Throughput optimization in

1

Throughput Optimization in
Mobile Backbone Networks
Emily M. Craparo, Member, IEEE, Jonathan P. How, Senior Member, IEEE,
and Eytan Modiano, Senior Member, IEEE

Abstract—This paper describes new algorithms for throughput optimization in a mobile backbone network. This hierarchical
communication framework combines mobile backbone nodes, which have superior mobility and communication capability, with
regular nodes, which are constrained in mobility and communication capability. An important quantity of interest in mobile backbone
networks is the number of regular nodes that can be successfully assigned to mobile backbone nodes at a given throughput level.
This paper develops a novel technique for maximizing this quantity in networks of fixed regular nodes using mixed-integer linear
programming (MILP). The MILP-based algorithm provides a significant reduction in computation time compared to existing methods
and is computationally tractable for problems of moderate size. An approximation algorithm is also developed that is appropriate for
large-scale problems. This paper presents a theoretical performance guarantee for the approximation algorithm and also demonstrates
its empirical performance. Finally, the mobile backbone network problem is extended to include mobile regular nodes, and exact and
approximate solution algorithms are presented for this extension.

Index Terms—Wireless sensor networks, mobile communication systems.

!

1 I NTRODUCTION not been adequately addressed. This paper provides tractable
solutions to the important problem of maximizing the number
D ETECTION and monitoring of spatially distributed phe-
nomena often necessitates the distribution of sensing
platforms. For example, multiple mobile robots can be used to
of regular nodes that achieve a desired level of throughput. It
also describes a new mobile backbone network optimization
problem formulation that models regular node mobility, and
explore an area of interest more rapidly than a single mobile
it provides tractable solutions to this problem, including the
robot [1], and multiple sensors can provide simultaneous
first known approximation algorithm with computation time
coverage of a relatively large area for an extended period of
that is polynomial in both the number of regular nodes and
time [2]. However, in many applications the data collected by
the number of mobile backbone nodes.
these distributed platforms is best utilized after it has been
aggregated, which requires communication among the robotic 2 BACKGROUND AND P ROBLEM S TATEMENT
or sensing agents. This paper focuses on a hierarchical network
architecture called a mobile backbone network, in which mo- Mobile backbone networks were described by Rubin et al.
bile agents are deployed to provide long-term communication [4] and Xu et al. [5] as a solution to the scalability issues
support for other agents in the form of a fixed backbone over inherent in mobile ad hoc networks. Noting that most commu-
which end-to-end communication can take place. Mobile back- nication bandwidth in single-layer large-scale mobile networks
bone networks can be used to model a variety of multi-agent is dedicated to packet-forwarding and routing overhead, they
systems. For example, a heterogeneous system composed of proposed a multi-layer hierarchical network architecture, as is
air and ground vehicles conducting ground measurements in currently used in the Internet. Srinivas et al. [6] defined two
a cluttered environment can be appropriately modeled as a types of nodes: regular nodes, which have limited mobility and
mobile backbone network, as can a team of mobile robotic communication capability, and mobile backbone nodes, which
agents deployed to collect streams of data from a network of have greater communication capability than regular nodes and
stationary sensor nodes. which can be placed at arbitrary locations in order to provide
communication support for the regular nodes.
Previous work has focused on optimal placement of mobile
Srinivas et al. [6] formulated the connected disk cover
backbone nodes in networks of fixed regular nodes, with the
(CDC) problem, in which many mobile backbone nodes with
objective of providing permanent communication support
fixed communication ranges are deployed to provide commu-
for the regular nodes [3]. Existing techniques, while exact,
nication support for a set of fixed regular nodes. The goal of
suffer from intractable computation times, even for problems
the CDC problem is to place the minimum number of mobile
of modest size. Furthermore, mobility of regular nodes has
backbone nodes such that each regular node is covered by
at least one mobile backbone node and all mobile backbone
• E. M. Craparo, J. P. How and E. Modiano are with the Department
of Aeronautics and Astronautics, Massachusetts Institute of Technology,
nodes are connected to each other. Thus, the CDC problem
Cambridge, MA 02139. takes a discrete approach to modeling communication, in that
E-mail: emilyc@alum.mit.edu two nodes can communicate if they are within communication
range of each other, and otherwise cannot.

2

This paper uses a more sophisticated model of communi- for the mobile backbone nodes to be “connected” to one
cation similar to that described by Srinivas and Modiano [3]. another. This assumption models the case in which mobile
We assume that the throughput (data rate) that can be achieved backbone nodes serve to provide a satellite uplink for regular
between a regular node and a mobile backbone node is a nodes; this is the case, for instance, in hastily-formed networks
monotonically nonincreasing function of both the distance that operate in disaster areas [7]. This assumption is also
between the two nodes and the number of other regular valid for the case in which the mobile backbone nodes are
nodes that are also communicating with that particular mobile known to be powerful enough to communicate effectively over
backbone node and thus causing interference. While our results the entire problem domain. For cases in which the problem
are valid for any throughput function that is monotonically domain is so large that mobile backbone nodes have difficulty
nonincreasing in both distance and cluster size, it is useful communicating with each other, it would be necessary to
to gain intuition by considering a particular example. One develop algorithms to ensure connectivity between the mobile
such example is the throughput function resulting from the backbone nodes (see [6], for example).
use of a Slotted Aloha communication protocol in which all We seek to provide the best possible throughput on a
regular nodes are equally likely to transmit. In this protocol, permanent basis; therefore, once the mobile backbone nodes
the throughput τ between regular node i and mobile backbone have been placed at their optimal positions, no improvement
node j is given by can be obtained by moving further. Thus, our model represents
1 1 (|A j |−1) 1 a “one-time” network design problem and is also suitable
τ(A j , di j ) = (1 − ) ( α ), (1) for cases in which mobile backbone nodes are deployable,
Aj Aj di j
but cannot move once they have been deployed. This is in
where A j is the number of regular nodes assigned to mobile contrast to the message ferrying problem, in which regular
backbone node j, di j is the distance between regular node i nodes have a finite amount of data available to transmit, and
and mobile backbone node j, and α is the path loss exponent. mobile backbone nodes must move around the network and
As noted in Ref. [3], one can use the fact that collect data [8]-[11].
1 1 We assume that the positions of regular nodes are known
(1 − )c−1 ≈ with complete accuracy, e.g., because the regular nodes are
c e
equipped with GPS. The problem of dealing with error in
for c > 0 to obtain a simpler expression for the Slotted regular node position estimates is a topic of future research.
Aloha throughput function. In this simplified expression, the As posed, the mobile backbone network optimization prob-
throughput is given by lem is quite difficult. Consider a simplification in which the
1 problem is decomposed into two parts: mobile backbone node
τ(A j , di j ) ≈ (2) placement and regular node assignment. Because the mobile
e · A j · diαj
backbone nodes can be placed anywhere in the plane, the
where e is the base of the natural logarithm. problem of finding an optimal placement is a hard nonconvex
Building upon this continuous throughput model, we pose optimization problem even when a simple heuristic technique
the mobile backbone network optimization problem as follows: is used to solve the assignment portion of the problem.
given a set of N regular nodes distributed in a plane, our Similarly, given a placement of mobile backbone nodes, the
goal is to place K mobile backbone nodes, which can occupy assignment portion of the problem is also non-trivial. An
arbitrary locations in the plane, while simultaneously assigning exhaustive enumeration of all K N possible assignments is
the regular nodes to the mobile backbone nodes, such that the impractical, and naive assignment techniques, such as that of
effectiveness of the resulting network is maximized. In this assigning each regular node to the nearest mobile backbone
work, the effectiveness of the resulting network is measured node, can perform quite badly [3].
by the number of regular nodes that achieve throughput at Although the problem considered in this paper is similar
least τmin , although other formulations (such as that which to that encountered in cellular network optimization, the
maximizes the aggregate throughput achieved by all regular approaches taken herein differ from those in the cellular liter-
nodes) are possible. The particular choice of objective in this ature. Some approaches to cellular network optimization take
work is motivated by applications such as control over a base station placement to be given, then optimize over user
network, in which a minimum throughput level is required, assignment and transmission power to minimize total overall
or sensing applications in which sensor measurements are of interference [12]-[15]. Others assume a simple heuristic for
a particular (known) size. Thus, our objective is to maximize the assignment subproblem and proceed to choose base station
the number of regular nodes that achieve throughput at least locations from among a restricted set [16], [17]. In contrast,
τmin . we seek to optimize the network simultaneously over mobile
Each regular node can be assigned to a single mobile backbone node placement and regular node assignment, with-
backbone node, and it is assumed that regular nodes assigned out assuming variable transmission power capability on the
to one mobile backbone node encounter no interference from part of the regular nodes and without limiting the placement
regular nodes assigned to other mobile backbone nodes (e.g., of the mobile backbone nodes.
because each “cluster” composed of a mobile backbone node A typical example of an optimal solution to the simultane-
and its assigned regular nodes operates at a different frequency ous placement and assignment problem is shown in Fig. 1 for
than other clusters). We also assume that there is no need a group of regular nodes denoted by ◦. The mobile backbone

