You are on page 1of 12

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 18, NO.

1, FEBRUARY 2010 41

Detection of Intelligent Mobile Target


in a Mobile Sensor Network
Jren-Chit Chin, Yu Dong, Wing-Kai Hon, Chris Yu-Tak Ma, and David K. Y. Yau

Abstract—We study the problem of a mobile target (the mouse) operational and functioning according to the planning stage.
trying to evade detection by one or more mobile sensors (we call Instead, with advances in robotics and unmanned aerial vehicles
such a sensor a cat) in a closed network area. We view our problem (UAVs) [12], it will be feasible for the zone to be monitored
as a game between two players: the mouse, and the collection of
cats forming a single (meta-)player. The game ends when the mouse
by a balance of mobile robots, UAVs, patrol vehicles, etc,
falls within the sensing range of one or more cats. A cat tries to de- according to the deployment conditions. It is then important
termine its optimal strategy to minimize the worst case expected to control and coordinate the patrol routes in order to achieve
detection time of the mouse. The mouse tries to determine an op- effective area monitoring.
timal counter movement strategy to maximize the expected de- We consider the use of a group of surveillance sensors, under
tection time. We divide the problem into two cases based on the possible coordination, to secure a network area against one or
relative sensing capabilities of the cats and the mouse. When the
mouse has a sensing range smaller than or equal to the cats’, we
more intruders. An intruder (e.g., an enemy vehicle) may lurk
develop a dynamic programming solution for the mouse’s optimal in the area to prepare for damage activities or gather intelli-
strategy, assuming high level information about the cats’ move- gence. It may also be able to sense its environment or plan its
ment model. We discuss how the cats’ chosen movement model movement to avoid detection. Similarly, the surveillance sen-
will affect its presence matrix in the network, and hence its payoff sors may be mobile. For example, they are carried by robots,
in the game. When the mouse has a larger sensing range than the UAVs, or a platoon on patrol schedules. Each sensor will have a
cats, we show how the mouse can determine its optimal movement
strategy based on local observations of the cats’ movements. We sensing range enabling it to detect an intruder within the range.
further present a coordination protocol for the cats to collabora- By moving, the sensors may be able to efficiently cover the net-
tively catch the mouse by: 1) forming opportunistically a cohort work area over time, although they do not have sufficient den-
to limit the mouse’s degree of freedom in escaping detection; and sity to ensure complete coverage all the time. The sensors may
2) minimizing the overlap in the spatial coverage of the cohort’s move independently or in a coordinated manner. The movement
members. Extensive experimental results verify and illustrate the may also be either deterministic or randomized. In particular,
analytical results, and evaluate the game’s payoffs as a function of
several important system parameters. stochastic movement is effective in overcoming unforeseen or
probabilistic events in the operating environment, such as the
Index Terms—Mobile sensor coverage, mobile target detection, failure of another sensor, or the unexpected presence of obsta-
mobility control, sensor coordination.
cles in the surveillance area.
In this paper, we model and analyze the game between the
I. INTRODUCTION sensors and the intruders. The sensors plan their movement to
detect the intruders as soon as possible, in order to initiate timely
C ONSIDER a strategic zone belonging to country and
bordering with another untrusted country
national safety, it is important for
. For its
to monitor the zone for
response against the presence of the intruders. The sighting of an
intruder could be reported by a wide-area cellular or 802.16 net-
work infrastructure, and the response action could be in the form
the presence of any intruder from . However, because the
of raising the alert level, or sending in troops to destroy/capture
strategic zone is vast and sometimes spans difficult terrains, it is
the intruders, although such actuation issues are not explicitly
impractical for to install an expansive static sensor network
considered in the paper. The intruders, on the other hand, plan
covering the whole area, and ensure that all the sensors are
their movement to avoid detection for as long as possible, in
order to prolong their own mission. In addition, an intruder it-
self may have a sensing range allowing it to see the presence
Manuscript received January 30, 2008; revised October 18, 2008; approved
by IEEE/ACM TRANSACTIONS ON NETWORKING Editor M. Liu. First published
or movement of the sensors without being detected. We analyze
September 29, 2009; current version published February 18, 2010. This work the best movement strategies for both the mobile intruders (we
was supported in part by the U.S. Oak Ridge National Lab/Office of Naval Re- call such an intruder a mouse) and the mobile surveillance sen-
search under Grant DE-AC05-00OR22725, the U.S. National Science Founda- sors (we call such a sensor a cat), under different conditions in
tion under Grant CNS-0305496, an IBM Fellowship awarded to Y. Dong, and
a Purdue Research Foundation Fellowship awarded to J. C. Chin. a closed network area. We assume that the network area is of
J.-C. Chin, C. Y.-T. Ma, and D. K. Y. Yau are with the Computer Science uniform interest to a mouse, and use the time until detection as
Department, Purdue University, West Lafayette, IN 47907 USA (e-mail: the primary performance metric of interest.
jcchin@cs.purdue.edu; ma18@cs.purdue.edu; yau@cs.purdue.edu).
Y. Dong is with the Silicon Valley Lab, IBM, San Jose, CA 95141 USA Our contributions in the paper are as follows:
(e-mail: yudong@us.ibm.com). 1) For a blind mouse whose sensing range is smaller than or
W.-K. Hon is with the Computer Science Department, National Tsing Hua equal to the cats’, we develop an optimal dynamic pro-
University, Hsinchu, Taiwan (e-mail: wkhon@cs.nthu.edu.tw).
Color versions of one or more of the figures in this paper are available online
gram for the mouse to maximize its expected detection
at http://ieeexplore.ieee.org. time, given statistical knowledge about the cats’ move-
Digital Object Identifier 10.1109/TNET.2009.2024309 ments in the form of a presence matrix. We also discuss
1063-6692/$26.00 © 2009 IEEE
42 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 18, NO. 1, FEBRUARY 2010

