Académique Documents
Professionnel Documents
Culture Documents
I.
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
Information nd Multi-Sensor Coordination
Greg Hager ad Hugh Durrant-Whyte*
GRASP Laboratory
Deparment of Computer and information Science
Universit of Pennsylvania
Philadelphia, PA 1910
Abstract
The contol ad integation of distibuted, multi-sensor per
ceptual systems is a complex ad challenging prblem. Te
obseations or opinions of diferent sensors ae often dis
paate, incompaable ad ar usually only paal views.
Sensor inforation is inherently uncertain, and in addition
the individual sensor may temselves be i eror with re
spect t the system a a whole. The successful operation of
a multi-sensor system must account for this uncerinty and
prvide for the aggegation of dispaate informion in a
intelligent and robust maner.
We consider te sensor of a multi-sensor system to be
member or agents of a tam, able t ofer opinions ad
bagain in group decisions. We will analyze te coordi
nation and contol of this stctur using a ter of team
decision maing. We present some new aalytic rsults on
multi-sensor aggegation ad detail a simulaton which we
use to investigate our idea. This simulaton prvides a ba
sis for the aalysis of complex agent stctes cooperating
in te presence of uncertaint. Te results of tis study
are discussed with reference to multi-sensor robot system,
distbuted AI ad decision making under uncerainty.
1 Introduction
The general problem of seeking, sensing, ad using perep
tual inforation is a complex ad, as yet unsolve problem.
Complications aise due t inherent uncernt of infora
tion fom perceptual souces,. icompleteness of inforation
fom partial views, ad questons of deployment coordina
tion ad fusion of multple data soures. Yet aoter dimen
sion of complexit result fom orgaizational ad compu
tational considerations. We feel that these three topics -
inforation, contol, and organization - are fundamental for
understading and constucting complex, intelligent robotics
systems. In this paper, we ae conceed with developing
useful aalytic methods for describing, aalyzing and com
paing the behavior of such constuctions based on tese
crteria.
Tis mateal is be on wok supor unde a National Science
Fundation Gadute Fllowship and by te National Science Fondation un
de Grant DMC-8411879 and DMC-12838. Any opinion, fndings, and
concluions or reomendtions epese in tis publication ae to of
te authos and do not nearily refet te views of t Naional Scienc
Fundation.
We assume fom the outet that such robotic systems
ae basically task-riented, goal-directed agent. The be
havior of a system is determined entirely by the goal it is
working toward, and the inforation it has about its envi
rnment. At any point in time, such an agent should use the
avaiable information t select some feaible ation. The
most preferable action should be tat which is expected to
lead the system closest t te curnt goal. In short we
wi consider te queston of driving robotics systems as a
lare ad complex problem i estmaton and contol. To
aopt te nomenclatre of decision thery[2], at ay point
in time a agent has a local informion structure refecting
the stat of the world, a set of feasible action to choose
fom ad a utilit which supplies a preference ordering of
ations wit respect t states of te world. We generally
asume tat a rationl deision make is one which, at any
point in tme, takes tat action which maximzes hs utility.
Our commitent, as a result of catng te problem in a de
cision theortic perpective, is t provide prncipled means
for specifing inforation stctures, actions, and (perhaps
most crucially) determination of utilit.
This monolitic forulation is certainly too naive and
general t successflly attack the prblem. The state of
te system is a complex entit which must be decomposed
ad analyzed t be understood. The rsulting proceures for
contol will undoubtedly be computtionally complex. Com
puter resources, like human problem solvers, have resource
limtations which bound te complexit of prblem that can
be solve by a single agent - oterwise kown as bouned
rationai [2]. Such computational considerations suggest
distbutng te workload t increae te prblem solving po
tential of the system. Fm a practical stadpoint te system
itself is composed of physically distinct devices, each with
its own special characteristics. Softar and hardwae mod
ules should be designed so that inforaton ad contol local
to subtasks is kept locally, and only information gerane t
other subtaks is made available. Ultatly, sensors and
subtasks could independent modules which can be added or
remove fom a system witout catastophic result. In this
ca we desire each subtask t have the ability to cooper
ate and coordinate it actions with a grup while maintaining
it own local processing intlligence, local control vaiables,
and possibly some local autonomy.
