I.

Introduction
In recent years there has been an increasing in-
Detection With \ terest in distributed sensor systems. This interest has
been sparked by the requirements of military sur-
Distributed Sensors veillance systems [1] and is reflected in the widespread
use of such terms as data fusion, correlation, and
multisensor integration.
The classical theory of optimal sensor signal pro-
cessing is based on statistical estimation and hy-
pothesis testing methods [2]. The motivation for this
theory has been the optimal detection and tracking of
targets using a single sensor such as a radar or sonar.
ROBERT R. TENNEY, Member, IEEE Thus all the sensor signals are implicitly assumed to
NILS R. SANDELL, JR., Member, IEEE be available in one place for processing.
The situation is substantially more complicated in
the case of a distributed sensor network. If it were
possible to transmit all sensor signals to some central
Abstract location with negligible delay, then the classical theory
is in principle applicable, even though there may be a
The extension of classical detection theory to the case of distributed
myriad of interesting problems associated with the
sensors is discussed, based on the theory of statistical hypothesis
disparate nature of the information sources.1
testing. The development is based on the formulation of a decen- However, because of such considerations as cost,
tralized or team hypothesis testing problem. Theoretical results con- reliability, survivability, communication bandwidth,
cerning the form of the optimal decision rule, examples, application compartment-alization, sensors on platforms under
to data fusion, and open problems are presented. emission control, or even simply the problem of
flooding the fusion center with more information than
it can process, there is never total centralization of in-
formation in practice. Thus extensions are needed to
the classical framework of detection theory if it is to
be relevant to the design of distributed surveillance
systems. This paper attempts a modest step in the
direction of a detection theory for distributed sensors
by considering decentralized hypothesis testing.
While hypothesis testing is a well understood tech-
nique, its extension to a decentralized formulation
yields behavior more complex than might be initially
expected. In the case of total decentralization (Fig. 1),
one would expect each detector to operate indepen-
dently and base its decision on the familiar likelihood
ratio test. While this is true in special cases, it is un-
true in general. When the detectors choose their deci-
sions to achieve a system-wide optimization, they
often use different strategies than in cases where the
joint cost associated with their decisions separates into
a cost for each.
Manuscript received November 13, 1979; revised September 25,
1980, and January 15, 1981. The problem of constructing decentralized hypo-
This work was supported in part by a National Science Foundation
thesis testing rules can be viewed in the framework of
graduate fellowship and in part by the Office of Naval Research decentralized optimal control theory.2 Unlike most
under Contract N00014-77-C0532. decentralized control problems, the hypothesis testing
Authors' addresses: R. Tenney, Massachusetts Institute of problem can be solved in a relatively straightforward
Technology, Laboratory for Information and Decision Systems, way. This is due principally to the fact that the deci-
Room 35-213, 77 Massachusetts Ave., Cambridge, MA 02139; N.R. sions made do not feed back into the system dynamics
Sandell, Jr., Massachusetts Institute of Technology, Department of
Electrical Engineering and Computer Science, Cambridge, MA
02139.
'For example it may be necessary to correlate a radar return
with a report from a Naval attache.

0018-9251/81/0700-0501 $00.75 1981 IEEE 2See [3] for a survey with 156 references.

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-17, NO.4 JULY 1981 501

Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:07 UTC from IE Xplore. Restricon aply.
 show that the general theory developed for the decen-
tralized hypothesis testing problem of Fig. 1 can be
applied to the situation of Fig. 2 using an appropriate
definition of the cost function. Thus the performance
of the distributed detection system can be quan-
titatively compared with that of the centralized system
to evaluate the performance versus the communica-
tions tradeoff.
II. Decentralized Hypothesis Testing
DECISION

Fig. 1. Decentralized detection. For the structure of Fig. 1, the problem of decen-
tralized (binary) hypothesis testing can be posed in its
Fig. 2. Fusion.
most general form as follows. The two hypotheses are
HI or H', with a priori probabilities

p(H0) = p° p(H') = p'. (1)

For each hypothesis the sensor observations have
known joint probability distributions

P (Yl , Y2 HO); p (yl , y2 HI). (2)

