Download as pdf or txt
Download as pdf or txt
You are on page 1of 23

An LP based algorithm for the

Privatized Rural Postman Problem Algorithm

Julián Aráoz, Elena Fernández


Departament d’EIO. Secció Informàtica
Universitat Politècnica de Catalunya
and
Oscar Meza
Dept. CTI, Simn Bolvar University

February 28, 2005


DR 2005/xx
An LP based algorithm for the
Privatized Rural Postman Problem
Julián Aráoz
Simón Bolı́var University∗, Venezuela
Technical University of Catalonia†, Spain
Elena Fernández
Technical University of Catalonia, Spain
Oscar Meza‡
Simón Bolı́var University, Venezuela.

Abstract
In this work we present an LP based algorithm for solving the
Privatized Rural Postman Problem. This problem has been defined
recently and is a generalization of other arc routing problems like, for
instance the Rural Postman Problem. The main difference is that
there are no required edges. Instead, there is a profit func-
tion on the edges that must be taken into account only the
first time that an edge is traversed. The problem is modeled
with an integer system of linear inequalities. From the solution of
a relaxed model we obtain upper bounds and generate vio-
lated cuts when possible. We also propose a heuristic to generate
feasible solutions that provide lower bounds. The numerical
results from a series of computational experiments both with
Rural Postman Problems and Privatized Postman Problems
are presented.

1 Introduction
In this work we present an algorithm to solve the Privatized Rural Postman
Problem (PRPP). This problem was defined in Aráoz, Férnandez and Zoltan

Retired Professor Dpt. Process and Systems.

Visiting Professor Dpt. Statistics and Operation Research.

Retired Professor Dpt. Computation and Information Technology

1
[2], and is part of the family of Arc Routing Problems (ARPs), that usually
aim to determine a least-cost traversal of a specified arc subset of a graph,
with or without constraints. Such problems arise in a variety of practical
contexts, like post delivery and garbage collection.
Typical constraints require the routes to begin and finish at a given
point (the depot), and guarantee the connectivity of the solutions
with the depot. We refer the reader to Dror [6] for a recent and compre-
hensive state on the art of such problems. Here we considered only Edge
Routing Problems (ERPs), which are the undirected cases of ARPs.

Like in other ERPs, in the PRPP we assume that the demand for service
is placed at the edges of a graph. However, there is no specific arc subset
to be traversed. Instead, we assume that giving service to an edge not only
would incur a cost (associated with displacement), but would also result
on a profit (associated with servicing edges). The displacement cost to an
edge will certainly account for the cost of all the edges that are traversed
in that same route, as many times as they are traversed. Similarly, the
profit associated with servicing an edge, should also take into account the
profit of the additional edges that are serviced in that same route. However,
the profit of each edge serviced in the route will be collected only once,
independently of the number of times the edge is traversed.
In the PRPP we look for traversals that maximize the total servicing
profit minus the displacement cost. They constitute a generalization of
most ERPs and of most Traveling Salesman Problems (TSPs).

The algorithm that we propose for RPP generates both an upper and
a lower bound. The upper bound is derived from solving a series of LP
relaxations of an integer system of linear inequalities that define the problem.
While possible we reinforce the relaxation by generating cuts violated by the
current solution. We present exact separation algorithms both for the family
of inequalities that guarantee the parity of the nodes and for the family of
inequalities that guarantee connectivity of the solutions with the depot.
The lower bound is obtained with a heuristic which is an adaptation of
the 3T heuristic of Fernndez, Meza, Garfinkel and Ortega [10] for the Rural
Postman Problem RPP.
In order to test the efficiency of the proposed algorithm, we
have run a series of computational experiments on benchmark
instances of both classical RPP’s and PRPP’s. The numerical
results assess the good behavior of the method.
The paper is structured as follows. In Section 2 we define the prob-

2
lem and recall from [2] some properties of its solutions. The polyhedral
model that we will use is presented in Section 3. Section 4 describes the
elements of the algorithm, including the separation procedure for each type
of cuts and the heuristic. In Section 5 we describe the computational ex-
periments and present the obtained results. The tables with the numerical
results are given in the Appendix. We finish the paper in Section 6 with the
conclusions and some remarks about future research.

2 The Privatized Rural Postman Problem


The Privatized Rural Postman Problem was defined in the work of Aráoz,
Férnandez and Zoltan [2]. We next repeat the definitions and results relevant
for our algorithm.

Definition 2.1 Privatized Rural Postman Problems (PRPPs).


Given a graph G(V, E) with a distinguished vertex d, the Depot, and two
functions from E in IR+ , the profit function b and the cost function c, the
Privatized Rural Postman Problem is to find one cycle C ∗ which maximizes
the value of X
(be − te ce )
e∈C

where C is a cycle in G passing trough d, and not necessarily simple, and te


is the number of times that edge e is traversed in C.
We denote a PRPP by (G, d, b, c).

That is, in a PRPP, we look for the most Profitable Subtour passing
through d. In [2] it was proven that the PRPP is NP–Hard.
In this work we model the PRPP by means of a linear system
with an exponential number of inequalities, based on the one de-
fined in [2], and we propose a solution approach to obtain lower
and upper bounds for the PRPP, that is similar to that in [10]
for the RPP. The approach combines the addition of violated cuts
to strengthen the LP formulation of the problem, and thus to
improve the associated upper bound, with a heuristic to obtain
feasible solutions, and thus lower bounds.
We define the functions ϕe = be − ce and ψe = be − 2ce = ϕe − ce . The
value ϕe is the net profit obtained when the edge e is traversed
once, whereas ψe is the the net profit when the edge e is traversed
twice.

3
In the PRPP the edges in R = {e ∈ E | ψe ≥ 0} play an important role,
similar to the set of required edges in a RPP (see [2]).

We denote by GR ≡ (V (R) ∪ {d}, R) the subgraph induced by the edge


set R and the depot. Let Ci , i ∈ P = {0, . . . , p} be the connected compo-
nents of the graph GR and we assume that d ∈ C0 . Using standard graph
nomenclature, when necessary, we denote by Vi = V (Ci ) and we differenti-
ate between γ(Vi ) in the original graph G and γR (Vi ) in the graph GR . That
is γR (Vi ) = γ(Vi ) ∩ R.

Let eR ∗
i be an arbitrarily selected edge in Ci , i ∈ P , and let C denote an
optimal solution. The following properties were proven in [2]:

2.2 Dominance 1. No edge is used more than twice in C∗ .

2.3 Dominance 2. If an edge e ∈ C∗ is used with value ce then it is used


twice in e ∈ C∗ .

2.4 Dominance 3. Let e ∈ C∗ , if for some connected component Ck of GR


we have that V (e) ∩ Vk 6= ∅ then all the truly positive edges of Ck are in C.
This imply that if e1 , e2 ∈ γR (Vk ) then either both edges are in C ∗ or none
of them is in C ∗ .

2.5 Preprocessing 1. Let Ck be one of the connected components of GR . If


e ∈ γ(Vk ) \ R then e is used at most once.

From a more general result in [2] we also have the following corollary:

2.6 Preprocessing 2. γR (V0 ) ⊆ C ∗ .

Remark 2.7 By Dominance 1 in any optimal solution to the PRPP no


edge is traversed more than twice. Thus, from now on, we will consider an
auxiliary graph G0 = (V, E 0 ∪ E 00 ), where both E 0 and E 00 are disjoint copies
of E with all the edges in E 0 corresponding to the first time that the original
edge is used and all the edges in E 00 corresponding to the second time that
the original edge is used. We denote by e an original edge in E and by e0
and e00 its corresponding copies in E 0 and E 00 respectively.
Hence for any e ∈ E we have be0 = be , ce0 = ce and be00 = 0, ce00 = ce ,
also eR 0 R
k ∈ E , k ∈ P . Throughout, ek is chosen to be one of the edges in Ck
with greatest φe .