Our solution is t view the systems as decomposed
99
into several distinct decision-maer. These modules a
t be organized and comunicate in such a maner as t
achieve the comon goal of the system. Oganizatons of
this tpe of are often refered t a team[8,16]. We pro
pose to consider a team theortic forulation formulation
of multi-sensor systms in the followig sense: Te aent
a considere as members of the tam, each observing the
environment and making local decisions bae on te infor
maton available t tem. A manaer (exeutive or coor
dinator) makes use of utilit considerations to converge te
opinions of the sensor system. Secton 2 will be devoted
t a review of team decision theor and present some new
analytic results[7].
One crticism of decision teor is tat optal solu
tions are ofen diffcult or impossible to fnd. In orde to
aid in analysis of these poblems, we have built a simula
ton environent We use the simulaton t exane varous
non-optimal and heuristc solutions t oterise intactble
problem, ad experent wit differnt loss functons t de
terine the charactr of the resultant decision metod. The
simulation is a generalizton of classic puruit and eva
sion games [141 to team of pursuer and evader. Eah
tam member ha loca sensor ad stat vaables. They
a coordinatd though a ta exeutive. Seton 3 will be
devotd t o a detile look at the simulaton ad our rsults
t dat.
We feel tat te team forulaton of snsor systm
ha implications for te broade study of Acial Intlli
gence. AI is rlevant t tis worl in at least t rspet:
Fisty, it is cely possible t consider the agents
of te systm as perforng som reaoning prcess.
Considerng AI system as deision-maer seems a
plausible apprah to te constcton of intlligent
distbuted systems. Thus, this wor ha commonali
tes wit Distibutd AI i tat bot a interestd in
questions of stuctrng inforaton ad comunica
tion betwen intellgent systm.
Seondly, we oftn wat t inerpret te information
available t t systm, and t comunicate infor
mation as intrrettons rater ta simple signas.
This is prmaily a prblem in representation of infor
mation. Again, AI has focussd on te intrprtation
of inforation, and te reprsentton of tat intr
pretation.
More generally, we would l t discover when systems like
tis ca be prftably pose a decision problem. Sction 4
will be devote t an in dept discussion of te general me
its and shoromings of te oraiztonal view, ad atempt
t defne when it is most appropriate o usfl.
2 A Team-Theoretic Formulation of
Multi-Sensor Systems
Team ter originated fom problem in game tery [26]
and multi-persn contol. Te bais for te aalysis of coop-
eration amongst stctures with different opinions or intr
est was formulated by Nah [20) i the well kown bagain
ing problem. Nash's soluton for te to person cooperative
game wa developed into the concept of information, group
ratonalit and multi-peron decisions by Savage [24]. Team
ter has since been extensively used by economist t a
alyze stct [16], inforaton [18] ad comunicaton.
Secton 2.1 intduces te tam stuctre and defnes te
fnction of te team membes ad manager. Diferent tam
orgaizatons a discussed ad t concepts of information
stcture, tam decision, team utility and cooperation ae
defned in Section 2.2. Secton 2.3 applies these techniques
t the multi-sensor team and a method for aggregating opin
ions is derved. Due to lack of space, we will assume some
familiaity wit probabilit ad decision teor.1
2.1 Team Preliminaries
A sensor or member of a ta of sensor is chaactized by
its information stucture and its decision function. Consider
a team comprsing of " membes or sensors each making
observations of te state of te envirnment. The infora
ton stcture of te i'" team member is a function 'i which
describes te character of te sensr obserations z; E J; in
trms of the state of te environment 8 E 8, and the other
sensor actions a; E A;. j = 1, , "
So that:
(1)
<ollectvely te n-tple 1 = (';,, 'n) is called te in
foraton stuctre of te team. The action a; of te it
"
team member is rlated to its inforaton z; by a decision
fnction 5; E /; as a; = 5;(z;). We may also allow ran
domized rles, in which case 6 associats inforation with a
distribution over the set of feasible actions. Collectively the
n-tuple 5 = (51,, 5n) is calle te tea decision func
tion. Rr an estimaton problem, the action space A; is the
same as te spac e of possible stts of natre 8: Our acton
is t choose a estimat a; = 8; E
8
.