Local Decision 2 (The subscripts indicate the sensor location to which a
variable is associated; we will consider only the case
of two locations.) The yi are random vectors resulting
from preprocessing of the original sensor signals.
The crux of the distributed hypothesis testing pro-
blem is to derive decision rules of the form
0, H0 is declared to have been detected
ui =
and thus affect the information of other decision 1, HI is declared to have been detected
makers. For example, we will be able to prove that in
the important special case where signals are indepen- where ui depends only on the observation yi. The
dent when conditioned on the hypotheses, each detec- decision rule can be defined by the conditional pro-
tor uses an independent, local likelihood ratio test, bability distribution function
but with thresholds determined via a coupled com-
putation. The computation of thresholds decouples p(ui = Olyi)
only in those special cases where the payoff function
separates. which defines a randomized decision rule.
However, even in the case of independent observa- The optimality criterion will be a function
tions, several types of unusual behavior can occur.
The threshold computations can yield locally optimal J: {O, 1} x (0, 1} x {H0,HI} - IR (3)
thresholds which are far from the globally optimal
values. Moreover, even when identical sensors and a with J(u1, u2, Hi) being the cost incurred for detector
symmetric cost function are considered, the optimal 1 choosing u1, and detector 2 choosing u2, when Hi is
thresholds need not be equal; the two detectors must true.
hedge their decisions in nonidentical ways. For example, the minimum error cost function Jm
As described above, the motivation for this study is
is for distributed systems in which data is being
generated at geographically dispersed sensors, and J(0, 0, H0) = J(1A 1, HI) = 0
local processing is to be used to reduce the com-
munications bandwidth required between the sensor J(0, 1, H) = J(1, 0, H1) = J(O, 1, H1)
stations and a fusion center (Fig. 2). The performance
of such a system is suboptimal in comparison with a = J(AI 0, IH) = 1
completely centralized processing scheme due to the
loss of information in the local processes. We will J(O,0,H') = J(1, 1,1H0) = 2. (4)
502 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-17, NO.4 JULY 1981

Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:07 UTC from IE Xplore. Restricon aply.
This function is separable, observations (y1, Y2) for each hypothesis (or eqUivalently,
the ratio of a posteriori likelihood of the hypotheses given
Jm(u,, U2, Hi) = J,(u,, H') + J2(U2, H9) (5) the observations) and t is a precomputed threshold; and
c) the threshold
where
t = [J(O, H1) - J(l, Hl)]/[J(1, HO) - J(O, H0)]
(1 1)
Au JO,
u= =O,H=H°oruj =1,H=IP provided
1, else
(6) J(1, H0) > J(O, H0).
is the usual minimum probability of error cost func- The decentralized solution can be derived in a
tion associated with detector j. manner parallel to that used in the centralized case.
The objective of the decision strategy will be to Begin by expanding (7) explicitly; the function to be
minimize the expected penalty incurred minimized is
min E [J(u,, uL, H] (7)
E f p(Ul,u2y1,Y2,H)J(ul,U2,H)
HUI,U2 Ynx
where the minization is over the (randomized) decision
rules of each detector.
=
H,uI,U2 Y1t2
-
fP(H)p(Y,Y2 H)p(Ully,)p(u2ly2)
In summary we have the following. * J(u,u2,H) (12)
Problem (Decentralized Binary Hypothesis
Testing): Given p°, pI, the distribution p(y1, Y2 Hi), i by invoking appropriate independence assumptions.
= 0, 1, and the cost function J, find the decision Explicitly summing over u, gives
strategies for each detector (expressed as functions
p(ui yi) of the corresponding observation only) If p(H)P(u2lY2)P(Yl,Y21H) (P(Ul = OjyJ)
H, U2 Y 1 V2
which minimize the expected cost.
The familiar centralized case is similar, but with a *J(O,u2,H) + (1 - p(U, = OJY1)) J(1,u2,H)]
crucial difference: (13)
Problem (Centralized Binary Hypothesis Testing): One obtains the following equivalent function to
Given pO, pI, the distributions p(yI, y2 H%), i = 0, 1,
and the cost function J, find the decision strategy (ex- minimize by ignoring a constant term:
pressed as a function p(u yy,, Y2) of both observations)
which minimizes the expected cost.
The solution of the centralized problem is, of course, f P(U1 = O1YO) 72 fP(H) P(t2lY2)P(yl, y21 H)
well known.
[J(O, u2,I) - J(1, u2, H)]. (14)
Theorem 1: The solution to the centralized binary
hypothesis testing problem is a) deterministic, so that the This is minimized by choosing
decision rule is a function
P(u, = Ojy,) =
y: Y1 x Y2 1{0, 1} (8)
0, if H,I-
u2 Y2
f P(H)p(u2Iy2) P(Yl,Y21H)
with u = i interpreted as choosing Hi; b) a likelihood
ratio test [J(O, U2,H) - HI1, U2,H)] > 0
0, if lr(y1,y2) > t 1i, else.
Y* (Yl,Y2) =
(15)
1, if lr(yl,Y2) < t (9)
where Notice that, regardless of the forms of p(y,, Y2 H),
J, orp(u2jY2),
Ir (Y1, Y2) = [P(Yi, Y2 H0) Po]/[P(Yl, Y2 I H') p'l p(U, = Ojyi) E (0, 1}
= p(Ho YI, Y2)/P(H' Y1, Y2) (10) so we have
is the ratio of the a priori likelihood of the received Lemma 1: The decision rule used by each detector
TENNEY/SANDELL: DETECTION WITH DISTRIBUTED SENSORS 503

Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:07 UTC from IE Xplore. Restricon aply.
is deterministic and can be expressed as a function yi: and its companion form (for the rule for u2) would in-
Yi -- {0, I } define volve solving for the entire functions yi (.) and Y2 (-)
through the coupled equations. Thus the optimal solu-
if P(Uj = °Iy') = 1 tion is not a likelihood ratio test in general.3
Fortunately matters simplify greatly if we make
1= else. assumption 2.
Assumption 2: The observations yl and Y2 are
This gives the decision rule for detector 1 as statistically independent:
I fY2 p(H) p(y1, Y21 H)p(u2
u2,H IY2) [J(O,U2,H) P(Y2 IY1, H) p(y21H)
=

- J(1,u2,H) 1>0 (17) P(YlIY2, H) = p(y, IH). (20)

where the notation This assumption is appropriate for the "known signal
in noise" case characteristic of most radar problems.
Aix) <O y It is inappropriate for the "unknown signal in noise"
case that arises in sonar problems, for example.
indicates Assumption 2 removes the dependence of the right-
hand side of (19) on yi, so it becomes
Ut = 1, if f(x) >y Ir (yi) >1 fl(Y2 (
choose either, if f (x) = y )^tl
U, = 0, if (x) < y.
which indicates that the likelihood ratio for Yi is com-
Expanding the sum over H, pared with a constant threshold t,, which can be com-
puted as a function f, of the decision rule Y2 (') for u2
[expressed as p(u2 1Y2)]. Thus we have the following
I f p(H0)
U2 Y2
HO) P(U2 [J(O, U2,H0)J lemma. p(y1,Y21 IY2)
- J(1, U2,H0)] > <0 f p(HI) Lemma 2: Assumptions I and 2 imply that the
u2 Y2
solution to the decentralized hypothesis testing pro-
* P(YV,Y2IH1) [J(0, U2JHA)] blem has each detector implementing a likelihood P(U2IY2)
ratio test, using a threshold derivable from the deci-
- J(1, U2, HI)]. (1v8) sion rule of the other detector.
By making the assumption 1, (20) becomes The final step towards solution of this problem is
taken by noticing that the decision function yi (-) is
[P0P(Y11 H )]/[P1P(Y11HI)]><O{EJ P(Y2 1Y1J characterized by the threshold ti and the likelihood
ratio. For example, knowing t2 one can find
* P(U2IY2) [J(O, U2,H1) - J(1,U2,)I)]I}/{I U2 P(U, = oly', = 37 f H Y2:lr{y2)>t2
p (y2l H)p(H)
* fY2 P(Y21y1,HO)P(U2IY2) VJ(1, U2,H°) and thus

