2014 - Vulnerability Assesment of Spatial Network - Models and Solutions
2014 - Vulnerability Assesment of Spatial Network - Models and Solutions
2014 - Vulnerability Assesment of Spatial Network - Models and Solutions
1. Introduction
Shortest path problems correspond to an old and very known class of problems
in combinatorial optimization. A variant of one of these basic problem consists
on analyzing the effects of removing arcs from a network. In [14] the problem of
removing k arcs that cause the greatest decrease in the maximum flow from a
source to a sink in a planar network is studied. This problem is a special case of
a broad class of network optimization problems known as interdiction problems.
Applied to the shortest s, t-path problem, the interdiction problem can be de-
fined in the following way. Given a graph G = (V, E) with a non-negative length
function on its arcs l : E R and two terminals s, t V , the goal is to destroy
all (or the best) paths from s to t in G by optimally eliminating as many arcs
of A as possible (usually respecting a so-called interdiction budget). Interdiction
problems are often used to measure the robustness of solutions of network op-
timization problems. In [9] several versions of these problems are studied; they
consider the case of total limited interdiction when a fixed number of k arcs can
be removed, and node-wise limited interdiction (for each node v V a fixed
number k(v) of out-going arcs can be removed). For a complete survey on early
interdiction problems with different underlying network properties the reader is
referred to [2]. For a more general discussion regarding network vulnerability
approaches we suggest to see [10].
Based on a well-known network interdiction model we formulate a framework
of combinatorial optimization problems whose solutions can be used for assess-
ing the vulnerability of spatial networks in the case of disruptions. We design
2
Let us assume that X is a finite set of points x in R2 and that R can take values
in R which is a finite subset of R>0 . Given a radius R R and a discrete set
X , we are interested in knowing what is the maximum length of a shortest
s, t-path across all possible locations x X of the disruption disk (x, R).
Knowing is threefold: (i) It tells us how severe a disruption can be by
comparing the value of with respect to `; in other words, the increase of the
transportation time between s and t induced by a failure located in the worst
location x = argxX {}. (ii) From the tactical point of view, preventive actions
can be taken in order to reduce the chances that a failure can be produced at x
3
(a) No failure disk. (b) A disk (x1 , R1 ). (c) Two disks (x1 , R1 ) and
(x2 , R2 ).
Fig. 1. Example of a network G = (V, E), nodes s and t, a region X and different
interdiction cases.
or the edges Ex can be reinforced to increase their reliability. And (iii) we can
know whether the network is so vulnerable that s and t might be disconnected,
which can be verify if = .
yx + ze 1, e E | d(e, x) R, x X (YZ.1)
X
yx ze 0, e E. (YZ.2)
xX |d(e,x)>R
Constraints (YZ.1) and (YZ.2) state that, for any x X , an edge e has to be
operative (ze = 1) if is not reached by the disruption disk (x, R). Since a single
disruption disk affects the network, we have that
X
yx = 1. (YZ.3)
xX
Using (YZ.1) and (YZ.2), for a given x X the edge costs lex can be written as
lex = le ze + (1 ze )M, e E. Hence, the MCSFSPP is as follows
n o
= max `x | (YZ.1), (YZ.2), (YZ.3) and (y, z) {0, 1}m+|X | . ()
xX
Problem (), as it is, is non-linear. To linearize it, we will convert the max min
objective into a pure max one; to do so, let us consider the dual of (`x ), which
is given by
`x = max {t s | j i lij zij + (1 zij )M, e : {i, j} E and Rn } . ()
= max t s (MCSF.1)
s.t (YZ.1), (YZ.2) and (YZ.3) (MCSF.2)
j i lij zij + (1 zij )M, e : {i, j} E (MCSF.3)
m+|X | n
(y, z) {0, 1} and R . (MCSF.4)
5
Note that in our approach we assume that (x, R) can be located in any
point x X without any stochastic characterization. That is, any point x X
is likely to host the center of the failure.
In the proposed setting we assume that if an edge e is disrupted by at least
one failure disk (x, R), then it becomes inoperative. However, one can easily
extend this to a more general case by defining a coefficient de 0 e E
representing the delay on edge e in case of interdiction (in our setting de = M
e E). The MCSFSPP can be redefined by replacing constraint (MCSF.3)
with
By solving (B) we can know how vulnerable the network is if the interdictors are
able to optimally use their resources to disrupt it. Models as the one presented
in [8, 7] are particular cases of (B) in which X coincides with the midpoint of
every edge e E and R = ( being infinitesimally small).