3

where A is the set of valid assignments (i.e., those in which
each regular node is assigned to at most one mobile backbone
node), and Fτ (R, M, A, τmin ) is the number of regular nodes
that achieve throughput level τmin , given node placements R
and M, assignment A, and throughput function τ.
As discussed in Section 2, this problem is quite difficult.
Fortunately, it is possible to solve a simpler problem that
nonetheless yields an optimal solution to the original problem.
A key insight discussed in Refs. [3], [18] is that although
the mobile backbone nodes can occupy arbitrary locations,
they can be restricted to a small number of locations without
sacrificing optimality. For throughput functions that are mono-
tonically nonincreasing in distance, each mobile backbone
node can be placed at the 1-center of its assigned regular
nodes in an optimal solution.
The 1-center location for a set of regular nodes is the
Fig. 1. A typical example of an optimal mobile backbone location that minimizes the maximum distance from the mo-
network. Mobile backbone nodes, indicated by ∗, are bile backbone node to any of the regular nodes in the set.
placed such that they provide communication support for Consider a feasible solution to the mobile backbone network
regular nodes, shown as ◦. Each regular node is assigned optimization problem, i.e., a solution in which K mobile
to one mobile backbone node. Dashed lines indicate the backbone nodes are placed anywhere in the plane, each regular
radius of each cluster of nodes. node is assigned to at most one mobile backbone node, and
nodes, denoted by ∗, have been placed, and regular nodes have each assigned regular node achieves throughput at least τmin .
been assigned to them such that the number of regular nodes Let Ak denote the set of regular nodes assigned to mobile
that achieve throughput τmin is maximized. This example is backbone node k, for k = 1, ..., K, and let rk denote the distance
typical of an optimal solution in that the clusters of regular from mobile backbone node k to the most distant regular node
nodes and mobile backbone nodes are of relatively small in Ak . By our assumption that the solution is feasible, we
size, and the regular nodes are distributed intelligently among know that τ(|Ak |, rk ) ≥ τmin . Now, modify the solution such
the mobile backbone nodes, with fewer regular nodes being that mobile backbone node k is placed at the 1-center of the
allocated to mobile backbone nodes with larger cluster radii. In set Ak , leaving the assignment Ak unchanged. By definition
this example, all regular nodes have been successfully assigned of the 1-center, the distance from every regular node in Ak
to mobile backbone nodes. to mobile backbone node k is no more than rk . In particular,
In the problem considered in Figure 1, regular nodes are if the distance from the mobile backbone node to the most
stationary, and their positions are given as problem data, as distant regular node in Ak is now denoted by rk , we know
has been assumed in previous work [3]. New solutions to that τ(|Ak |, rk ) ≥ τ(|Ak |, rk ) ≥ τmin , since τ is a nonincreasing
this problem are described in Section 3. In Section 4, we function of distance. Thus, the modified solution in which the
consider an extension to this problem in which the placement mobile backbone node is placed at the 1-center of its assigned
of regular nodes is also a decision variable. That is, the goal regular nodes is feasible and has the same objective value as
in Section 4 is to place both N regular nodes and K mobile the original solution. Repeating the argument for the remaining
backbone nodes, while simultaneously assigning regular nodes mobile backbone nodes 1, ..., K, we can see that restricting the
to mobile backbone nodes, such that the effectiveness of the feasible set of mobile backbone node locations to the set of 1-
overall network is maximized. center locations of all subsets of regular nodes does not reduce
the maximum objective value that can be obtained.
Fortunately, although there are 2N possible subsets of N
3 S TATIONARY REGULAR NODES
regular nodes, there are only O(N 3 ) distinct 1-center lo-
This section describes the mobile backbone network optimiza- cations [19]. Although a particular 1-center location may
tion problem in which regular nodes are stationary. Denoting correspond to multiple subsets of regular nodes, it is uniquely
the problem data as Ri (the locations of the regular nodes, defined by the regular nodes that are most distant from it
i = 1, . . . , N), τmin (the desired minimum throughput level), in all of these sets. Each 1-center location either coincides
and τ (the throughput function); and the decision variables with a regular node, lies at the center of the diameter de-
as Mi (the selected locations of the mobile backbone nodes, scribed by two regular nodes, or lies at the circumcenter of
i = 1, . . . , K), and A (the assignment of regular nodes to mobile N
three regular nodes [20]. Thus, there are at most +
backbone nodes), this optimization problem can be stated as: 1
N N
max Fτ (R, M, A, τmin ) + distinct 1-center locations, and they can be
M, A 2 3
efficiently enumerated through enumeration of the possible
subject to Mi ∈ R2 , i = 1, . . . , K
sets of “defining” regular nodes. Moreover, the maximum
A∈A communication radius associated with each possible mobile

4

N+1
backbone node location is easy to compute. This radius, which
1
we will denote as the 1-center radius, is the distance from N+2
the 1-center location to any of the defining regular nodes. y1c1
2
For 1-center locations defined by the diameter between two N+3
y2c2
nodes or the circumcircle of three nodes, the 1-center radius
y3c3

…
is simply the radius of the associated circle. For 1-center

…
t
locations defined by a single regular node, the associated 1- s
i yjcj
center radius is zero.1 N+j