how the cats can optimize their presence matrix to min- We consider general movement strategies for the mouse and
imize the expected detection time, assuming that each the cats. They can be autonomous or reactive, and they can
position is equally likely to be the mouse’s starting po- be deterministic or probabilistic. In particular, when the mouse
sition. The optimal cat and mouse strategies are in Nash has a larger sensing range than the cats (i.e., ), the
equilibrium. mouse can see some of the cats’ movements while remaining
2) For a seeing mouse having a larger sensing range than undetected. Such movement information can be exploited by the
that of the cats, we show how the mouse can use its mouse to avoid or delay detection. The cats, on the other hand,
local observations of the cats’ movements to maximize can opportunistically coordinate their movement to maximize
the expected detection time. their ability to catch the mouse. First, when two or more cats fall
3) We show how a network of cats, in playing against the within the sensing range of each other, the cats may minimize
seeing mouse, can coordinate their movement to maxi- the overlap in their spatial coverage by moving away from each
mize their ability to catch the mouse. First, we show that other. Second, if the cats can additionally communicate within
two cats who fall within each other’s sensing range can a wireless range of , they can run a coordination protocol to
attempt to minimize the overlap in spatial coverage by maximally detect the mouse in a collaborative manner.
moving away from each other, and that such a strategy
can reduce the detection time. Second, if the cats can III. RELATED WORK
additionally communicate within a wireless range, the
cats can opportunistically form a cohort to minimize Meguerdichian et al. [10] derive an optimal path for a mobile
the mouse’s degree of freedom in escaping, while maxi- target to minimize its exposure to a set of static sensors in getting
mizing the barrier coverage by the cohort members. We to a given destination. They do not assume that the target can
show that the communication-enabled coordination pro- observe the sensors, and implicitly study the case of the blind
tocol can perform significantly better than the approach mouse only. Our strategic goal for the target is also different.
enabled by the sensing only. Rather than minimize its exposure in getting to some destina-
4) We present extensive experimental results to evaluate tion, the target’s goal is to maximize the sensors’ expected de-
how the players’ strategies can affect the detection time. tection time among all possible paths.
Network coverage by mobile sensors has also been addressed
by Liu et al. [9]. Their work and ours differ in the following
II. PROBLEM FORMULATION aspects. First, we consider a closed network area with explicit
boundaries, whereas they consider the network to be an un-
We study the problem of mice trying to evade detection bounded infinite space. Second, they implicitly study the case of
by cats in a closed network area. The network is modeled as the blind mouse only. Third, they assume infinitely many cats
an rectangular region, where and are in distance spatially distributed with a given density. We assume a given
units. When there are multiple mice (i.e., ), we assume number of cats, and consider how their movement strategy can
that each mouse will try its best to escape detection, independent impact their steady state spatial distribution or their ability to
of the actions by the other mice. Hence, without loss of gener- detect the mouse quickly.
ality, we will consider the case of a single mouse. We will view Our cat-and-mouse problem can be considered an instance of
our problem as a game between two players: the mouse, and the pursuit-evasion games. Significant results about various forms
collection of all cats forming a single (meta-)player. The mouse of the full information game (i.e., both pursuers and evaders
has a sensing range of and a speed of . For simplicity, know each other’s moves completely) are discussed in [6]. For
we assume that each cat has a sensing range of and a speed extending the theory to games of incomplete information, they
of . The assumption can be relaxed to include the case when describe a princess-and-monster problem for which it is stated
different cats have different sensing ranges and different speeds that there is no known solution. The princess-and-monster is
of movement. The game ends when the mouse falls within the similar to our blind mouse problem, but it is not identical. In par-
sensing range of one or more cats. ticular, we assume that the mouse knows the cat’s presence ma-
Given that the mouse is initially located at position , the trix, while the princess has no such knowledge about the mon-
expected detection time of the mouse is denoted by , ster. In general, instances of pursuit-evasion games and their so-
which is also the mouse’s payoff in the game. For the cats, they lutions differ fundamentally based on the amount and kind of
do not know the mouse’s initial position, so that they simply as- information available to the players.
sume that the mouse’s initial position is uniformly distributed in More recently, for partial-information games, Trummel and
the network. Hence, the cats try to minimize the “hypothetical” Weisinger [13] show that the optimal solution for a pursuer to
expected payoff of the mouse; i.e., the game’s payoff for the cats catch a stationary target in the least time, given a probability dis-
is , where denotes the expectation under the tribution of the target’s position, is NP hard. Their problem uses
above assumption. Notice that because the initial position of the a general discrete-time graph model, in which any pair of ver-
mouse is known to the mouse, but not to the cats, the game is not tices may be connected by an edge. The NP hardness result is in
a simple zero-sum game. A player’s strategy in the game speci- contrast to our dynamic program solution for the blind mouse.
fies how the player should move in the network area. The mouse We are able to obtain an efficient optimal algorithm because in
and the cats will optimize their strategies in order to maximize our surveillance area, each cell is a neighbor of only up to eight
their own payoffs.1 geographically adjacent cells. Based on a line-of-sight capture
1Liu et al. [9] proposed and studied a similar game-theoretic problem formu- model (i.e., an evader is captured if it is in the line-of-sight of
lation, but for a different network and mobility model. a pursuer), Isler et al. [7] show that in a connected polygonal
CHIN et al.: DETECTION OF INTELLIGENT MOBILE TARGET IN MOBILE SENSOR NETWORK 43

