Omega 31 (2003) 269–273
www.elsevier.com/locate/dsw
On the Maximum Bene"t Chinese Postman Problem
W.L. Pearn
a;∗, K.H. Wang
baDepartment of Industrial Engineering and Management, National Chiao Tung University, 1001 Ta Hsueh Road, Hsin Chu 30050, Taiwan
bDepartment of Applied Mathematics, National Chung Hsing University, Taichung, Taiwan Received 2 March 2001; accepted 28 March 2003
Abstract
The Maximum Bene"t Chinese Postman Problem (MBCPP) is a practical generalization of the classical Chinese Postman Problem (CPP), which has many real-world applications. In this paper, we consider the MBCPP on undirected networks, and show that the MBCPP is more complex than the Rural Postman Problem (RPP). We present a su6cient condition for the MBCPP solution to cover the whole network, and provide an upper bound. Based on the upper bound, we propose an e6cient solution procedure to solve the MBCPP approximately. The proposed algorithm applies the minimal spanning tree and the minimal-cost matching algorithms, which performs well on problems satisfying the su6cient condition.
? 2003 Elsevier Science Ltd. All rights reserved.
Keywords: Chinese Postman Problem; Rural Postman Problem; Upper bound
1. Introduction
There are many generalizations of the well-known Chi-nese Postman Problem (CPP). Examples include the
Ru-ral Postman Problem (RPP) [1], the Hierarchical Postman
Problem (HPP) [2], the k-person Chinese Postman Problem
(k-CPP) [5], the Capacitated Arc Routing Problem (CARP)
[3], and many others.
The Maximum Bene"t Chinese Postman Problem (MBCPP) is another interesting generalization of the CPP, in which each edge on the network is associated with a service cost for the traversal with service, a deadhead cost for the traversal with no service, and a set of bene"ts. Each time an edge is traversed a bene"t is generated. The objec-tive of the MBCPP is to "nd a postman tour traversing a selected set of edges with the total net bene"t maximized. Such a generalization re?ects the real-world situations more closely than the classical CPP. Applications directly related
∗Corresponding author. Tel.: +886-35-714261; fax:
+886-35-722392.
E-mail address:roller@cc.nctu.edu.tw(W.L. Pearn).
to the MBCPP include routing of street sweepers, snow-plows, spraying roads with salt, inspection of streets for maintenance, and reading of electric meters.
The problem may be brie?y de"ned as follows. Given an undirected network G(V; E), with V representing the set of nodes, and E representing the set of edges. For each edge
(i; j) ∈ E, we are given a non-negative service cost cs
ij for
the edge traversal with service, and a non-negative deadhead
cost cd
ij for the edge traversal with no service, which we
expect cs
ij¿ cdij. We are also given a set of non-negative
bene"ts bijrij from node i to node j for the rijth traversal,
where rij= 1; 2; : : : ; nij.
To re?ect real situations more closely, we assume that
the bene"t bijrij is non-increasing in rij. The net cost of the
rijth traversal of the edge (i; j), therefore, can be explicitly
expressed as cijrij= csij− bijrijfor rij= 1; 2; : : : ; nij, with the
index nij= max{rij| bijrij¡ cdij}, and bij(nij+1)= 0. The net
cost for rij¿ nij + 1, is therefore cijrij = cdij. That is, for
the traversal of the deadhead edges, no bene"t is generated. Then, the MBCPP is to "nd a postman tour, starting from the depot, traversing a set of edges in E, and returning to the same depot with total net cost minimized (or total net bene"t is maximized).
0305-0483/03/$ - see front matter ? 2003 Elsevier Science Ltd. All rights reserved. doi:10.1016/S0305-0483(03)00051-3
Malandraki and Daskin [4] have investigated the MBCPP on directed networks. Their cost/bene"t structure, however, is more restricted than one we consider here.
2. Problem complexity
In the following, we present a linear transformation, which converts the Rural Postman Problem (RPP) into a special case of the MBCPP. Hence, the MBCPP is more complex than the RPP.