The insight that mobile backbone nodes can be restricted

…
to a relatively small number of locations without sacrificing yMcM

…
optimality of the overall solution allows the mobile backbone
N
network optimization problem to be simplified. The problem
N+M
becomes
max Fτ (R, M, A, τmin ) Fig. 2. The network design problem corresponding to
M, A
the joint placement and assignment problem for mobile
subject to Mi ∈ C(R), i = 1, . . . , K backbone networks. Unlabeled arc capacities are equal
A∈A to one.
where C(R) denotes the set of 1-center locations of the regular
nodes, and |C| = O(N 3 ).
Slotted Aloha function described by Eq. (2), one can simply
take the inverse of the expression with respect to cluster size,
3.1 MILP approach
evaluate the inverse at the desired minimum throughput level
A primary contribution of this work is the development of τmin , and take the floor of the result to obtain an integer value
a single optimization problem that simultaneously solves the for ci . For the throughput function given by Eq. (2), we have
mobile backbone node placement and regular node assignment
1
problems. This is accomplished through the formulation of a ci = , (3)
network design problem. In network design problems, a given e · τmin · riα
network (represented by a directed graph) can be augmented where ri is the 1-center radius associated with 1-center location
with additional arcs for a given cost, and the goal is to i. If the throughput function cannot easily be inverted with re-
intelligently “purchase” a subset of these arcs in order to spect to cluster size, as is the case with the exact Slotted Aloha
achieve a desired flow characteristic [21]. throughput function given in Eq. (1), one can perform a search
The network design problem that produces an optimal for the largest cluster size ci ≤ N such that τ(ci , ri ) ≥ τmin .
solution to the mobile backbone network optimization problem For example, a binary search for ci would involve O(log(N))
is constructed as follows. A source node, s, is connected to evaluations of the function τ. In either case, the resulting value
each node in the set of nodes N= {1, . . . , N} (see Fig. 2). of ci is the maximum number of regular nodes that can be
N represents the set of regular nodes. The arcs connecting assigned to the mobile backbone node at location i, such that
s to i ∈ N are of unit capacity. Each node i ∈ N is in turn each of these regular nodes achieves throughput at least τmin .
connected to a subset of the nodes in M= {N + 1, . . . , N + M}, Denote the set of nodes for the network design problem
where M is O(N 3 ). M represents the set of possible mobile by N and the set of arcs by A . If K mobile backbone
backbone node locations, i.e., the 1-center locations of the nodes are available to provide communication support for N
subsets of regular nodes. Node i ∈ N is connected to node regular nodes at given locations, and a throughput level τmin is
N + j ∈ M if, and only if, regular node i is within the 1-center specified, the goal of the network design problem is to select
radius of location j. The arc connecting i to N + j is of unit K arcs incident to the sink and a feasible flow x such that the
capacity. Finally, each node in M is connected to the sink, net flow through the graph is maximized.
t. The capacity of the arc connecting node N + i ∈ M to the The network design problem can be solved via the following
sink is the product of a binary variable yi , which represents mixed-integer linear program (MILP), which we denote as the
the decision of whether to place a mobile backbone node at
Network Design MILP:
location i, and a constant ci , which is the maximum number of
N
regular nodes that can be assigned to a mobile backbone node
max ∑ xsi (4a)
at location i at the desired throughput level τmin . The quantity x,y
i=1
ci can be efficiently computed in one of two ways. For an M
easily-inverted throughput function such as the approximate subject to ∑ yi = K (4b)
i=1
1. In practice, we consider the 1-center radius of such locations to have
a small positive value ε in order to assure boundedness of the throughput ∑ xi j = ∑ xli i ∈ N {s,t} (4c)
function; this modification does not impact the solution as long as ε is such j:(i, j)∈A l:(l,i)∈A
that the throughput achieved by the regular node does not cross the threshold xi j ≥ 0 ∀ (i, j) ∈ A (4d)
of τmin , and no additional regular nodes fall within a distance of ε of the
mobile backbone node. xi j ≤ 1 ∀ (i, j) ∈ A : j ∈ N {t} (4e)

5

x(N+i)t ≤ yi ci i ∈ {1, . . . , M} (4f) TABLE 1
Average computation times for the MILP-based and
yi ∈ {0, 1} i ∈ {1, . . . , M} (4g) search-based algorithms, for various numbers of regular
xi(N+ j) ≤ y j ∀ (i, N + j) ∈ A : i ∈ {1, . . . , N} (4h) (N) and mobile backbone nodes (K) in the maximum fair
placement and assignment (MFPA) problem. All models
The objective of the Network Design MILP is to maximize
were formulated in GAMS 22.9 and solved using ILOG
the flow x through the graph (Eq. (4a)). The constraints state
CPLEX 11.2.0 on a 3.16 GHz Intel Xeon CPU with 3.25
that K arcs (mobile backbone node locations) must be selected
GB of RAM.
(4b), flow through all internal nodes must be conserved (4c),
arc capacities must be observed (4d - 4f), and yi is binary for N K MILP Algorithm Search-based Algorithm
all i (4g). Constraint 4h is a valid inequality that decreases 3 2 .92 sec 3.12 sec
computation time by reducing the size of the feasible set 4 2 1.24 sec 14.97 sec
in the LP relaxation [22]. Note that, for a given specification 5 2 1.35 sec 45.15 sec
of the y vector, all flows x are integer in all basic feasible 6 2 1.37 sec 1 min 28 sec
solutions of the resulting (linear) maximum flow problem. 6 3 1.45 sec 5 min 44 sec
This network design problem exactly corresponds to the 8 3 1.95 sec 33 min
mobile backbone network optimization problem as posed in 10 3 2.23 sec 2 hr 6 min
this section. The geometry of the mobile backbone network 15 5 5.06 sec 5 hr 4 min
problem is described by the arcs connecting node sets N
and M, while both the throughput function and the desired O(log(N)) times for different throughput values in order to
minimum throughput level are captured in the capacities of find the maximum throughput value such that all regular
the arcs connecting nodes in M to the sink, t. Any feasible nodes can be assigned. There are at most O(N 4 ) possible
placement of mobile backbone nodes and assignment of regu- values for the minimum throughput achieved by any
lar nodes is associated with a feasible solution to the network regular node; searching among these throughput values
design problem with the same objective value; likewise, any via binary search would require O(log(N)) solutions of
integer feasible solution to the network design problem yields the MILP problem.
a feasible placement and assignment in the mobile backbone Remark 2: It should be noted that the worst-case complexity of
network optimization problem, such that the number of regular mixed-integer linear programming is exponential in the num-
nodes assigned is equal to the volume of flow through the ber of binary variables. However, this approach performs well
network design graph. The equivalence of these two problems in practice, and simulation results indicate that it compares
is formally stated in Theorem 1: very favorably with the search-based approach developed in
Ref. [3] for the MFPA problem (See Table 1). Note that while
Theorem 1 Given an instance of the mobile backbone net- the computation time of the search-based algorithm increases
work design problem, the corresponding Network Design very rapidly with the problem size, the MILP-based algorithm
MILP yields an optimal solution to the mobile backbone node remains computationally tractable for problems of practical
placement and regular node assignment problems. scale.
Proof: The proof of Theorem 1 can be found in Ap-
pendix A. 3.2 Approximation algorithm
A solution to problem (4) provides both a placement of Table 1 indicates that the MILP formulation described by Eq. 4
mobile backbone nodes and an assignment of regular nodes provides an optimal solution in tractable time for moderately-
to mobile backbone nodes. Mobile backbone nodes are placed sized problems. However, this method is demonstrated to scale
at locations for which yi = 1, and regular node i is assigned poorly with problem size. Moreover, we have shown that the
to the mobile backbone node at location j if xi(N+ j) = 1. The network design problem on a network of the general form
number of regular nodes assigned is equal to the volume of shown in Figure 2 is NP-hard [23]. Therefore, an approxima-
flow through the graph. tion algorithm with computation time that is polynomial in the
We make the following observations about this algorithm: number of regular nodes and the number of mobile backbone
Remark 1: If K mobile backbone nodes are available and the nodes is desirable. This section describes such an algorithm.
goal is to assign as many regular nodes as possible such that The primary insight on which the approximation algorithm
a desired minimum throughput is achieved for each assigned is based is the fact that the maximum number of regular
regular node, the MILP problem in Eq. 4 needs only to be nodes that can be assigned is a submodular function of the
solved once for the desired throughput value and with a fixed set of mobile backbone node locations selected. Given a finite
value of K. However, we also note that the Network Design ground set D = {1, . . . , d}, a set function f (S) defined for all
MILP can be used as a subroutine in solving the maximum subsets S of D is said to be submodular if it has the property
fair placement and assignment (MFPA) problem considered that
in Ref. [3], in which the objective is to maximize the f (S ∪ {i, j}) − f (S ∪ {i}) ≤ f (S ∪ { j}) − f (S)
minimum throughput achieved by any regular node, such
that all regular nodes are assigned. To solve the MFPA for all i, j ∈ D, i = j and S ⊂ D {i, j} [24]. In the context
problem, it is necessary to solve the Network Design MILP of the network design problem, this means that the maximum