space, a pursuer is guaranteed to capture an evader with proba-


bility approaching one, even if the evader knows where the pur-
suer is all the time. The line-of-sight model does not consider
the finite sensing range of cats as in our problem.
For multiple coordinated pursuers, Hespanha et al. [4] study
a greedy (non-optimal) policy for a swarm of pursuers to find
an evader. They show that, assuming a similar greedy strategy
by the evaders, the greedy pursuer policy guarantees capture
in finite time. Isler et al. [8] study the question of how many
pursuers are needed to capture a limited-visibility evader with
high probability. Their result uses the same discrete-time graph
model as in [13]. Vidal et al. [14] propose local-max and
global-max heuristic pursuit strategies to coordinate a group
of aerial/ground pursuers in catching a number of randomly
moving evaders. They observe that the dependency between the
expected capture time and the pursuit policy is “very complex,”
which motivates their proposal of efficient sub-optimal policies
with good practical performance. Hollinger and Singh [5] make 2
Fig. 1. The presence matrix with 10 10 cells for a single cat moving under
the RWP algorithm in the network area. The mean of the probabilities is 0.01,
similar observations about the hardness of the optimal coordi- while the standard deviation is 0.012.
nated pursuer problem, due to exponential growth of the search
space as the number of pursuers increases. For the cat coordina-
tion problem, we similarly study the empirical performance of
heuristic strategies that can be efficiently implemented. Part of single cat. In particular, let be the presence probability of one
our focus is, however, the contrast between sensing- and com- cat in cell . Then, the presence probability for at least one of
munication-based coordination (SBC and CBC) approaches. the cats appearing in cell is . Also note
SBC coordination has been less explored in the literature, but that and .
can be attractive because it does not assume any extra system Dynamic Programming Solution for the Best Strategy: Given
costs/complexities beyond the sensing resources. the presence matrix , what strategy should the mouse use
to maximize the expected detection time? One simple greedy
IV. THE BLIND MOUSE: strategy is for the mouse to continually move to the neighboring
cell having the lowest presence probability, and stop moving
In this section, we focus on the case where the cats have a when all the current cell’s neighbors are more dangerous (i.e.,
longer sensing range than the mouse. In this case, the mouse is they have a higher presence probability). The intuition is that
always detected before it can see the cat; hence, the mouse can the expected detection time will immediately increase when-
be considered a “blind” mouse. ever the mouse visits a new cell along the path. However, the
greedy strategy may not always find the best path for the mouse,
A. Strategy of the Mouse since it prevents the mouse from considering those paths that
Since a blind mouse cannot sense a cat’s actual movement, to temporarily access a more dangerous neighboring cell but will
avoid detection, the mouse is assumed to have some high level eventually lead to a safe network location. For example, in
information on the cats’ movements. Specifically, we assume Fig. 2, the greedy strategy suggests that the mouse should move
that the mouse knows the cats’ statistical movement model and from cell to cell along the illustrated path, and finally stop
their sensing range . Using the information, we estimate the at . However, under most of the mouse’s and cat’s speeds, the
probability of finding at least one cat for any position in the mouse’s optimal path is to move from cell to cell .
network, and design the best strategy for the mouse accordingly. To avoid missing the optimal path, we apply the dynamic
With a sensing range , each cat controls a circular region program DP_ALGO as shown in Fig. 3. In the figure, there
of area . Roughly speaking, each cat controls a square re- are three groups of variables, namely , ,
gion—which we call a cell—of size , with . and . The value represents the expected
We then consider the network area to consist of by detection time when the mouse starts at cell and uses the
disjoint cells, with each cat controlling the cell in which it is best path among the paths considered so far. This value will
currently located. The presence matrix is defined such that be updated as the algorithm proceeds, and will eventually hold
denotes the probability that at least one cat is present in the desired expected detection time when the mouse uses the
cell at any instant (when the context is clear, we use best path among all possible paths. The value denotes
as a shorthand for ); note that can be obtained since the expected detection time when the mouse starts at cell and
the cats’ movement model is known. For instance, Fig. 1 shows stays there forever. Finally, the value denotes the
the presence matrix for a single cat (i.e., ) moving under expected detection time when the mouse starts at cell , moves
the random waypoint (RWP) algorithm [1] in a network area of to a neighbor cell , and follows the best strategy once it reaches
10 10 cells. cell .
Remark: If and the cats are moving independently In the algorithm, we initialize to for
under the same movement model, it is easy to obtain the pres- each cell , and insert the cell into a max-heap with key value
ence matrix for cats based on the presence matrix of a (Lines 1–3). For each iteration (Lines 4–14), we first
44 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 18, NO. 1, FEBRUARY 2010

Fig. 2. A greedy movement strategy may not be optimal.

either to stay, or to move to the best adjacent cell (if it yields a


longer expected detection time). If the latter case is true, such
an adjacent cell, say , must be one of the best cells,
so that of is set to the optimal value upon ’s
extraction (Lines 6–13). Thus in both cases, the th best cell
must have the largest among all remaining cells at
the th iteration, and will be extracted. Also, the optimal action
and the corresponding are correct as desired.
It remains to show how to compute and .
To do so, we use a concept called the cell sojourn time, which
is the length of time that a cat stays in the current cell before it
moves to a neighboring cell. The expected sojourn time, denoted
by , can be calculated since the cats’ movement model is
known. For the purpose of estimation, we may assume that the
Fig. 3. DP_ALGO: A dynamic program for the mouse’s best strategy.
status of whether any cat is present in a certain cell is unchanged
during the time interval , for and any
nonnegative integer . Then, we can estimate by
extract the cell with the largest key value from the heap, so
that the value will not be further updated. Then, for
(1)
each neighboring cell of , we update to
if the latter value is larger, meaning that the mouse starting at
cell can benefit from moving to cell . If there was an update, To calculate based on the optimal , we
we reinsert cell into the heap with the updated key value let be the time taken by the mouse to move from (the center
(Lines 6–9), and mark cell to be the next cell to move to when of) cell to (the boundary of) cell . Then, the mouse will either
the mouse is in cell (Line 10). be caught in cell , or it can reach cell so that it is expected
Theorem 1: DP_ALGO is an optimal algorithm for the mouse, to be caught after another time of . The probability
against any presence matrix . that it can reach cell without being caught can be estimated
Proof: We shall prove by induction the following: the by . Using an approach similar to the one for
th best cell is extracted in Line 5 of the th iteration, and the estimating , we estimate by
mouse’s optimal action at that cell is to move to the location
specified by “next step.”2 In addition, will be cor-
rectly set to the expected detection time under any optimal
mouse strategy.
In the first iteration, the cell with the largest is It can be shown that the dynamic program has time com-
extracted. This cell must be the one with the largest , so plexity , where . We omit the analysis
that it must be the best cell. Here, we determine that the mouse due to limited space.
should “stay” (i.e., “next step” is itself), and such an action
is clearly optimal. Next, to complete the induction, suppose that B. Strategy of the Cats
the best cells are extracted in the first iterations. Based on the strategy in the previous section, the mouse may
For the th best cell , we can see that its optimal action is eventually move to a safer position that has a lower cat presence
2The j th best cell is the one, when chosen as initial position, with the j th probability than its current position, thereby maximizing the ex-
longest expected detection time, under an optimal mouse strategy. pected detection time. Hence, the best strategy for the cats seems
CHIN et al.: DETECTION OF INTELLIGENT MOBILE TARGET IN MOBILE SENSOR NETWORK 45

~
Fig. 4. The mouse’s choice of to maximize . (a) V = V~ 0 V~ . (b) To
maximize .