4
3 Polyhedral Model
The linear equations that we use to define the feasible solutions are based
in the ones given in the paper of Aráoz, Férnandez and Zoltan [2]. The
model is defined on the graph G0 of Remark (2.7) as defined above. Each
edge e0 ∈ E 0 is associated with the variable xe0 whereas each edge e00 ∈ E 00
is associated with the variable ye00 . Whenever there is not confusion we use
xe and ye instead of xe0 and ye00 . The PRPP Linear Program (PRPPLP)
model is the following:

X X
max z(x, y) = ϕe xe − ce ye (1)
e∈E 0 e∈E 00
x(δ 0 (v) \ F 0 ) + y(δ 00 (v) \ F ) ≥ x(F 0 ) + y(F 00 ) − |F 0 ∪ F 00 | + 1,
00

v ∈ V, F 0 ⊆ δ 0 (v), F 00 ⊆ δ 00 (v), |F 0 ∪ F 00 | odd. (2)


0 00
x(δ (S)) + y(δ (S)) ≥ 2xe ,
S ⊆ V \ {d}, {e ∈ γ 0 (S) ∪ δ 0 (S)|V (e) ∩ V (R) = ∅} (3)
0 00
x(δ (S)) + y(δ (S)) ≥ 2xeR ,
k

S ⊆ V \ {d}, k 6= 0, Vk ∩ S 6= ∅ (4)
0 0
xe = xeR , e ∈ γ (Vk ) ∩ R , k ∈ P (5)
k

ye00 ≤ xe0 , e∈E (6)


0 0 0
xe ≤ xeR , e ∈ ((γ (Vk ) ∪ δ (Vk )) \ R ), k ∈ P (7)
k

ye = 0, e ∈ γ 00 (Vk ) \ R00 , k ∈ P (8)


0 0
xe = 1, e ∈ γ (V0 ) ∩ R (9)
xe0 , ye00 ∈ {0, 1}, e∈E (10)
P
Notice that we use the standard compact notation f (A) ≡ e∈A fe when
A ⊆ E, and f is a vector or a function defined on E. Also, by extension,
if H is a graph or a path we denote f (H) ≡ f (E(H)). Finally, we use the
prime and double prime in edge sets to denote which copies we are using,
like, for instance, δ 0 (V ) = δ(V ) ∩ E 0 , δ 00 (V ) = δ(V ) ∩ E 00 .

• Inequalities (2) ensure even degree at every vertex, and they are im-
plied by the so-called co-circuit inequalities in Barahona and Grötschel
[3] in a matroid context (see also Aráoz et al. [1]).

• Inequalities (3) and (4) ensure connectivity with the depot of any edge
that is selected, that is any edge e0 such that xe0 = 1. By Dominance

5
2 we do not need to check the edges in E 00 and by Dominance 4 if
e0 ∈ Ck we only use eR
k.

• Equalities (5) correspond to Dominance 4.

• Inequalities (6) correspond to Dominance 2.

• Inequalities (7) correspond to Dominance 3 and Dominance 4.

• Equalities (8) correspond to Preprocessing 1.

• Equalities (9) correspond to Preprocessing 2.

• Binary conditions (10) correspond to Dominance 1.

4 Cutting Plane Algorithm


4.1 Initial Linear Relaxation
The cutting plane algorithm starts with the linear relaxation that is de-
scribed below. At each iteration of the algorithm we add cutting planes, to
be described later on. The initial linear relaxation is the following:

(P RP P LP 0)
X X
max z(x, y) = ϕe xe − ce ye (11)
e∈E 0 e∈E 00
x(δ 0 (v)) + y(δ (v)) ≥ 2xe ,
00
v ∈ V, e ∈ δ 0 (v). (12)
0 00
x(δ (Vk )) + y(δ (Vk )) ≥ 2xeR , k ∈ P \ {0} (13)
k

xe = xeR , e ∈ E 0 ∩ γR (Vk ) and k ∈ P \ {0} (14)


k

ye00 ≤ xe0 , e∈E (15)


0 0 0
xe ≤ xeR , e ∈ ((γ (Vk ) ∪ δ (Vk )) \ R ), k ∈ P (16)
k
00 00
ye = 0, e ∈ γ (Vk ) \ R , k ∈ P (17)
0 0
xe = 1, e ∈ γ (V0 ) ∩ R (18)
0 ≤ xe0 , ye00 ≤ 1, e∈E (19)

This initial program has the same functional (1) and all the inequalities
(5)-(9) of PRPPLP. However integrality conditions (10) are relaxed and it
only includes those inequalities of the type (2) that correspond to the sets
with |F 0 | = 1 and |F 00 | = 0. Note that when F 0 = {e} we have

6
x(δ 0 (v) \ e) + y(δ 00 (v)) ≥ xe ≡
x(δ 0 (v)) − xe + y(δ 00 (v)) ≥ xe ≡
x(δ 0 (v)) + y(δ 00 (v)) ≥ 2xe (see (12))

No inequality (3) is included but (PRPPLP0) has the inequalities of


type (4) that correspond to the sets S = Vk with e = eR
k.
Note that, since some of the original constraints are omitted, the
solutions to this system may be integer despite not being feasible to the
PRPPLP.

4.2 Cutting Plane Algorithm


The cutting plane algorithm is as follows:

1. Let LP R0 be the linear program relaxation (11) to (19). Set k := 0

2. Let z̄ = ∞; z = 0; optimum = f alse; bestxy = (0, 0), end = f alse

3. While end = f alse do:

a. Find a solution (x∗ , y ∗ ) of the current linear program relaxation


LP Rk .
b. if z̄ > z(x∗ , y ∗ ) then z̄ = z(x∗ , y ∗ )
c. If (x∗ , y ∗ ) is feasible to PRPPLP then
bestxy = (x∗ , y ∗ ); end = true. z = z̄
else
i. Identify inequalities of types (2), (3), (4) violated by (x∗ , y ∗ ).
ii. Add the valid inequalities identified in step (3.c.i) to LP Rk .
iii. If no new inequality has been add in step (3.c.ii) then
end = true

4. a. Compute the heuristic lower bound and solution z h = (xh , y h )


b. if z < z(xh , y h ) then z = z(xh , y h ); bestxy = (xh , y h )

5. if z = z̄ then
output z; bestxy
else
output gap = z̄ − z; z; bestxy

7
4.3 Separation of Cuts
At a given iteration of the algorithm Step 3.c.i consists of two parts: first we
find violated inequalities of type (2) and second we find violated inequalities
of type (3)or (4), if it is possible.