6

flow through the network is a submodular function of the set Algorithm 1
of arcs incident to the sink that are selected. S←0 /
It has been shown [25] that for maximization of a nonde- max f low ← 0
creasing submodular set function f , where f (0) = 0, greedy
/ for k=1 to K do
selection of elements yields a performance guarantee of for m=1 to M do
1 − (1 − P )P > 1 − 1 , where P is the number of elements
1
e
if f (S ∪ {m}) ≥ max f low then
to be selected from the ground set and e is the base of the max f low ← f (S ∪ {m})
natural logarithm. This means that if an exact algorithm selects m∗ ← m
P elements from the ground set and produces a solution of end if
value OPT , a greedy selection of P elements (i.e., selection end for
via a process in which element i is selected if it is the S ← S ∪ {m∗ }
element that maximizes f (S ∪ {i}), where S is the set of end for
elements already selected) produces a solution of value at least return S
1
(1 − (1 − P )P ) · OPT .
For the network design problem considered in this paper,
P = K (the number of mobile backbone nodes that are to be until |S| = K. Algorithm 1 then returns set S. The performance
placed), and OPT is the number of regular nodes that are of Algorithm 1 is described by Theorem 2.
assigned in an optimal solution. Note that greedy selection of
K arcs amounts to solving at most O(N 3 K) linear maximum Theorem 2 Algorithm 1 returns a solution S such that f (S) ≥
flow problems on graphs with at most N + K + 2 nodes. Thus, (1 − 1 ) · f (S∗ ) , where S∗ is the optimal solution to the
e
the computation time of the greedy algorithm is polynomial network design problem on G.
in the number of regular nodes and the number of mobile
Proof: This follows from the fact that all maximum flows
backbone nodes. Furthermore, each of the maximum flow
through G are integer, and from the observation that the
problems solved by the greedy algorithm is solved over a
maximum flow that can be routed through G is a submodular
bipartite graph with node sets N∪{t} and {s}∪K, where K
function of S, the set of arcs that are selected.
is the set of nodes from M whose outgoing arcs are selected.
Because maximum flow problems can be solved even more
efficiently in bipartite networks than in general networks 3.3 Experimental evaluation of approximation algo-
[21], the greedy algorithm is thus highly efficient. Further rithm
computational efficiency can result from exploitation of max As described in Section 3.2, greedy selection of mo-
flow/min cut duality [23]. bile backbone node locations results in assignment of at
The submodularity of the network design objective is for- 1
least (1 − (1 − K )K ) · OPT ≥ (1 − 1 ) · OPT regular nodes,
e
mally stated in Lemma 1: where K is the number of mobile backbone nodes that are to
be placed and OPT is the number of regular nodes assigned
Lemma 1 If G is a graph in the network design problem by an exact algorithm (such as the MILP algorithm described
described in Section 3.1, the maximum flow that can be routed in Section 3.1) [25]. While such worst-case performance
through G is a submodular function of the set of arcs selected. guarantees are quite useful, it is also worthwhile to examine
the typical performance of the approximation algorithm on
Proof: The proof of Lemma 1 can be found in Ap- many problems.
pendix B. To this end, we have performed computational experiments
Lemma 1 implies that greedy selection of mobile backbone on a number of problems of various degrees of complexity.
node locations (i.e., selection via a process that maximizes the Regular node locations were generated randomly in a finite
total number of regular nodes assigned for each incremental 2-dimensional area, and a moderate throughput value was
selection of a new mobile backbone node location) yields prov- specified (i.e., one high enough such that there was no trivial
ably good solution to the overall placement and assignment selection of mobile backbone node locations that would result
problem. in assignment of all regular nodes). Results were averaged
Given a network design graph G, K mobile backbone nodes over a number of trials for each problem dimension.
and M possible mobile backbone node locations, and denoting Fig. 3 shows the performance of the approximation algo-
by f the maximum flow through G as a function of the set of rithm relative to the exact (MILP) algorithm. In Fig. 3(a),
mobile backbone node locations selected, this greedy selection the average percentage of regular nodes assigned by the
process is described by Algorithm 1. Algorithm 1 begins with exact algorithm that are also assigned by the approximation
an empty set of selected mobile backbone node locations, S, algorithm is plotted, along with the theoretical lower bound
and incrementally adds K elements to this set. In each of of (1 − 1 ) · OPT , for various problem sizes. In this fig-
e
K rounds of selection, M − |S| maximum flow problems are ure, a data point at 100% would mean that, on average,
solved on graphs consisting of nodes s, t, N, S and each of the approximation algorithm assigned as many regular nodes
the M − |S| nodes belonging to set MS. In each round of as the exact algorithm for that particular problem size. As
selection, the node from set MS that maximizes the total the graph shows, the approximation algorithm consistently
flow through G is added to set S, and this process continues exceeds the theoretical performance guarantee and achieves

7

nearly the same level of performance as the exact algorithm
for all problem sizes considered.
Fig. 3(b) shows the computation time required for each of
these algorithms, plotted on a logarithmic axis. As the figure
shows, the computation time required for the approximation
algorithm scales gracefully with problem size. The average
computation time of the approximation algorithm was about 15
seconds for N = 100 and K = 14, whereas the MILP algorithm
took nearly twelve minutes to solve a problem of this size. The
significant improvement in computation time achieved by the
approximation algorithm makes it appropriate for some real-
time applications, while the exact algorithm is a promising
candidate for one-time design problems involving significant
costs.
Both the MILP algorithm and the approximation algorithm
(a) Performance of the approximation algorithm developed in this paper,
were formulated in GAMS 22.9 and solved using ILOG relative to an exact solution technique, in terms of number of regular nodes
CPLEX 11.2.0 on a 3.16 GHz Intel Xeon CPU with 3.25 GB assigned at the given throughput level.
of RAM.

4 J OINT PLACEMENT OF REGULAR AND MO -
BILE BACKBONE NODES
As described in Section 3, existing problem formulations in
the study of mobile backbone networks have assumed that the
locations of regular nodes are fixed a priori and that only
the locations of mobile backbone nodes are variable [18], [3],
[6]. This assumption is reasonable for some applications, such
as scenarios that involve mobile agents extracting data from a
fixed sensor network. However, in many applications the loca-
tions of both regular nodes and mobile backbone nodes can be
controlled. For example, a heterogeneous system composed of
air and ground vehicles conducting ground measurements in an
urban environment can be appropriately modeled as a mobile
(b) Computation time of the approximation algorithm and the exact (MILP)
backbone network: the ground vehicles are well positioned to algorithm for various problem sizes. Due to the large range of values
make observations of phenomena at ground level, but their represented, a logarithmic scale is used.
movement and communication are hindered by surrounding
obstacles. Air vehicles, on the other hand, are poorly equipped Fig. 3. Comparison of the exact and approximation
to observe events on the ground but can easily move above algorithms developed in this section. Although the MILP-
ground obstacles and communicate. based exact algorithm developed in this section signif-
This section develops a modeling framework and solution icantly outperforms existing techniques in terms of re-
technique that are appropriate for problems of this nature. We quired computation time, our experiments indicate that
assume that L candidate regular node locations are available a the greedy approximation algorithm achieves nearly the
priori, perhaps selected by heuristic means or due to logistical same level of performance with an even greater reduction
constraints. Each of N regular nodes (N ≤ L) must occupy one in computation time.
of these locations, and no two regular nodes can be assigned
to the same location. Given an initial location and a mobility
constraint, each regular node is capable of reaching a subset reachable from each of the initial regular node locations); and
of the other locations. There are K mobile backbone nodes the decision variables as Fi (the final locations of the regular
(K ≤ N) that can be placed anywhere, a throughput function τ nodes), Mi (the selected locations of the mobile backbone
is specified, and a desired minimum throughput τmin is given. nodes, i = 1, . . . , K), and A (the assignment of regular nodes
Given these assumptions, the goal of this section is to place to mobile backbone nodes), this optimization problem is:
both the regular nodes and mobile backbone nodes while max Fτ (F, M, A, τmin )
simultaneously assigning regular nodes to mobile backbone F, M, A
nodes in order to maximize the number of regular nodes that subject to Fi ∈ r(Ii ), i = 1, . . . , N
are successfully assigned and achieve the desired minimum Mi ∈ C(I), i = 1, . . . , K
throughput level τmin , under the given throughput function τ.
A∈A
Denoting the problem data as Ii (the initial locations of
the regular nodes, i = 1, . . . , N) and and r(Ii ) (the locations where Fτ (F, M, A, τmin ) is the number of regular nodes that