Fig. 5. The black dots show the positions where the mouse is detected in 50 000
simulation runs. The mouse is caught at the the network corners and boundaries
to maximize the minimum presence probability among all the in most cases.
cells in .
The maximin strategy implies that when the cats are moving
independently under the same movement model, the best choice Theorem 3: Nash equilibrium is achieved when the mouse
for each cat is to move in a way such that the presence proba- applies DP_ALGO and the cats apply the maximum-uniform
bility is the same in each cell. (We call the resulting presence presence matrix.
matrix in which all the entries have equal values a uniform pres- To yield a uniform presence matrix in the single-cat case,
ence matrix.) With a uniform presence matrix, there is no par- one simple example movement is to sequentially and circularly
ticularly safe place for the mouse to stay, which reduces the scan all the network cells. However, the deterministic nature of
(worst-case) expected detection time of the mouse. In fact, the the cat’s movement may allow the mouse to accurately predict
uniform presence matrix is an optimal strategy for the cats, as where the cat will be and therefore easily avoid it. Hence, we
shown in the following theorem: will present in Section V-B a probabilistic movement strategy
Theorem 2: When the mouse applies DP_ALGO, an optimal that can achieve a presence matrix close to being uniform.
strategy for the cat is to set its presence matrix uniform. As mentioned, when there are cats, it is better for them to
Proof: Let denote an arbitrary presence matrix, and move in disjoint areas, as the sum of in the presence matrix is
denote the uniform presence matrix. When the mouse applies equal to , which is greater than the sum of if the cats move
DP_ALGO, we let and denote the expected time for the independently. This suggests that if the cats are allowed to move
cat to catch the mouse when using and , respectively. To in a coordinated way, we should always assign them to monitor
prove the theorem, we now show that . To link the two disjoint parts of the network. Accordingly, the best strategy for
quantities, we consider a further expected time , which corre- the cats is to divide the network into equally sized partitions,
sponds to the case that the cat uses presence matrix , and the in which one cat moves within one partition optimally to yield
mouse simply never moves. It is easy to see that , since a maximum-uniform presence matrix.
if the mouse never moves, it will never improve (i.e., lengthen)
the detection time. On the other hand, when the network has
cells V. THE SEEING MOUSE , INDEPENDENT CATS
In this section, we discuss the case when the mouse has a
larger sensing range than the cats. In this case, once a cat enters
the mouse’s sensing range, the mouse can know the cat’s move-
ment in advance without being detected. We focus on the case
that the mouse’s speed is less than the cats’ (i.e., );
so that . The above inequality holds since for we believe that the other case is not as interesting since intu-
any , . itively, the faster mouse can always avoid being detected. In the
The presence matrix can be generalized for cats, in which following, we first propose a strategy for the mouse to escape
case each cell is associated with a probability that at least one from the cats based on such advance knowledge. We then dis-
cat is present in the cell. The sum of these probabilities can be cuss some possible strategies that the cat may use to reduce the
as large as , which is achievable when the cats are always in detection time, knowing that the mouse may run away when it
disjoint cells. When the network has cells, and each cell is sees the cat. We assume in this section that the cats act indepen-
associated with a probability , we call this presence ma- dently.
trix maximum-uniform. Repeating the arguments in Theorem 2,
one can easily show the following result: A. Strategy of the Mouse
Corollary 1: When the mouse applies DP_ALGO, an optimal
strategy for cats is to set the presence matrix maximum- We first study the special case in which there is only one cat
uniform. in the network. Then, we generalize the strategy to the case of
Then by Theorem 1 and Corollary 1, we have the following. multiple cats.
46 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 18, NO. 1, FEBRUARY 2010

2
Fig. 6. The presence matrix with 10 10 cells for a single cat moving under
the bouncing strategy in the network area. The matrix is close to the uniform
presence matrix.
Fig. 7. The CBC approach.

Avoiding a Single Cat: When the mouse tries to escape from between and the vector from the mouse to the cat (i.e.,
a cat, it is better if it can move in a direction such that the ) to be as large as possible. As shown in Fig. 4, we have
minimum distance between the mouse and the cat (assuming
that the cat does not change its speed and moving direction) in
the future is as large as possible. (5)
Let and be the current positions of the cat and the mouse,
respectively. Let be the moving direction of the cat. Then, the Avoiding Multiple Cats: We have shown how the mouse can
position of the cat at time , denoted by , can be expressed choose the optimal direction when there is only one cat. In
as the case of multiple cats, we need to find a that maximizes
the minimum distance to all the cats in the future. Using similar
(2) notations, we let denote the minimum distance between
the mouse and the th cat in the future, assuming that the mouse
Similarly, if the mouse chooses to move in a direction at is moving in the direction . In other words, if there are cats
a speed of , the position of the mouse at time , denoted by within the sensing range of the mouse, we have
, can be expressed as
(6)
(3)
Notice that in the case of multiple cats, a direction that
Then, the distance between the cat and the mouse at time is maximizes the minimum distance for one cat does not neces-
equal to . By differentiating sarily yield a short minimum distance for another cat. Finding
with respect to , we can find the minimum distance between the optimal may require us to consider all the intersections
the mouse and the cat for . We denote such a distance by between any two of the curves , so as to decide the best
. In other words, to maximize the distance from the cat in direction; in the worst-case, there can be such intersec-
the future, the mouse should find a direction such that tions. An alternative way is to obtain a close estimation of ,
by choosing a small angle and computing all values
for , and then finding based
(4) on only these values. The drawback of the above estimation is
that we may miss the optimal when it is not a multiple of
Remark: In fact, by a simple geometric argument, the optimal . However, there are only values to compute, and the
for the single-cat case can be easily obtained. Essentially, the simplicity of the method will usually give a good enough ef-
cat is moving in a vector , while ficiently in practice.
the mouse choosing a direction is moving in a vector Degree of Freedom and a Revised Strategy: In general, it is
. Equivalently, we may assume that not necessary for the mouse to choose the optimal direction
the cat is stationary, while the mouse is moving in a direction to avoid being caught. It is because when choosing a direction
towards the cat. To maximize the distance between , a mouse will not be detected as long as its minimum distance
the cat and the mouse, we want (the absolute value of) the angle to all the cats in the future is larger than the cats’ sensing range
CHIN et al.: DETECTION OF INTELLIGENT MOBILE TARGET IN MOBILE SENSOR NETWORK 47

Fig. 8. Coordinated waypoint selection to minimize coverage overlap.

