You are on page 1of 7

European Journal of Operational Research 214 (2011) 588–594

Contents lists available at ScienceDirect

European Journal of Operational Research


journal homepage: www.elsevier.com/locate/ejor

Production, Manufacturing and Logistics

Freight transportation in railway networks with automated terminals:


A mathematical model and MIP heuristic approaches
D. Anghinolfi, M. Paolucci, S. Sacone, S. Siri ⇑
Department of Communications, Computer and System Sciences, University of Genova, Italy

a r t i c l e i n f o a b s t r a c t

Article history: In this paper we propose a planning procedure for serving freight transportation requests in a railway
Received 10 December 2010 network with fast transfer equipment at terminals. We consider a transportation system where different
Accepted 11 May 2011 customers make their requests (orders) for moving boxes, i.e., either containers or swap bodies, between
Available online 17 May 2011
different origins and destinations, with specific requirements on delivery times. The decisions to be taken
concern the route (and the corresponding sequence of trains) that each box follows in the network and
Keywords: the assignment of boxes to train wagons, taking into account that boxes can change more than one train
Freight transportation
and that train timetables are fixed.
Optimal planning
Mathematical programming
The planning procedure includes a pre-analysis step to determine all the possible sequences of trains
MIP heuristics for serving each order, followed by the solution of a 0–1 linear programming problem to find the optimal
assignment of each box to a train sequence and to a specific wagon for each train in the sequence. This
latter is a generalized assignment problem which is NP-hard. Hence, in order to find good solutions in
acceptable computation times, two MIP heuristic approaches are proposed and tested through an exper-
imental analysis considering realistic problem instances.
Ó 2011 Elsevier B.V. All rights reserved.

1. Introduction matter of decision. The orders coming from customers regard the
movement of a set of boxes (containers or swap bodies) from an
Freight transportation is a very important activity in our society origin to a destination within a deadline. We aim at generating
and it is also a complex domain to be studied. Freight transporta- transportation plans for orders considering the boxes similarly to
tion planning has been often faced by mathematical programming passengers planning a train trip: for each box of each order we
and optimization approaches and dealt with in many research determine the sequence of trains and train changes at terminals
works (see for instance the reviews [1,2]). In this paper, a kind of so that the box can reach its destination in time.
long haul transportation of containerized goods [3] is considered, The planning problem faced in this paper combines two main
characterized by a railway network whose terminals are auto- decisions. The former concerns the routing of boxes in the network
mated, i.e., equipped with transhipment devices allowing fast and the selection of train changes at terminals; the latter decision
train–train transfer operations. The study of this transportation deals with the assignment of boxes to wagons of the selected
service derives from a research project carried out with an Italian trains. These two decisions have been generally considered sepa-
company that is developing a new automated container handling rately in the literature. The box routing and train selection are of-
system named Metrocargo [4]. The adoption of Metrocargo in the ten related to aggregate flows of goods represented as a set of
national railway network would provide a very innovative trans- commodities, as in the general formulation provided in the review
portation service for the Italian logistic context. Our main interest [5]. In this paper, instead, we determine the routes and the se-
in this project is the definition of optimization methods for the rel- quences of trains so that the boxes of a same order can reach their
evant off-line planning. common destination independently from each other, i.e., even
We consider the point of view of a central decision maker who through different routes and with different sequences of trains.
provides the transportation service in the railway network and has The decision on wagon assignment has already been treated in
to satisfy transportation requests (orders) from customers. The the literature usually considering a single terminal. In [6] the
timetable and stop sequence of trains in the railway network are authors assume a rail–rail container terminal with a rapid trans-
fixed by the provider of the transportation service and are not shipment yard and they define different models and algorithms
for determining the optimal placement of containers on arriving
⇑ Corresponding author. Tel.: +39 019 21945137; fax: +39 019 21945212. trains and on departing trains, in order to minimize the move-
E-mail address: silvia.siri@unige.it (S. Siri). ments of containers in the terminal. Analogously, in [7] the

