Integer Programming Formulations for the Elementary Shortest Path Problem

Abstract

Given a directed graph with arbitrary arc costs, the Elementary Shortest Path Problem (ESPP) consists of finding a minimum-cost path between two nodes s and t such that each node of G is visited at most once. If negative costs are allowed, the problem is -hard. In this paper, several integer programming formulations for the ESPP are compared. We present analytical results based on a polyhedral study of the formulations, and computational experiments where we compare their linear programming relaxation bounds and their behavior within a branch-and-cut framework. The computational results show that a formulation with dynamically generated cutset inequalities is the most effective.

Full text

Integer programming formulations for the elementary
shortest path problem
Leonardo Taccari
Dipartimento di Elettronica, Informazione e Bioingegneria, Politecnico di Milano, Italy
Abstract
Given a directed graph G= (V, A)with arbitrary arc costs, the Elementary
Shortest Path Problem (ESPP) consists of finding a minimum-cost path be-
tween two nodes sand tsuch that each node of Gis visited at most once. If
negative costs are allowed, the problem is NP-hard. In this paper, several inte-
ger programming formulations for the ESPP are compared. We present analyt-
ical results based on a polyhedral study of the formulations, and computational
experiments where we compare their linear programming relaxation bounds and
their behavior within a branch-and-cut framework. The computational results
show that a formulation with dynamically generated cutset inequalities is the
most effective.
Keywords: integer programming, elementary shortest path, branch-and-cut,
extended formulations, subtour elimination constraints, generalized cutset
inequalities
1. Introduction
Given a directed graph G= (V, A)and arc costs cij for each (i, j)∈A,
the shortest path problem consists of finding a minimum-cost path between two
nodes sand t.
Often, an implicit assumption is that such path has to be elementary. A
path is elementary if it does not visit any node more than once, i.e., if it does
not contain subtours. When the costs cij induce no negative cycles on G, the
problem can be solved efficiently via ad-hoc polynomial time algorithms, like
Bellman-Ford’s or Dijikstra’s algorithm (if cij ≥0). However, if negative cy-
cles do arise, subtours must be explicitly prevented, leading to the so-called
Elementary Shortest Path Problem (ESPP).
The ESPP is clearly NP-hard due to a simple reduction from the Hamilto-
nian path problem. Its equivalent maximization counterpart, where one seeks a
longest path over a graph with positive cycles, has been vastly discussed in the
Email address: leonardo.taccari@polimi.it (Leonardo Taccari)
Preprint submitted to Elsevier December 21, 2015

literature, usually referred to as the Longest Path Problem (LPP). Björklund et
al. [7] prove that the LPP is hard to approximate on unweighted directed graphs
within a n1−for any , unless P=NP. For undirected graphs, approximation
algorithms are described in, e.g., [27] and [17].
Related to the ESPP are variants of the Travelling Salesman Problem (TSP)
with profits, such as the Prize Collecting TSP [3], the Orienteering Problem [36]
and the Capacitated Profitable Tour Problem [26], that involve prizes on each
node, which can be visited at most once.
While an interesting problem in its own right, the ESPP also arises in the
pricing subproblems of branch-and-price algorithms [1]. Often, the pricing phase
in Vehicle Routing Problems (VRP) involves resource-constrained variants of the
elementary shortest path problem (ESPPRC) [23]. These problems are usually
solved with fast dynamic programming-based label algorithms, e.g., see [32, 15,
8]. It is possible to adapt this kind of approach to the unconstrained ESPP by
considering an artificial resource for each node which is consumed when the node
is visited, and imposing that no more than one unit of each resource is used, as
already proposed in [6, 8]. However, this is a rather weak constraint, in the sense
that it allows for very long paths, so that approaches based on label algorithms
become very time consuming, as noted by Drexl and Irnich [14]. This limitation
is already highlighted for the ESPP with a capacity constraint by Jepsen et
al. [25], that propose a branch-and-cut algorithm that significantly outperforms
label algorithms. In the context of a branch-and-price algorithm, where the
ESPP is solved repeatedly in the pricing phase, another desirable feature of
an integer programming approach is its flexibility, that allows one to easily
incorporate general branching decisions or valid inequalities (e.g., the subset-row
inequalities in [24]) that would change the structure of the pricing subproblem.
These reasons motivate the study of integer programming techniques for the
ESPP.
Several branch-and-cut approaches can be found in the literature for related
problems [5, 19, 16, 26]. However, not much previous work has appeared on
integer programming approaches tailored specifically for the ESPP. Ibrahim et
al. [22] provide computational results on the linear programming (LP) bounds
of a flow-based extended formulation, but give no details on its behavior in
a branch-and-bound algorithm. Another extended formulation is proposed by
Haouari et al. [21]. Drexl and Irnich [14] describe a branch-and-cut approach
and compare its efficiency with the extended formulation in [22], while Drexl [13]
studies the efficient separation of subtour elimination constraints for the ESPP.
In the context of integer programming, the choice of a formulation is crucial
for the effectiveness of methods based on branch-and-cut. In this article we
present a thorough comparison between different formulations for the ESPP,
which are described in detail in Section 2. We include integer programming for-
mulations with exponentially many subtour elimination constraints, and mixed-
integer programming extended formulations with a polynomial number of vari-
ables and constraints. In Section 3 we provide some analytical results, including
a proof of equivalence between the polyhedra described by the two strongest for-
mulations. Section 4 reports computational experiments where we compare the
2

LP relaxation bounds and branch-and-cut results. Finally, in Section 5, we give
some concluding remarks and discuss further research topics.
2. Integer programming formulations
Let us consider a directed graph G= (V, A)with the set of nodes Vand the
set of arcs A. Let nand mbe the cardinality of Vand A, respectively. A path
is a sequence of nodes v1, . . . , vk, and is said to be elementary if no node appears
in the path more than once. A cycle, or tour, is a path with v1=vk. We denote
by δ+(i)and δ−(i)the set of outgoing and incoming arcs of node i, by δ+(S)
and δ−(S)the arcs leaving/entering the set S⊆V, and by A(S)the set of arcs
with both ends in S⊆V. Let us also define Vi:= V\{i},Vij := V\{i, j}, and,
for any arc set B⊆A, we define x(B) := Pb∈Bxb. In all the formulations, it
is assumed w.l.o.g. that |δ−(s)|=|δ+(t)|= 0.
A standard integer programming formulation to determine a shortest path
from node sto node tis the following:
min X
(i,j)∈A
cij xij (1)
X
(i,j)∈δ+(i)
xij −X
(j,i)∈δ−(i)
xji =