4.3.1 Finding violated inequalities of type (2)


To separate the inequalities (2) we use a modification of a heuristic proposed
by Ghiani and Laporte [11] for the RPP. This modification results in an
exact separation method for singletons although the separation is
still a heuristic for general vertex sets.
Given a graph G and a vector x : E → [0, 1], the cocycle inequalities
for vertex v are:

x(δ(v) \ F ) ≥ x(F )− | F | +1, ∀F ⊆ δ(v), F odd (20)


The separation problem for these inequalities is: Given an vector x∗ ,
0≤
x∗ ≤ 1 determine if there is one of the inequalities violated by this vector o
determine that none exists.
4.1 Algorithm:
1. For each vertex v do
(a) Let F = {e ∈ δ(v)|x∗e ≥ 0.5}.
(b) If F is even then let x∗e1 = min{x∗e |e ∈ F } and x∗e2 = max{x∗e |e ∈
delta(v) \ F }. If x∗e1 − 0.5 ≤ 0.5 − x∗e2 delete e1 from F otherwise
add e2 to F .
(c) If the Inequality (20) is violated by this set F we have a cut.
Theorem 4.2 The separation Algorithm (4.1) is exact and the order of the
algorithm is |E|.
Proof: Consider rewriting the inequalities (20):
X
x∗ (δ(v) \ F ) + (1 − x∗e ) ≥ 1, |F | odd. (21)
e∈F
It is clear that an edge e contributes to the left hand side of the
above expression with value (1 − x∗e ) if it is in F and with value x∗e if it
is not. Therefore Step (1a) gives the set F with minimum value of the left
hand side and Step (1b) gives the minimum correction to get an odd set.
EF: Yo quitaria el siguiente comentario.If there is not cut for any
vertex, there is no inequality (20) that separates x∗ . 2

8
Let (x∗ , y ∗ ) be a solution of the current linear program relaxation LPR.
Applying Algorithm (4.1) to this solution we obtain an odd set (F 0 ∪ F 00 ).
If
X X
x∗ (δ 0 (v) \ F 0 ) + y ∗ (δ 00 (v) \ F 00 ) + (1 − x∗e ) + (1 − ye∗ ) − 1 < 0
e∈F 0 e∈F 00

then the following inequality is violated by (x∗ , y ∗ ):

x(δ 0 (v) \ F 0 ) + y(δ 00 (v) \ F 00 ) ≥ x(F 0 ) + y(F 00 ) − |F 0 ∪ F 00 | + 1

4.3.2 Finding violated inequalities of type (3) or (4)


Inequalities of type (3) can be classified in two types. The in-
equalities of the first type are of the form x(δ(S)) + y(δ(S)) ≥ 2xe ,
for e ∈ δ(S), S ⊆ V \{d}. We will call then set-parity inequalities (as
opposed to inequalities (2) which are referred to as node-parity).
The other type of inequalities (3) are of the form x(δ(S)) +
y(δ(S)) ≥ 2xe , for e ∈ γ(S), S ⊆ V \ {d}. We call them connectivity
inequalities.
The inequalities (4) can also be classified in these two classes: set-parity
inequalities (xeR ∈ δ(S)) and connectivity inequalities (xeR ∈ /δ(S)).
k k
For separating any of the above inequalities (connectivity or set-parity)
we use the algorithm of Gusfield [12] to calculate the tree of min-cuts of the
graph G00 , obtained from the graph G0 by eliminating all edges in E 0 or E 00
with value zero in the current solution (x∗ , y ∗ ), and by assigning a cost
to each edge with value equal to the value of the variable associated with
that edge.
The connectivity inequalities are separated with the exact algorithm of
Belenguer and Benavent [4, p.261] EF: Aqui cambiaria la referencia y
pondria directamente la de la referencia original de Belenguer y
Benavent que incluyo al final. The algorithm does the following:
EF: Lo que sigue tenemos que ponerlo tambin en trminos del
grafo G’, no? Hago un intento, pero no se si acierto del todo
For each edge e = {u, v} ∈ E 0 with v and w different from the depot and
x∗e > 0 the minimum cut δ 0 (S) ∪ δ 00 (S) such that e ∈ γ 0 (S) is easily obtained
from the min-cut tree. If the value of the min cut is smaller than 2x∗e then
the following inequality is violated by (x∗ , y ∗ ): x(δ 0 (S)) + y(δ 00 (S)) ≥ 2xe0 .
Los dos parrafos que siguen es como estaba hasta ahora: For
each edge e = {u, v} ∈ E 0 with v and w different from the depot
and x∗e > 0 the minimum cut δ(S) such that e ∈ γ(S) is easily

9
obtained from the min-cut tree: If the value of the min cut is
smaller than 2x∗e then the following inequality is violated by (x∗ , y ∗ ):
x(δ(S)) + y(δ(S)) ≥ 2xe0 .
EF: Para las de set parity pasa lo mismo. Hago un intento de
ponerlo bien, pero dejo detras exactamente lo que hay ahora
The set parity inequalities are separated exactly with a very simple al-
gorithm: For each edge e = {u, v} ∈ E 0 we determine the minimum cut
δ 0 (S) ∪ δ 00 (S) that separates v and u. If the value of this min cut is less
than 2xe then the following inequality is violated by (x∗ , y ∗ ): x(δ 0 (S)) +
y(δ 00 (S)) ≥ 2xe .
lo que habia The set parity inequalities are separated exactly
with a very simple algorithm: For each edge e = {u, v} ∈ E 0 we
determine the minimum cut δ(S) that separates v and u. If the
value of this min cut is less than 2xe then the following inequality
is violated by (x∗ , y ∗ ): x(δ(S)) + y(δ(S)) ≥ 2xe .

4.4 Heuristic Bound


The algorithm used to obtain a lower bound is the following heuristic:
1. Let (x∗ , y ∗ ) be the solution obtained with the linear relaxation de-
scribed in the previous section.
2. Transform the PRPP into a RPP. The graph of the RPP will be the
original graph G of PRPP plus a fictitious edge {1, 10 }. Edge {1, 10 }
is defined as required to guarantee that the solution passes through
the depot. All other edges of PRPP with value x∗e greater or
equal to ² are also defined as required edges for RPP are (² is
a parameter). The cost of edge {1, 10 } is zero, and the cost of
any other edge is the same as the costs of the edge in PRPP.
3. Find a feasible solution to RPP with the 3T heuristic of Fernández,
Meza, Garfinkel and Ortega [10].
4. The feasible solution to PRPP results from eliminating the parallel
edges {1, 10 } from the feasible solution to RPP obtained in (3).

5 Computational Experiments
In order to evaluate the performance of the proposed algorithm
we have run a series of computational experiments. We next de-
scribe such experiments and give the obtained numerical results.

10
Programs have been coded in C using CPLEX 7.0 library. All instances
were run on a Sun ULTRA 10, model 440, 1x440 MHZ, 1GB DRAM. In this
section use the Acronyms in Table 1 for the identification of columns in the
tables of results. Since there are no available benchmark instances,
we have generated PRPP instances from well known sets of benchmark
RPP instances. To generate PRPP instances from RPP instances we have
kept the cost function c and we have assigned random profits b to the edges.
In all cases the depot has been taken as vertex 1.
The RPP benchmark instances are divided in five groups. The first
group contains two problems, ALBAIDAA and ALBAIDAB, obtained
from the Albaida, Spain Graph (see Corberán and Sanchis [8]). The second
group contains the 24 instances (problems labeled P) of Christofides
et al. [7]. The last three groups contain instances from Hertz et
al [13]: 36 instances with vertices of degree 4 and disconnected
required edge sets (labeled DEGREE), 36 grid instances (labeled GRID),
and 20 randomly generated instances (labeled RANDOM). Tables 2 and 3
depict information on these instances.
From these 118 RPP instances we have generated two different
sets of PRPP instances. The first set of instances was obtained in such a
way that the optimal solutions were known, so we could better evaluate the
obtained results. From [2] we know that this can be achieved by keeping in
PRPP the cost function (c) of RPP and by defining a profit function (b) in
PRPP that assigns high enough profits to the edges that are required in RPP
and zero profit for the edges that are non-required in RPP. Along this line,
the first set of test problems was generated by taking be = 1000∗ce , ∀e ∈ ER ,
and be = 0, ∀e ∈ E \ ER .
The results with this first set of instances are given in Tables 4 and 5.
As can be seen the obtained results are very good. From the optimal/best
known solutions to the RPP instances, we know when the solutions obtained
by the heuristic are/are not optimal for the PRPP. In particular, the heuris-
tic provided the optimal solution for 85 of the 118 instances. In 22 of these
instances, we could prove the optimality of the obtained solution (0% per-
cent gap between the upper and the lower bounds) and for the remaining
instances these percent gaps never exceed is 0.33%. In 66 instances the per-
cent gap does not exceed 0.1%, in 10 instances this gap is in (0.1, 0.2], and
only in ten instances the percent gap is in (0.2, 0.33]. So we can conclude
that both the upper and lower bounds were very good for these instances.
In general, the number of LP iterations of the algorithm is small. For 78 in-
stances this number does not exceed 20, and only 2 instances required more
than 100 iterations to terminate. In general, the instances that require more

