic probability num-
Bayesian approach suffers from major deficiencies in a hi- ber and represents the exact belief in the proposition de-
erarchical context, when fully specified likelihoods are not picted by A. A function Bel is a belief function if it assigns
available. Pearl’s rules3 can be applied but may not always a number Bel (A) to each subset AP2Q in such a way that:
lead to an appropriate estimation of the posterior probabili- Bel ~B!50, Bel ~Q!51 and for every positive integer n and
ties associated with certain nodes of a hierarchy. Other every collection, A 1 , A 2 , . . . , A n of subsets of Q,
problems associated with the Bayesian approach are the
coding of ignorance and the strict requirements on the be-
lief of a hypothesis and its negation.
Consequently, the Bayesian approach may not be the
Bel ~ A 1 ø•••øA n ! > (i Bel ~ A i ! 2 (
i, j
Bel ~ A i ùA j !
ideal technique to fuse uncertain information in the context
of identity estimation by multiple dissimilar sources. We 1•••1 ~ 21 ! n11 Bel ~ A 1 ù•••ùA n !
thus investigate a generalization of the Bayesian approach
that does not arbitrarily allocate probabilities to the children > (
I, $ 1...n %
~ 21 ! u I u 11 Bel ~ ù A i ! ....
and parents of a node when this node is updated with new IÞB
iPI
information. This technique is based on the Dempster-
Shafer theory of evidence.
The quantity Bel (A) is called the degree of belief about
3.1 Dempster-Shafer Theory of Evidence proposition A. A belief function assigns to each subset of Q
The Dempster-Shafer technique does not require prior a measure of the total belief in the proposition represented
probabilities nor does it need to know the capability of each by the subset. Here u I u stands for the cardinality of the set I.
source. Evidence is not committed to any specific event or The simplest belief function is the one where Bel ~Q!51
set of events until evidence is gained. Also, it is not re- with Bel (A)50 for all AÞQ; this function is called the
quired that belief not committed to a given proposition vacuous belief function. A belief function can be obtained
should be committed to its negation, nor that belief com- from the basic probability assignment:
mitted to a given proposition should be committed more
specifically.4 The Dempster-Shafer technique has the capa-
bility of expressing ignorance explicitly. The technique ac-
tually focuses on the probability of a collection of points
Bel ~ A ! 5 (
B#A
m ~ B ! , for all A#Q. ~1!
belonging to the sample space, whereas the classical prob-
ability theory is interested in the probability of the indi-
vidual points.5 A subset A of Q is called a focal element of a belief func-
In the Dempster-Shafer evidential theory, the terminol- tion Bel over Q if m(A).0. The union of all the focal
ogy is slightly different from that used in probability elements of a belief function is called its core. A belief
theory. The frame of discernment Q is defined as an ex- function Bel is called a simple support function S if there
haustive set of mutually exclusive events or propositions of exists a nonempty subset A#Q and a number 0<s<1 such
a particular experiment. We denote 2Q the set of all subsets that:
S ~ B ! 5Bel ~ B ! 5 H 0
s
1
if B does not contain A
if B contains A but BÞQ
if B5Q.
~2!
and Bel2’s focal elements are C 1 , . . . , C n , the total portion
of belief exactly committed to A (AÞB) is given by the
orthogonal sum m5m 1 % m 2 :
selves; the computational time will be reduced from expo- frames, one for each set of mutually exclusive hypotheses.
nential to polynomial. This approach, however, is appealing Of course, once the elements have been separated into mul-
only when one is interested in final conclusions about the tiple frames, items from different frames can no longer be
elements of Q; in the study of identity declaration fusion, compared since they then pertain to different belief func-
we are mostly interested in subsets of Q. tions. Shafer’s first approach combines belief functions
Barnett8 demonstrated that if each piece of evidence each of which is carried by the field of subsets generated by
consists of simple support functions focused on singleton the children of a particular node. His second approach com-
propositions and their negations, computational time will bines simple support functions focused on a subset of Q or
be reduced from exponential to linear. Another option was its complement.11 The latter uses the same type of evidence
proposed by Gordon and Shortliffe.9 It is based on the as- as considered by Gordon and Shortliffe,9 while avoiding
sumptions that ~1! each piece of evidence consists of some of its shortcomings. This approach is detailed in Ref.
simple support functions focused for or against subsets of 13.
Q instead of singletons, and that ~2! the subsets of Q can be
structured in a strict hierarchical tree. This method builds 4 Statistical Decision Making
on Barnett’s approach while permitting hierarchical rela-
tionships among hypotheses; its aim is similar to that of An important element to take into consideration in the de-
Pearl.2 sign of an identity declaration fusion function is the deci-
However, combining negative evidence leads to compu- sion making process required to select the identity declara-
tational difficulties because the intersection of the comple- tion which best supports the combined declarations.
ments of nodes may fail to correspond to a node or its Statistical decision making is necessary since, after fusion,
complement. In such a case, an approximation is suggested the resulting hierarchical structure may contain many iden-
by the authors but this approximation restricts the useful- tity declarations with a non-null confidence value. Because
ness of the plausibility measure. The algorithm can be decisions are subjective, the decision maker will undoubt-
implemented in a form that is linear in the number of nodes edly depend on his or her own judgment as well as infor-
in the tree. mation collected from various sources. The fusion function
Cleckner12 offers a comparative study in terms of should still suggest to the decision maker the ‘‘best’’ can-
memory requirements and computational complexity in didate or candidates according to predetermined decision
which the standard Dempster-Shafer combination rule is rules. A number of factors are involved in the choice of the
compared with Barnett’s algorithm and the alternative due decision rules to be used:14
to Gordon and Shortliffe.9 Cleckner concluded that when 1. rule complexity ~if the computational time required
dealing with simple support functions focused on singleton to make a decision is too high, the process becomes
hypothesis, Barnett’s technique is the least computer inten- useless!
sive of the three.