1if i=s
−1if i=t
0else ∀i∈V(2)
X
(i,j)∈δ+(i)
xij ≤1∀i∈V(3)
xij ∈ {0,1} ∀(i, j)∈A, (4)
where cij ∈Rare the arc costs, and xij are binary arc variables that take value
1 if the arc (i, j)belongs to the path. Constraints (2) are flow conservation
constraints, while Constraints (3) ensure that the outgoing degree of each node
is at most one. When the costs cij induce negative cycles on G, i.e., there is a
subtour such that the total cost of its arcs is negative, this system of inequalities
is not sufficient to guarantee the elementarity of the path. Thus, additional
constraints (and possibly variables) are necessary to prevent subtours.
Notice that the crucial difference with respect to problems in which the path
has to be Hamiltonian is the absence of the degree constraints:
X
(i,j)∈δ+(i)
xij =X
(j,i)∈δ−(i)
xji = 1 ∀i∈Vst.
We now describe different sets of constraints and variables that can be added
to Formulation (1)–(4) to obtain a valid integer programming formulation for
the ESPP.
3

2.1. Dantzig-Fulkerson-Johnson (DFJ)
For the TSP, success has been achieved with strong formulations with ex-
ponentially many constraints. It is possible to write a formulation for the
ESPP based on the classical Dantzig-Fulkerson-Johnson subtour elimination
constraints [10], adding to the basic formulation (1)–(4) the following inequali-
ties:
X
(i,j)∈A(S)
xij ≤ |S| − 1∀S⊆Vst,|S| ≥ 2.(5)
In each subset S, subtours are prevented ensuring that the number of arcs in S
which are selected is smaller than the number of nodes in S. This formulation
includes O(m)variables and O(2n)constraints.
Observation. For Hamiltonian path problems, due to the degree constraints, the
DFJ subtour eliminations constraints can be equivalently written in the cutset
form:
X
(i,j)∈δ+(S)
xij ≥1∀S⊂V, |S| ≥ 2.(6)
For the ESPP, Constraints (6) are valid only for subsets Swith s∈Sand
t /∈S. Moreover, they are not sufficient to prevent all subtours.
Example 1. Consider the solution depicted in Figure 1, assuming it is a com-
plete graph and that only the arcs with xij = 1 are drawn. The solution does
not violate any inequality (6) for any set Scontaining s, since x(δ+(S)) = 1 for
any such S, although it contains a (disconnected) subtour. On the other hand,
notice that the inequality (6) is not valid for S0, although it does not contain
subtours, due to x(δ+(S0)) = 0.
SS0
st
Figure 1: An example where Constraints (6) are not sufficient to prevent subtours.
2.2. Generalized cutset inequalities (GCS)
DFJ Constraints (6) can be adapted to the ESPP by replacing the constant
right-hand side with a variable expression. This approach is used for a sym-
metric version of ESPPRC by Jepsen et al. [25] and applied to the asymmetric
ESPP by Drexl and Irnich [14]. Similar subtour elimination constraints are also
4

used in branch-and-cut algorithms for the VRP [29] or variants of the TSP with
profits [16, 26]. We refer to them as generalized cutset inequalities (GCS):
X
(i,j)∈δ+(S)
xij ≥X
(k,j)∈δ+(k)
xkj ∀k∈S, ∀S⊆Vst,
|S| ≥ 2.(7)
Constraints (7) prevent subtours by ensuring that, for each subset S, the number
of selected arcs leaving Sis not smaller than the number of selected arcs outgoing
from any node in S. In an integer solution, this means that the cut induced
by Smust contain at least one arc if at least one node in Sbelongs to the s-t
path, while, if Sdoes not contain any node in the s-tpath, the constraint is the
trivial inequality. The number of variables in the formulation is O(m), while
the number of constraints is O(n2n).
Constraints (7) can be shown to be equivalent to a strengthened version of
(5).
Proposition 2. Constraints (7) can be rewritten as:
X
(i,j)∈A(S)
xij ≤X
i∈S\{k}X
(i,j)∈δ+(i)
xij ∀k∈S, ∀S⊆Vst,
|S| ≥ 2.(8)
Proof. For all k∈S,x(δ+(k)) ≤x(δ+(S)) = x(δ+(S)) + x(A(S)) −x(A(S)) =
Pi∈Sx(δ+(i)) −x(A(S)).
These inequalities can be interpreted as imposing that the number of selected
arcs in a subset Sis strictly smaller than the number of nodes in Sthat belong
to the s-tpath.
2.3. Sequential formulation (MTZ)
To derive an extended formulation à la Miller, Tucker and Zemlin [28] (here-
after MTZ) it is enough to introduce, for each node, an auxiliary variable that
can be viewed as the position of the node along the path and a constraint for
each arc:
tj≥ti+ 1 + (n−1)(xij −1) ∀(i, j)∈A,
i6=s, j 6=t. (9)
For the Asymmetric TSP (ATSP), this formulation is well-known to give poor
linear relaxation bounds. However, it is very compact, as it requires only O(m)
additional constraints and O(n)auxiliary variables.
2.4. Reformulation-linearization based formulation (RLT)
From the following nonlinear reformulation of the MTZ formulation:
tjxij = (ti+ 1)xij ∀(i, j)∈A, i 6=s(10)
tjxsj =xsj ∀(s, j)∈δ+(s)(11)
1≤ti≤n−1, i ∈Vs(12)
5

Haouari et al. [21] apply a partial Sherali-Adams reformulation-linearization
technique [33] to obtain the following stronger formulation for the ESPP:
αij =βij +xij ∀(i, j)∈A(13)
xsj +X
(i,j)∈δ−(j)
i6=s
αij −X
(j,i)∈δ+(j)
βji = 0 ∀j∈Vt,
(s, j)∈δ+(s)(14)
X
(i,j)∈δ−(j)
αij −X
(j,i)∈δ+(j)
βji = 0 ∀j∈Vst,
(s, j)/∈δ+(s)(15)
αij ≤(n−1)xij ∀(i, j)∈A, i 6=s(16)
xij ≤βij ∀(i, j)∈A, i 6=s(17)
αij ≥0, βij ≥0∀(i, j)∈A, (18)
where the bilinear terms are linearized introducing the variables αij := tjxij and
βij := tixij and Constraints (16)–(17). On a given selected arc (i, j)∈A, the
variables βij and αij can be interpreted respectively as the position of the nodes
iand jalong the path. This extended formulation requires O(m)constraints
and O(m)auxiliary variables.
2.5. Single-flow formulation (SF)
A formulation similar to the single-flow ATSP formulation of Gavish and
Graves [18] can be obtained introducing an auxiliary flow qto be delivered to
the nodes belonging to the s-tpath. In addition, variables zkare added to the
formulation:
qij ≤(n−1)xij ∀(i, j)∈A(19)
X
(s,j)∈δ+(s)
qsj =X
k∈Vs
zk(20)
X
(i,k)∈δ−(k)
qik −X
(k,j)∈δ+(k)
qkj =zk∀k∈Vs(21)
X
(i,k)∈δ−(k)
xik =zk∀k∈Vs(22)
qij ≥0∀(i, j)∈A(23)
zk∈ {0,1} ∀k∈Vs.(24)
Constraints (19) impose that the auxiliary flow is positive only over the arcs
where xij = 1. The auxiliary flow leaving from the node shas value equal
to the number of nodes that are reached by the s-tpath. Constraints (21)
ensure that the balance of the auxiliary flow on each node is equivalent to zk,
which, according to Constraint (22), is either 1, if node kis in the s-tpath, or
0 otherwise.
6