11
iterations are the ones with a larger number of connected componentes in
the graph induced by the requires edges of RPP. These are the instances
that also tend to be harder to solve as RPP instances. To some extent the
difficulty for solving the instances is also reflected on the number of cuts
that were added: both in the total number of cuts and on the number of
cuts per iteration. In the vast majority of the cases there are more con-
nectivity cuts than parity cuts (either set parity or node parity). However,
in our opinion this number is small in all the cases taking into account the
difficulty of the problems. The efficiency of the algorithm can also be ap-
preciated in the required cpu times. As was expected, most of the cpu time
is consumed for obtaining the lower bound. There are four instances that
required more than 200 seconds, although in general these times were small,
and 79 instances required less than one second. The times of the heuristic
are very small: they only exceed one second in 14 instances and the largest
time is below 10 seconds.

In the second group of instances the profits are not as large as in the first
group. In particular, in this set the profits are generated as follows (U[a,b]
denotes the Uniform distribution in the interval [a,b]):

• be ∈ U[ce , 3ce ], if e a required edge of the RPP instance.

• be ∈ U[0, ce ], if e a non-required edge of the RPP instance.

The results of the instances of this second group are given in Tables
6 and 7. Now ve have been able to prove optimality for 43 of the 118
instances. However, for the remaining instances the percent gaps are con-
siderably larger than for the first set of instances, and there are two instances
(GRID19M and PM18M) for which the percent gap between the upper and
the lower bound is too large. This is due to the fact that the profits in the
current set of instances are similar in magnitude to the costs. This results
in much smaller values of the objective function, so that small absolute dif-
ferences between the upper and the lower bound result in much larger the
relative gaps. However, the total number of instances for which we have
proven optimality (43) is considerably larger for this set of instances than
for the set with very large profits. In addition, we can see that the absolute
values of the difference between the upper bounds and the lower bounds are
in general quite similar to those of the first set of instances and in very few
cases the absolute difference of the upper and the lower bound is notably
larger for the set with smaller profits. Since now the optimal values of these

12
instances are unknown, we cannot assess the optimality of the heuristic so-
lutions unless the percent gap is zero, but according to the results obtained
with the first set of instances we conjecture that there are more instances
for which the solution provided by the heuristic is in fact optimal. Unfortu-
nately, in quite a few cases the percent gaps between the upper and the lower
bounds are too large. In terms of the number of iterations and the number
of cuts added, the results with this set of instances with smaller profits are
similar to those with the set of instances with larger profits: these numbers,
tend to increase for the instances that are harder to solve and again the
number of connectivity constraints exceeds that of parity constraints. As
for the cpu times they are again very, small both for obtaining the lower
bound and for the heuristic.
Thus, we can conclude that the results provided by the algorithm with
the set of instances with large profits are very satisfactory, both in terms of
the quality of the solutions and of the required computation effort. For the
set of instances with smaller profits, it is difficult to summarize the obtained
results. In terms of quality of the solutions, on the one hand, for more than
36% of these instances the obtained lower bound allowed to prove optimality
of the solution provided by the heuristic. On the other hand, for some of
the instances were optimality could not be proven, the resulting percent gap
was too big. In terms of the required computational effort, the proposed
algorithm has proven to be efficient since, in general, it requires a small
number of iterations and does not generate an excessive number of cuts,
which results in small cpu times.
A second group of PRPP instances were generated from the RPP in-
stances by taken vertex 1 as the depot and for each RPP we assign uniformly
random generated benefits:

6 Conclusions
In this work we have presented an algorithm to obtain upper and lower
bounds for the Privatized Rural Postman Problem that was introduced in
[2]. The lower bound is obtained with an iterative LP-based cutting plane
algorithm and the upper bound is obtained with a heuristic that takes as
starting point the solution of the last LP relaxation. We have presented
procedures to solve exactly the separation problem for the generated cuts.
In order to evaluate the performance of the algorithm we have generated
two sets of benchmark instances. The results with the first of instances, for
which the optimal solutions were know are very satisfactory both in terms of

13
quality and of efficiency. An important number of instances of the second set
(more than 36%) were optimally solved, although for the reamining instances
some of the percent gaps were too large.

7 Acknowledgements
The work of the first and third authors has been partially supported by
grants from FONACIT Project USB S1- 2000000438, Venezuela. The re-
search of the first author has been partially financed by the Spain “Secre-
tarı́a de Estado de Educación y Universidades”. The research of the second
author has been partially supported by grant TIC 2000 1750 006/3 of the
Inter-Ministerial Spanish Commission of Science and Technology. These
supports are gratefully acknowledged.

References
[1] J. Aráoz , W. Cunningham , J. Edmonds, and J. Green–Krotki. Reduc-
tions to 1–matching polyhedra. Networks, 13:455–473, 1983.

[2] J. Aráoz, E. Fernández, C. Zoltan. “The Privatizad Rural Postman


Problem”. Reporte DR-2003/12, EIO Departament , Universitat Politc-
nica de Catalunya, Espaa (2002). Accepted in Computers and Opera-
tions Research.

[3] F. Barahona and M. Groeschel. On the cycle polytope of a binary ma-


troid.J. Comb. Theory, 40:40–62, 1986.

[4] E. Benavent, A. Corberan, and J.M. Sanchis. Linear programming


based methods for solving arc routing problems. In Arc Routing: The-
ory, Solutions and Applications M. Dror (edt), Kluwer Academic Pub-
lishers, 2000.

[5] C. Berge Graphs and Hypergraphs, North-Holland, Amsterdam, 1973.

[6] M. Dror Edt. Arc Routing: Theory, Solutions and Applications. Kluwer
Academic Publishers, 2000.

[7] N. Christofides, V. Campos, A. Corberan, and E. Mota. An algorithm


for the rural postman problem. Imperial College Report IC.O.R., 81.5,
1981.

14
[8] A. Corberán, J. M. Sanchis. A polyhedral approach to the rural post-
man problem. Eur. J. Oper. Res. 79, 95-114, 1994.

[9] A. Corberán, J. M. Sanchis. The General Routing Problem: facets from


the RPP and GTSP polyhedra. Eur. J. Oper. Res. 108, 538-550, 1998.

[10] E. Fernández, O. Meza, R. Garfinkel, and M. Ortega. On the undirected


rural postman problem: Tight bounds based on a new formulation.
Operations Research, 51:281–291, 2003.

[11] G. Ghiani and G. Laporte. A branch-and-cut algorithm for the undi-


rected rural postman problem. Mathematical Programming, 87:467–481,
2000.

[12] D. Gusfield. Very Simple Methods for All Pairs Network Flow Analysis.
SIAM Journal of Computing 19, 143-155, 1990.

[13] A. G. Hertz, P. Laporte, H. Nanchen. Improvement procedures for the


undirected rural postman problem. INFORMS J. Comput. 1, 53-62,
1999.