Fig. 9. Average detection time (in s) of the mouse versus the cat’s sensing range
. Hence, the mouse may choose a direction from a feasible R (in m); V V = = 10 m/s.
set such that

(7) TABLE I
AVERAGE DETECTION TIME (IN s) FOR DIFFERENT CAT AND MOUSE
2
MOVEMENT STRATEGIES IN A 500 m 500 m NETWORK.
Thus, the larger the size of is, the more directions the mouse V = V = 10 m/s, R = 25 m, AND THE MOUSE
can choose from, so that it is more likely for a mouse to escape IS INITIALLY LOCATED AT THE CENTER OF THE NETWORK
successfully. Note that the mouse may have fewer choices when
it is located at the boundaries or the corners of the network area.
In fact, if the cats move under the RWP algorithm and the mouse
uses the above strategy to escape (so that it chooses to move at
an angle of when one or more cats are around, and stops
moving when no cat is around), the results in Fig. 5 show that
the mouse will likely be “pushed” to the corners/boundaries of Remark: In the seeing mouse case, the cats have no informa-
the network, and be caught there. tion about the mouse during the game. An optimal cat strategy
This leads to a revised strategy for the mouse to always move can be defined in the minimax sense—one that minimizes the
towards the center region of the network whenever there is no hypothetical expected detection time over all possible mouse
cat detected within its sensing range. We call this the centric strategies. However, given that there are infinitely many pos-
strategy. Our results in Section VII show that when the cats sible mouse strategies, and that the cats have no information
have a uniform presence matrix, the centric strategy works well about what the mouse might choose, we conjecture that finding
among other strategies evaluated. But when the cats use RWP, the optimal cat strategy is undecidable. Hence, we propose only
the mouse’s degree of freedom at the center may be offset by heuristic cat strategies in the above discussion.
the predominant cats’ presence near the center.
VI. THE SEEING MOUSE AND COORDINATED CATS
B. Strategy of the Cats In this section, we consider the seeing mouse, and a network
Since a cat knows that the mouse may see it before it sees the of cats who may coordinate their movement to maximize their
mouse, the cat may assume that the mouse will try to run away ability to detect the mouse. We consider two coordination ap-
from it in advance. One logical choice of strategy for the cats is proaches. The first one is enabled by the cats’ ability to see each
then to visit more frequently the center region of the network, other within their sensing range. The second one is enabled by
where the mouse has more freedom to move and escape. As the cats’ additional ability to communicate over a wireless range
shown in Fig. 1, a cat can achieve the goal of visiting the center of .
region more often by playing the RWP strategy. On the other
hand, the cats may also benefit from a movement strategy that A. Sensing-Based Coordination (SBC)
will yield a uniform presence matrix, so that they will eliminate When a cat moves close to other cats, their sensing regions
safe havens for the mouse to hide in. As both kinds of strategy overlap, which results in inefficient use of the sensing resources.
have their merits, we will compare their empirical performance Such an overlap can be detected partially by the cats involved.
in Section VII. Specifically, they can see each other when they fall within the
We use the following simple random algorithm, which we sensing range of each other. When that happens, the cats can try
call the bouncing strategy, for the cats to approximate closely to move away from each other in order to reduce the overlap in
the uniform presence matrix. In the algorithm, a cat moves in a spatial coverage. We call such a coordination approach SBC.
straight line until it hits the boundary or a corner of the network In SBC, a cat moves according to the usual algorithm (e.g.,
area. Whenever it reaches the boundary/corner, it selects a di- bouncing or RWP) until it detects one or more other cats within
rection randomly and uniformly from all the feasible directions its sensing range. When that happens, the first cat will try to
and proceeds to move in the selected direction. Fig. 6 shows maximally avoid the other cat(s), i.e., to minimize the expected
the presence matrix for a cat moving under the bouncing model. future overlap in coverage by the cats. This can be done by
Alternative algorithms are known that can provably achieve uni- using the same algorithm that a seeing mouse uses to avoid a
form presence for the cats [2], [11]. detected cat, namely the geometry-based algorithm presented
48 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 18, NO. 1, FEBRUARY 2010

Fig. 10. Average detection time (in s) versus the number of cats N in a
2
500 m 500 m network area; V =V = 10 m/s.
2
Fig. 12. Escape paths found by DP_ALGO for a 500 m 500 m network area
V
divided into 10 by 10 grids, for different mouse’s speeds. (a)V= =
V
10 m/s. (b) V
= 10 m/s, = 15 m/s.

formed as follows. A cat, say , that is acting independently


(i.e., not in any cohort currently) will continuously attempt to
establish wireless communication with another cat, which hap-
pens when the other cat, say , comes within a distance of
of . With a probability of , will ignore the commu-
nication and repeat looking for a new cat for communication.
Otherwise (i.e., with a probability of ), asks about its
current destination waypoint. will then abandon its current
waypoint and join by adopting a new waypoint that is “close
to” ’s waypoint. (The “close to” notion will be made precise
in the following.) Once decides to join and is on its way
to the new waypoint, it considers itself committed and will not
Fig. 11. Average detection time (in s) versus the speeds (in m/s) of the cat and attempt to establish communication with other cats. Once has
2
the mouse in a 500 m 500 m network; R = 15 m. reached the new waypoint, however, it will repeat the procedure
of entering another cohort.
Notice that while a more effective barrier is formed by a larger
in Section V-A. According to the algorithm, the cat computes cohort, a larger cohort size will result in a fewer number of co-
the optimal in order to optimally move away from the other horts since the number of cats is fixed. A small number of co-
cat(s). horts limits the ability to spread out the cats over different parts
of the surveillance area where the mouse may be found. Hence,
B. Communication-Based Coordination (CBC) a tradeoff between the cohort size and the number of cohorts is
In SBC, the cats coordinate by observing each other’s move- useful. The parameter aims to provide this tradeoff, in which
ments only. If the cats can communicate over a wireless channel larger is more likely to produce larger cohorts.
of range , they can further coordinate their strategies to re- Notice also that, while the teaming of cats forms a useful bar-
duce the expected detection time of the mouse. First, if rier against the escape of the mouse, it is also important to avoid
, the coordination can occur sooner and therefore be more the inefficient coverage overlap that can happen with uncon-
proactive. Second, the cats can actively exchange information trolled teaming. This is the reason when decides to follow
about their strategies (e.g., the future movements), instead of in the above example, it does not try to go to the same way-
using passive observations only. We call such a coordination point as , but selects a new waypoint close to ’s. Specifically,
approach CBC. the new waypoint is computed as follows: selects a waypoint
The CBC algorithm we use (Fig. 7) is motivated by the obser- that is away from ’s and whose direction from ’s is
vation in Section V-A about the importance of the mouse’s de- perpendicular to ’s movement vector. The waypoint selection
gree of freedom in enabling its escape strategy. Specifically, we strategy is illustrated in Fig. 8. Notice that two cats having the
aim to allow the cats to team up and form a cohort in searching same speed are likely to arrive at the new waypoints at about
for the mouse. The concerted efforts of the cohort members the same time, and will form a maximum barrier against the
will then form a larger barrier that limits the mouse’s degree of mouse’s escape.
freedom in passing the barrier. The algorithm in Fig. 7 allows In general, we have conjectured about the hardness of the op-
cohorts to form in an opportunistic manner. We assume that each timal independent cat strategy against the seeing mouse. The
cat’s plan of movement consists of a sequence of trips, and the problem becomes more difficult when coordinated cats are con-
destination of one trip is the starting point of the next trip, and so sidered. A number of related heuristic algorithms for coordina-
on. We call the destination of a trip a waypoint. Cohorts are then tion are known [4], [8], [14]. Our goal in this paper is not to
CHIN et al.: DETECTION OF INTELLIGENT MOBILE TARGET IN MOBILE SENSOR NETWORK 49