2.6. Multicommodity-flow formulation (MCF)
An extension of the single-flow formulation is obtained by disaggregating the
auxiliary flow into n−1unitary flows. Subtours are prevented by enforcing one
unit of a distinct auxiliary flow from sto each node that belongs to the s-tpath:
qk
ij ≤xij ∀k∈Vs,
(i, j)∈A(25)
X
(i,j)∈δ+(i)
qk
ij −X
(j,i)∈δ−(i)
qk
ji =




zkif i=s
−zkif i=k
0else
∀i∈V,
∀k∈Vs(26)
X
(i,k)∈δ−(k)
xik =zk∀k∈Vs(27)
X
(s,j)∈δ+(s)
xsj = 1 (28)
X
(i,t)∈δ−(t)
xit = 1 (29)
qk
ij ≥0∀k∈Vs,
(i, j)∈A(30)
zk∈ {0,1} ∀k∈Vs.(31)
The formulation includes O(nm)additional variables and constraints. This
extended formulation is introduced, for the ESPP, by Ibrahim et al. [22], and
it is very similar to classic multi-commodity flow formulations for the ATSP
proposed by Wong [37] and Claus [9].
2.7. Overview
In Table 1 we summarize the presented formulations. To the best of our
knowledge, formulations MTZ and SF have not been previously considered for
the ESPP, although similar ones are well known for TSP or VRP variants.
Table 1: A summary of the considered formulations.
number of
variables constraints description
DFJ O(m)O(2n)Dantzig-Fulkerson-Johnson (5)
GCS O(m)O(n2n)generalized cutsets (7)
MTZ O(m)O(m)Miller-Tucker-Zemlin (9)
RLT O(m)O(m)reformulation-linearization (13)–(18)
SF O(m)O(m)single-flow (19)-(24)
MCF O(nm)O(nm)multi-commodity flow (25)-(31)
7

3. Polyhedral results
Let us describe some analytical results for the considered ESPP formulations.
Proposition 3. Formulation MCF is stronger than formulation SF.
Proof. Constraints (19)–(21) can be obtained from MCF simply aggregating
Constraints (25)–(26) over k∈Vs, and then substituting Pk∈Vsqk
ij with qij .
The example in Figure 2 shows that the inclusion is strict.
Proposition 4. Formulation GCS is stronger than formulation DFJ.
Proof. The result follows by considering GCS as stated in (8), whose right-
hand side is obviously smaller or equal to |S| − 1, right-hand side in (5), and
the inclusion is strict by the example in Figure 2.
s
t
a
b
c
−20 −10
1
1
−10
−10
Figure 2: Example proving strict inclusion for Proposition 3 and 4. With GCS and MCF, the
LP optimal solution is the one with xst = 1 and optimal value −20. With SF, the optimal
solution has value −26, with xst =xca =1
4,xsa =xct =3
4and xab =xbc = 1. With DFJ, the
solution has value −40, with xst = 1 and a disconnected subtour with xab =xbc =xca =2
3.
Showing that formulation MCF is as tight as GCS requires to calculate the
projection of the MCF extended formulation into the space of the xvariables.
We will make use of a strong result presented by Padberg and Sung in [31], and
follow a similar approach to the equivalence proofs therein.
Theorem 5. The projection of the MCF-polytope onto the x-space is equivalent
to the GCS-polytope.
Proof. Recall that |δ−(s)|=|δ+(t)|= 0. The variables zkcan be projected out
8

of (25)–(31) so that MCF can be rewritten as:
qk
ij ≤xij ∀k∈Vs,
∀(i, j)∈A(32)
X
(i,j)∈δ+(i)
qk
ij −X
(j,i)∈δ−(i)
qk
ji = 0 ∀i∈Vs, i 6=k,
∀k∈Vs(33)
X
(s,j)∈δ+(s)
qk
sj =X
(i,k)∈δ−(k)
xik ∀k∈Vs(34)
X
(j,k)∈δ−(k)
qk
jk =X
(i,k)∈δ−(k)
xik ∀k∈Vs(35)
X
(i,k)∈δ+(k)
xik =X
(i,k)∈δ−(k)
xik ∀k∈Vst (36)
X
(s,j)∈δ+(s)
xsj = 1 (37)
X
(i,t)∈δ−(t)
xit = 1 (38)
qk
ij ≥0, xij ≥0∀(i, j)∈A, k ∈Vs.(39)
In order to compare MCF and GCS, we need to project out also the q-variables
of the MCF formulation. Let us define the sets:
X={x∈Rm|xsatisfies (36), (37) and (38)},
PGCS ={x∈X|xsatisfies (7)},
PMC F ={(x, q)∈Rmn |(x, q)satisfies (32)–(39)},
P rojx(PM CF ) = {x∈X| ∃ qs.t. (x, q)∈PM CF },
where PGCS is the GCS -polytope, PM CF is the MCF -polytope and P rojx(PM CF )
is its projection onto the x-space. It is convenient to rewrite Constraints (32)–
(35) in matrix form as follows:
Bx +M q = 0 (40)
−Dx +Iq ≤0(41)
x, q ≥0.(42)
Equation (40) corresponds to (33)–(35), while (41) corresponds to (32). The
matrices B,M,Dand Ican be decomposed according to the index k. Each
block Mkrepresents the node-arc incidence matrix of the graph G.Ikare
identity matrices of dimension m×m. Each submatrix Bkhas zeros everywhere,
except for the row corresponding to node s, with entries of value −1for each arc
in δ−(k), and the row corresponding to node k, with +1 entries for each arc in
δ−(k). Each row of Dkcorresponds to a variable qk
ij and has zeros everywhere,
except for a +1 in the column associated with variable xij.
9