0377-2217/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2011.05.013
D. Anghinolfi et al. / European Journal of Operational Research 214 (2011) 588–594 589

so-called load planning problem (assignment of containers to train route to destination by using the available (i.e., scheduled) train
slots) is faced for an intermodal terminal where containers are services.
transferred between trucks and trains. Differently, our purpose in The transportation demand to be satisfied is given by a set of or-
this work is optimally matching the whole transportation demand ders. For each order we consider the origin and destination, the re-
for the railway network with the overall available transportation lease time (i.e., the earliest time from which the boxes are available
offer (the fixed train timetables and stop sequences). Specifically, at the origin), the deadline (i.e., the latest time by which the boxes
we determine the minimum cost assignment of the boxes to wag- must be delivered at destination), and the number of boxes, each
ons of the different trains in order to satisfy train and wagon capac- one characterized by length and weight. The railway network is
ity constraints. Since the train timetables and stop sequences are composed of railway terminals and railway links connecting the
fixed input data, this paper does not deal with train routing and terminals. We assume that for each pair of terminals in the net-
scheduling, as done for instance in [8,9], or with other classical rail- work, i.e., for each origin–destination pair, the available routes (de-
way problems, such as the definition of railroad blocking plans fined as sequences of railway links) are specified a priori. Such
[10]. Even though box routing and train selection, as well as wagon routes are identified by the transportation service provider among
assignment, have been already treated in the literature, the main all the possible link sequences connecting an origin–destination
novelty of this paper stands in considering them jointly. pair, as the most effective ones to provide the transportation ser-
As detailed in the following, the considered planning problem (a vice. Note that this is usually done in order to eliminate possible
preliminary version can be found in [11,12]) basically consists in a but not reasonable/acceptable routes. Moreover, the available
generalized assignment problem (which is known to be NP-hard routes connecting each origin–destination pair are a priori
[13]) that we formulate as a 0–1 linear programming problem. This ordered by the transportation service provider according to a given
kind of problems can be faced by means of general purpose mixed priority.
integer programming (MIP) solvers that seek for the optimal solu- The goal of the proposed planning procedure is to make the
tion by exploiting branch-and-bound or branch-and-cut algo- planning of transportation operations for all the orders within a
rithms. However, most MIP problems are very difficult to solve specified time horizon. The problem is to determine for each box
when their dimensions become too large. For this reason, complex of each order (a) the route that it must follow from its origin to
combinatorial optimization problems from both academic research its destination, (b) the sequence of trains that it must use along
and real world applications have been tackled by specialized heu- such route so that the order time constraints are satisfied, and
ristics or metaheuristics. Recently, new approaches have been pro- (c) the wagons used to transport it for each train in the sequence.
posed to combine ideas from metaheuristics with MIP solver The problem data are listed in the following.
algorithms [14–16], and they are generally referred to as matheu-
ristics [17]. In many cases, matheuristics have been adopted for
large scale problems coming from real applications in order to N set of railway terminals
reduce the solution search space and to speed up the computation R set of trains
(e.g., in production planning [18], in production scheduling [19], in RN n set of trains passing in terminal n; RN n # R
supply chain management [20] and in vehicle routing and trans- dn fixed unitary cost for handling boxes at
portation logistics [21]). In this paper, we propose two matheuris- terminal n
tic approaches for the considered planning problem; the first one is qn hourly unitary cost for storing boxes at
an ad hoc procedure integrated with the MIP solver to drive and terminal n
restrict the solution space exploration, whereas the second one is hn maximum number of handling operations
a general purpose heuristic method to solve MIP problems by iter- (loading and unloading) allowed for each
atively calling the MIP solver as a black box tool. This latter train at terminal n
approach shares some concepts with the MIP heuristics recently L set of railway links, L ¼ fði; jÞ : i; j 2 N ; i–jg;
appeared in literature as Local Branching [14], Relaxation Induced we denote also a railway link as l = (i, j),
Neighborhood Search [15] and Variable Neighborhood Decomposi- where i is the tail of l and j is its head
tion Search [22], but, differently from them, it exploits only a ran- nlr number of railway links covered by train r
domization mechanism to guide the MIP solver. Sr schedule for train r, specifying the nlr links
This paper is organized as follows. In Section 2 the main aspects covered by r and the arrival and departure
of the considered problem are introduced; in Section 3 the pre- times at the visited terminals
analysis and the mathematical programming formulation adopted kr cost related to the use of train r
in the planning procedure are described in detail. Then, in Section 4 gr maximum bearable weight for train r
the two proposed MIP heuristic approaches are introduced. The Wr set of wagons of train r
experimental results are discussed in Section 5 and some conclu- f r;w length of wagon w of train r
sive remarks are finally reported in Section 6. gr;w maximum bearable weight for wagon w of
train r
2. Problem description O set of orders
t so release time of order o at the origin terminal
The Metrocargo system is a new technology, now at a prototype t do deadline for order o at the destination
level, allowing to load and unload cargo units from trains in a hor- terminal
izontal way and under the electric feeding line, thanks to some npo number of alternative available routes
shuttles and a properly devised storage area where all the opera- connecting the origin and destination
tions are realized automatically. With this technology rail–rail terminals of order o
transfer operations at terminals become fast. The planning proce- nlo,p number of railway links of route p of order o
dure proposed in this paper refers to a railway network with auto- lo,p vector 1  nlo,p indicating the sequence of
mated terminals equipped with Metrocargo system or any other railway links of route p of order o
innovative system allowing fast transfers. With such systems, the po,p cost of route p of order o (this cost is
boxes can be considered as passengers that change trains in their weighted according to the route priority)
590 D. Anghinolfi et al. / European Journal of Operational Research 214 (2011) 588–594