2.1. The transformation
Consider a (totally) undirected RPP network G(V; E; ER)
with ER representing the set of required edges. For each
edge (i; j) with distance dij, (1) if (i; j) ∈ ER, de"ne the
service cost and the deadhead cost as cs
ij= cdij= dij, bene"t
bij1= 2dij+ , bij2= dij− , and bijrij= 0, for rij¿ 3, where
is a su6ciently small number less than one, and (2) if
(i; j) ∈ E −ER, de"ne the service cost and the deadhead cost
as cs
ij=cdij=dij, bene"t bij1=dij−=(4m), bij2=dij−=(2m),
and bijrij= 0 for rij¿ 3, where m = |E| = the number of
edges in E.
Clearly, for (i; j) ∈ ER the net cost can be calculated as
cij1= −(dij+ ), cij2= , and cijrij= dij, for rij¿ 3. On the
other hand, for (i; j) ∈ E −ER, the net cost can be calculated
as cij1= =(4m), cij2= =(2m), and cijrij= dij for rij¿ 3.
If we denote the transformed network as G(V; E; E
R), then,
we can show the following theorem.
Theorem. The optimal RPP solution over the original
net-work G(V; E; ER) is equivalent to that of the MBCPP over
the transformed network G(V; E; E
R) with equal solution
values.
We "rst note that minimizing the total net cost cijrij =
cs
ij− bijrij is obviously equivalent to maximizing the total
net bene"t −cijrij = bijrij− csij. Let |ER| be the number of
edges in ER. If the MBCPP solution over the transformed
network G(V; E; E
R) traverses all the edges in ER exactly
twice without traversing any edge in E−ER, then the total net
bene"t is(i;j)d(i; j)+|ER| for (i; j) ∈ ER. If the MBCPP
solution over the transformed network G(V; E; E
R) traverses
all the edges in ER exactly twice without traversing any
edge in E − ER, then the total net bene"t is(i;j) d(i; j) +
|ER| −|ER| =(i;j)d(i; j) for (i; j) ∈ ER. Since traversing
one edge in ER for the second time reduces the total net
bene"ts by , any MBCPP solution would attempt not to
traverse the edges in ERfor more than once.
Further, traversing over any edge in E − ERwould only
reduce the total net bene"t by either =(4m) for the "rst time traversal, or =(2m) for the second time traversal, any MBCPP solution would attempt to minimize the cost
in-curred from traversing the edges in E − ERin order to
max-imize the total net bene"t. Noting that the sum of the edge
cost (corresponding to the "rst time traversal) from ERis a
constant, it is then clear that the optimal MBCPP solution
over the transformed network G(V; E; E
R) is an Euler
cy-cle covering the edges in ERwith minimal distance (cost),
hence is equivalent to the optimal RPP solution over the
original network G(V; E; ER).
3. A special case CPP
Consider a special case of MBCPP with the traversal cost
of edge (i; j) de"ned as cs
ij= cdij= dij (length of the edge
(i; j)), and bene"t de"ned as bij1= dij+ c0; bij2= dij− c1,
and bijrij = 0, for rij¿ 3, where c0; c1 are constants with
c0¿ c1¿ 0. The net bene"t for traversing each edge for the
"rst time is c0, and the second time is −c1.
If the postman traverses each edge on the network G(V; E) exactly once, then the total net bene"t can be calculated as
|E|c0. On the other hand, if the postman traverses all the
edges in the network exactly twice, then the total net bene"t
is |E|(c0−c1). Since traversing any edge in E for the second
time reduces the total net bene"t, the MBCPP solution would attempt not to traverse the edges in E for the second time. Consequently, minimizing the cost from traversing the edges for the second time would certainly maximize the total net bene"t. As a result, the MBCPP with the de"ned cost/bene"t structure reduces to the classical CPP.
4. Solution properties
We note that in the special case of MBCPP presented
above, the condition of bij1+bij2¿ cij1+cij2is satis"ed for
all edges. Such condition is su6cient for the MBCPP solu-tion to cover all the edges on G(V; E), but not vice versa. In fact, if the MBCPP solution does not cover the whole network, then adding two copies of the edges currently not in the solution would certainly increase the total net bene"t.
If the condition of bij1+ bij2¿ cij1+ cij2is not satis"ed,
the MBCPP solution may or may not cover the whole
net-work, as shown in the following examples. Fig.1depicts an
MBCPP with solution covering the whole network. Fig.2
depicts an MBCPP with solution not covering the whole
net-work. In both cases, the condition of bij1+ bij2¿ cij1+ cij2
is not satis"ed for some edges.