The projection onto the x-space of the polytope PM CF defined by (40)–(41)
can be obtained as follows (see, e.g., [4]):
P rojx(PM CF ) = {x∈X|(uB −vD −w)x≤0
∀(u, v, w)∈C},(43)
where Cis the cone defined as:
C={(u, v, w)|uM +vI ≥0, v ≥0, w ≥0}.
The result allows us to carry out the comparison between PGCS and P rojx(PM CF )
simply by finding a system of generators for the cone C.
From the inequalities w≥0we obtain extreme rays of the form u= 0,
v= 0,w=ei, where eiis the i-th standard basis vector of Rm, that yield the
nonnegativity constraints
xij ≥0∀(i, j)∈A. (44)
This allows us to restrict our following study to the cone C0defined as:
C0={(u, v)|uM +vI ≥0, v ≥0}.
Exploiting the decomposition of M, we can work on the even smaller cones:
Ck={(uk, vk)|ukMk+vk≥0, vk≥0}.(45)
Due to (43), once we have the system of generators (uk, vk)for each cone Ck,
the constraints in the x-space are obtained by calculating (ukBk−vkDk)x≤0
for each k∈Vs.
According to Proposition 6 in [31], a full system of generators of a cone Ck
defined as in (45), where Mkis a node-arc incidence matrix of a digraph, is
given by:
- a basis of its lineality space, of the form:
uk=±e, vk= 0,
where eis the all-ones vector, that in our case translate to the trivial equality
0 = 0, and
- the extreme rays, given by all the positive multiples of the vector (uk, vk)such
that:
(i)uk
i= 0 ∀i∈V, vk
ij =(1for one (i, j)∈A
0otherwise
(ii)uk
i=(1∀i∈S,
0otherwise vk
ij =(1∀i∈¯
S, j ∈S,
0otherwise
(iii)uk
i=(−1∀i∈S,
0otherwise vk
ij =(1∀i∈S, j ∈¯
S,
0otherwise
10

for any S⊆V, where ¯
S=V\S. The extreme rays of the form (i)give rise to
nonnegativity constraints.
From the extreme rays given by (ii)and (iii)we obtain the inequalities:
−x(δ−(S)) + ukx(δ−(k)) −usx(δ−(k)) ≤0(46)
−x(δ+(S)) −ukx(δ−(k)) + usx(δ−(k)) ≤0(47)
where ui= 1 if i∈S, and 0 otherwise. For both (46) and (47), we can
distinguish four cases depending on whether sand kare in S, thus whether us, uk
are 0 or 1. If both sand kare in S, or neither of them is, the inequality is implied
by the nonnegativity constraints (44). If only the coefficient with negative sign
is nonzero, the corresponding inequality is, again, redundant. Therefore, the
only meaningful cases are the following:
x(δ−(S)) ≥x(δ−(k)) ∀S⊆V, s /∈S, ∀k∈S(48)
x(δ+(S)) ≥x(δ−(k)) ∀S⊆V, s ∈S, ∀k /∈S. (49)
We have established so far that P rojx(PM CF )is fully described by the nonneg-
ativity constraints and Constraints (48)–(49). This set of inequalities can be
shown to be equivalent to:
x(δ+(S)) ≥x(δ+(k)) ∀S⊆Vst, k ∈S. (50)
Constraints (48) and (49) are equivalent, due to the fact that x(δ+(S)) =
x(δ−(¯
S)). Let us then consider only (48). For k=t, the inequality is trivially
satisfied by all x∈X, thus redundant. For k6=tand t /∈S, we obtain exactly
the inequalities in (50), since by (36)–(38), we have that x(δ−(k)) = x(δ+(k))
and x(δ−(S)) = x(δ+(S)) for any Scontaining neither snor t. If k6=tand
t∈S, it suffices to observe that, since δ+(t) = 0, the inequality x(δ−(S)) ≥
x(δ−(k)) is implied by x(δ−(S\{t})) ≥x(δ−(k)), which, again, can be rewritten
in the form (50).
Hence, the projection of PM CF onto the x-space is given by
P rojx(PM CF ) = {x∈X|xsatisfies (36)–(38) and (50)},
and it follows that P rojx(PMCF ) = PGCS .
From this result and Proposition 4, it also follows that formulation MCF is
stronger than formulation DFJ.
4. Computational comparison
Let us now compare the described formulations with respect to their LP re-
laxation bounds and their behavior within an exact branch-and-cut framework.
Formulation DFJ is not included in the tests, as it is clearly dominated by GCS.
Four types of instances are considered in the tests. The first benchmark set
consists of instances from the pricing phase of the unsplittable flow problem
11

in [1, 34] on small-sized networks from the SNDlib [30], namely, the topologies
atlanta,france,geant,germany and nobel-us.
The second one is a set of small to medium-sized random-cost graphs, either
sparse (rnd-s) or dense (rnd-d). The graphs for the instances in rnd-s are
generated by building a connected component including all the nodes, and then
randomly adding arcs until the desired density is reached. The instances in
rnd-d are the dense instances in [13], with random arc costs on a complete
graph.
The third benchmark set (prc) contains the pricing instances in [13]. It
consists of small and medium-size pricing problems from a column generation
algorithm for the asymmetric m-salesmen TSP at the first (f), penultimate (p)
and last (l) pricing iterations.
The fourth set (rome99) contains a part of the directed road network of the
city of Rome, Italy, used in the 9th DIMACS Implementation Challenge on
Shortest Paths [11]. Since all the arcs have a positive cost, representing the
distance in meters, we flip their sign (i.e., we solve a longest path problem over
the original graph). To generate distinct instances over the same graph, we
sample randomly 30 (s, t)pairs from V.
Table 2 summarizes the features of the test instances. For each subset, we
have 30 instances, for a total of 690.
Table 2: Description of the instances.
n m range of arc costs number of
instances
nobel-us 14 42 [−10000,10000] 30
atlanta 15 44 [−10000,10000] 30
geant 22 72 [−10000,10000] 30
france 25 90 [−10000,10000] 30
germany 50 176 [−10000,10000] 30
rnd-s 50/100/200 164/660/2654 [−1000,1000] 30/30/30
rnd-s 500/1000 16634/66601 [−1000,1000] 30/30
rnd-d 25/50/100 600/2450/9900 [−1000,1000] 30/30/30
prc-f 27/52/102 702/2652/10302 [−108,−9.5·107] 30/30/30
prc-p 27/52/102 702/2652/10302 [−4·104,5.2·106] 30/30/30
prc-l 27/52/102 702/2652/10302 [−4·104,5.2·106] 30/30/30
rome99 3353 8870 [−13000,−1] 30
4.1. Linear programming relaxation bounds
The LP relaxation bounds are computed constructing the complete model for
the extended formulations MTZ,RLT and SF. For formulations GCS and MCF
we use a Min Cut-based separation procedure (its implementation details are left
to the next section). Note that we use a delayed row-generation algorithm also
to solve MCF since its size is rather large, although polynomial, and preliminary
experiments indicated this is an effective strategy. The tests are carried out with
IBM Ilog Cplex 12.6 on an Intel Xeon E5645 @2.40GHz.
12