B set of boxes po,s route associated with train sequence s of order o


ob order including box b; ob 2 O co,s cost of sequence s of order o
fb length of box b
gb weight of box b
In addition, the following two quantities are computed:

3. The planning approach uno;s 2 f0; 1g denoting whether the boxes of o following
sequence s involve a train change at terminal
The proposed planning procedure is divided in two sequential n2N
phases, corresponding to a pre-analysis and to the solution of a Hno;s storage time (in hours) required for boxes of o
mathematical programming problem, as described in the following. following sequence s at terminal n 2 N

3.1. The pre-analysis The cost co,s (cost of serving an order o with sequence s) is deter-
mined as the priority po;po;s of order o on route po,s multiplied by the
The pre-analysis consists of an algorithm that computes all the sum of unloading and storage costs over all the involved terminals,
sequences of trains available for serving each order, taking into ac- i.e.:
count the network structure, the timetables and the stop se- X  
quences of trains, the origin, the destination and the time
co;s ¼ po;po;s n2N
uno;s  dn þ Hno;s  qn ; ð1Þ
requirements of the order. Note that we refer to sequences of trains
because each box is generally transported by more than one train 3.2. The mathematical programming formulation
from its origin to its destination (the maximum number of trains
that a box can change is given by the number of railway links of The mathematical programming problem is defined to opti-
the route). The pre-analysis algorithm is applied to one order at a mally assign each box to a train sequence and to a wagon of the
time without taking into account capacity constraints but only trains in the sequence, by considering the sequences of trains for
considering feasibility in terms of time delivery and train each order found in the pre-analysis and taking into account the
connections. physical characteristics of boxes and wagons, i.e., considering
More in detail, this algorithm allows to compute all the feasible capacity and other operational constraints. The decision variables
train sequences (for each order o and for each route p a priori spec- are the following.
ified) as those available to connect the order origin and destination.
The algorithm proceeds backward analysing the railway links in yb;s 2 f0; 1g; b 2 B; s 2 S ob , equal to 1 if box b is assigned to
lo,p, starting from the last link of the route, then verifying which sequence s, 0 otherwise
trains, among the ones scheduled to cover the links, can be used xb;s;r;w 2 f0; 1g; b 2 B; s 2 S ob ; r 2 Rob ;s ; w 2 W r , equal to 1 if box b
to serve the order according to their timetable. In particular this is assigned to wagon w of train r in sequence s, 0 otherwise
is done for the last railway link in lo,p by finding all the trains arriv- v b 2 f0; 1g; b 2 B, equal to 1 if box b is not served, 0 otherwise
ing at the link head not after the deadline t do and leaving from the zr 2 f0; 1g; r 2 R, equal to 1 if train r is used, 0 otherwise
link tail not before t so . Then, the algorithm proceeds backward
along the route, selecting among the trains travelling on the con- The planning problem is stated with the following 0–1 linear
sidered link the ones allowing a feasible connection with the feasi- programming formulation.
ble trains already determined for the subsequent railway links. For
example, in a path including three links we may find that trains 1, Problem 1.
2 and 3 cover the last link in a feasible way for an order. Proceeding X X X X
backward, we can find that train 1 is also feasible for the second min cob ;s  yb;s þ kr  zr þ M  vb; ð2Þ
link and it allows a feasible connection with train 2 but not with b2B s2S ob r2R b2B