Minimum Simultaneity for Complete Vulnerability: Critical k One
might be interested in knowing the minimum number of failures (the critical
k or k c ) that should occur simultaneously in order to have at least one set
(x1 , R), . . . , (xk , R) that damages the network so that s and t cannot be con-
nected anymore or the shortest length between them is greater than a threshold
.
The value k c and the corresponding collection {x1 , . . . , xkc } will enable a
decision maker to perform more general preventive actions to endure the network
not in a single but a in several areas. In many practical contexts, the possibility
of multiple and synchronized failures might be the rule, so knowing k c might
play a strategical role. Clearly, for a given R, the larger k c is the more robust
the network is. Mathematically, one can formulate the search for k c as
kc = min {k | (YZ.1), (YZ.2k), (YZ.3k), (MCSF.3), (MCSF.4) ,
s t and k Z0 } (kc )
3. Computational Results
such that an edge is added if lij (euclidean distance) satisfies lij / n and the
planarity of the network is still preserved; (iv) the set X is created by randomly
located K points within the are area defined by points (x1 , y1 ), (x2 , y1 ), (x1 , y2 )
and (x2 , y2 ).
For experiments we have considered n {500, 1000}, = 1.5, = 1.6,
(x1 , x2 , y1 , y2 ) = (0.3, 0.7, 0.0, 1.0) (X1 ) and (x1 , x2 , y1 , y2 ) = (0.1, 0.9, 0.1, 0.9)
(X2 ), and K = 100.
In Figure 2(a) it is shown an example of an instance with 500 nodes and X
contained in (0.3, 0.0), (0.7, 0.0), (0.3, 1.0) and (0.7, 1.0).
In the case of groups US and Bangladesh we consider the geographical co-
ordinates of the most populated cities in each case [see 13] to define the set V .
Then, we used an approximation of their highway and interurban road system
with the information available in [6] to approximate the set of edges E. The set
X is created by randomly located K points within the are area defined by points
(x1 , y1 ), (x2 , y1 ), (x1 , y2 ) and (x2 , y2 ). In Figures 2(b) and 2(c) we show the
networks used to generate the instances US and Bangladesh respectively. In the
case of US, the area X is given by placing 100 points in the so-called south area.
With this we intend to represent possible cases of failure produced by hurricanes
and other natural disasters. For the Bangladesh instances, we have created X
by placing 100 points in squared area in the very center that covers around the
15% of the total area.
In the case of instances ND, nodes s and t are selected as those with the
longest euclidean distance. In the case of instances US we have used s
{NY:New York, CH:Chicago} and t {LA:Los ngeles, HS:Houston}; likewise,
in the case of instances Bangladesh we have used s = Rajshahi and t = Silhat.
Solver Setting Models (MCSF.1)-(MCSF.4), (MCMF) and (k c ) were solved
using CPLEX 12.5 (all CPLEX parameters were set to their default values).
The experiments were performed on a Intel Core i7-3610QM machine with 8 GB
RAM.
8
n = 500 n = 1000
X1 X2 X1 X2
R % t[sec] R % t[sec] R % t[sec] R % t[sec]
0.01 2.17 38.92 0.01 0.00 32.93 0.01 2.48 115.39 0.01 0.14 144.66
0.02 3.93 46.46 0.02 1.64 36.83 0.02 2.93 153.58 0.02 0.69 215.16
0.03 3.93 65.63 0.03 1.64 49.73 0.03 5.38 235.95 0.03 1.52 240.88
0.04 5.15 80.79 0.04 1.64 63.06 0.04 7.31 258.85 0.04 1.52 265.00
0.05 5.15 103.01 0.05 1.64 79.67 0.05 7.17 395.46 0.05 1.52 259.80
0.10 5.15 97.83 0.10 1.64 112.76 0.10 8.69 917.78 0.10 3.87 373.61
0.15 - 53.70 0.15 10.64 111.65 0.15 9.93 587.27 0.15 6.19 843.45
Table 1. Solutions for the MCSFSPP considering different values of R (Instances ND)
500 1000
X1 X2 X1 X2
R k % t[sec] R k % t[sec] R k % t[sec] R k % t[sec]
0.01 1 2.17 24.15 0.01 1 0.00 22.51 0.01 1 2.48 88.16 0.01 1 0.01 88.64
2 3.93 25.07 2 0.00 21.96 2 3.03 94.80 2 0.09 96.35
3 4.74 25.02 3 0.00 22.11 3 3.03 93.90 3 1.63 93.57