For the special case considered in Malandraki and Daskin
[4], the condition assumed (with bene"t bij2= 0 for the
second time traversal) does not guarantee the MBCPP so-lution to cover all the edges in the network (see Example
2). Hence, the MBCPP does not reduce to the CPP in that
setting.
Example 1. Consider the MBCPP depicted in Fig.1with
six nodes and six edges. The edge traversal cost is de"ned as cs
depot
4
5
6
2
3
1
Fig. 1. An MBCPP example with solution covering the whole network.
depot
5
6
2
1
3
4
7
Fig. 2. An MBCPP example with solution not covering the whole network.
dij, rij¿ 3, for all (i; j) ∈ {(2; 3); (3; 4); (4; 5); (5; 6); (6; 2)}
except for the edge (1, 2) with edge traversal cost de"ned as cs
12= cd12= 1, and bene"t b12rij= 0 for all rij. The optimal
solution covers the whole network with total net bene"t 8.
We note that the condition bij1+ bij2¿ cij1+ cij2 is not
satis"ed for edge (1, 2).
Example 2. Consider the MBCPP depicted in Fig.2with
seven nodes and eight edges. The edge traversal cost is de-"ned as cs
ij= cdij = 1, and bene"t bij1= 2; bij2= 0, for
(i; j) ∈ {(2; 3); (3; 4); (4; 5); (5; 6); (6; 2)}, and bij1=2; bij2=
1, for edge (1, 2). For edges (4, 7) and (5, 7), the traversal
cost is cs
12= cd12= 1, and the bene"t is bijrij= 5=4 for rij¡ 2,
and bijrij= 0 for rij¿ 2. The optimal postman tour is (1, 2,
3, 4, 5, 6, 2, 1), with a total net bene"t of 6, while the edges (4, 7) and (5, 7) are not traversed in the postman tour.
We note that the condition of bij1+ bij2¿ cij1+ cij2is
not satis"ed for edges (4, 7) and (5, 7) in Example2. If all
the edges in the network must be covered, then the postman tour becomes (1, 2, 3, 4, 7, 5, 4, 5, 6, 2, 1), with a total net bene"t of 6 + 1=4 + 1=4 − 1 = 6 − 1=2. The total net bene"t obtained in this case, is obviously not maximal.
4.1. An upper bound
If the postman only services the edges (traversals) with
positive net bene"t, bijrij− csij¿ 0, without traversing any
other edges with traversal yielding bijrij− cijs ¡ 0, then the
total net bene"t obtained is maximal. Those traversals may not form a complete postman tour, but the total net bene"t generated certainly provides an upper bound on the solution. It is straightforward to verify that for the MBCPP described
in Example1, an upper bound on the total net bene"t can
be found as 10. For the MBCPP described in Example2, an
upper bound can be found as 6 + 1=2.
To form a complete postman tour (Euler cycle), additional edge traversals may be required, which results in a reduc-tion on the total net bene"t. If such net bene"t reducreduc-tion is signi"cant, the MBCPP solution may choose not to service
some of those edge traversals with bijrij− csij¿ 0, in order
to maximizes the total net bene"t. This is true, in particular, for cases where those edges (traversals) with positive net bene"t do not form a connected network.
5. A solution algorithm
Since the MBCPP is more complex than the RPP which is NP-complete, then it must be di6cult to solve the problem exactly. Based on the solution properties discussed above, we present an e6cient algorithm, which solves the MBCPP approximately. The algorithm expands the original network by replacing each edge with a set of edges of positive net bene"t. Minimal spanning tree (MST) and matching algo-rithms are then applied to generate a complete postman tour. 5.1. The algorithm
Step 1: (Network expansion) Replace each edge (i; j) by
a set of new edges with net cost cijrij= cijs − bijrij, where
rij= 1; 2; : : : ; qij, to qij= max{rij|cijrij= cijs − bijrij¡ 0}, to
obtain an expanded network G∗.
Step 2: (Minimal spanning tree) If G∗ is connected,
then proceed to Step 3. Otherwise, let G∗ be a set of
disconnected components {Ct}. De"ne the distance
be-tween every pair of components as D(Ct(i); Ct(j)) =
minx;y{spl(x; y)|x ∈ Ct(i); y ∈ Ct(i); }, where spl(x; y) is the
least-cost path over the network GT consisting of edges
with traversal cost cij(qij+1) for edges in G∗ and cijrij = cdij
for edges not in G∗. Let E
T be the minimum spanning tree
(MST) solution over GT. Note that if G∗is connected, then
ET= ?.