- J(, U2,H°)]} (19)~~lt, ==

(19) f P(YyHII) {[J(O,1,II) - J(A,1,IP)]
Assumption 1: J(1, u2, H°) > J(O, u2, H°), or it is
J(1,OH,I) + P(u2 =0°Y2) [J(O,0,H') -

more costly for detector 1 to err than for it to be cor- - J(0,1,IH) + J(A,l,H)]}/ (21)
rect when H° occurs, regardless of the decision of
detector 2.
The form of the decision rule expressed in (19) is rJP(Y2 l1Ho) {[J(1,1,Ho) - J(0,1,H°)]
interesting as the left-hand side is the likelihood ratio
for Yi. However the right-hand side is not a simple + p(u2=0Oy2) [J(1,0,H°) - J(0,0,H°)
threshold constant due to the existence of the condi-
tional density p(y2 Lyi, Hi), and it requires knowledge - J(l,1,1H°) + J(0,1,IH°)]} =fi(t1) (22)
of the entire decision function for u2, namely, p(u2L2)
[or equivalently y2 (yv2)] in order to be evaluated. This 30r equivalently, the optimal solution is a likelihood ratio test
causes significant difficulty, as any solution of (19) but with a data dependent threshold.

504 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-17, NO.4 JULY 1981

Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:07 UTC from IE Xplore. Restricon aply.
Equation (22) and its companion i = ai' + 05 a>O, PER (26)
t2 = f )2
(t) (23) which do not affect the solution, to
provide simultaneous equations, not for the solution J(O, 0, HI) = J(1, 1, HI) = 0 (O errors)
for the functions y, (.), but for the parameters t, and
t2 which characterize them. J(O, 1, H") = J(l, 0, H0) = J(O, 1, H) = 1(1, 0, H')
It is important to note that (22) and (23) are
necessary conditions which must be satisfied if a pair = 1 (1 error)
(t,, t2) is to be an optimal solution to this problem,
but are not sufficient. They define locally optimal J(l, 1, 10) = J(O, 0, HI) = k (2 errors). (27)
solutions in that no threshold can be changed
unilaterally to improve the solution. There may be
several local minima; each must be checked to assure Notice that, when k = 2, this becomes the minimum
that the global minimum is found. error cost function JmP Also, assumption 1 of
The foregoing development can be summarized by Theorem 2 requires
the following theorem.
J(1, 0, H0) > J(0, 0, If)
Theorem 2: The solution to the decentralized J(1, 1,R 0) >J(0, 1,I ) (28)
binary hypothesis testing problem is a) deterministic
(Lemma 1) the latter of which requires k > 1. This means simply
that double errors are penalized at least as much as
yi: Yl {0, 1}; Y2: Y2 -' {O, 1} single errors.
Substituting this cost function into (21), the equa-
b) a likelihood ratio test for each detector; tion for t,, yields
0I if Ir, (yi) > t, tl = fP(Y2 H) [1 + (k - 2)p(u2 = 0 Y2YI/
Y*~(yi) = (24) Y2
1, if Iri (yi) < ti .fp(y2 IF) [(k 1) (k 2 U

with lri (ye) = pop(y1I H')/p1p(y,I HI) (Lemma 2), pro-

= 0 y2)J (29)
vided assumptions I and 2 hold; and c) the thresholds from which Theorem 3 can be immediately deducted.
t1 and t2 satisfy

tl= f (ti2); Theorem 3: If J = Jm, then for any observation

t2 = f (ti) (25)
distributions p (yi Hi), the threshold computations
where f' is given in (21) andfisa symmetricform decouple and yield the thresholds which result from
obtained by interchanging the roles of y, and y2 and each detector using the optimal centralized minimum
of and U2.
u1 error decision rule.
Note the similarities of Theorems 1 and 2. The key Proof: If J = Jm, then the coefficient of p(u2 = 01
difference is that in the centralized case a single deci- y2) in (27) goes to zero, eliminating the dependence of
sion is made, and it is based on both observations. In t1 on y, (-). This produces t, = t2 = 1 as the solution,
the decentralized case two decisions are made, each which is also the solution for each detector using the
based on only one observation. It is important to em- cost function Ji (ui, H) defined in (6).
phasize that the decentralized decision rules are not in The importance of this theorem is that Jm is a
general the same as those that would be derived using rather special case; in general the computations do not
classical theory independently for each detector. The decouple.
computation of thresholds couples the choice of local For the class of cost functions introduced above, it
decision rules so that system-wide performance is op- would be expected that as k decreases from 2 to 1, the
timized, rather than the performance of each in- thresholds would change in a way which increases the
dividual detector. There is a case, however, in which probability of error, as double errors are discounted
the detectors can be treated separately. relative to single ones. As k increases from 2, double
Begin by defining the cost function j to be sym- errors become prohibitively expensive, so it is to be
metric in u, and u, and to penalize all cases of n er- expected that some mechanism will emerge to reduce
rors, n = 0, 1, 2, by the same amount. Any such their likelihood.. Thus in the following, the limiting
function can be reduced, using affine transformations cases k = 1 and k = X are of special interest.