Thee are a number of different for that the infor
mation stctur ca t which in tu chaacterizes te
tpe of prblem t be slved. If for all tam member 'i
is defned ony on e ('i! e -}) te resulting strcture
is called a static team [16]. When 'i also depends on the
other tam members' actions, then the stcture is called a
dynic tea (13]. Clealy a eah team membe ca not
mae decisions ad be awa of te result simultaneously,
te general for of inforation stuctur for a dynamic team
must induce a causal relation on the team member actions
a
;
. We can apply a preedence stuctre on te time instat
a member maes a decision, so that if member i maes a
decision prior to meme j ten the inforation stcture
I
i will not be a function of a;
. Indexing the tam members
by thei decision maing precedene order we ca rewite
te inforation stuctre a:
z; = 7;(8, a1 , a;-1}
(zn))] (2)
a;EA;
If te clas of group decision rles D includes all jointly
randomized rles then L will awas be convex. If we re
ally believed in an altuistic team, we must use this clas
ad be subject to these result. Considerable work ha been
done on fnding solutions to equation 2.3 under these con
ditons [16,13,12,11], particulaly as regards the effect of
inforation stctue on distbuted contol problems.
We ae primaly interested in team of observers-sen
sor making obserations of the state of te environment. In
tis case the team members ca be considered as Bayesian
estimators, and te tam decision is to come to a consensus
view of te obsred state of nature. The static team of esti
mators is often called a Multi-Bayesian system [28]. These
system have many of te same chaacteristics a more gen
eral team decision problem. Weerahadi [27] has shown
tat te set of non-randomized decision rles is not complete
in these system. If to tea members using decision rles
6 = [61,6:] have utlities u(8) = u1(61,l) and u2(62,8),
ten the team utilit fnction L(8) = L(u(8)) will only
admt a consensus i it satisfes te inequalit:
E[L(u(8)));L(E[u(8))) (3)
This is te Jensen inequalit, and it is well kown that tis
will be satisfied i and only i the function L( u( l)) and
te risk set ae convex. Generally, tis will only be true
when the set D of decision rles includes jointly randomized
decision rules.
Consider te team utilit L a a function of the team
member utilities so tat L = L(u1, , un) = L(u). The
group rationality principles described above restict te func
tions L that ae of interest t tose that have the following
properties[ I]:
1. Unanimity: . > 0 Vi
lu;
,
2. No dictator: If 'i: u
;
= 0, there is no Uf such that
L= Uj.
3. Indifference: If Vi, 351,62 such tat u
;
(61,) =
u
;
(62, ),then L(61) = L(62)
If the tam utilit function L satisfies these properties, we
will say tat the team is rational. The function L is often
called an "opinion pool". Two comon examples of opinion
pools are the generalized Nash product:
n
L(8;61, ... ,6n)=ciu;(6
;
,8) a;;: o
i=l
ad the logaithmic or linea opinion pool:
n
L(8; 81, .
.
. , On)= L >;u;(8;, e), >.; ;:0,
i=l
101
The value of te generaize Nah product ca be seen by
noting tat if u;(6;(Z),8) = f(z; !8) and a; = 1 ten L
i the posteior densit of 8 with respect t te obseratons
z;. A crtcism leveled at the geealize Nah prduct is
that it asumes indeendence of opinions, however tis may
be acountd for tugh te weight a;. A cticism of
te linea opinion pool is tat tee is no reinforcement of
opinion.
Suppose we now restct gup decision rules 6 E D
t non-randomzed deisions. Ts alows team mmber
t disagre in the following sense: If te tam rsk set
u = [u1(61,8), un(6n,8)] i convex for non-radomzd
6, then equaton 3 holds a a consnsu may be reache. If
however u is concave i at leat one "" ad if randome
rles ae disallowed, it is bettr (in t of utlit) for te
associat tam mmber t dsag: a i tey wee atng
as a an atagonistc t. It shuld be clea fm t ex
ample that te differne beten atonistc an altuistc
tam i te abilit t obtin a convex "opinion" sa.
If all the U; a convex fnctons, ten L will aways be
convex on t clas of non-radomzd decisions. However
i locaton estimation o Mult-Bayesian system, te u; will
oftn be concave s that L(u) will be guaateed convex
only i te class of radomizd rles. Thus L( u) will always
be convex for an altistic team. Fr an antagonistic team L
will only be convex when agrement ca be reached (in te
class of non-randomized deisions), oterise if opinions
diverge sufciently then L will be concave. Concavit will
geneally te te for of sepaating team membes int
convex goups of opinions coaltions which may overlap.