8

N+1 N+L+1 N+2L+1
placement and assignment of regular and mobile backbone
1 nodes is associated with a feasible solution to the network
N+2 N+L+2
N+2L+2 design problem; likewise, any feasible solution to the network
y1 c1 design problem yields a feasible placement and assignment in
2
N+2L+3 the mobile backbone network optimization problem.
y2 c2
Denote the set of nodes in the network design graph by

…
…
y3 c3 N and the set of arcs by A . If K mobile backbone nodes
…

…
N+i N+L+i t are available and a minimum throughput level is specified, the
s
i yjcj goal of the network design problem is to select K arcs from
N+2L+j {N + 2L + 1, . . . , N + 2L + M} and a feasible flow xi j , (i, j) ∈
A such that the s −t flow is maximized. This problem can be
…

yM cM solved via the following MILP:
…
…

… N
max ∑ xsi (5a)
N x,y
i=1
M
N+L N+2L N+2L+O(N3) subject to ∑ yi = K (5b)
i=1
Fig. 4. The network design problem corresponding to ∑ xi j = ∑ xli i ∈ N {s,t} (5c)
the joint placement and assignment problem for mobile j:(i, j)∈A l:(l,i)∈A
backbone networks, with regular node mobility. Unlabeled xi j ≥ 0 ∀ (i, j) ∈ A (5d)
arc capacities are equal to one. For clarity, not all arcs and
xi j ≤ 1 ∀ (i, j) ∈ A : j ∈ N {t} (5e)
nodes are shown.
x(N+2L+i)t ≤ yi ci i ∈ {1, . . . , M} (5f)
yi ∈ {0, 1} i ∈ {1, . . . , M} (5g)
achieve throughput level τmin , given node placements F and
M, assignment A, and throughput function τ. A denotes the x(N+L+i)(N+2L+ j) ≤ y j (N + L + i, N + 2L + j) ∈ A : i ∈ {1, ..., L}
set of valid assignments, while C denotes the 1-centers of (5h)
candidate regular node locations. where the constraints state that K arcs (mobile backbone node
locations) must be selected (5b), flow through all internal
4.1 Network design formulation nodes must be conserved (5c), arc capacities must be observed
(5d- 5f), and yi is binary for all i (5g). Constraint 5h is again
Optimal simultaneous placement and assignment of regular
a valid inequality included to reduce computation time by
nodes and mobile backbone nodes is again achieved through
strengthening the LP relaxation [22].
the solution of a network design problem.
The network design graph over which this optimization Fig. 5 shows an example of a solution to the simultaneous
takes place is schematically represented in Fig. 4. This graph is placement and assignment problem with regular node move-
similar to that shown in Fig. 2, with the exception that sets of ment. The regular nodes, initially in positions indicated by •,
nodes L= {N + 1, . . . , N + L} and L = {N + L + 1, . . . , N + 2L} are able to move to other locations (◦) within their radii of
are added. These nodes represent locations to which regular motion, indicated by shaded circles. This initial configuration
nodes may move. Node i ∈N is connected to node N + j ∈L if, is shown in Fig. 5(a). In an optimal solution to this problem,
and only if, regular node i can reach sensing location j under shown in Fig. 5(b), the regular nodes have moved such that
its mobility constraint. Each node N + i ∈L is connected to they are grouped into compact clusters for which the mobile
node N + L + i ∈L via an arc of unit capacity. This enforces backbone nodes can provide an effective communication in-
the constraint that at most one regular node can occupy each frastructure. The clusters are relatively balanced, in that the
location. Finally, node N + L + i ∈L is connected to node clusters with larger radii tend to have fewer regular nodes,
N + 2L + j ∈M if, and only if, sensing location i is within while the more compact clusters can accommodate more regu-
the 1-center radius of 1-center j. Finally, each node in M is lar nodes and still achieve the desired minimum throughput. In
connected to the sink t, and the capacity of the arc connecting this example, all regular nodes have been successfully assigned
N + 2L + i ∈M to t is again the product of yi and ci . to mobile backbone nodes.
This network design problem exactly describes the mobile This algorithm is designed maximize the number of regular
backbone network optimization problem with mobile regular nodes that are assigned at throughput level τmin . If, instead,
nodes. The arcs connecting node sets N and L reflect the the goal is to achieve the best possible minimum throughput
mobility constraints of the regular nodes, while the geometric such that all regular nodes are assigned to a mobile backbone
aspects of the mobile backbone node placement problem are node (i.e., to solve the MFPA problem), it is necessary to
described by the arcs connecting node sets L and M. As solve the MILP problem in Eq. 5 O(log(L)) times for different
in Section 3, both the throughput function and the desired throughput values (which result in different values for the ci ’s
minimum throughput level are captured in the capacities of in the network design problem).
the arcs connecting nodes in M to the sink, t. Any feasible In this paper, all candidate regular node locations are

9

(a) Initial regular node placement, with radius of motion for (b) An optimal placement of regular and mobile backbone
each regular node. nodes.
Fig. 5. A small example of mobile backbone network optimization with limited regular node movement. Open circles
represent possible regular node locations, and filled circles are the initial locations of the regular nodes. Shaded circles
in the left figure indicate the possible radius of motion of each regular node. In the right figure, mobile backbone nodes
are placed such that they provide communication support for the regular nodes. Each regular node is assigned to at
most one mobile backbone node. Dotted lines indicate regular node motion in this optimal solution. Dashed circles
indicate the communication radius of each cluster of nodes. In this example, all regular nodes have been successfully
assigned to mobile backbone nodes.

considered to be equally valuable – that is, a regular node Algorithm 2
that transmits to a mobile backbone node from location i S←0 /
contributes as much to the objective as a mobile backbone max f low ← 0
node transmitting from location j. However, this formulation for k=1 to K do
can easily be modified to model variably valuable sensing for m=1 to M do
locations by replacing the objective function with if f (S ∪ {m}) ≥ max f low then
max f low ← f (S ∪ {m})
L
m∗ ← m
max ∑ vi xN+i,N+L+i
x,y
i=1
end if
end for
where vi is the value of location i. S ← S ∪ {m∗ }
end for
4.2 Approximation Algorithm return S

While the MILP-based algorithm described in Section 4.1
provides an exact solution to the mobile backbone network
Lemma 2 motivates consideration of a greedy algorithm
optimization problem, its worst-case computation time is again
for the problem of maximizing the number of regular nodes
exponential in the number of binary variables. Fortunately, the
that achieve throughput level τmin . Given a network design
submodularity property described in Section 3.2 also holds for
graph G of the form shown in Figure 4, K mobile backbone
graphs of the form shown in Figure 4:
nodes and M possible mobile backbone node locations, and
denoting by f the maximum flow through G as a function of
Lemma 2 If G is a graph in the network design problem the set of mobile backbone node locations selected, this greedy
described in Section 4.1, the maximum flow that can be routed algorithm is described by Algorithm 2. Theorem 3 describes
through G is a submodular function of the set of arcs selected. the performance of Algorithm 2:
Proof: The proof of Lemma 2 is similar to that of
Lemma 1, with the exception that the reduction to a bipartite Theorem 3 Algorithm 2 returns a solution S such that f (S) ≥
matching problem is replaced with a reduction to a node- (1 − 1 ) · f (S∗ ) , where S∗ is the optimal solution to the
e
disjoint path problem on a tripartite graph. network design problem on G.