TENNEY/SANDELL: DETECTION WITH DISTRIBUTED SENSORS 505

Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:07 UTC from IE Xplore. Restricon aply.
 Using the identity
30

fY2 P(U2 = 0 Y2) P(Y2 H) = p(U2 = 0 1 H) (36)

2h- when

P(U2 = 0I Y2) = P(U2 = 0 I Y2, H) (37)

IL
3
as is the case here, (28) can be written as
I0(?)lw
Ir~
I

".1
3 4I ti = [1 + (k - 2)p(u2 = 0 1 HI)I/(k - 1)
k

Fig. 3. Solutions to threshold computations. - (k - 2)p(u2 = Oj H°)] (38)

which when combined with (35) yields
Example 1
We consider a simple case of exponentially t, = fl ((2) = (4t2 + k - 2)/[4(k - 1) td
distributed observations. - 2(k - 2) t2] (39)
p(y, H3J) = Aexp (- Ay1) a symmetric form forf2 (.).
and
Equation (39) and its companion can be solved for
p(yi I H) = 2Aexp (- 2Ayi) A>0,i= 1,2 (30) the following conditions on t, and t2.
Case A: t1 = t2 = t: the common value of the
for yi > 0, and threshold must satisfy
p(yi H') = 0, i = 1, 2, j = 0, 1 (31) 4(k - l)t3 - 2kt2 - (k- 2) = 0 (40)
elsewhere. Notice that the observations are indepen- Case B: t, t2: The unequal thresholds satisfy
dent when conditioned on H, and are identically
distributed for each detector. We consider the t1t2 (2 k)/2k; tl + t2
hypotheses to be equally likely, p0 = pl = 1/2.
= - = (k - 1) (2 - k)/2k.
The likelihood ratio for each detector is (41)
(yi) Case C: ti 1/2: If one threshold, say t,, is com-
Ir, =
p0 Aexp (- Ay,)/p'2Aexp (- 2Ayi)
puted to be
=

= (1/2)exp (AyE) (32) setti2) = to < 1/2

so the optimal decision rule for each is
setting it to 1/2 yields the other as
Yi >1 (I1/l) In(2ti). (33) t2 =k- 1. (42)
Notice that lr,(y,) > 1/2 for any yi possible, so that
any threshold less than 1/2 is equivalent to one of ex-
The solutions to the above equations are shown in
actly 1/2. Thus attention can be restricted to cases Fig. 3. (Remember case C is important as (35), and
where hence (41) is valid only for ti > 1/2.) The various por-
tions of the solution curves are as follows:
ti >. 1 /2. (34) Case A: For 1 < k < 2 three solutions exist, one
of which (1) is the global minimum and thus the op-
With the statistics of yi as given by (30), and the timal threshold pair. One solution is always negative,
decision rule (33), the value of p(ui = 0 H) can be and hence corresponds to the equivalent solution t =
found as t2 = 1/2 (2), and the third is greater than 1/2 for 1I
k < 1.54 (3). The latter two are person by person op-
p(ui = 0 | H°) = exp {-A[(1 /A) In(2ti)]} = (2ti)-' timal in that if only one of the thresholds is varied,
performance degrades; it improves only if both are
p(ui = 0 HI) = exp {-2A[(1/A) In(2ti)]} = varied together. For 2 < k only one real solution ex-
ists; it is optimal for 2 < k < 3.21 (4), but becomes
(2ti) .
(35) suboptimal for 3.21 < k as case B yields a better solu-