Ou itrst in these result centes on fnding when
agement can be reached ad in calculating te value of the
consnsus. We sume tese concepts in te following:
Result 1: Consider a team wit member utlities u;(6;19)
and ta utlity satsfying the grup ratonait conditons.
Then:
1.1. Conenus: Cooperation will only occur when te set
of risk poit L(611 ,6n) E R
n
is convex.
1.2. Altruistic: If 6 E [ is te clas of all randomized
decision rules then L will always be convex.
1.. Anagonistic: If Vi, L u; then L will be convex in
te clas of non-randomizd deision rles.
1.4. Diagreemn: When L is concave ther is no best
decision and agreement cannot be reached.
The point at which L beomes concave for each member
is called the disagreement point te value of a member's
utility at tis point is called te security level.
2.3 Multi-Sensor Teams
The fusion of sensor obserations requires tat we have a
metod for comparing information fom disparate sources.
We consider each sensor to be a member of an antagonistc
team in the following sense: Eah sensor comes up wit un
ceran partial views !f te state of the envionment te goal
of the executve is t integrate the vaous sensor opinions ,
by oferng incentives ad intrprtations for comining dis
paat viewpoints. Te atagonistic ta stuctur allows
member t disagree if for some reaon tey have made a
mste or canot rconcile their views with tose of the
other ta members. A altistc tam could not tae this
aton.
We suggest tat te compaison of diverse obseratons
can be intepreted in tes of a compason of the utility of
a consensus decision. Suppose we have two observatons z1
ad z2 which ae not direty compaable. Each observation
contbutes t some higher level descrption of the environ
met , and each i depedent on te other. We can interret
ay deision 6 aout te environment i t of it utilt to
te obsevations: u1(6(z1)18) and u2(6(z2)18). Although
z1 ad z2 canot b comae diretly, teir contibutons
t pacula decisions can be evaluated in a comon utilit
fawork. The team teretc comparisn of utilities a
mts a masur of disagement ad allows for te evaluation
of sesr inforaton in a consistnt maner.
Define e t be te st of stats of natre and consider
a robot system with snsors 81, J = 1
1 1m, ting se
quences Of ObseratiOnS Z; = . { zf,
I
zn Of featureS in
te environment. We will rstct interest t te static team
stctue so tat Z; = f;(6). Lcally, sensors can make
decisions basd on local obsevatons as 8 = 6;(z;) frm
COmpaable sequenceS Z = {zl 1
I
zn, With respect tO
a comon utilit u;(6;(Z;)1 9). Jointly the sensor team ha
a utilit L = L( 8; 611 1 6n ), which can be considered
a a functon of te individual utilites L = L(u11 , u2)
satsfying the group rationalit conditions.
If the obseratons from different sensors a incompa
rable, they must be interpretd in sme comon famework.
This will be the cae when te snsor ae located in differ
ent locations for examle. Let
D
; interrt S;'s obseratons
in sme comon descrption famework. Ten the team loss
can be writen as:
L( u1(ct[D1(z1)
]1 8)1
1 u
,.
(
o[D,
(
z,
)
]1
8)
]
)
By selectng L and aalyzing its convexit, we will establish
te character of te sensor team.
The rationality axioms derved fm utilit theory re
quire that we be able t put a prference ordering on deci
sions ci(). It seems reasonable tat the preference ordering
admitted by a observation Z; will be the same ordering as
tat obtined by a maximum likelihood estmator (unbiased
Bayes rationality). In this cae, te utility fnction of an
obseration will be coincident wit its likelihood function.
Thus te Gaussian distibution N( Zi 1 A;) associated wit the
observation
z
; can aso be considered as the preference or
dering or posterior utility function of z; on any resulting
estimate 8. In this framework. to observations Z; ad z1
will have a bais for agreement only if their combined utilit
exceeds their individual utility, that is a consensus can only
be reached if te set of observation utilities form a convex
set.
102
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
To f tis, defe u;(z,8) = f;(ztl8) " N(z;, A;) as
te loss t te obseration Z; of te estmat 8, ad let
denote te ta utit. Ten, in tr of expected utiit, a
consensus ca only be reached i L satisfes Equaton 3, i.e
te fnction L i convex.