15
A TABLES APPENDIX
In this Appendix we concentrate the tables showing the results of our ex-
periments, the analysis of these results are in Section (5). Si cambiamos

Acronyms Parameter Name


PN Problem Name
NV Number of Vertices
NE Number of Edges
NER Number of Required Edges
NENR Number of non Required Edges
SERC Sum of Required Edges Costs
NRC Number of connected components in the graph induced by the required edges
LRV Linear Relaxation Value
HV Heuristic Value
DIF (LRV-HV)
PDRH Percent GAP Between LRV and HV=100*(LRV-HV)/LRV
TBHS Total Profit of the Heuristic Solution
TCHS Total Cost of the Heuristic Solution
LRT Linear Relaxation Time (sec)
HT Heuristic Time (sec)
LRIN Linear Relaxation Number
LRCR Number of Connectivity Constraints in Linear Relaxation
RSPRN Number of Set Parity Constraints in Linear Relaxation
RVPRN Number of Vertex Parity Constraints in Linear Relaxation
ORN Objective Rounding Number Yo esta la quitaria y no diria nada

Table 1: Table Column Acronyms

las leyendas, tambien habra que cambiar alguno de los acronimos.

16
PN NV NE NER NENR SERC NRC
ALBAIDAA 102 160 99 61 7295 10
ALBAIDAB 90 144 88 56 5803 11
P01 11 13 7 6 25 4
P02 14 33 12 21 80 4
P03 28 58 26 32 73 4
P04 17 35 22 13 55 3
P05 20 35 16 19 69 5
P06 24 46 20 26 70 7
P07 23 47 24 23 93 3
P08 17 40 24 16 93 2
P09 14 26 14 12 57 3
P10 12 20 10 10 45 4
P11 9 14 7 7 14 3
P12 7 18 5 13 13 3
P13 7 10 4 6 12 3
P14 28 79 31 48 145 6
P15 26 37 19 18 180 8
P16 31 94 34 60 139 7
P17 19 44 17 27 63 5
P18 23 37 16 21 61 8
P19 33 55 29 26 141 7
P20 50 98 63 35 282 7
P21 49 110 67 43 288 6
P22 50 184 74 110 499 6
P23 50 158 78 80 380 6
P24 41 125 55 70 292 7
DEGREE0 16 32 3 29 119 2
DEGREE1 16 31 6 25 247 3
DEGREE2 16 31 9 22 322 4
DEGREE3 16 32 8 24 343 3
DEGREE4 16 31 8 23 336 4
DEGREE5 16 31 12 19 490 4
DEGREE6 16 32 11 21 405 5
DEGREE7 16 31 12 19 464 4
DEGREE8 16 31 16 15 614 4
DEGREE9 36 72 12 60 366 8
DEGREE10 36 72 10 62 235 7
DEGREE11 36 72 17 55 629 11
DEGREE12 36 72 17 55 514 8
DEGREE13 36 72 22 50 725 5
DEGREE14 36 72 30 42 958 9
DEGREE15 36 72 32 40 913 6
DEGREE16 36 72 34 38 1066 4
DEGREE17 36 72 38 34 1235 6
DEGREE18 64 128 28 100 717 15
DEGREE19 64 128 29 99 729 11
DEGREE20 64 128 27 101 662 11
DEGREE21 64 128 47 81 1151 10
DEGREE22 64 128 47 81 1132 12
DEGREE23 64 128 51 77 1174 9
DEGREE24 64 128 68 60 1723 6
DEGREE25 64 128 62 66 1474 9
DEGREE26 64 128 75 53 1868 5
DEGREE27 100 200 50 150 903 22
DEGREE28 100 200 55 145 1038 19
DEGREE29 100 200 50 150 902 20
DEGREE30 100 200 86 114 1552 13
DEGREE31 100 200 90 110 1784 10
DEGREE32 100 200 81 119 1478 16
DEGREE33 100 200 121 79 2162 9
DEGREE34 100 200 118 82 2207 9
DEGREE35 100 200 116 84 2182 9

Table 2: RPP Problems Data

17
PN NV NE NER NENR SERC NRC
GRID0 16 24 3 21 3 3
GRID1 16 24 5 19 5 5
GRID2 16 24 4 20 4 4
GRID3 16 24 8 16 8 5
GRID4 16 24 7 17 7 5
GRID5 16 24 7 17 7 3
GRID6 16 24 13 11 13 4
GRID7 16 24 8 16 8 4
GRID8 16 24 9 15 9 4
GRID9 36 60 11 49 11 7
GRID10 36 60 13 47 13 9
GRID11 36 60 15 45 15 9
GRID12 36 60 26 34 26 7
GRID13 36 60 23 37 23 6
GRID14 36 60 25 35 25 7
GRID15 36 60 35 25 35 5
GRID16 36 60 30 30 30 7
GRID17 36 60 34 26 34 5
GRID18 64 112 24 88 24 10
GRID19 64 112 27 85 27 12
GRID20 64 112 27 85 27 14
GRID21 64 112 46 66 46 8
GRID22 64 112 47 65 47 8
GRID23 64 112 50 62 50 11
GRID24 64 112 68 44 68 4
GRID25 64 112 61 51 61 5
GRID26 64 112 66 46 66 6
GRID27 100 180 41 139 41 19
GRID28 100 180 49 131 49 20
GRID29 100 180 44 136 44 20
GRID30 100 180 73 107 73 13
GRID31 100 180 77 103 77 18
GRID32 100 180 82 98 82 11
GRID33 100 180 113 67 113 4
GRID34 100 180 107 73 107 9
GRID35 100 180 109 71 109 6
RANDOM0 20 37 3 34 6249 3
RANDOM1 20 47 4 43 9722 4
RANDOM2 20 47 4 43 8393 3
RANDOM3 20 75 7 68 17478 4
RANDOM4 20 60 6 54 16776 3
RANDOM5 30 70 7 63 9307 4
RANDOM6 30 111 10 101 21510 4
RANDOM7 30 70 7 63 14411 5
RANDOM8 30 111 11 100 22903 5
RANDOM9 30 111 11 100 22507 6
RANDOM10 40 130 13 117 20655 8
RANDOM11 40 103 10 93 15215 6
RANDOM12 40 82 8 74 11346 5
RANDOM13 40 203 18 185 35160 9
RANDOM14 40 203 18 185 42031 7
RANDOM15 50 203 20 183 36166 8
RANDOM16 50 162 15 147 19389 12
RANDOM17 50 130 13 117 15961 6
RANDOM18 50 203 19 184 35661 7
RANDOM19 50 203 19 184 31691 8

Table 3: RPP Problems Data (continuation)