Table 3 reports the average gap of the linear relaxation bounds with respect
to the optimal integer values, computed as 100 |LB−Opt|
|LB|, the number of optimal
integer solutions, and the average computing time. The results confirm that the
LP bounds of MCF and GCS are equivalent, and show that they are by far the
tightest formulations. MCF and GCS have gap 0 on 40% of the instances and
find an optimal integer solution on almost 30% of the instances. The remaining
formulations have a relaxation with gap 0 in less than 10% of the considered
instances, and find an optimal integer solution in less than 2% of the cases.
Formulations RLT and SF provide similar bounds. Considering the computing
time, MTZ is solved more quickly than all the other formulations, while RLT
and SF are challenging for large networks. On the largest instances, the LP
relaxation of the MCF formulation could not be solved within the time limit of
1200 seconds.
It is worthwhile to point out that, even in the cases where the bounds are
very good with all the formulations (e.g., on the prc-f instances), weaker for-
mulations provide solutions with many more fractional values, as reflected by
the smaller number of optimal integer solutions.
Table 3: Average LP relaxation gaps (%) and number of optimal integer solutions. Missing
values are due to time limit. rome99 instances are not included, since optima are not available.
GCS MTZ RLT SF MCF
%gap opt time %gap opt time %gap opt time %gap opt time %gap opt time
nobel-us 26.5 19 0.0 72.0 0 0.0 69.1 0 0.0 69.7 0 0.0 26.5 19 0.0
atlanta 9.9 23 0.0 42.8 0 0.0 36.5 0 0.0 37.5 0 0.0 9.9 23 0.0
geant 5.5 14 0.0 32.0 0 0.0 30.9 0 0.0 31.1 0 0.0 5.5 14 0.0
france 4.3 22 0.0 64.5 0 0.0 59.9 0 0.0 60.4 0 0.0 4.3 22 0.0
germany 2.0 2 0.4 31.5 0 0.0 31.3 0 0.0 31.3 0 0.0 2.0 2 0.4
rnd-s-50 1.7 4 0.1 13.1 0 0.0 12.5 0 0.0 12.6 0 0.0 1.7 4 0.5
rnd-s-100 0.2 2 0.2 1.5 0 0.0 1.5 0 0.0 1.5 0 0.0 0.2 2 7.1
rnd-s-200 0.0 0 0.5 0.4 0 0.0 0.4 0 0.5 0.4 0 0.4 0.0 0 345.5
rnd-s-500 0.0 0 5.6 0.1 0 0.8 0.1 0 68.5 0.1 0 29.7 – – –
rnd-s-1000 0.0 2 51.9 0.0 0 37.4 0.0 0 792.5 0.0 0 319.8 – – –
rnd-d-100 0.0 5 0.2 0.0 3 0.1 0.0 3 4.1 0.0 3 3.5 0.0 5 84.2
rnd-d-25 0.2 15 0.0 0.6 3 0.0 0.5 4 0.0 0.5 3 0.0 0.2 15 0.0
rnd-d-50 0.0 6 0.0 0.1 3 0.0 0.1 3 0.1 0.1 3 0.2 0.0 6 2.4
prc-f-25 0.0 13 0.0 0.0 0 0.0 0.0 0 0.0 0.0 0 0.0 0.0 13 0.7
prc-f-50 0.0 3 0.2 0.0 0 0.0 0.0 0 0.2 0.0 0 0.2 0.0 3 44.9
prc-f-100 0.0 0 2.5 0.0 0 0.0 0.0 0 11.0 0.0 0 4.1 – – –
prc-p-25 10.1 15 0.1 91.5 0 0.0 87.1 0 0.0 87.4 0 0.0 10.1 15 2.6
prc-p-50 8.2 6 1.1 80.6 0 0.0 77.5 0 0.2 77.6 0 0.2 8.2 6 89.3
prc-p-100 1.8 0 14.0 60.3 0 0.0 42.1 0 11.6 42.3 0 4.1 – – –
prc-l-25 1.4 24 0.2 86.5 0 0.0 80.8 0 0.0 81.1 0 0.0 1.4 24 2.8
prc-l-50 2.3 8 2.1 55.6 0 0.0 49.8 0 0.2 49.9 0 0.2 2.3 8 53.6
prc-l-100 2.2 0 12.1 60.4 0 0.0 42.1 0 11.0 42.4 0 4.0 – – –
4.2. Branch-and-cut
Since we aim at integer solutions, let us compare the behavior of a state-of-
the-art MIP solver using the considered formulations. The formulations and the
13

separation procedures are implemented in C++ with IBM Ilog Cplex/Concert
12.6, using default settings for the branch-and-cut.
For the polynomial-size extended formulations MTZ,RLT and SF, the full
model is built.
For GCS, we report results obtained with two different separation routines.
In the approach denoted by GCS-StrongComp, the separation is carried out, on
both fractional and integer solutions, identifying the strongly connected compo-
nents in the support graph induced by the variables xij . This can be done in a
O(n+m)running time with Tarjan’s algorithm [35]. Once a strong component
Shas been found, it is enough to check if Constraint (7) is violated for any of the
nodes in S. This separation procedure is efficient, but not guaranteed to find all
the violated inequalities on fractional solutions. Correctness is preserved by the
fact that the procedure is exact for integer solutions. In the approach denoted
by GCS-MinCut, the separation is carried out on fractional solutions by solving
a sequence of Min Cut (or Max Flow) problems between each node and t. Solv-
ing n−1Min Cut problems yields an overall worst-case complexity of O(n3√m)
using Goldberg-Tarjan’s highest-label preflow-push algorithm [20]. This way, all
violated inequalities are identified, although with a higher computational cost.
Note that, on integer solutions, the faster strong component-based procedure is
sufficient, and, on fractional solutions, it is computationally convenient to try
the strong components procedure first, and resort to the Min-Cut separation
only if the heuristic finds no violated inequality. The same separation proce-
dures are also used for MCF: when a violation is found for a node k, we add
the full set of Constraints (25)–(27) corresponding to that node.
To solve the Min Cut problems and identify the strongly connected compo-
nents, we use the efficient implementations in the open-source LEMON Graph
Library 1.3 [12]. We refer the interested reader to [13] for additional considera-
tions on the separation of subtour elimination constraints for the ESPP.
A remark is in order. In a branch-and-cut algorithm, the generation of the
cutting planes must be balanced with respect to the branching: adding too
many inequalities may hinder the solution of the LPs in the nodes, although
better bounds result in fewer explored nodes. We use two parameters to control
the trade-off between the quality of the lower bounds and the computing time to
solve the LPs. Specifically, given a solution x, we only consider the inequalities
with a violation not smaller than ε(correctness is preserved on integer solutions
for ε < 1), and we add at most νof them (in particular, we select the first
νmaximally violated). In Tables 4 and 5 we summarize a tuning procedure
that is carried out on a subset of 120 medium-size instances to identify the best
parameters for GCS-StrongComp and GCS-MinCut. We report the geometric
mean of the time to optimality, the number of nodes and the number of added
cuts. The values are normalized, for each instance, with respect to the results
with ε= 0.001 and ν= 1. The tables indicate that, in both cases, it is
convenient to add all the inequalities that are violated by the given tolerance
ε. According to these results, in the following experiments, we use ε= 0.8and
m=all for GCS-MinCut, and ε= 0.2, m =all for GCS-StrongComp.
Concerning the row generation for MCF, recall that for every violation we
14