The last category of options, which also assumes that 2. confidence level ~decision rules are made on a proba-
evidence is hierarchically structured, was proposed by bilistic basis!
Shafer.10 He suggested that belief functions being com- 3. fusion technique used to combine uncertain informa-
bined should be carried by a partition P of Q, which has tion ~for example the Dempster-Shafer approach pro-
fewer elements than Q. This is done by separating the duces belief and plausibility values instead of single
frame of discernment into smaller, more manageable probability values!
lief functions focused on the same nodes B and F interesting issue that distinguishes this study from other
@ m( $ B % )50.99, m( $ F % )50.01 with m( $ B % )50.02, studies on identity fusion is the fact that two different
m( $ F % )50.95, m~Q!50.03#. Dempster’s rule calculates a frames of discernment are introduced to estimate the iden-
small normalizing constant ~K50.06!, indicating conflict. It tity of the observed objects. The first frame of discernment
is interesting to note that this evidence is dichotomous is the hierarchical tree of surface and air classifications, as
since $ B % 5$ F̄ % . The small evidence for $ B % slightly modi- shown in Figs. 3 and 4, and the second one deals with the
fies the belief and plausibility of $ B % ; the big uncertainty is threat categories as described in Fig. 5. The purpose of this
distributed amongst the other nodes. approach is to circumvent the problem whereby the threat
The plausibility threshold T Pl is chosen equal to 0.6. At category of a detected object is directly inferred from its
each step, the decision technique creates a graph ~Fig. 8! identity. For example, the Exocet missile can be automati-
from which the decision maker can select the best alterna- cally assumed friendly. During the Falklands war, however,
tive according to his or her needs. If no nodes other than
the Exocet missile was definitely not considered friendly to
the direct children of Q appear in the graph, then no alter-
native has a plausibility value higher than the threshold T Pl the British Navy. By eliminating false automated infer-
and no decision can be made. The contradictory evidence at ences, this approach allows more freedom in the decision
step 5 greatly modifies the belief of $ B % and $ F % . The de- making.
cision technique reproduces the effect of this contradictory The most basic concept to which the study adheres is the
evidence. fact that information sources are self-contained and that
each one represents a local decision node capable of iden-
5 Example of the Identity Declaration in a tifying a detected object. The aim is therefore to combine
Military Context local decisions in the hope of obtaining the correct identity
The following example illustrates the use of the Shafer- with a high probability. Dasarathy15 calls this type of fusion
Logan algorithm for the specific problem of combining ‘‘Decision In-Decision Out Fusion’’ because both the input
identity declarations using two frames of discernment. An and output are decisions.
Table 2 Evidence received from various sources. tion received can consist of a general classification ~surface
combatant!, a specific type of ship ~frigate!, a specific class
Frame of discernment Q1 Frame of discernment Q2 ~city class! or a unique identity ~Ville de Québec!. In Fig. 1
we labeled these ‘‘identity declarations.’’ Identity declara-
m ($fighter%)50.4 m ($unknown%)50.2
tions also include the threat designation, as in Fig. 5. The
m (Q)50.6 m (Q)50.8
belief and plausibility values at each node of the two hier-
m ($carrier%)50.7 m ($friend%)50.6 archical trees are 0. The evidence received from various
m (Q)50.3 m (Q)50.4 information sources are given in Table 2 ~assuming that the
m ($fixed-wing%)50.3 m ($hostile%)50.7 order of the received evidences is unimportant!.
m (Q)50.7 m (Q)50.3 The evidence items are the only ones received concern-
m ($noncombatant%)50.8 m ($neutral%)50.3 ing the object; a decision could be made after each evi-
dence if the plausibility of at least one node is greater than
m (Q)50.2 m (Q)50.7
T Pl . However, we have chosen to combine all the evidence
m ($air%)50.5 m ($hostile%)50.8
before the decision making process. In this example, T Pl is
m (Q)50.5 m (Q)50.2
set to 0.5. Evidence m( $ air% )50.5 and m~$ air%!50.1 are
m ($air%)50.1 contradictory but not enough for the Dempster’s combina-
m (Q)50.9 tion rule to produce irregular results. Figures 9 and 10 show
m ($MiG-25%)50.1 the belief and plausibility measures for each node after the
m (Q)50.9 combination of evidence for the two frames of discernment,
m ($helicopter%)50.5 respectively.
m (Q)50.5 Figures 11 and 12 give the graphs that are available to
m ($MiG-19%)50.6
the decision maker. It seems that the object is airborne and
hostile; there is a fairly good chance that it is a fixed wing
m (Q)50.4
fighter. According to the mission, the decision maker will
m ($missile %)50.7 choose the best alternative and take appropriate action if
m (Q)50.3 necessary.
m ($MiG-25%)50.4
m (Q)50.6
6 Conclusion
This paper concerns the use of the Dempster-Shafer theory
of evidence for the fusion of identity declarations within a
The scenario is as follows: the commander of a CPF- naval environment. It proposes to hierarchically structure
type ship receives a series of identity declarations concern- the identity declarations according to NATO’s STANAG
ing one object; he or she must determine the identity of the 4420 charts, which provide a better base for achieving in-
object and take action. For example, the types of informa- teroperability in information exchange between nations