Step 3: (Minimal cost matching) Identify the set of
odd-degree nodes, S, on the G∗ ∪ E
T, and construct a
matching network GM on S with distance between the
nodes de"ned as the least-cost path over the network with
edge traversal cost cij(qij+1)for edges in G∗− ET, cij(qij+2)
for edges in G∗− E
Find the minimal-cost matching solution EMover GM. The
resulting network, G∗∪ E
T∪ EM, is the MBCPP solution.
Step 4: (Bene"t maximization) Find cycles with negative
net bene"t if they exist. Remove those cycles from G∗∪
ET∪EMif the removal does not separate the remaining graph
into disconnected components.
If we apply the algorithm to the two MBCPP examples described above, both optimal solutions can be obtained,
which is 8 for the problem in Example1, and 6 for the
prob-lem in Example2. Note that in Example2, the cycle (7, 5, 4)
with negative net bene"t −1=2 is found and removed from the solution (1, 2, 3, 4, 7, 5, 4, 5, 6, 2, 1) generated from Step 3. For problems with cost/bene"t satisfying the
su6-cient condition of bij1+ bij2¿ cij1+ cij2, the algorithm
pro-ceeds with Step 1 and Step 3 (without the MST segment), which generates a solution covering the whole network. Hence, the algorithm is expected to work well on problems with cost/bene"t satisfying the su6cient condition. In
par-ticular, if the expanded network G∗is an even graph, the
solution found is optimal. 5.2. Discussion
The cycles with negative net bene"t are removed from the
solution G∗∪ E
T∪ EMgenerated in Step 3, to maximize the
total net bene"t. Those cycles can be found by repeatedly
proceeding with (1) identifying a path in ET∪ EM, and (2)
"nding a path in G∗with minimal net bene"t, between the
two end nodes of the identi"ed path in ET∪ EM, to form
closed cycles.
In general, if the net bene"t reduction in traversing the
edges in ET∪ EMis signi"cant, then the improvement made
by removing the cycles (Step 4) would be signi"cant. The MBCPP solution, in that case, chooses not to service some of
the edges in G∗(with positive net bene"t) in order to
max-imize the total net bene"t. EKective strategies in identifying
paths in ET∪ EM, and choosing the corresponding paths in
G∗to form closed cycles, are essential to the maximization
of the total net bene"t, and will be investigated further. 6. Computational examples
Consider the MBCPP depicted in Fig.3with 15 nodes
and 26 edges. The depot is at node 1. Table 1 displays
the edge traversal costs for the edges. For simplicity of the
computation, we assume that cs
ij=cijd=cij. We proceed with
Step 1 of the algorithm to obtain the expanded network G∗,
with the edge traversal net bene"ts shown in Table2. Since
the expanded network G∗ is connected, we proceed with
Step 3 to identify the set of odd-degree nodes S on G∗by
checking Table2to obtain S ={6; 7; 11; 12; 13; 15}, and "nd
the minimal-cost matching solution. The network G∗∪ E
M
obtained in Step 3 is shown in Fig.4.
Following Step 4 we "nd three cycles with negative net bene"t, including the cycle (7, 4, 6, 4, 7) with a net bene"t
depot 1 2 3 4 5 6 7 8 10 9 11 12 15 13 14
Fig. 3. The 15-node example.