4 5.56 24.87 4 0.00 22.21 4 3.03 95.08 4 1.63 95.08
5 6.79 24.52 5 0.00 22.14 5 3.03 94.15 5 1.63 94.4
0.1 1 5.15 59.94 0.1 1 1.64 69.61 0.1 1 8.68 889.61 0.1 1 3.87 222.36
2 - 26.43 2 13.82 142.93 2 17.89 3448.93 2 8.16 1012.06
3 - 98.05 3 13.92 297.95 3 21.24 75319.9 3 12.45 5434.01
4 - 526.47 4 - 63.01 4 28.14 26123.3 4 21.99 13130.3
5 - 147.39 5 - 149.54 5 - 16576.5 5 - 742.63
Table 2. Solutions for the MCMFSPP considering different values of R and k (In-
stances ND)
n X R kc t[sec] n X R kc t[sec]
500 X1 0.10 2 170.76 1000 X1 0.10 5 628.76
0.15 1 43.09 0.15 2 805.32
X2 0.03 10 39.05 X2 0.02 19 219.38
0.04 9 109.70 0.03 13 621.60
0.05 7 161.06 0.04 11 933.85
0.10 4 274.20 0.05 8 1759.32
0.15 3 162.47 0.10 5 1277.24
0.15 3 1214.14
Table 3. Solutions of (kc ) considering different values of R (Instances ND)
From the algorithmic point of view, we can notice in Tables 1, 2 and 3 that
the search for an alternative path in a disrupted network is not for free. In all
cases we see an increase of the algorithmic effort (time) needed to find such a
path (if exists). This is due to the high combinatorial nature of the problem
when more edges are subject to be interdicted (when R increases and/or when
k is either greater than 1 or when it is a variable).
In the case of USA Instances, we report in Table 4 results of the MCSFSPP
considering different pairs of s and t and different values of R. In this case, we
can see that different combinations of s and t yield to different levels of vulner-
ability in the system. For instance, the network is considerably more vulnerable
when it is intended to host a path from Chicago to Los ngeles than when the
path should be established from Chicago to Houston. This is due to the fact
that, in our instance, the system of roads connecting the north of the Midwest
with the south of the West Coast is composed by relatively few elements. Hence,
a single disruption disk (that is optimally placed) is enough to interrupt the
communication between the cities. In this case the values of % are particu-
10
larly important from the tactic point of view; if it is up to the decision maker to
decide where to establish both the source and the target of the transportation
system, then it might preferable to have New York - Houston than, for instance,
Chicago - Los ngeles. However, this analysis is valid only when a single failure
occurs. For an approximate equivalence to real distances in kilometers, R should
be multiply by 1700.
In Figure 3(a) we show the solution of the shortest path problem between
New York and Houston when there is no disruption. In Figure 3(b) is shown the
solution of the MCMFSPP when 5 disruption disks with R = 0.01 are optimally
located. In Figure 3(b) is shown the solution of the MCMFSPP with k = 1 and
R = 0.10. These figures show how different the optimal s, t-paths can be when
the network is disrupted by failures of different magnitude.
Fig. 3. Solutions for the MCMFSPP for different k and R (Instances USA)
R = 0.01 R = 0.1
R % t[sec] R kc t[sec] k % t[sec] k % t[sec]
0.01 3.76 8.72 0.05 3 10.06 1 3.76 7.69 1 24.17 8.11
0.02 2.93 7.44 0.10 2 28.52 2 5.48 8.75 2 - 38.28
0.03 3.76 8.99 0.15 1 24.01 3 5.48 8.42 3 - 26.75
0.04 3.76 9.53 4 5.48 8.41 4 - 28.32
0.05 4.78 19.03 5 5.48 7.64 5 - 10.55
0.10 24.17 36.97
0.15 - 11.19
Table 5. Solutions for MCSFSPP, MCSFMPP and kc , s =Rajshahi and t =Silhat
(Instances Bangladesh)
0.05. For greater values, the network can be dramatically damaged. This later
observation is reinforced by the results reported in columns 4-6 in the same table:
a critical k can be found only if R 0.05. When looking at the results of the
MCSFMPP (columns 7-9 for R = 0.01 and 10-12 for R = 0.1) we can see that
the network resists well (% 5%) several failures with R = 0.01; however,
if R = 0.1 then the network collapses even if k = 2.