10

1
Proof: This follows from the fact that all maximum flows

fraction of regular nodes assigned
through G are integer, and from the observation that the 0.95 approximation algorithm
maximum flow that can be routed through G is a submodular theoretical lower bound
function of S, the set of arcs that are selected. 0.9

Thus, Algorithm 2 is an approximation algorithm with 0.85
approximation guarantee 1 − 1 . Additionally, because each
e
round of greedy selection consists of solving a polynomial 0.8

number of maximum flow problems on graphs with at most
0.75
N + 2L + K + 2 nodes, and there are K rounds of selection,
the running time of Algorithm 2 is polynomial in the number 0.7
of regular nodes, the number of locations, and the number
of mobile backbone nodes. Furthermore, all network flow 0.65

problems solved by Algorithm 2 are formulated on bipartite
graphs, for which highly efficient algorithms exist [21]. N=40, K=8 N=50, K=9 N=60, K=10 N=70, K=11 N=80, K=12 N=90, K=13
problem size
The performance of Algorithm 2 relative to the exact
(MILP) algorithm developed in this section is shown in Fig- (a) Performance of the approximation algorithm developed in this section,
relative to an exact solution technique, in terms of number of regular nodes
ure 6. Again, the approximation algorithm exhibits excellent assigned at the given throughput level.
empirical performance, achieving results comparable to those
4
of the exact algorithm with a great reduction in computation 10

time. The results shown in Figure 6 were obtained from models approximation algorithm
formulated in GAMS 22.9 and solved using ILOG CPLEX
computation time (seconds)
exact algorithm
11.2.0 on a 3.16 GHz Intel Xeon CPU with 3.25 GB of RAM. 10
3

5 C ONCLUSIONS 10
2

This work has described new algorithms for solving the
problem of mobile backbone network optimization. Exact 1
10
MILP-based techniques and the first known approximation
algorithms with computation time polynomial in the number
of regular nodes and the number of mobile backbone nodes
0
were described. 10
N=40, K=8 N=50, K=9 N=60, K=10 N=70, K=11 N=80, K=12 N=90, K=13
Based on simulation results, we conclude that the MILP- problem size
based approach provides a considerable computational advan- (b) Computation time of the approximation algorithm and the exact (MILP)
tage over existing techniques for mobile backbone network algorithm for various problem sizes. Due to the large range of values
represented, a logarithmic scale is used.
optimization. This approach has been successfully applied to
a problem in which a maximum number of regular nodes are Fig. 6. Comparison of the exact and approximation algo-
to be assigned to mobile backbone nodes at a given level rithms developed in this section. On average, the approx-
of throughput, and to a related problem from the literature imation algorithm greatly exceeded its performance guar-
in which all regular nodes are to be assigned to a mobile antee, achieving nearly the same level of performance as
backbone node such that the minimum throughput achieved the exact algorithm for all problem sizes considered. In all
by any regular node is maximized. cases, L = 1.5N.
For cases in which a MILP approach is impractical due
to constraints on computation time, the greedy approximation
algorithms developed in this paper present viable alternatives. R EFERENCES
The approximation algorithms carry the benefit of a theoret-
ical performance guarantee, and simulation results indicate [1] W. Burgardt, M. Moorstt, D. Fox, R. Simmons and S. Thrun, “Col-
laborative Multi-Robot Exploration,” Proceedings of the 2000 IEEE
that they perform very well for the problem of assigning a International Conference on Robotics and Automation, San Francisco,
maximum number of regular nodes such that each assigned CA, April 2000.
regular node achieves a minimum throughput level. [2] R. Balasubramanian, S. Ramasubramanian and A. Efrat, “Coverage
Time Characteristics in Sensor Networks,” in 2006 IEEE International
Conference on Mobile Adhoc and Sensor Systems (MASS 2006), October
2006.
[3] A. Srinivas and E. Modiano, “Joint node placement and assignment for
ACKNOWLEDGMENT throughput optimization in mobile backbone networks,” To appear in
Proc. IEEE INFOCOM’08, Apr. 2008.
This work was funded in part by AFOSR grant FA9550-04- [4] I. Rubin, A. Behzadm R. Zhang, H. Luo, and E. Caballero, “TBONE: a
Mobile-Backbone Protocol for Ad Hoc Wireless Networks,” Proc. IEEE
1-0458, NSF grant CCR-0325401, NSF grant CNS-0915988, Aerospace Conference, 6, 2002.
and an NSF Graduate Fellowship. [5] K. Xu, X. Hong, and M. Gerla, “Landmark Routing in Ad Hoc