Table 1
Traversal costs of the edges on the network
Edge cij Edge cij
(1, 2) 18 (7, 9) 10 (1, 4) 15 (8, 10) 13 (2, 3) 16 (8, 9) 4 (2, 4) 7 (8, 11) 10 (3, 5) 11 (9, 11) 10 (3, 9) 15 (9, 15) 5 (4, 5) 17 (10, 12) 8 (4, 6) 18 (11, 12) 18 (4, 7) 8 (11, 14) 8 (4, 8) 9 (12, 13) 11 (5, 9) 9 (12, 14) 13 (6, 10) 11 (13, 14) 6 (7, 8) 10 (14, 15) 18 Table 2
The edge traversal net bene"ts of the edges on G∗
Edge bijrij− csij Edge bijrij− csij
(1, 2) 9, 7, 4, 1 (7, 9) 11, 8, 4 (1, 4) 8, 6, 4, 2 (8, 10) 14, 7, 3 (2, 3) 9, 6, 3, 2 (8, 9) 4, 3 (2, 4) 3, 2 (8, 11) 9, 4, 1 (3, 5) 6, 3, 1 (9, 11) 8, 5, 2 (3, 9) 12, 10, 5 (9, 15) 5, 3 (4, 5) 11, 7, 4, 2 (10, 12) 3, 1 (4, 6) 10, 6, 3, 2 (11, 12) 15, 10, 5 (4, 7) 7, 3, 2 (11, 14) 7, 4 (4, 8) 6, 3, 1 (12, 13) 11, 8, 3 (5, 9) 10, 8, 4 (12, 14) 9, 6, 2 (6, 10) 10, 7, 3 (13, 14) 6, 3 (7, 8) 6, 5, 2 (14, 15) 3, 2, 1
1 2 3 4 5 6 7 8 10 9 11 12 15 13 14 depot Fig. 4. The network G∗∪ EM.
Table 3
Number of edge traversals in the solution
Edge Traversal Edge Traversal
(1, 2) 4 (7, 9) 3 (1, 4) 4 (8, 10) 3 (2, 3) 4 (8, 9) 2 (2, 4) 2 (8, 11) 3 (3, 5) 3 (9, 11) 2 (3, 9) 3 (9, 15) 1 (4, 5) 3 (10, 12) 2 (4, 6) 3 (11, 12) 3 (4, 7) 3 (11, 14) 2 (4, 8) 3 (12, 13) 2 (5, 9) 3 (12, 14) 3 (6, 10) 3 (13, 14) 2 (7, 8) 3 (14, 15) 3
of −22, the cycle (12, 13, 12) with a net bene"t of −8, and the cycle (11, 9, 15, 9, 11) with a net bene"t of −10. The resulting network obtained by removing the three cycles
from the network G∗∪ E
M is the desired solution, which
has a total net bene"t of 403.
The number of traversals on each edge is tabulated in
Table3, and the total net bene"t generated on each edge
is displayed in Table4. We note that the solution does not
cover the edge traversals (4, 6), (4, 7), (11, 12), (12, 13), and (14, 15) with positive net bene"ts 2, 2, 5, 3, and 1, re-spectively. The corresponding solution covering those edge traversals has a total net bene"t of 363.
Table 4
Total net bene"ts on each edge in the solution
Edge Bene"t Edge Bene"t
(1, 2) 21 (7, 9) 23 (1, 4) 20 (8, 10) 24 (2, 3) 20 (8, 9) 7 (2, 4) 5 (8, 11) 14 (3, 5) 10 (9, 11) 13 (3, 9) 27 (9, 15) 5 (4, 5) 22 (10, 12) 4 (4, 6) 19 (11, 12) 30 (4, 7) 12 (11, 14) 11 (4, 8) 10 (12, 13) 19 (5, 9) 22 (12, 14) 17 (6, 10) 20 (13, 14) 9 (7, 8) 13 (14, 15) 6 7. Conclusions
In this note, we considered the MBCPP, an interesting generalization of the classical Chinese postman problem, on the undirected networks. We showed that the MBCPP is more complex than the RPP by presenting a linear transfor-mation converting the RPP into a special case of the MBCPP. Hence, the MBCPP includes the CPP and the TSP as special cases.
We also discussed some solution properties, and presented a su6cient condition for the MBCPP solution to cover all the edges on the network. Based on those solution properties, we proposed an e6cient algorithm to solve the MBCPP approximately, and presented an example to illustrate the proposed algorithm.
References
[1] Christo"des N, Campos V, Corberan A, Mota E. An algorithm for the rural postman problem. Imperial College Report IC-OR, 1981.
[2] Dror M, Stern H, Trudeau P. Postman tour on a graph with precedence relation on arcs. Networks 1987;17:283–94. [3] Golden B, Wong R. Capacitated arc routing problems.
Networks 1981;11:305–15.
[4] Malandraki C, Daskin MS. The maximum bene"t Chinese postman problem and the maximum bene"t traveling salesman problem. European Journal of Operational Research 1993;65:218–34.
[5] Pearn WL. Solvable cases of the k-person Chinese postman problem. Operations Research Letters 1994;16:241–4.