train 3. Considering the first link we can found that train 1 is still
subject to
feasible and also train 4 allows a feasible connection with train X
1. After the completion of the backward analysis, a forward proce- yb;s þ v b ¼ 1; b 2 B: ð3Þ
dure is applied to determine all the feasible train sequences for an s2S ob

order o on a route p. In the considered example, we find that the


X
order can be served on that route by four feasible train sequences, xb;s;r;w ¼ yb;s ; b2B s 2 S ob r 2 Rob ;s : ð4Þ
i.e., 1-1-1 (train 1 on the three links), 1-1-2 (train 1 on the first two w2W r
links and train 2 on the third link), 4-1-1 (train 4 on the first link X X X
and train 1 on the following two links) and 4-1-2 (train 4 on the g b  xb;s;r;w 6 gr  zr ; n2N r 2 RN n :
 
first link, train 1 on the second link and train 2 on the last link). b2B w2W r
s2 S Lo T
;n;r [S ob ;n;r
The following quantities are obtained from the pre-analysis and b

are used as input data in the mathematical programming problem: ð5Þ


So set of train sequences found for order o X X
Ro;s set of trains in sequence s of order o
g b  xb;s;r;w 6 gr;w ; n2N r 2 RN n w 2 Wr :
b2B
 
S Uo;n;r set of indices of sequences for order o which involve the s2 S Lo ;n;r [S To ;n;r
b b
unloading of boxes at terminal n from train r
set of indices of sequences for order o which involve the
ð6Þ
S Lo;n;r
loading of boxes at terminal n to train r X X
set of indices of sequences for order o which involve a fb  xb;s;r;w 6 f r;w ; n2N r 2 RN n w 2 Wr :
S To;n;r  
b2B
transfer operation at terminal n with train r (i.e., s2 S Lo [S To
b ;n;r b ;n;r
neither loading nor unloading operations are executed)
ð7Þ
D. Anghinolfi et al. / European Journal of Operational Research 214 (2011) 588–594 591

X X X
xb;s;r;w 6 hn ; n2N r 2 RN n : ð8Þ The partially fixed MIP sub-problems are solved by imposing
b2B
  w2W r
s2 S Lo ;n;r [S U
both a maximum computation time and a maximum number of
b ob ;n;r
nodes for the MIP solver. Other heuristic details are here omitted
for the sake of brevity, as, for example, the use of a tabu list in order
yb;s 2 f0; 1g b2B s 2 S ob ; ð9Þ to avoid variables yb,s entering and leaving set F too frequently.