Fig. 13. Minimum distance d ( ) between six cats and one mouse as a func- Fig. 15. Statistics of the maximin distance of the mouse from any of the cats, as
a function of N . Both the cats and the mouse move according to the bouncing
tion of the mouse’s movement direction . The cats and the mouse are initially
2
randomly placed in a 100 m 100 m network. 2
strategy in a 100 m 100 m network area.

to the case when the mouse determines its movement by the


dynamic programming strategy in Section IV-A. The mouse is
initially located at the center of the network area. The column
labeled “Stay” in Table I is used for baseline comparison, and
refers the to case when the mouse simply stays at its starting
position (i.e., it does not move at all afterward). The experi-
ments use one cat. Its movement strategy is shown as the rows of
Table I, and is chosen to be: 1) the uniform scan strategy; 2) the
bouncing strategy in Section V-B; and 3) the random waypoint
(RWP) strategy, in three sets of runs. From the table, notice that
the mouse can achieve a significantly higher average detection
time by playing the dynamic programming strategy, as the anal-
ysis in Section IV-A shows.
2) Uniform Presence Matrix Benefits the Cat: The “DP”
column in Table I also shows how a uniform presence matrix
may benefit the cat. From the results, notice that the simple
Fig. 14. The fraction of feasible solutions as a function of R . Ten cats are scan strategy can greatly reduce the detection time compared
2
randomly placed in a 100 m 100 m network. R = . 1 with the nonuniform RWP strategy. However, as discussed in
Section IV-B, the deterministic nature of the scan strategy may
allow the mouse to predict where the cat will be in advance
extensively cover the design space of CBC, but to contrast with and thus effectively avoid the cat. To avoid the problem, the
the SBC and CBC approaches, and evaluate how the cats may bouncing strategy can closely approximate the uniform pres-
use coordination to improve against effective mouse strategies ence matrix without being deterministic.
identified in our previous discussions. 3) Effects of the Cat’s Sensing Range: In this experiment,
there are one cat and one mouse moving in a 500 m 500 m net-
work area. The mouse plays the dynamic programming strategy.
VII. EXPERIMENTAL RESULTS The cat uses: 1) the RWP strategy; and 2) the bouncing strategy,
in two different runs. We measure the average detection time
A. The Blind Mouse:
of the mouse for different sensing ranges, , of the cat. Fig. 9
This section evaluates the case of a blind mouse. Unless oth- shows that as increases, the average detection time decreases
erwise specified, we report average results over 100 simulation as an inversely proportional function of , showing that the
runs, each lasting 200 000 s. We omit the standard deviations, cat’s ability to detect the mouse increases proportionally to its
because they are very small compared with the means. In the sensing capability.
discussion, when a node deterministically cycles through all 4) Number of Cats: In this experiment, we measure the av-
the cells in the network, we refer to the movement as the scan erage detection times when the number of cats is varied. The net-
strategy. work area is 500 m 500 m, and the sensing range of each cat
1) Benefits of Dynamic Programming Solution for the Mouse: is 25 m. Similar to the previous experiment, the mouse plays the
We compare the performance of different movement strategies dynamic programming strategy, while the cats independently
for the mouse in Table I. In the table, the column “DP” refers play either the RWP or the bouncing strategy. Fig. 10 shows that
50 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 18, NO. 1, FEBRUARY 2010

Fig. 16. Statistics of the detection times as a function of (a) N , (b) R , and (c) V . The cats randomly move in a 100 m 2 100 m network. Both the cats and
the mouse use the bouncing strategy.

Fig. 18. Average detection time comparison of the SBC and CBC coordination
approaches, for different combinations of the mouse and basic cat movement
algorithms.
Fig. 17. Average of the mouse detection time (in s) under different strategies for
2
the cats and the mouse in a 100 m 100 m network. V = 5 m/s, V = 10 m/s,
R = 5 m, and R = 10 m. The paths when the mouse has a speed of 10 m/s are shown
in Fig. 12(a). The paths when the mouse has a higher speed of
15 m/s are shown in Fig. 12(b). In the figure, cells and are cell
as the number of cats increases, the average detection time de- ’s neighbor. Cell is safer than , and both of them are safer
creases, approximately like inversely proportional to , where than . Assume that the mouse is currently in cell . Notice that
is a constant slightly larger than one. when the mouse has the lower speed, it will move to the less safe
5) Effects of and on Dynamic Programming Solution: neighbor because is closer in distance than . Choosing the
In this experiment, there are one cat and one mouse. We illus- closer neighbor allows the mouse to leave the more dangerous
trate how the expected detection time of the mouse varies with cell soon (considering that the mouse moves slowly). When
different speeds and of the cat and the mouse, respec- the mouse has the higher speed, it can afford to move a longer
tively. The mouse plays the dynamic programming strategy. distance before leaving cell . Hence, it will choose to move di-
Fig. 11 shows that the average detection time is reduced when rectly to the safer neighbor although is farther away.
the cat moves faster (i.e., is higher). This is because a faster
cat can cover a larger area in the same amount of time. Notice B. The Seeing Mouse , Independent Cats
also that the detection time increases when the mouse moves at We now evaluate the case of a seeing mouse, when the cats do
a higher speed. This is because a faster mouse can move from not coordinate. In the experiments, unless otherwise specified,
its current position to a safe position more quickly, and benefit the cats play the bouncing strategy and and
from staying in the safe position longer. We also find that the . We report maximum, minimum, and average results over
mouse can benefit more from moving at a higher speed, if the 5000 simulation runs within a 100 m 100 m area.
cat itself is moving at a higher speed. This shows that a fast cat 1) Movement Direction for the Mouse: In this experiment,
will force the mouse to be fast to avoid detection. we report the optimal direction calculated by (6) to maximize
We further show that the path of movement computed by the the minimum distance between the cats and the mouse. We use
dynamic program in Fig. 3 is dependent on the speed of the . There are six cats and one mouse. Fig. 13,
mouse. In Fig. 12(a) and (b), the gray level represents the cat’s shows the minimum distance of the mouse from each cat, as a
presence probability (the darker an area, the lower the cat’s pres- function of the mouse’s movement direction . In the figure, the
ence probability in the area). The cat has a speed of 10 m/s. thick line shows the minimum distance of the mouse from any
The arrows in the figure show the escape paths of the mouse cat. Of the thick line, the dark segment shows the range of the
in a 500 m 500 m network area divided into 10 10 cells. movement angles over which the minimum distance (from any
CHIN et al.: DETECTION OF INTELLIGENT MOBILE TARGET IN MOBILE SENSOR NETWORK 51