Table 4: Tuning of εand νfor GCS-StrongComp.
ε0.001 0.1 0.2 0.4 0.8
νtime/nodes/cuts time/nodes/cuts time/nodes/cuts time/nodes/cuts time/nodes/cuts
1 1.00/1.00/1.00 1.00/1.05/1.00 0.99/1.17/0.97 0.98/1.22/0.92 1.00/1.70/0.90
5 0.90/0.68/1.11 0.90/0.73/1.11 0.87/0.81/1.10 0.89/0.94/1.08 0.86/1.20/1.06
10 0.83/0.60/1.06 0.79/0.61/1.06 0.79/0.66/1.05 0.80/0.74/1.02 0.79/1.01/1.01
20 0.85/0.58/1.08 0.78/0.60/1.07 0.77/0.66/1.05 0.77/0.74/1.04 0.79/0.95/1.02
all 0.77/0.62/1.08 0.75/0.63/1.06 0.73/0.65/1.05 0.77/0.76/1.05 0.75/0.96/1.00
Table 5: Tuning of εand νfor GCS-MinCut.
ε0.001 0.1 0.2 0.4 0.8
νtime/nodes/cuts time/nodes/cuts time/nodes/cuts time/nodes/cuts time/nodes/cuts
1 1.00/1.00/1.00 0.94/1.32/1.01 0.89/1.61/0.98 0.83/1.80/0.90 0.80/2.93/0.85
5 0.82/0.76/1.19 0.78/1.00/1.19 0.73/1.22/1.13 0.70/1.43/1.04 0.68/2.02/1.02
10 0.79/0.75/1.21 0.71/0.90/1.20 0.67/1.04/1.15 0.64/1.24/1.02 0.61/1.68/0.95
20 0.76/0.76/1.16 0.63/0.88/1.10 0.63/1.03/1.11 0.60/1.20/1.01 0.59/1.58/0.95
all 0.72/0.78/1.20 0.65/0.97/1.20 0.60/0.93/1.11 0.62/1.23/1.05 0.56/1.60/0.94
need to add a set of O(m)constraints. For this reason, we decide to add only
the inequalities corresponding to the node with maximum violation. The choice
of the separation algorithm and of εdoes not have a significant impact; in
the following experiments, we set ε= 0.2and use the strong component-based
separation procedure.
In Table 6 we summarize the results over the whole set of test instances. Col-
umn “opt” reports the number of instances that are solved to proven optimality
within the time limit of 1200 seconds. Column “time” reports the average com-
puting time. Column “nodes” reports the average number of explored nodes in
the branch-and-bound tree. Column “cuts” reports the average number of GCS
inequalities added by the separation procedures.
First of all, we can observe that the small-size pricing instances from the
SNDlib are easy for all the formulations. It is also worth noting that, for graphs
of similar size, the rnd instances are typically much easier than the prc ones. For
the TSP, instances with random costs are known to be rather easy to solve [2],
and a similar effect may take place for the ESPP. None of the approaches is able
to solve to proven optimality any of the instances in rome99.
GCS-StrongComp solves to optimality all the instances in the set, except for
those in rome99, and proves to be by far the best choice for the largest instances.
The GCS-MinCut approach has difficulties on large-size instances, where most
of the computing time is spent in the separation phase. In particular, optimality
cannot be proven on 4 instances of rand-1000, and 2 instances of prc-l-100.
Compared to the extended formulations, GCS-StrongComp is a clear winner
on all benchmark sets, where it is often more than one order of magnitude
faster. Despite the typically better linear relaxation bounds, formulation RLT
is usually not faster than SF, that allows the solver to reach proven optimality
15

for a larger number of instances. Interestingly, while rather ineffective overall,
formulation MTZ yields good results on the rnd instances. This might be due
to the fact that, on the rnd set, all the extended formulations have similarly
good linear relaxation bounds, and it probably pays off to have an LP of smaller
size.
Formulation MCF, despite the tight linear relaxation, appears to be too
heavy to be of practical interest. On small-sized instances, very few B&B nodes
are necessary. However, this is not enough to overcome the computational load
required by solving the linear relaxation: even with a row-generation approach,
the size of the LP grows quickly. On graphs with 100 or more nodes, only a
small fraction of the instances can be solved within the time limit. On the largest
graphs, the instances in rnd-s-1000 and rome99, the solver quickly reaches the
memory limit of 16 GB.
Figure 3 summarizes the computational experiments with a performance
profile. On the y-axis, we report the fraction of all the instances that are solved
to optimality within the time on the x-axis (in logarithmic scale).
GCS-StrongComp is the topmost curve, solving more than 85% of the in-
stances within 10 seconds. GCS-MinCut is not far behind, although it is gen-
erally slower. Both solve around 95% of the instances within 1200 seconds.
Formulations RLT,SF and MTZ yield similar results on the easiest instances,
although, overall, only less than 70% of the instances are solved to optimality
with MTZ, while SF and RLT reach, respectively, 90% and 87%. With the
MCF approach (bottom curve), Cplex is already significantly slower on the eas-
iest instances, and solves the smallest fraction of the instances (around 65%).
101100101102103
10
20
30
40
50
60
70
80
90
100
GCS-StrongComp
GCS-MinCut
SF
RLT
MTZ
MCF
Figure 3: Fraction of instances solved to optimality within a given time (seconds). The x-axis
is in logarithmic scale.
16