11

Networks with Mobile Backbones,” Journal of Parallel and Distributed Let Ak denote the set of regular nodes assigned to mobile
Computing, 63, 2, 2003, pp. 110-122. backbone node k, for k = 1, ..., K. Calculate the 1-center of
[6] A. Srinivas, G. Zussmanm and E. Modiano, “Construction and Mainte-
nance of Wireless Mobile Backbone Networks,” IEEE/ACM Trans. on set Ak and denote its location by lk and its 1-center radius by
Networking, Vol. 17, No. 1, pp. 239 - 252, Feb. 2009. rk . Note that although some 1-center locations may coincide,
[7] B. Steckler, B. Bradford, and S. Urrea, “Hastily Formed Networks For each will be distinct in the set M. By our assumption that
Complex Humanitarian Disasters - After Action Report and Lessons
Learned from the Naval Postgraduate School’s Response to Hurricane we are given a feasible solution to the original problem, we
Katrina,” http://faculty.nps.edu/dl/HFN/documents/ know that Ai ∩ A j = 0 ∀ i, j ∈ {1, ..., K}, i = j. Therefore, the
/
NPS_Katrina_AAR-LL_04-MAY-06.pdf, Sept 2005 (accessed defining regular nodes of each 1-center must be distinct, and
Sept 2009).
[8] D. Jea, A. A. Somasundara, and M. B. Srivastava, “Multiple controlled li is distinct from l j in the set M.
mobile elements (data mules) for data collection in sensor networks,” Assume without loss of generality that the network design
In Proc. IEEE/ACM DCOSS 05, Jun. 2008. graph is labeled such that nodes N + 1, ..., N + K correspond
[9] R. Shah, S. Roy, S. Jain, and W. Brunette, “Data MULEs: Modeling
a three-tier architecture for sparse sensor networks,” In Proc. IEEE to the locations l1 , ..., lK . Set binary variables y1 , ..., yK equal
SNPA03, May 2003. to 1, and set the remaining binary variables yK+1 , ..., yM equal
[10] M. M. Bin Tariq, M. Ammar, and E. Zegura, “Message ferry route to 0. Note that constraints 4b and 4g in the Network Design
design for sparse ad hoc networks with mobile nodes,” In Proc. ACM
MobiHoc 06, May 2006. MILP are now satisfied.
[11] W. Zhao, M. Ammar, and E. Zegura, “A message ferrying approach for Next, for each regular node i that is a member of set Ak
data delivery in sparse mobile ad hoc networks,” In Proc. ACM MobiHoc for some k, set xsi equal to 1. Set xs j equal to 0 for each
04, May 2004.
[12] S. V. Hanly, “An algorithm for combined cell-site selection and power regular node j for which k : j ∈ Ak . For k = 1, ..., K, and
control to maximize cellular spread spectrum capacity,” IEEE Journal for all regular nodes i such that i ∈ Ak , set xi(N+k) equal to
on Selected Areas in Communications, Vol. 13, No. 7, pp. 1332-1340, 1 (recall that an arc exists between every such pair of nodes
Sept. 1995.
[13] G. J. Foschini and Z. Milzanic, “A simple distributed autonomous power (i, N + k) by definition of the 1-center and by construction
control algorithm and its convergence,” IEEE Transactions on Vehicular of the network design problem). Set x j(N+k) equal to 0 for all
Technology, Vol. 40, pp.641-646, 1993. regular nodes j such that j ∈ Ak (if an arc exists between node
/
[14] R. Yates and C. Y. Huang, “Integrated power control and base station
assignment,” IEEE Transactions on Vehicular Technology, Vol. 44, No.
j and node N + k). For all arcs terminating at nodes N + K +
3, pp.638-644, 1995. 1, ..., N + M, set the flow x equal to 0. Note that flows for all
[15] R. Mathar and T. Niessen, “Integrated power control and base station arcs terminating in node sets N and M have now been assigned
assignment,” Wireless Networks, Vol. 6, No. 6, pp. 421-428, Dec. 2000.
[16] C. Glaßer, S. Reith and H. Vollmer, “The complexity of base station
such that constraints 4d and 4e are satisfied. Furthermore, flow
positioning in cellular networks,” Discrete Applied Mathematics, Vol. conservation (constraint 4c) is now satisfied for nodes 1, ..., N:
148, No. 1, pp. 1-12, 2005. regular nodes that are assigned to a mobile backbone node
[17] E. Amaldi, A. Capone, F. Malucelli, “Radio planning and coverage
optimization of 3G cellular networks,” Wireless Networks, Vol. 14, pp.
have one unit of incoming flow and one unit of outgoing flow
435-447, 2008. (since every regular node is assigned to at most one mobile
[18] E. Craparo, J. How and E. Modiano, “Optimization of Mobile Backbone backbone node by our assumption that the original solution
Networks: Improved Algorithms and Approximation,” Proceedings of
the American Control Conference, June 2008.
was feasible), while regular nodes that are not assigned to a
[19] F. Preparta and M. Shamos, Computational Geometry: An Introduction, mobile backbone node have no incoming or outgoing flow.
Springer-Verlag, New York, 1985. Finally, consider the arcs connecting nodes N + 1, ..., N + M
[20] P. Agarwal and M. Sharir, “Efficient Algorithms for Geometric Opti-
mization,” ACM Comput. Surveys, 30, 1998, pp. 412-458.
to the sink. For k = 1, ..., K, the arc from node N + k to node
[21] A. Ahuja, T. Magnanti and J. Orlin, Network Flows: Theory, Algorithms, t has capacity ck , since yk = 1. The remaining arcs have zero
and Applications. Prentice Hall, Englewood Cliffs, New Jersey, 1993. capacity, since yk = 0 for k = K +1, ..., M. Set the flows x(N+k)t
[22] G. Cornu` jols, “Valid Inequalities for Mixed Integer Linear Programs,”
e equal to |Ak | for k = 1, ..., K, and set the flows x(N+k)t equal to 0
Mathematical Programming B, vol. 112, 2006.
[23] E. Craparo, Cooperative Exploration under Communication Constraints, for k = K + 1, ..., M. By definition of ck and by our assumption
Doctoral thesis, Massachusetts Institute of Technology, Cambridge, MA, that all assigned regular nodes achieve throughput at least τmin
USA, September 2008. in the original solution, constraint 4f is satisfied. Constraint 4d
[24] D. Bertsimas and R. Weismantel, Optimization over Integers, Dynamic
Ideas, 2005, p. 88. is also now satisfied for all arcs. Furthermore, constraint 4c is
[25] G. Nemhauser and L. Wolsey, “Maximizing submodular set functions: satisfied for nodes N + 1, ..., N + M; node N + k has |Ak | units
Formulations and Analysis of Algorithms,” Studies on Graphs and of flow incoming and outgoing for k = 1, ..., K, and zero units
Discrete Programming, P. Hansen, ed., 1981, pp. 279-301.
[26] A. S. Asratian, T. M. J. Denley and R. H¨ ggkvist, Bipartite Graphs and
a of flow incoming and outgoing for k = K + 1, ..., M. Thus, all
their Applications. Cambridge University Press, 1998. constraints are satisfied, and the objective value of this solution
is equal to the number of regular nodes that were assigned in
the original solution, i.e., ∑K |Ak |.
A PPENDIX A k=1
Now, assume that we are given a feasible solution to the
P ROOF OF T HEOREM 1 network design problem and its associated MILP. Furthermore,
Proof: Given a feasible solution to the original problem assume that all flows in this solution are integer. (By virtue
(i.e., a solution in which K mobile backbone nodes are placed of the total unimodularity of the constraint matrix in the
anywhere in the plane, each regular node is assigned to at maximum flow problem and the integrality of the right hand
most one mobile backbone node, and every assigned regular side vector, all basic feasible solutions of the linear program
node achieves throughput at least τmin ), a feasible solution to induced by any feasible choice of the y vector are integer.)
the corresponding network design problem and its associated Since all flows are integer, flows along arcs terminating in
Network Design MILP (4) can be constructed as follows: node sets N and M are either 0 or 1. Again, assume without

12

1
loss of generality that the nodes are labeled such that y1 , ..., yK N+1
are equal to 1, and yK+1 , ..., yM are equal to 0. Thus, no flow
c1=1
traverses nodes N + K + 1, ..., N + M. 2

To construct a feasible solution to the mobile backbone
s N+2 c2=3
network optimization problem, first place mobile backbone t
3
nodes at the K locations l1 , ..., lK corresponding to nodes
c3=2
N + 1, ..., N + K in the network design graph. Next, for each
mobile backbone node k = 1, ..., K, let Ak be the set of regular N+3
nodes for which xi(N+k) = 1. Assign the regular nodes in set N=4
Ak to regular node k. Note that each regular node is assigned (a) A graph over which a maximum flow problem can be
to at most one mobile backbone node, since at most one unit formulated. Unlabeled arc capacities are equal to one.
of flow can traverse each node 1, ..., N. Furthermore, note that
1
each assigned regular node achieves throughput at least τmin by N+1
definition of the arc capacity ck . The number of regular nodes 2

assigned in this solution is equal to ∑K |Ak |. For each mobile
k=1
backbone node k, |Ak | is equal to the flow traversing nodes
N+2
N + k. Since no flow traverses nodes N + K + 1, ..., N + M, the 3

total flow through the graph is equal to ∑K |Ak |. Thus, the
k=1
two solutions have the same objective value.
We have shown that for every feasible solution to the orig- N=4
N+3