18
PN LRV HV DIF PDRH TBHS TCHS LRT HT LRIN LRCR RSPRN RVPRN

ALBAIDAA 719057 718901 156 0,02 729500 10599 4,16 1,48 6 126 23 38
ALBAIDAB 571851 571671 180 0,03 580300 8629 2,38 1,23 6 227 22 38
P01 2424 2424 0 0,00 2500 76 0,01 0,18 3 11 1 2
P02 7858 7848 10 0,13 8000 152 0,02 0,04 2 9 2 2
P03 6899 6898 1 0,01 7000 102 0,10 0,05 5 24 1 4
P04 5419 5416 3 0,06 5500 84 0,03 0,02 3 34 1 7
P05 6780 6776 4 0,06 6900 124 0,04 0,05 3 11 4 3
P06 6900 6898 2 0,03 7000 102 0,06 0,03 3 19 6 7
P07 9170 9170 0 0,00 9300 130 0,04 0,01 3 29 1 10
P08 9179 9178 1 0,01 9300 122 0,06 0,02 5 0 0 13
P09 5619 5617 2 0,04 5700 83 0,03 0,01 4 9 2 6
P10 4428 4420 8 0,18 4500 80 0,01 0,03 2 7 2 0
P11 1377 1377 0 0,00 1400 23 0,01 0,00 1 0 0 0
P12 1281 1281 0 0,00 1300 19 0,01 0,00 2 4 0 0
P13 1165 1165 0 0,00 1200 35 0,30 0,00 1 0 0 0
P14 14312 14298 14 0,10 14500 202 0,06 0,18 2 28 4 11
P15 17562 17559 3 0,02 18000 441 0,06 0,15 4 56 8 6
P16 13702 13697 5 0,04 13900 203 0,17 0,53 4 64 5 8
P17 6193 6188 5 0,08 6300 112 0,03 0,08 2 17 4 2
P18 5956 5954 2 0,03 6100 146 0,05 0,05 3 38 6 8
P19 13545 13544 1 0,01 13800 256 0,13 0,06 4 62 2 12
P20 27815 27796 19 0,07 28200 404 0,49 3,21 6 100 1 30
P21 28443 28430 13 0,05 28800 370 0,35 0,69 5 172 9 23
P22 49285 49279 6 0,01 49900 621 0,42 0,10 4 131 0 27
P23 37528 37525 3 0,01 38000 475 0,62 1,04 6 208 0 38
P24 28798 28795 3 0,01 29200 405 0,39 0,11 4 112 1 17
DEGREE0 11628 11628 0 0,00 11900 272 0,02 0,00 5 10 1 0
DEGREE1 24006 23927 79 0,33 24700 773 0,07 0,02 11 76 21 1
DEGREE2 31551 31498 53 0,17 32200 702 0,02 0,01 2 9 4 1
DEGREE3 33546 33546 0 0,00 34300 754 0,08 0,00 12 72 36 2
DEGREE4 32759 32680 79 0,24 33600 920 0,06 0,03 8 60 14 0
DEGREE5 48106 48037 69 0,14 49000 963 0,03 0,04 4 18 3 4
DEGREE6 39648 39591 57 0,14 40500 909 0,03 0,03 4 40 7 2
DEGREE7 45464 45366 98 0,22 46400 1034 0,03 0,12 3 29 3 3
DEGREE8 60471 60387 84 0,14 61400 1013 0,03 0,11 3 12 4 6
DEGREE9 35586 35563 23 0,06 36600 1037 0,19 0,05 8 130 50 3
DEGREE10 22761 22761 0 0,00 23500 739 1,18 0,02 26 559 168 2
DEGREE11 61812 61812 0 0,00 62900 1088 0,18 0,08 5 117 35 1
DEGREE12 50315 50315 0 0,00 51400 1085 0,15 0,03 5 86 46 6
DEGREE13 71466 71396 70 0,10 72500 1104 0,12 0,10 3 54 18 2
DEGREE14 94409 94384 25 0,03 95800 1416 0,13 0,03 3 32 11 8
DEGREE15 89919 89895 24 0,03 91300 1405 0,14 0,08 4 33 9 11
DEGREE16 105221 105153 68 0,06 106600 1447 0,20 0,09 4 77 17 14
DEGREE17 121855 121743 112 0,09 123500 1757 0,32 0,11 5 6 2 16
DEGREE18 70450 70385 65 0,09 71700 1315 22,20 0,62 54 1654 252 3
DEGREE19 71549 71507 42 0,06 72900 1393 25,18 0,21 15 359 140 7
DEGREE20 64957 64937 20 0,03 66200 1263 3,46 0,09 15 773 155 11
DEGREE21 113516 113486 30 0,03 115100 1614 1,33 0,32 6 213 43 23
DEGREE22 111532 111445 87 0,08 113200 1755 1,16 0,45 5 183 55 17
DEGREE23 115681 115626 55 0,05 117400 1774 1,53 0,13 8 184 43 27
DEGREE24 170142 170105 37 0,02 172300 2195 0,91 0,12 5 52 12 31
DEGREE25 145445 145386 59 0,04 147400 2014 1,07 0,54 6 125 22 32
DEGREE26 184434 184332 102 0,06 186800 2468 0,86 0,26 5 135 21 42
DEGREE27 88575 88507 68 0,08 90300 1793 56,90 1,93 28 1600 377 17
DEGREE28 101996 101918 78 0,08 103800 1882 208,33 2,81 74 5242 737 22
DEGREE29 88510 88394 116 0,13 90200 1806 298,31 6,68 134 8757 1388 19
DEGREE30 152955 152890 65 0,04 155200 2310 17,50 0,80 17 1669 140 43
DEGREE31 175918 175883 35 0,02 178300 2417 43,10 1,31 35 3048 176 47
DEGREE32 145719 145636 83 0,06 147800 2164 4,21 3,93 5 244 90 26
DEGREE33 213394 213254 140 0,07 216200 2946 7,44 9,99 7 393 29 53
DEGREE34 217857 217737 120 0,06 220600 2863 11,02 1,21 13 1360 41 52
DEGREE35 215469 215354 115 0,05 218200 2846 4,41 0,55 5 231 29 55