5. Conclusions
We have analyzed integer programming formulations for the ESPP, including
formulations not yet appeared in the literature.
The polyhedral results in Section 3 provide a partial hierarchy among the
ESPP formulations, and prove that the strong extended formulation MCF has
a projection on the space of the arc variables which is equivalent to the poly-
tope of the GCS formulation, that has exponentially many subtour elimination
constraints.
It is also important to understand how effective the formulations are from a
computational point of view. In this regard, we report a set of extensive com-
putational experiments, suggesting that the extended formulations are inferior
for all practical purposes, and the dynamic separation of subtour elimination
constraints appears to be the best option when tackling the ESPP as an integer
program. The GCS approach with the strong component-based separation pro-
cedure is able to solve small-sized instances in a few seconds and medium-sized
instances, with up to 1000 nodes, within a minute, but it is still not sufficient to
solve large-scale problems. When the implementation of a separation procedure
is not possible or convenient, formulation SF is probably the best choice.
It seems likely that the development of good primal heuristics and additional
strong valid inequalities, possibly extended from of the ATSP (e.g., 2-matching
or comb inequalities), might further speed up the computing times and allow the
solution of larger-sized instances. It might also be useful to borrow techniques
from the typical approaches used for the ESPPRC, such as a preprocessing phase
with the aim of reducing the search space.
Acknowledgments
The author would like to thank the two anonymous referees for the useful
comments that helped improve the quality of the manuscript.
References
[1] Amaldi, E., Coniglio, S., Taccari, L., 2014. Maximum throughput network
routing subject to fair flow allocation. In: ISCO. Vol. 8596 of Lecture Notes
in Computer Science. Springer, pp. 1–12.
[2] Applegate, D. L., Bixby, R. E., Cook, W. J., Chvátal, V., 2006. The travel-
ing salesman problem: a computational study. Princeton University Press.
[3] Balas, E., 1989. The prize collecting traveling salesman problem. Networks
19 (6), 621–636.
[4] Balas, E., 2005. Projection, lifting and extended formulation in integer and
combinatorial optimization. Annals of Operations Research 140 (1), 125–
161.
17

[5] Bauer, P., Linderoth, J. T., Savelsbergh, M. W., 2002. A branch and cut
approach to the cardinality constrained circuit problem. Mathematical Pro-
gramming 91 (2), 307–348.
[6] Beasley, J., Christofides, N., 1989. An algorithm for the resource con-
strained shortest path problem. Networks 19 (4), 379–394.
[7] Björklund, A., Husfeldt, T., Khanna, S., 2004. Approximating longest
directed paths and cycles. In: Automata, Languages and Programming.
Springer, pp. 222–233.
[8] Boland, N., Dethridge, J., Dumitrescu, I., 2006. Accelerated label setting
algorithms for the elementary resource constrained shortest path problem.
Operations Research Letters 34 (1), 58–68.
[9] Claus, A., 1984. A new formulation for the travelling salesman problem.
SIAM Journal on Algebraic Discrete Methods 5 (1), 21–25.
[10] Dantzig, G., Fulkerson, R., Johnson, S., 1954. Solution of a large-scale
traveling-salesman problem. Operations Research 2 (4), 393–410.
[11] Demetrescu, C., Goldberg, A. V., Johnson, D. S., 2009. The Shortest Path
Problem: Ninth DIMACS Implementation Challenge. Vol. 74. AMS.
[12] Dezső, B., Jüttner, A., Kovács, P., 2011. LEMON – an open source C++
graph template library. Electronic Notes in Theoretical Computer Science
264 (5), 23–45.
[13] Drexl, M., 2013. A note on the separation of subtour elimination constraints
in elementary shortest path problems. European Journal of Operational
Research 229 (3), 595–598.
[14] Drexl, M., Irnich, S., 2014. Solving elementary shortest-path problems as
mixed-integer programs. OR Spectrum 36 (2), 281–296.
[15] Feillet, D., Dejax, P., Gendreau, M., Gueguen, C., 2004. An exact algo-
rithm for the elementary shortest path problem with resource constraints:
Application to some vehicle routing problems. Networks 44 (3), 216–229.
[16] Fischetti, M., Gonzalez, J. J. S., Toth, P., 1998. Solving the orienteering
problem through branch-and-cut. INFORMS Journal on Computing 10 (2),
133–148.
[17] Gabow, H. N., 2007. Finding paths and cycles of superpolylogarithmic
length. SIAM Journal on Computing 36 (6), 1648–1671.
[18] Gavish, B., Graves, S. C., 1978. The travelling salesman problem and
related problems. Working Paper GR-078-78, Massachusetts Institute of
Technology.
18

[19] Gendreau, M., Laporte, G., 1998. A branch-and-cut algorithm for the undi-
rected selective traveling salesman problem. Networks 32, 263–273.
[20] Goldberg, A. V., Tarjan, R. E., 1988. A new approach to the maximum-flow
problem. Journal of the ACM (JACM) 35 (4), 921–940.
[21] Haouari, M., Maculan, N., Mrad, M., 2013. Enhanced compact models
for the connected subgraph problem and for the shortest path problem in
digraphs with negative cycles. Computers & Operations Research 40 (10),
2485–2492.
[22] Ibrahim, M., Maculan, N., Minoux, M., 2009. A strong flow-based for-
mulation for the shortest path problem in digraphs with negative cycles.
International Transactions in Operational Research 16 (3), 361–369.
[23] Irnich, S., Desaulniers, G., 2005. Shortest Path Problems with Resource
Constraints. Springer, Ch. 2, pp. 33–65.
[24] Jepsen, M., Petersen, B., Spoorendonk, S., Pisinger, D., 2008. Subset-
row inequalities applied to the vehicle-routing problem with time windows.
Operations Research 56 (2), 497–511.
[25] Jepsen, M. K., Petersen, B., Spoorendonk, S., 2008. A branch-and-cut algo-
rithm for the elementary shortest path problem with a capacity constraint.
Tech. Rep. 08/01, Department of Computer Science, University of Copen-
hagen.
[26] Jepsen, M. K., Petersen, B., Spoorendonk, S., Pisinger, D., 2014. A branch-
and-cut algorithm for the capacitated profitable tour problem. Discrete
Optimization 14, 78–96.
[27] Karger, D., Motwani, R., Ramkumar, G., 1997. On approximating the
longest path in a graph. Algorithmica 18 (1), 82–98.
[28] Miller, C. E., Tucker, A. W., Zemlin, R. A., 1960. Integer programming
formulation of traveling salesman problems. Journal of the ACM (JACM)
7 (4), 326–329.
[29] Naddef, D., Rinaldi, G., 2002. Branch-and-cut algorithms for the capaci-
tated VRP. SIAM Monographs on Discrete Mathematics and Applications
Philadelphia, PA, Ch. 3, pp. 53–84.
[30] Orlowski, S., Wessäly, R., Pióro, M., Tomaszewski, A., 2010. SNDlib 1.0 -
survivable network design library. Networks 55 (3), 276–286.
[31] Padberg, M., Sung, T.-Y., 1991. An analytical comparison of different for-
mulations of the travelling salesman problem. Mathematical Programming
52 (1-3), 315–357.
19