506 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-17, NO.4 JULY 1981

Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:07 UTC from IE Xplore. Restricon aply.
tion (5). In this region changing one threshold alone
still produces degradation; improvement results only TABLE I
from varying the two thresholds in opposite direc- Cost Functions for "and" and "or" Fusion
tions.
J(u,,U2,H)
Case B: These equations (41) have real solutions u, u2 H "and" "or"o
only for 3.21 < k and represent the optimal solutions
(7). Notice the paradox here: each detector is identical O 0 HoI coo coo
to the other, and the cost function is symmetric, yet o o H' cot Coi
unequal thresholds are optimal. O i HO coo Cio
O I H1 Coi Cl
Case C: Since case B supplies solutions with t1, t2 1 0 HI coo C,O
> 1/2 only in the region 3.21 < k < 3.42, the optimal 1 0 HI co, Cl
solution becomes ti = 1/2, tj = k - 1 for 3.42 < k, i 1 0 H° c,o CIO
# j (7). This behavior continues for any higher value 1l 1 HICIl C,,

of k. Note: J(uo, H) = c.u

The apparently paradoxical behavior of (3) and (4)
is due to hedging by the decision rules as the cost of
double errors increases. In the limit as k - 00, double For a fixed p(u,, u,, u2), this can be placed in the
errors must be avoided at any cost. One way to do previous framework by defining
this is to have one detector constantly choose u, = 0
and the other choose uj = 1, regardless of the data. J(U1, u2, H) = p(U0 = 0 | ul, u2) J(0, H) + p(u0
While this degenerate behavior is undesirable and easy
to see here, it may arise in more subtle ways resulting = 11 u1, u2)J(1,H). (43)
from certain combinations of payoff function and
observation distribution in other situations. Two particular interesting cases exist when a) the
center chooses u0 = 1 only if both ul and u2 are I
III. Application to Data Fusion (the "and" rule), and b) the center chooses u. = 1 if
either ul or u2 is 1 (the "or" rule). In these cases,
So far only the structure of Fig. 1 has been con- J(u,, u2, H) is given by Table I. Notice that this cost
sidered, where each detector operates independently function is not separable.
and there is no communication at all. We now con- Approaching the problem of restricted com-
sider Fig. 2 and show that the results of the previous munications capacity in this manner, it is possible to
section apply, provided the activity of the fusion analyze its effect in terms of the overall system per-
center is fixed a priori. formance. By considering various cases of p(yi H)
The situation of Fig. 2 arises when there are con- and observing the effect of varying the system struc-
straints on the amount of information that can be ture for a fixed systemwide signal-to-noise ratio
sent to the data fusion center because of channel (SNR), the tradeoff of decentralized versus centralized
capacity or similar limitations. In the absence of these performance can be explored.
limitations the optimal strategy for each detector is to
simply transmit its observation yi for centralized deci- Example 24
sion making. However, strategies requiring less com-
munication are clearly of interest. The simplest such We consider a case of two independent Gaussian
scheme is for the detectors to just report detections. observations. Let those be denoted
This approach can be implemented with greatly re-
duced channel capacities for two reasons. First, a Yi = m, + vi (44)
report of a detection is a simpler message than a sen-
sor observation, and thus can presumably be coded where
into a smaller number of bits. Second, most observa-
tions need not be reported at all since they do not cor- 09 under Ho
respond to a detection. The question associated with Mi (45) =

this scheme is: what are the optimal local detection ai, under HI
rules?
and
Problem (Decentralized Dectection with Fusion):
Given pO, p', the distributions p(YO, Y2 Hi), and the vr\jN(0, 1) (46)
fusion rule expressed as p(uo ui, u2), find the deci-
sion strategies for each detector p(ui y,) to minimize
the expected system cost J(uo, H). 'See [41 for the details of this analysis.