2
Fig. 19. Effects of communication range on average detection time. Network size is 100 m 100 m, V = V = 10 m/s, R = 5 m, R = 10 m, N = 10,
p = 0:7. (a) Bouncing mouse algorithm. (b) Centric mouse algorithm. (c) RWP mouse algorithm.

cat) is maximized, and thus gives the range of the optimal max- bouncing strategy, the best strategy for the mouse is centric. We
imin solutions computed. Any angle that falls within the range further quantify how deviations from the center position by the
can be used by the mouse as its optimal movement direction. mouse will impact the average detection times in this case in our
2) Effects of and : In this experiment, we use 10 cats, technical report [3].
and show how can affect the range of maximin solutions The above results show an interesting interplay between
available to the mouse as it tries to find an optimal angle of the cats’ and the mouse’s movements. When the cats use the
movement. In Fig. 14, we report the fraction of the movement bouncing strategy, the mouse benefits by moving toward the
angles that are in the feasible set (defined by (7)) as a function center of the network, where it has a higher degree of freedom
of . The results show that the fraction of feasible solutions de- as discussed in Section V-A. However, this higher degree of
creases like exponentially as increases (up to ), showing freedom is offset, in the case of this experiment, when the cats
that the mouse’s degree of freedom in choosing a solution is se- use the RWP strategy to achieve a higher presence in the center
verely restricted as the cat’s sensing capability increases. area. The static strategy performs the worst in all of the cases.
We next consider the effects of , the number of cats. In
Figs. 15 and 16(a), we show that as increases, both the op- C. Seeing Mouse and Coordinated Cats
timal minimum distance (of the mouse) from any of the cats and
the average detection time decrease. We evaluate the performance of the coordination approaches
3) Effects of and : We show the impact of the mouse’s in Section VI for the cats playing against a seeing mouse. We
sensing range on the detection time in Fig. 16(b). As use 10 cats in a surveillance area of 100 m 100 m; the cats and
increases, the mouse can detect the cats earlier and are better the mouse all have the same speed of 10 m/s; the cat and mouse
able to make the necessary moves to avoid the cats. Hence, the sensing ranges are 5 and 10 m, respectively. When there is no
detection time increases significantly at first. As increases cat within the sensing range of a seeing mouse, the mouse plays
beyond 9 m, however, the detection time does not change much. the bouncing, centric, or RWP strategy, and we correspondingly
This shows that information about the cats that are far away (as call it the bouncing, centric, or RWP mouse. For each set of pa-
captured by a large ) may not be that useful to the mouse in rameters, we report the average detection times of 1000 simu-
deciding its immediate course of action. lation runs in Fig. 18. The error bars show the 95% confidence
Fig. 16(c) shows that the detection time generally increases interval. For CBC coordination, and are set to be 10 m
as the mouse’s speed increases. This is because a faster mouse and 0.7, respectively, unless otherwise stated.
can avoid the cats more quickly and have more choices in the The results in Fig. 18 show that the coordination approaches
optimal movement direction . are highly useful for the RWP cats, but have relatively small ef-
4) Comparison of Different Strategies: We discussed fects for the bouncing cats. This is because the bouncing cats
the bouncing and centric strategies for the seeing mouse in are evenly spread out in the surveillance area as evidenced by
Section V-A. These strategies define what the mouse should their uniform presence matrix. For SBC, they get few chances
do when it sees no cat within its sensing range. An alternative to meet in the first place, and therefore do not need help from
strategy in such a situation is for the mouse to simply stay SBC to move away from each other. For CBC, the cats simi-
where it is, presumably to conserve energy, and we call it the larly have few chances to get close and attempt the opportunistic
static strategy. In this experiment, we compare the performance clustering if the communication range is small. If the communi-
of these strategies for the mouse, while 10 cats in the network cation range becomes larger, the pairing opportunities increase.
play either the bouncing or the RWP strategy. However, the cats who decide to cluster may arrive at the coor-
The average detection time results are shown in Fig. 17. As a dinated waypoints far apart from each other, which reduces the
baseline case for comparison, we also show a “stay” strategy effectiveness of the barrier formed. For the RWP cats, SBC co-
in which the mouse simply stays at its starting position for- ordination reduces the average detection time by at least 161%
ever (and never moves). Fig. 17 shows that bouncing is the best (against the centric mouse), by decreasing the overlap in spa-
strategy for the cats, where their presence matrix is approxi- tial coverage relative to the uncoordinated cats. Compared with
mately uniform. For the mouse, its best strategy depends on SBC, CBC coordination reduces the average detection time by
the cats’ strategy. When the cats play the RWP strategy, the at least 384% (against the bouncing mouse). This shows the ad-
best strategy for the mouse is bouncing. When the cats play the ditional benefits of teaming up the cats to reduce the mouse’s
52 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 18, NO. 1, FEBRUARY 2010