[32] Righini, G., Salani, M., 2006. Symmetry helps: bounded bi-directional dy-
namic programming for the elementary shortest path problem with resource
constraints. Discrete Optimization 3 (3), 255–273.
[33] Sherali, H. D., Adams, W. P., 1990. A hierarchy of relaxations between
the continuous and convex hull representations for zero-one programming
problems. SIAM Journal on Discrete Mathematics 3 (3), 411–430.
[34] Taccari, L., 2015. Mixed-integer programming models and methods for
bilevel fair network optimization and energy cogeneration planning. Ph.D.
thesis, Politecnico di Milano, Milano.
[35] Tarjan, R., 1972. Depth-first search and linear graph algorithms. SIAM
Journal on Computing 1 (2), 146–160.
[36] Vansteenwegen, P., Souffriau, W., Oudheusden, D. V., 2011. The orienteer-
ing problem: A survey. European Journal of Operational Research 209 (1),
1–10.
[37] Wong, R., 1980. Integer programming formulations of the travelling sales-
man problem. In: Proceedings of the IEEE international conference on
circuits and computers. pp. 149–152.
20

Table 6: Branch-and-cut results. Missing values are due to the memory limit being reached.
GCS-StrongComp GCS-MinCut MTZ RLT SF MCF
opt time nodes cuts opt time nodes cuts opt time nodes opt time no des opt time nodes opt time nodes cuts
nobel-us 30 0.0 0.6 10.5 30 0.0 3.1 9.6 30 0.0 26.2 30 0.0 19.1 30 0.0 9.0 30 0.0 1.0 397.3
atlanta 30 0.0 0.8 10.9 30 0.0 1.1 10.7 30 0.0 22.0 30 0.0 7.4 30 0.0 1.8 30 0.0 0.6 415.4
geant 30 0.0 1.1 12.8 30 0.0 2.5 11.5 29 0.0 20.9 30 0.0 9.6 30 0.0 4.7 30 0.0 1.1 633.7
france 30 0.0 0.7 26.4 30 0.0 2.8 24.7 30 0.1 344.3 30 0.1 51.1 30 0.1 21.1 30 0.1 1.0 1549.7
germany 30 0.1 25.6 89.9 30 0.1 40.0 81.4 30 3.5 5947.7 30 0.4 352.5 30 0.7 403.0 30 1.6 10.6 4595.3
rnd-s-50 30 0.1 22.3 46.2 30 0.1 33.1 43.4 30 0.2 82.9 30 0.2 79.8 30 0.2 59.0 30 3.2 18.3 6611.3
rnd-s-100 30 0.4 37.3 56.1 30 0.5 40.2 56.7 30 0.9 189.5 30 4.3 403.6 30 2.0 151.0 24 557.9 72.3 32582.7
rnd-s-200 30 0.8 22.1 66.7 30 1.8 25.7 75.0 30 9.0 308.8 30 38.5 322.1 30 19.7 218.1 1 1193.8 1.9 87376.9
rnd-s-500 30 4.2 27.6 77.4 30 32.3 29.4 77.1 30 218.1 496.5 6 1104.1 217.3 26 557.1 373.6 0 1200 0.0 2.5·105
rnd-s-1000 30 21.3 29.6 139.2 26 352.4 25.6 142.4 3 1158.3 51.8 0 1200 0.0 0 1200 0.0 0 – – –
rnd-d-25 30 0.0 2.2 6.5 30 0.0 2.4 6.6 30 0.1 31.6 30 0.1 13.6 30 0.1 10.5 30 0.5 4.6 2365.5
rnd-d-50 30 0.1 8.0 14.7 30 0.1 8.3 15.3 30 1.7 460.9 30 1.3 39.7 30 1.0 18.2 30 69.0 14.6 18769.7
rnd-d-100 30 0.4 14.0 24.2 30 0.8 16.1 24.8 30 3.5 70.0 30 30.0 113.8 30 22.1 107.7 6 996.0 2.5 67352.1
prc-f-25 30 0.0 4.0 25.0 30 0.0 3.9 25.7 30 2.8 3059.3 30 1.0 317.1 30 0.4 49.6 30 3.6 13.9 7978.8
prc-f-50 30 0.2 19.9 68.1 30 0.3 16.3 70.8 23 332.8 1.3·10530 29.9 1346.4 30 16.4 1047.8 28 459.5 34.2 64155.9
prc-f-100 30 11.0 2068.9 246.8 30 40.5 1649.6 236.6 3 1142.7 30600.6 27 328.0 3465.5 30 174.6 1633.7 0 1200 0.0 4.6·105
prc-p-25 30 0.1 7.4 54.5 30 0.2 30.8 49.2 16 688.2 8.1·10530 1.4 1455.7 30 0.9 426.6 30 5.4 4.1 11687.6
prc-p-50 30 0.6 26.8 103.6 30 0.8 112.4 103.3 11 872.3 4.5·10530 5.5 1224.6 30 13.4 1113.0 30 288.4 11.8 65648.2
prc-p-100 30 12.7 1243.0 325.3 30 88.1 3003.6 428.2 0 1200 90700.8 29 165.2 19837.3 30 75.6 3432.7 0 1200 0.0 4.3·105
prc-l-25 30 0.2 12.3 64.3 30 0.2 50.9 59.5 10 808.5 1.4·10630 5.3 1552.5 30 1.3 175.6 30 5.2 4.1 12078.0
prc-l-50 30 0.9 77.3 169.7 30 1.2 188.4 164.0 5 1027.5 5.1·10530 24.2 1068.5 30 14.8 764.6 30 248.5 23.4 59021.8
prc-l-100 30 14.5 1597.2 340.4 28 209.8 6144.9 444.6 0 1200 88407.3 30 134.7 15209.0 30 84.2 4188.2 0 1200 0.0 4.3·105
rome99 0 1200 16944.9 5668.8 0 1200 728.0 13271.4 0 1200 3126.2 0 1200 780.4 0 1200 447.8 0 – – –
21