TENNEY/SANDELL: DETECTION WITH DISTRIBUTED SENSORS 507

Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:07 UTC from IE Xplore. Restricon aply.
 and using

Yo O to (48)
obtained from (13). The SNR is given as

PD
SNRD (u1, u2) = (a' + af)/\ a? + af
= 2a,+ a2 (49)
and reflects the entire information available to the
system.
For this example, construction of the receiver
operating characteristic (ROC) curve plotting

PD = Pr (Uo = I IH1) (50)

and

Fig. 4. ROC for case A. PF = Pr (uO = I HO)

Fig. 5. ROC for case B. as functions of one another is done by varying the
ratio c0jc,0 as some factor aE (0, °°). As a decreases,
P-
F
both the probability of detecting HI (PD) and the pro-
4 .5 .6 .7 bability of false alarm (PF) should decrease as the set
of y, which map into ui 1, and therefore uo = 1,
=

shrinks [4].
The ROC curves for the "and" decision rule
demonstrate a) how decentralized performance com-
pares with centralized performance for identical detec-
P
D
tors over a range of system SNR, and b) how dif-
ferences in local SNR affect system performance for
fixed system SNR.
The curves for case a) are shown in Fig. 4 for two
values of system SNR in each case. Clearly increasing
SNR leads to improvement in both cases, but it is im-
portant to notice that the performance advantage of
the centralized curve over the decentralized one in-
creases with SNR. This is a reasonable result of the
centralized decision rule making more efficient use of
the increased information.
independent of everything else. Notice that the op- The curves for case b), shown in Fig. 5, are for
timal centralized detection rule for this case is ob- constant system SNR of 2.0, but with local SNR vary-

tained by forming ing from identical to 0 for one and 2 for the other.
This leads to two interesting conclusions, 1) as the im-
yo - aly, + a2y2 balance between detectors increases, performance im-
proves, and 2) in the limit, as the SNR of one detec-
which is described as tor goes to zero, performance reaches that of a cen-
tralized system.
yo = aO + Vo (47) These can be understood by examining the
thresholds as the relative SNR changes. As it goes to
where zero, the threshold for the most noisy detector goes
to 0, leading it to declare ui = I (detect) regardless
-

vo r
N(0,a? + a-) of yi. The "and" decision of the fusion center then
just sets uo = u;, so all decision making is really being
under I
done by the detector with high SNR locally.
aoa= Curves similar to the above can be computed for
a12 + a2 under I (47) more complicated examples to characterize the perfor-

508 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-17, NO.4 JULY 1981

Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:07 UTC from IE Xplore. Restricon aply.
mance tradeoff between centralized and distributed DISTRIBUTED IMPLEMENTATION
detection. Then the communications requirements can
be compared to choose a preferred architecture. Z2
In affecting such a comparison, it is important to IL I
realize that the centralized decision rule can be im- .t2- A ----
plemented in several ways. Consider Fig. 5 in which m =
the optimal decision region for the previous example
is the area above and to the right of the diagonal line. (Z.,Z2) e Ir object detected
Observe that the optimal centralized decision rule can- (Z1,Z2) E m =:object not detected
(Z 1,Z2) e 1,lILm ambiguous situation,
not be implemented by separate operations on yi and must communicate

Y2 However it is possible to locally test for Fig. 6. Distribution implementation of optimal centralized detec-
tion rule.
Y1 kt
Y2 t may be locally optimal but not globally so, 4) these
solutions involve hedging and consequently counterin-
and to send reports of threshold crossings only to the tuitive results, such as two identical detectors having
fusion center. The fusion center can then determine in different thresholds even for symmetric cost func-
which of the four regions shown the composite signal tions, 5) the theory applies to the fusion problem
(Yv, Yy) lies. Only in the case of regions II and IV is through appropriate definition of a payoff function,
further communication necessary to determine and, 6) comparison of centralized and distributed
whether (y,, Y,) lies above the threshold line. decision rules shows the tradeoff between communica-
This communication can be implemented in several tion and performance.
ways. For example, the fusion center can request the Of course the present analysis can be extended in
measurement from the sensor that reports a detection. many directions. For instance, the case of m-ary
It can then compute a lowered threshold for the sen- hypothesis testing with n local detectors could be con-
sor that did not request a detection, and request that sidered, as could the case of observations that are
its observation be reprocessed. stochastic processes. In this latter situation, sufficient
The advantage of such a distributed implementa- statistics must be determined; it is more or less clear
tion of the optimal decision rule is that communica- that the classical approach [5, pp. 229-2321 can be us-
tions are event driven. Thus the average data rate ed in the known signal-to-noise case when the noise
needed to support optimal centralized decision making processes for each sensor are independent. The case of
can be reduced from that implied by transmission of observations that are dependent is in many ways the
all observations to the fusion center. most interesting and is essentially wide open. It is
possible to consider decentralized estimation as well as
IV. Conclusions and Suggestions for Future Research detection; see [6] for some initial results.
A decentralized optimal hypothesis testing problem
has been introduced to provide a theoretical Acknowledgment
framework for detection with distributed sensors.
Most of the results were obtained for the case of Some of the ideas for this paper are the result of a
statistically independent observations and we were discussion between N.R. Sandell, Jr., and Dr. J.
able to show that 1) each local detector implements a Lawson of Naval Electronics System Command.
likelihood ratio test as its optimum decision rule, 2) The numerical computation of the optimal
the computations of the thresholds for these tests are thresholds and corresponding ROCs for example 2
generally coupled, 3) solutions to these computations were carried out by T. Jones [4].