Te functon L will be convex i ad only if it matx
of second order derivatives is non-negative defnit. If L
satsfes te grup rationalit principles, tis requies that
0 fori= 1, , n. Differentiating u; gives
au
0
8
3' = 1- ,- .\,
,-.\]n(..\
Fr tese to be positve, and hence te set u e R
n
t be
convex, we ae must fnd a 8 which satsfes:
(0- z;
)
' A/
1
(8- z;
)
S 1 (4)
Fr all i = 1, , n obserations.
Consider any two observatons z; ad ZJ They ca
form a consensus i we ca fnd a 9 that satisfies equaton
4 for bot z; ad ZJ. To compae obserations, we interret
tem in a comon framework a
D;(z;) and D1(z1). If
J; and Ji
ae te jaobias of
D; ad Di respetvely [6],
define
.
E;= Ji1 Ai1J;
'
. This i te inforaton matx
of te obseration Zt tansfored to the comon frame of
reference by the transforation D.
Since the left hand side of equation 4 is always positive,
we must .nd a 8 which satisfes
i(8-D;(z,)?.E, (9- D;(z;)) +
(0-Dj(Z)i)T.Ej (0-Dj(Zi)) s 1 {5)
The value of 8 which maes the left hand side of this equa
to a mnimum (ad which i also the consensus when
it exists) is given by the usua combinaton of noral
observations[2]:
8 = (.E;+ Dri)
-l (D!;D;(z,)
+
DriDi(zi))
Substituting this into equation 5 gives:
i(D;(z;
)-Di(zi))'Dr,(Dr, +DEi) -1
DEi (D;(z;)-Di(zi)) S 1 (6)
We will say tat z; ad Zi
admt a Bayesian (non
randomized) consensus i ad only i they satisfy equation
6. The left side of Euation 6, which we will denot as
d
i
is called te generalized Mahalanobis distance (a re
sticted form of this is derved in [27]) ad is a measure of
disagreemnt between to observations. Figure 2.3 shows
plot of u; against U
j
for vaious values of a11 and which
clealy demonstate tat te convexity of the set [u;, Uj]
corresponds t requiig tat d
i
S 1. This measure ca be
further extended t consider more than two observations at a
time. Fr example, if each obseration z;, i = 1, , n has
v.
v,
Figure 1: Plot of Mahalanobis distances
te same vaiance-covariance matix A, then a consensus be
obtained only i:
1
n n
d <l
2n2 L.L. 'J -
i=l i=l
(7)
It is clear that a set of obserations that satisfy Equation 6
pair-wise wll also satsfy Equation 7
In most real situations, it is unlikely that we will kow
te varace-covarance matces exactly. In tis case, any
estimates of the A; act as i tey were tesholds in te sense
tat te larer te A; that i used, te more disagreement will
be toleratd.
3 Simulation Studies
To ti point, we have discussed the theoretical aspect of
estmation in the tam framework. Our goal is t even
tually pose problems of multi-sensor control in coordina
tion and solve tem in a simila manner. However, fnding
and analyzing solutions to decision, contol, or game prob
lems, especially in the face of anyting less ta perfect
inforation, can be extremely difcult. Fom a technical
perspectve, solutions under even relatively simple losses
ae complex optimization problems. Other heuristic or ad
hc approaches must oftn be considered. Methodologi
cally, tere is a queston as to what proper loss functons
ae for different prblems. Ideally, the loss function should
refect the actual state of affais under consideration since
it refects te preferences of the decision maker. Whereas
in te economics literature, losses ae usually derived fom
utilit considerations based on monetry rewads, we have a
mch wider set of competg critria to consider. This com
plicates matters to the point that we need to gain intuition
about te issues involved before hypothesizing a solution.