inal problem, there is a feasible solution to the corresponding
(b) A bipartite matching problem that is equivalent to the
network design problem and its associated MILP with the maximum flow problem above.
same objective value. Thus, although the restriction of mobile
backbone nodes to the 1-center locations of regular nodes Fig. 7. An example of conversion from a maximum flow
may exclude an uncountable number of optimal solutions to problem to an equivalent bipartite matching problem, for
the original placement and assignment problem, it does not N = 4, M = 3.
exclude all optimal solutions. Furthermore, we have shown
that for every integer solution to the network design problem
and its associated MILP (including the optimal solution), there original node N + i. Thus, each node N + i in the maximum
is a feasible solution to the original problem with the same flow problem becomes a set of nodes N + i in the bipartite
objective value. Thus, the MILP formulation cannot produce matching problem, and the cardinality of this set is equal to
a solution with a higher objective value than is possible in ci . An example of this reformulation is shown in Fig. 7.
the original problem. Therefore, the MILP formulation can be For any feasible flow in the original graph, there is a corre-
used to obtain a solution to the original problem that achieves sponding matching in the bipartite graph with cardinality equal
the same optimal objective value that was possible in the to the volume of flow; likewise, for any feasible matching in
original problem. the bipartite graph, there is a corresponding flow of volume
equal to the cardinality of the matching. Therefore, the volume
of the maximum flow through the original graph is equal to
A PPENDIX B the cardinality of a maximum matching in the bipartite graph.
P ROOF OF L EMMA 1 The graphs expressing the relation in Eq. (6) are shown in
Proof: We begin by restating the submodularity condition the top row of Fig. 8: the sum of the maximum flows through
as follows: the left two graphs must be less than or equal to the sum of
the maximum flows through the right two graphs.
f ∗ (S ∪ {i, j}) + f ∗ (S) ≤ f ∗ (S ∪ {i}) + f ∗ (S ∪ { j}) (6) Converting these maximum flow problems into their equiv-
alent bipartite matching problems, we obtain the condition
where f ∗ is the maximum flow through G, as a function of the
that the sum of the cardinalities of maximum matchings in
set of selected arcs. Next, we note that for a fixed selection
bipartite graphs G1 and G2 in Fig. 8 is at most the sum of the
of arcs S, the problem of finding the maximum flow through
cardinalities of maximum matchings in G3 and G4 .
G can be expressed as an equivalent maximum matching
Consider a maximum matching M1 in graph G1 , and denote
problem on a bipartite graph with node sets L and R2 . This is
its cardinality by Ns . This means that Ns nodes from set S are
accomplished as follows: node set L in the bipartite matching
covered by matching M1 . Note that M1 is a feasible matching
problem is simply node set N in the maximum flow problem.
for G2 as well, since all arcs in G1 are also present in G2 .
Node set R is derived from node set M in the maximum flow
It is a property of bipartite graphs that if a matching Q is
problem, with one modification: if the arc from node N +i ∈ M
feasible for a graph H, then there exists a maximum matching
to t has outgoing capacity ci , then R contains ci copies of node
Q∗ in H such that all of the nodes covered by Q are also
N +i, each of which is connected to the same nodes in L as the
covered by Q∗ [26]. Denote such a maximum matching for
2. A set of edges in a graph is a matching if no two edges share a common matching M1 in graph G2 by M2 , and note that Ns nodes from
end node. A maximum matching is a matching of maximum cardinality [26]. set S are covered by M2 . Denote the number of nodes covered

13

1 1 1 1
c1 c1 c1 c1
2 S 2 S 2 S 2 S

…

…

…

…
… …

… …

… …

… …
…
…

…

…
cT
ci

…

…

…

…
ci
cj cj
i i
N N j N N j

1 1 1 1

2 2 2 2
S S S S

…

…

…
…

…

…

…
… …

… …

… …
… …
…

…

…

…
i |i|=ci i |i|=ci
N N j |j|=cj N N j |j|=cj
|S|=∑k=1,…,T ck
G1 G2 G3 G4

Fig. 8. Schematic representation of the graphs involved in the proof of the submodularity condition. The top graphs
relate to the original maximum ﬂow problem, while the bottom graphs are their equivalent reformulations in the bipartite
matching problem. For clarity, not all arcs are shown.

by M2 in node sets i and j by Ni and N j , respectively. Then,
the total cardinality of these maximum matchings for graphs
G1 and G2 is equal to 2Ns + Ni + N j .
Now consider the matching obtained by removing the edges
incident to node set j from M2 . Note that this matching is
feasible for graph G3 , and its cardinality is Ns + Ni . Likewise,
the matching obtained by removing the edges incident to node
set i from M2 is feasible for graph G4 , and its cardinality is
Ns +N j . Since these matchings are feasible (but not necessarily
optimal) for G3 and G4 , the sum of the cardinalities of maxi-
mum matchings for these graphs must be at least 2Ns +Ni +N j .
This establishes the submodularity property for the matching
problem as well as for the maximum ﬂow problem.

14

Emily M. Craparo is a graduate of the De- Eytan Modiano Eytan Modiano received his
partment of Aeronautics and Astronautics at B.S. degree in Electrical Engineering and Com-
the Massachusetts Institute of Technology (MIT). puter Science from the University of Connecticut
She received the S.B., S.M. and Ph.D. degrees at Storrs in 1986 and his M.S. and PhD degrees,
in Aeronautics and Astronautics in 2002, 2004 both in Electrical Engineering, from the Univer-
and 2008, respectively. Her research interests sity of Maryland, College Park, MD, in 1989 and
include design and management of complex 1992 respectively. He was a Naval Research
systems, particularly networked systems that Laboratory Fellow between 1987 and 1992 and
operate in information-rich environments. She a National Research Council Post Doctoral Fel-
is currently a National Research Council Post low during 1992-1993. Between 1993 and 1999
Doctoral Fellow. he was with MIT Lincoln Laboratory where he
was the project leader for MIT Lincoln Laboratory’s Next Generation
Internet (NGI) project. Since 1999 he has been on the faculty at MIT;
where he is an Associate Professor. His research is on communication
networks and protocols with emphasis on satellite, wireless, and optical
networks. He is the co-recipient of the Sigmetrics 2006 Best paper
award for the paper ”Maximizing Throughput in Wireless Networks
via Gossiping,” and the Wiopt 2005 best student paper award for the
paper ”Minimum Energy Transmission Scheduling Subject to Deadline
Constraints.”
He is currently an Associate Editor for IEEE/ACM Transactions on
Networking. He had served as Associate Editor for IEEE Transactions
on Information Theory, and as guest editor for IEEE JSAC special issue
on WDM network architectures; the Computer Networks Journal special
issue on Broadband Internet Access; the Journal of Communications
and Networks special issue on Wireless Ad-Hoc Networks; and for IEEE
Journal of Lightwave Technology special issue on Optical Networks. He
was the Technical Program co-chair for IEEE Wiopt 2006, IEEE Infocom
2007, and ACM MobiHoc 2007.

Jonathan P. How is the Richard Cockburn
Maclaurin Professor of Aeronautics and Astro-
nautics in the Department of Aeronautics and
Astronautics at the Massachusetts Institute of
Technology (MIT). He received a B.A.Sc. from
the University of Toronto in 1987 and his S.M.
and Ph.D. in Aeronautics and Astronautics from
MIT in 1990 and 1993, respectively. He then
studied for two years at MIT as a postdoctoral
associate for the Middeck Active Control Ex-
periment (MACE) that ﬂew on-board the Space
Shuttle Endeavour in March 1995. Prior to joining MIT in 2000, he
was an Assistant Professor in the Department of Aeronautics and
Astronautics at Stanford University. Professor How was the planning and
control lead for the MIT DARPA Urban Challenge team that placed fourth
in the 2007 race at Victorville, CA. Current research interests include
robust coordination and control of autonomous vehicles in dynamic
uncertain environments.
He is currently an Associate Editor for the IEEE Control Systems
Magazine and has served as an Associate Editor for the AIAA Journal
of Guidance, Control, and Dynamics. He was the recipient of the 2002
Institute of Navigation Burka Award, a recipient of a Boeing Special
Invention award in 2008, was the Raymond L. Bisplinghoff Fellow for
MIT Aeronautics/Astronautics Department (2006-2008), is an Associate
Fellow of AIAA, and a Senior Member of IEEE.

Throughput optimization in

More Related Content

What's hot (16)

Viewers also liked (6)

Similar to Throughput optimization in (20)

More from ingenioustech (20)

Recently uploaded (20)

Throughput optimization in