Table 4: Experiments RPP

19
PN LRV HV DIF PDRH TBHS TCHS LRT HT LRIN LRCR RSPRN RVPRN

GRID0 288 288 0 0,00 300 12 0,05 0,01 8 80 18 0


GRID1 486 486 0 0,00 500 14 0,05 0,02 6 30 11 0
GRID2 388 388 0 0,00 400 12 0,06 0,01 8 54 11 0
GRID3 784 784 0 0,00 800 16 0,02 0,01 2 7 8 0
GRID4 684 684 0 0,00 700 16 0,05 0,02 5 54 10 1
GRID5 687 686 1 0,15 700 14 0,04 0,02 7 58 14 1
GRID6 1281 1280 1 0,08 1300 20 0,03 0,06 3 10 5 1
GRID7 784 784 0 0,00 800 16 0,17 0,01 3 24 9 1
GRID8 884 882 2 0,23 900 18 0,04 0,04 5 50 11 2
GRID9 1076 1076 0 0,00 1100 24 1,78 0,02 28 511 221 3
GRID10 1271 1270 1 0,08 1300 30 0,85 0,05 14 408 66 2
GRID11 1471 1470 1 0,07 1500 30 2,37 0,12 16 243 111 4
GRID12 2559 2556 3 0,12 2600 44 0,43 0,42 8 246 13 11
GRID13 2262 2260 2 0,09 2300 40 0,75 0,10 16 347 65 9
GRID14 2460 2458 2 0,08 2500 42 0,35 0,34 6 138 21 8
GRID15 3453 3450 3 0,09 3500 50 0,12 0,23 3 25 7 10
GRID16 2957 2954 3 0,10 3000 46 0,15 0,26 4 113 10 6
GRID17 3352 3352 0 0,00 3400 48 0,11 0,03 3 67 6 11
GRID18 2353 2350 3 0,13 2400 50 15,06 0,21 35 1797 227 7
GRID19 2650 2648 2 0,08 2700 52 33,97 0,23 70 2939 312 5
GRID20 2648 2646 2 0,08 2700 54 38,84 0,15 72 2269 383 7
GRID21 4533 4530 3 0,07 4600 70 17,35 0,26 34 1753 130 20
GRID22 4633 4630 3 0,06 4700 70 2,04 0,26 8 401 37 13
GRID23 4926 4924 2 0,04 5000 76 0,88 0,40 5 250 33 16
GRID24 6711 6710 1 0,01 6800 90 1,83 0,23 8 75 14 27
GRID25 6017 6016 1 0,02 6100 84 1,27 0,20 5 148 13 27
GRID26 6517 6516 1 0,02 6600 84 1,02 0,13 5 225 12 21
GRID27 4023 4020 3 0,07 4100 80 140,89 0,68 71 3504 521 16
GRID28 4816 4814 2 0,04 4900 86 202,78 0,59 87 7680 438 25
GRID29 4320 4316 4 0,09 4400 84 360,29 1,51 192 8563 1208 18
GRID30 7193 7190 3 0,04 7300 110 25,91 0,72 21 1730 270 31
GRID31 7588 7584 4 0,05 7700 116 6,79 7,02 7 438 100 19
GRID32 8083 8078 5 0,06 8200 122 3,44 5,05 6 51 23 29
GRID33 11154 11152 2 0,02 11300 148 3,38 0,33 4 62 6 39
GRID34 10559 10556 3 0,03 10700 144 6,82 2,01 10 150 9 41
GRID35 10760 10756 4 0,04 10900 144 4,52 2,30 6 162 14 44
RANDOM0 595047 595047 0 0,00 624900 29853 0,15 0,01 16 154 30 1
RANDOM1 937883 937883 0 0,00 972200 34317 0,17 0,00 15 128 35 0
RANDOM2 814332 814332 0 0,00 839300 24968 0,22 0,01 20 126 76 1
RANDOM3 1713746 1713746 0 0,00 1747800 34054 0,08 0,00 8 61 24 4
RANDOM4 1651523 1651344 179 0,01 1677600 26256 0,14 0,01 11 98 22 2
RANDOM5 894942 894942 0 0,00 930700 35758 0,69 0,01 23 517 84 5
RANDOM6 2112266 2111642 624 0,03 2151000 39358 1,26 0,00 42 495 108 2
RANDOM7 1405960 1405960 0 0,00 1441100 35140 0,42 0,01 26 163 95 1
RANDOM8 2241571 2241571 0 0,00 2290300 48729 1,71 0,01 42 784 149 2
RANDOM9 2208128 2208128 0 0,00 2250700 42572 0,29 0,01 11 148 70 4
RANDOM10 2026759 2026759 0 0,00 2065500 38741 2,63 0,01 30 530 154 1
RANDOM11 1479373 1478872 501 0,03 1521500 42628 1,18 0,02 31 473 129 4
RANDOM12 1084255 1083201 1054 0,10 1134600 51399 1,66 0,05 32 809 87 3
RANDOM13 3462002 3461133 869 0,03 3516000 54867 0,45 0,05 9 94 32 3
RANDOM14 4138844 4137158 1686 0,04 4203100 65942 1,25 0,12 16 305 124 4
RANDOM15 3554511 3552188 2323 0,07 3616600 64412 11,83 0,10 53 1765 124 10
RANDOM16 1888551 1888551 0 0,00 1938900 50349 2,14 0,06 29 440 233 2
RANDOM17 1552828 1552828 0 0,00 1596100 43272 7,46 0,12 49 1199 305 3
RANDOM18 3509230 3509230 0 0,00 3566100 56870 5,35 0,01 38 1129 160 9
RANDOM19 3124737 3123439 1298 0,04 3169100 45661 0,35 0,04 5 66 42 2

Table 5: Experiments RPP (continuation)

20
PN LRV HV DIF PDRH TBHS TCHS LRT HT LRIN LRCR RSPRN RVPRN ORN

ALBAIDAAM 6324 5909 415 6,6 14400 8491 70,81 0,20 75 7607 88 50 0
ALBAIDABM 4405 4330 75 1,7 10723 6393 3,53 0,08 8 449 60 35 0
P01M 3 3 0 0,0 9 6 0,01 0,00 2 6 0 0 0
P02M 66 66 0 0,0 202 136 0,02 0,00 4 24 0 3 0
P03M 57 49 8 14,0 160 111 0,10 0,02 4 57 7 7 0
P04M 46 41 5 10,9 123 82 0,04 0,01 4 34 1 8 0
P05M 35 35 0 0,0 158 123 0,03 0,01 3 18 12 4 0
P06M 60 60 0 0,0 151 91 0,06 0,01 3 11 4 8 0
P07M 91 89 2 2,2 217 128 0,05 0,01 3 38 6 8 0
P08M 90 86 4 4,4 185 99 0,06 0,01 6 23 3 15 1
P09M 48 43 5 10,4 119 76 0,02 0,00 3 11 6 4 1
P10M 42 38 4 9,5 126 88 0,02 0,01 3 7 3 4 0
P11M 9 9 0 0,0 32 23 0,01 0,15 2 4 0 0 0
P12M 10 10 0 0,0 26 16 0,01 0,00 3 4 0 2 0
P13M 5 5 0 0,0 29 24 0,30 0,00 1 0 0 0 0
P14M 129 106 23 17,8 277 171 0,17 0,16 5 66 15 17 1
P15M 46 39 7 15,2 261 222 0,10 0,00 6 102 45 8 1
P16M 113 106 7 6,2 291 185 0,26 0,04 5 56 8 20 1
P17M 43 36 7 16,3 148 112 0,06 0,02 5 33 11 9 1
P18M 22 10 12 54,5 149 139 0,12 0,01 8 117 22 7 1
P19M 90 80 10 11,1 269 189 0,23 0,01 6 49 17 14 0
P20M 255 222 33 12,9 573 351 0,44 0,19 5 57 3 22 1
P21M 259 245 14 5,4 591 346 1,02 0,10 10 292 29 32 1
P22M 475 438 37 7,8 1103 665 0,50 0,04 5 120 4 35 0
P23M 360 360 0 0,0 762 402 0,34 0,04 4 160 8 30 0
P24M 238 226 12 5,0 586 360 0,68 0,06 9 195 15 28 1
DEGREE0M 109 109 0 0,0 223 114 0,01 0,00 1 0 0 0 0
DEGREE1M 115 115 0 0,0 873 758 0,05 0,00 8 71 19 4 0
DEGREE2M 274 274 0 0,0 948 674 0,02 0,00 2 10 8 2 0
DEGREE3M 172 172 0 0,0 627 455 0,04 0,00 5 30 18 4 0
DEGREE4M 210 210 0 0,0 1065 855 0,04 0,00 4 33 7 5 0
DEGREE5M 316 268 48 15,2 1113 845 0,04 0,01 5 27 9 7 0
DEGREE6M 166 166 0 0,0 921 755 0,03 0,00 4 31 10 4 0
DEGREE7M 275 235 40 14,5 1231 996 0,07 0,00 7 64 0 9 0
DEGREE8M 478 457 21 4,4 1187 730 0,05 0,00 5 27 11 10 1
DEGREE9M 160 160 0 0,0 757 597 0,44 0,00 14 250 86 9 0
DEGREE10M 0 0 0 0,0 0 0 0,23 0,00 11 192 34 4 0
DEGREE11M 409 392 17 4,2 1362 970 0,77 0,01 14 441 60 8 0
DEGREE12M 286 271 15 5,2 1146 875 0,29 0,02 8 161 103 10 0
DEGREE13M 728 662 66 9,1 1754 1092 0,42 0,02 9 207 61 11 0
DEGREE14M 810 810 0 0,0 2148 1338 0,20 0,02 4 70 33 12 0
DEGREE15M 722 649 73 10,1 1887 1238 0,31 0,02 7 99 15 20 1
DEGREE16M 1022 965 57 5,6 2405 1440 0,24 0,04 5 153 21 15 1
DEGREE17M 1122 1010 112 10,0 2580 1570 0,24 0,04 6 97 10 23 0
DEGREE18M 533 508 25 4,7 1677 1169 1,33 0,04 8 324 118 8 1
DEGREE19M 517 472 45 8,7 1623 1151 3,02 0,05 12 483 254 14 1
DEGREE20M 470 440 30 6,4 1553 1113 8,19 0,03 22 1156 166 12 1
DEGREE21M 1017 939 78 7,7 2540 1601 1,70 0,15 8 333 111 22 1
DEGREE22M 1014 875 139 13,7 2571 1696 1,68 0,40 7 195 62 18 0
DEGREE23M 995 878 117 11,8 2537 1659 4,11 0,09 13 528 70 31 1
DEGREE24M 1504 1384 120 8,0 3338 1954 0,85 0,07 4 125 34 30 0
DEGREE25M 1328 1313 15 1,1 3219 1906 1,02 0,05 5 251 42 35 1
DEGREE26M 1674 1499 175 10,5 3757 2258 1,57 0,09 6 44 8 36 1
DEGREE27M 573 497 76 13,3 2175 1678 51,76 0,13 36 2781 843 32 1
DEGREE28M 845 678 167 19,8 2413 1735 228,67 0,70 89 6004 938 32 1
DEGREE29M 582 424 158 27,1 2149 1725 42,34 1,07 25 1900 302 31 1
DEGREE30M 1414 1350 64 4,5 3485 2135 4,68 0,13 5 233 54 44 1
DEGREE31M 1549 1350 199 12,8 3721 2371 21,10 2,53 14 1240 141 54 0
DEGREE32M 1087 909 178 16,4 3119 2210 4,08 0,27 5 195 72 36 0
DEGREE33M 2090 1957 133 6,4 4880 2923 7,22 4,44 8 452 50 51 0
DEGREE34M 1828 1640 188 10,3 4256 2616 24,13 0,12 19 2085 62 61 0
DEGREE35M 1942 1783 159 8,2 4513 2730 11,52 0,56 9 378 34 62 0