TENNEY/SANDELL: DETECTION WITH DISTRIBUTED SENSORS 509

Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:07 UTC from IE Xplore. Restricon aply.
References [41 Jones, T. (1979)
Analysis of decentralized detection networks.
[1] Lawson, J. (1978) B.S. thesis, Massachusetts Institute of Technology, Cam-
An elementary theory of surveillance. bridge, May 1979.
Naval Electronics System Command, Technical Report, [5] Wozencraft, J.M., and Jacobs, I.M. (1967)
Aug. 1978. Principle of Communication Engineering.
[2] Van Trees, H.L. (1968) New York: Wiley, 1967.
Detection, Estimation, and Modulation Theory, vol. 1. [6] Barta, S.M. (1978)
New York: Wiley, 1968. On Linear Control of Decentralized Stochastic Systems.
[3] Sandell, N.R., Jr., Variaya, P.C., Athans, M. and Safonov, Ph.D. dissertation, Massachusetts Institute of
M.G. (1978) Technology, Cambridge, July 1978.
A survey of decentralized control methods for large scale
systems.
IEEE Transactions on Automatic Control, 1978, AC-23,
108-128.

Robert R. Tenney (S'77, M'79) received the B.S. degree in computing science and
engineering, the M.S. and Ph.D. degrees in electrical engineering in 1976 and
1979, respectively, all from the Massachusetts Institute of Technology, Cam-
bridge. He has been employed as a consultant with Honeywell's Systems and
Research Center, the Institute for Defense Analysis, and Alphatech, Inc., and in
1979 he became an Assistant Professor in the Department of Electrical Engineer-
ing and Computer Science at M.I.T. His current research interests include the
development of practical methods for controlling large, stochastic, dynamic
systems, methods for representing system structure which blend systems theory
with concepts of artificial intelligence, organizational structures for distributed
control, and estimation theory.
Dr. Tenney is a member of Sigma Xi, Eta Kappa Nu, and Tau Beta Pi. In
1975 he received a National Science Foundation Graduate Fellowship in Com-
puter Science.

Nils R. Sandell, Jr., (S'70-M'74) was born in Brooklyn, NY on June 13, 1948.
He received the B.E.E. degree from the University of Minnesota, Minneapolis, in
1970 and the M.S. and Ph.D. degrees from the Massachusetts Institute of
Technology, Cambridge, in 1971 and 1974, respectively.
In 1974 he joined the faculty of the Department of Electrical Engineering and
Computer Science, Massachusetts Institute of Technology, Cambridge, as Assis-
tant Professor and he was promoted to Associate Professor in 1976. While at
M.I.T. he served as a consultant to various organizations including the Analytic
Sciences Corporation, Adcole Corporation, the Lincoln Laboratories of M.I.T.,
and the Honeywell Systems and Research Center. In 1977 he received the Donald
P. Eckman Award of the American Automatic Control Council for his contribu-
tions to control theory and applications. Since 1979 he has been President of
Alphatech, Inc., Burlington, MA, and a part-time lecturer at M.I.T.
Dr. Sandell is a member of Eta Kappa Nu, Tau Beta Pi, Sigma Xi and the
American Institute of Aeronautics and Astronautics.
510 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES-17, NO.4 JULY 1981
Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:07 UTC from IE Xplore. Restricon aply.