degree of freedom in evading the cats. The benefits are particu- Jren-Chit Chin received the B.Sc. degree in com-
larly obvious for the centric mouse and the RWP mouse, as both puter engineering from Iowa State University, Ames,
in 2005.
kinds of mouse tend to concentrate in the center and thus benefit He is a Ph.D. degree candidate in the Department
greatly from the high freedom of movement if not countered by of Computer Science, Purdue University, West
the cats’ coordination strategy. Lafayette, IN. His current area of research includes
target localization and sensor coverage in mobile
Fig. 19 shows the impact of on the average detection sensor network.
time of CBC, for the mouse strategies of bouncing, centric, and
RWP, respectively, when is fixed at 0.7. As explained be-
fore, CBC coordination is very useful for the RWP cats, but has
less effect for the bouncing cats. For the RWP cats, the per-
formance of CBC is highly dependent on the communication
Yu Dong received the B.E. degree in industry
range. In general, it increases quickly as increases, up to a automation control from the University of Science
value of , after which further increment in has little im- and Technology Beijing (USTB), Beijing, China,
pact. Also, RWP/CBC cats are most effective against the cen- in 1995; the M.E. degree in information system
engineering from Osaka University, Osaka, Japan, in
tric mouse. This is because CBC approach is specifically de- 2001; and the M.S. and Ph.D. degrees in computer
signed to counter the degree of freedom for the mouse to es- science from Purdue University, West Lafayette, IN,
cape at the center. The design is so effective that whereas RWP in 2003 and 2006, respectively.
cats without coordination have far worse performance than the He was the recipient of an IBM Ph.D. fellowship
and is now with IBM Silicon Valley Laboratory,
bouncing cats, RWP/CBC cats can outperform the bouncing cats San Jose, CA. His research interests are networking,
(with or without coordination). For the centric mouse, when databases, and multimedia systems.
the communication range is large enough, RWP/CBC detects
the mouse twice as fast as uncoordinated bouncing, and about
1.5 times faster than bouncing/CBC. In our technical report [3], Wing-Kai Hon received the Ph.D. degree from the
we also systematically explore the effects of and on University of Hong Kong in 2005.
the average detection time, for 10 bouncing/CBC cats and 10 He visited Purdue University, West Lafayette,
RWP/CBC cats. IN, during 2004–2006. He is an Assistant Professor
with the Department of Computer Science, National
Tsing Hua University, Taiwan. His research interests
REFERENCES include data compression, indexing, and algorithm
[1] J. Broch, D. A. Maltz, D. B. Johnson, Y. C. Hu, and J. Jetcheva, “A design.
performance comparison of multi-hop wireless ad hoc network routing
protocols,” in Proc. ACM MobiCom, 1998, pp. 85–97.
[2] T. Camp, J. Boleng, and V. Davies, “A survey of mobility models for
ad hoc network research,” WCMC, vol. 2, no. 5, 2002, Special Issue on
Mobile Ad Hoc Netw.: Res., Trends Appl..
[3] J. C. Chin, Y. Dong, W. Hon, C. Y. T. Ma, and D. K. Y. Yau, “De-
tection of intelligent mobile target in a mobile sensor network,” West Chris Yu-Tak Ma received the B.Eng. degree in
Lafayette, IN, Purdue CS Tech. Rep., 2009. computer engineering and the M.Phil. degree in
[4] J. P. Hespanha, H. J. Kim, and S. Sastry, “Multiple-agent proba- computer science and engineering from the Chinese
bilistic pursuit-evasion games,” in Proc. IEEE CDC, 1999, vol. 3, pp. University of Hong Kong in 2004 and 2006, respec-
2432–2437. tively.
[5] G. Hollinger and S. Singh, “Proofs and experiments in scalable, near- He is a Ph.D. student with the Department of Com-
optimal search by multiple robots,” Robot.: Sci. Syst., 2008 [Online]. puter Science, Purdue University, West Lafayette, IN.
Available: http://www.roboticsproceedings.org/rss04/p27.html His research interests include performance study of
[6] R. Isaacs, Differential Games. New York: Wiley, 1965. wireless networks and mobile sensor networks.
[7] V. Isler, C. Belta, K. Daniilidis, and G. J. Pappas, “Hybrid control
for visibility-based pursuit-evasion games,” in Proc. IEEE/RSJ IROS,
2004, vol. 2, pp. 1432–1437.
[8] V. I. S. Kannan and S. Khanna, “Randomized pursuit-evasion with lim-
ited visibility,” in Proc. ACM-SIAM SODA, 2004, pp. 1060–1069.
David K. Y. Yau received the B.Sc. degree from the
[9] B. Liu, P. Brass, O. Dousse, P. Nain, and D. Towsley, “Mobility im-
Chinese University of Hong Kong in 1989, and the
proves coverage of sensor networks,” in Proc. ACM MobiHoc, 2005,
M.S. and Ph.D. degrees from the University of Texas
pp. 300–308.
at Austin in 1992 and 1997, respectively, all in com-
[10] N. Megiddo, S. L. Hakimi, M. R. Garey, D. S. Johnson, and C. H.
puter science.
Papadimitriou, “The complexity of searching a graph,” J. ACM, vol.
From 1989 to 1990, he was with the Systems and
35, no. 1, pp. 18–44, 1988.
Technology Group of Citibank, NA. He is currently
[11] P. Nain, D. Towsley, B. Liu, and Z. Liu, “Properties of random direction
an Associate Professor of computer science with
models,” in Proc. IEEE INFOCOM, 2005, vol. 3, pp. 1897–1907.
Purdue University, West Lafayette, IN. His research
[12] R. Teo, J. S. Jang, and C. J. Tomlin, “Automated multiple UAV
interests are in protocol design/implementation,
flight—The Stanford dragonfly UAV program,” in Proc. IEEE CDC,
wireless/sensor networks, and network security.
2004, vol. 4, pp. 4268–4273.
Dr. Yau was the recipient of an NSF CAREER Award for research in quality
[13] K. E. Trummel and J. R. Weisinger, “The complexity of the optimal
of service provisioning. He serves on the Editorial Board of the IEEE/ACM
searcher path problem,” Oper. Res., vol. 34, pp. 324–327, 1986.
TRANSACTIONS ON NETWORKING. He was TPC Co-Chair (2006) and Steering
[14] R. Vidal, O. Shakernia, H. J. Kim, D. H. Shim, and S. Sastry, “Prob-
Committee Member (2007–09) of IEEE IWQoS, and Vice General Chair (2006)
abilistic pursuit-evasion games: Theory, implementation, and exper-
and TPC Co-Chair (2007) of IEEE ICNP.
imental evaluation,” IEEE Trans. Robot. Autom., vol. 18, no. 5, pp.
662–669, Oct. 2002.

You might also like