In order t deal wit tese issues, we have constucted
a simulation on a Symbolcs Lisp MachineThe simulation
takes the for of a game of pursuit and evaion simila in
103
chaactr t te classic diferntal game kow a te h
mcid chufeu [14}. That classical for of tis game
consists of a puruer ad evader moving at constat velocit
in a plane. Both player have pefect inforaton about te
oter's stat, and atempt use ths information t intrcept
or evae thei opponent rspectvely. Te payoff stct
of the game is te te untl captre. The major chages we
have made a tat we have euippe te playes with im
perfet sesing devices (i.e. te playes use imeect stt
inforation), ad we allow multple pusues ad evaes
gruped into teams coordinatd by a tam executive. It is
importnt t note tat te motvaon for using te pusuit
evaion famework is prmay t provide eah ta wit
a well-defined method for comparing stuctrs and contol
policies. The game is not of intnsic value by itelf, but
forms a stct, fexible, closed systm in which sensor
models, organiztonal stctrs and decision metods may
be implementd ad eaiy evaluatd
The simlaton i constct so tat we ca vmy t
stcte of ta member, as wel as oveall ta stctre,
a quickly evauat the effect of te chage basd on te
chaatr of te simulate gam tat ensues. We have in
mind t allow varaton in such fator a dyamcs, sen
sors, informtion intgration polcies, incentive stctures,
an uncertaint of inforation, and obsere what tpes of
policies lead t the aequat peforance in tes circum
stnces. We expect t tasport what we lea fm the
simulaton t real-wrld prblem of mult-sensor rbot sys
tm curently being developed in the Grasp labortr[21}.
We imagine a situaton wh tis simlaton provides an
environment in which distbut exper cordinaton and
contl prblems ca be investgatd befor implemntation
and converely tat applicatons of the sensor systm unde
development will suggeSl what diections, snsor models and
dynamics would most fitul t explor in te simulation.
The remaider of tis scton details te curnt stctre of
te simlation envirnment ad outlines our initial exper
ences with it.
3.1 The Game Environment
The simulation tes place on a plana feld possibly litred
with obstacles. The baic cycle execution involves team
members taing sensor readings, exeutives intgratng in
formation ad offering incentives, and fnaly ta mem
bers making decisions. The stat variables a updated a
te game moves t a new step. A game terminats when
and if the pursuit robot;which ae euipped wit a simple
ballistics system, capture all te evaders. This is a medium
level of granulaity wit emphasis on the general behavior of
tams, not the precise performace issues of ta members.
Some time-constaint issues can be investigate by includ
ing time paameters in te payoff functions, but computa
tional complexity issues ad investigations of aynchronous
behavior are outide the scope of our considerations. Fr
instnce, if some decision policy is computationally more
complex than another, differences in performance will not
refect tat complexit.
3.2 The Stucture of Team Members
The chaactr of individual team members is detrned by
te modules:
1. The kinematics ad dynamics of motion on te plane,
2. what sensor a avalable ad the noise chaactr
istics of tose sensors, ad their kinematics and dy
namics, and
3. the ballistics which detrine te termination of the
game.
The tm memes a constat velocit, vaiable di
rction unit opeating i a plae wit stat variables z, y,
ad 8. Since t robot move wit constat velocit, te
only dircty contlled varable is 8. The only dynamical
consideraton involved is how we alow te robot t chage
its curnt heading t some new desired heading 8 4 - the
single contl. Cunty, we asume that when reorienting
each agent can move with some fed (possibly infinite) ve
locit, w. The has te effect of defining a minima tring
rdius. Pursuers generally have some fnite w, while evaders
have innit w - i.e. they tm instteously.
The sensor model we a curently using is a range ad
dirction sensor. The sensor has a limitd cone of dat gat
eing, and a liit rage. It ha a single contol vaiable
a which t robot ca selet t point te snsor. We a
sume tat snsr tically rtr noisy data, s we have
difent noise models which we "wap aund" te sensor
t make it more closly approximat real dat gateing de
vices. Te induces decision problem in dealing with both
te noise and rage limitations of devices. Te fat te te
snsor ae distbute intoduces issues in intgrating noisy
obserations fom different fames of reference[6]. Finally,
since sensors are tansported by the robot tere ae issues
involved in resolving the confl ict betwen action for pur
suit or evasion, and actions which will allow more effcient
gaterng of information.
Terination of the game occurs when all evaders are
eliminatd. We define a captue region which delineates
how close a pursuer must come to eliminat an evader.
However, when information is noisy, te aea in which the
evader ca be located will have a associatd uncertainty.
We sometimes equip each pursuer wit some mehaism t
"shoot" evaders, allowing the possibility of uncertainty in
obseation t make it "mss". Pa of the payoff strcture
of the game can include csts for using projectiles and miss
ing; thereby adding incentive t localize the evader t te
best degre possible.
3.3 Information Structures, Organization, and
Control
The intesting issues ae how the robot systems ae contolled,
ad how team members interact Each team member must
104
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I