Table 6: Experiments PRPP

21
PN LRV HV DIF PDRH TBHS TCHS LRT HT LRIN LRCR RSPRN RVPRN ORN

GRID0M 0 0 0 0,0 0 0 0,16 0,00 1 0 0 0 0


GRID1M 0 0 0 0,0 0 0 0,01 0,00 2 6 0 0 0
GRID2M 0 0 0 0,0 6 6 0,02 0,00 4 6 6 0 0
GRID3M 2 2 0 0,0 6 4 0,03 0,00 5 16 14 0 0
GRID4M 0 0 0 0,0 0 0 0,04 0,00 5 50 4 0 0
GRID5M 4 4 0 0,0 14 10 0,05 0,00 9 63 13 1 0
GRID6M 9 8 1 11,1 24 16 0,04 0,01 6 23 23 1 0
GRID7M 1 1 0 0,0 17 16 0,05 0,01 6 56 9 1 0
GRID8M 4 4 0 0,0 14 10 0,05 0,00 8 62 17 0 0
GRID9M 2 2 0 0,0 20 18 0,63 0,01 19 270 108 0 0
GRID10M 0 0 0 0,0 0 0 1,23 0,00 36 303 140 0 0
GRID11M 4 4 0 0,0 28 24 0,59 0,01 17 324 117 3 0
GRID12M 16 11 5 31,3 45 34 0,33 0,13 7 171 19 5 1
GRID13M 13 10 3 23,1 44 34 0,60 0,05 13 275 72 5 0
GRID14M 15 12 3 20,0 50 38 1,20 0,35 14 310 58 8 1
GRID15M 26 22 4 15,4 64 42 0,21 0,12 5 94 5 14 1
GRID16M 20 17 3 15,0 51 34 0,49 0,11 10 236 22 9 1
GRID17M 24 21 3 12,5 57 36 0,30 0,09 5 59 14 13 1
GRID18M 8 8 0 0,0 46 38 14,46 0,02 36 1579 289 4 0
GRID19M 8 2 6 75,0 36 34 32,05 0,07 71 2213 390 2 1
GRID20M 9 8 1 11,1 48 40 62,40 0,04 123 3250 573 12 0
GRID21M 33 30 3 9,1 92 62 11,66 0,54 30 1339 332 15 1
GRID22M 32 25 7 21,9 81 56 42,95 0,54 28 1434 129 16 1
GRID23M 33 31 2 6,1 93 62 25,48 0,20 12 496 129 13 1
GRID24M 58 56 2 3,4 132 76 1,72 0,13 6 207 46 23 0
GRID25M 48 44 4 8,3 112 68 5,11 1,32 18 1055 80 27 0
GRID26M 58 55 3 5,2 129 74 2,00 0,39 7 385 44 15 0
GRID27M 14 9 5 35,7 67 58 62,13 0,13 43 2191 461 9 1
GRID28M 24 20 4 16,7 96 76 365,76 0,69 113 6705 1151 16 0
GRID29M 15 10 5 33,3 84 74 191,26 2,15 82 3565 1040 11 1
GRID30M 51 47 4 7,8 137 90 6,81 2,81 7 418 204 20 1
GRID31M 54 52 2 3,7 138 86 16,20 0,93 13 656 162 22 0
GRID32M 60 55 5 8,3 155 100 3,60 5,23 5 206 52 24 0
GRID33M 94 90 4 4,3 210 120 6,35 0,91 7 425 30 34 0
GRID34M 86 81 5 5,8 199 118 7,67 2,23 10 580 66 36 1
GRID35M 90 84 6 6,7 200 116 7,46 0,99 7 395 57 42 1
RANDOM0M 1742 1742 0 0,0 15682 13940 0,06 0,00 10 49 10 4 0
RANDOM1M 4297 3853 444 10,3 13786 9933 0,07 0,00 8 83 11 7 1
RANDOM2M 5638 5638 0 0,0 34427 28789 0,08 0,00 9 81 21 1 0
RANDOM3M 18453 18453 0 0,0 49481 31028 0,09 0,00 8 60 34 3 0
RANDOM4M 17316 17316 0 0,0 43923 26607 0,06 0,00 6 48 15 5 0
RANDOM5M 298 298 0 0,0 30621 30323 0,63 0,01 22 469 107 12 0
RANDOM6M 12478 12478 0 0,0 53765 41287 1,00 0,01 35 536 98 5 0
RANDOM7M 9405 9405 0 0,0 48056 38651 0,27 0,01 15 184 94 10 0
RANDOM8M 14847 14847 0 0,0 65952 51105 0,56 0,01 19 438 53 4 0
RANDOM9M 17523 17523 0 0,0 56830 39307 0,27 0,01 10 191 46 6 0
RANDOM10M 17405 17405 0 0,0 57193 39788 1,24 0,01 19 448 89 7 0
RANDOM11M 7125 7125 0 0,0 35769 28644 0,58 0,00 20 228 132 4 0
RANDOM12M 1493 1493 0 0,0 33011 31518 0,60 0,00 19 331 39 6 0
RANDOM13M 32619 30323 2296 7,0 89523 59200 0,37 0,01 8 93 55 8 1
RANDOM14M 30757 27063 3694 12,0 75649 48586 0,50 0,01 8 172 53 12 0
RANDOM15M 27953 27481 472 1,7 101508 74027 2,38 0,02 19 688 49 18 1
RANDOM16M 10566 10533 33 0,3 68246 57713 1,25 0,02 15 392 117 6 0
RANDOM17M 4290 3296 994 23,2 10263 6967 3,51 0,00 49 1106 79 16 0
RANDOM18M 28510 27239 1271 4,5 84083 56844 0,58 0,02 8 254 67 13 1
RANDOM19M 26873 26873 0 0,0 77701 50828 0,22 0,01 3 1 2 5 0

Table 7: Experiments PRPP (continuation)

22

You might also like