xb;s;r;w 2 f0; 1g b2B s 2 S ob r 2 Rob ;s w 2 Wr ; ð10Þ 4.2. The randomized neighborhood search heuristic

zr 2 f0; 1g r 2 R; ð11Þ The second method used to face Problem 1 is called Randomized
Neighborhood Search (RANS) and it is a general purpose MIP heuris-
v b 2 f0; 1g b 2 B: ð12Þ tic that simply operates similarly to an iterated local search using
randomization to generate the explored neighborhood. The RANS
The objective function (2) is the minimization of the costs algorithm starts from a feasible first incumbent solution for the
associated with train sequences (computed in the pre-analysis), original MIP problem and proceeds iteratively following three
train costs, and the penalty for not serving orders (here M de- main steps:
notes a very large constant). Constraints (3) impose that, if
served, each box is assigned to one and only one train sequence, 1. Solution destruction: a subset V # G of k binary and integer vari-
while (4) establish that if a box is assigned to a train sequence, ables is randomly selected and the partially fixed MIP sub-
then it must be assigned to one wagon of each train belonging problem (obtained from the original one by fixing the variables
to that sequence. Constraints (5) impose for each train that in GnV to their values in the incumbent solution) is defined.
the maximum bearable weight is not exceeded. Similarly, con- Note that k is a parameter initialized as k = 0.1  jGj and auto-
straints (6) and (7) ensure that boxes assigned to wagons are matically tuned at each iteration (step 3).
compatible with the wagon length and weight limitations. 2. Solution construction: this is a local search step consisting in
Constraints (8) impose that the maximum number of handling solving the partially fixed MIP sub-problem by calling a MIP sol-
operations to be performed for each train at a given terminal ver, having imposed a maximum computation time (note that
is not exceeded. Finally, constraints (9)–(12) are the definitions also this time is a parameter automatically determined by the
of the variables. algorithm depending on the time needed to solve the linear
relaxation of the MIP problem). Then, if a new best solution is
found, the incumbent is updated.
4. Two MIP heuristic approaches
3. Parameter adjustment and iteration: if the MIP sub-problem is
optimally solved in the available time, then k = k  1.1, otherwise
In order to find good quality solutions to Problem 1 in an
k = k  0.9, and a new iteration is started. In this simple way the
acceptable computation time, we devised and experimented two
algorithm controls the dimension of the explored neighborhood
MIP heuristic approaches: an ad hoc procedure integrated with
(i.e., the number of binary/integer variables), depending on the
the MIP solver and a general purpose heuristic method that itera-
experienced difficulty in solving sub-problems.
tively calls the MIP solver as a black box tool.

Similarly to the ad hoc heuristic, RANS rationale is to face the


4.1. The ad hoc heuristic
solution of a MIP problem by iteratively solving problems of smal-
ler dimensions. RANS generates the MIP sub-problems by hard fix-
Problem 1 is a generalized assignment which involves two main
ing a subset of binary and integer variables whose cardinality is
nested assignment decisions: the assignment of boxes to se-
automatically adjusted in order to find an optimal solution in the
quences, through yb,s variables, and the assignment of boxes to
maximum time allotted. Note that this corresponds to perform a
train wagons according to the selected sequence, through xb,s,r,w
local search in a neighborhood of the incumbent solution of vari-
variables. The method focuses on the yb,s binary variables, as they
able dimension.
determine the highest level assignment decisions, so mainly influ-
encing the dimension of the solution space. Let Y denote the set of
all yb,s variables. The heuristic iteratively solves a sequence of MIP 5. Experimental results
sub-problems obtained from the original Problem 1 having fixed a
subset F # Y of yb,s binary variables by imposing both their lower We ran a set of experimental tests on a set of randomly gener-
and upper bound equal to 0. The method then analyzes the reduced ated problem instances in order to evaluate the performance of the
costs for the fixed variables from the solution of the branch-and- planning procedure and the solution approaches presented above.
cut node producing the optimal integer solution at an iteration, Both the pre-analysis algorithm and the solution schemes for Prob-
and determines for which variables it is convenient to reset the ori- lem 1 were implemented in C]; in particular, the 0–1 linear optimi-
ginal upper bound to 1. This method is called Ad hoc MIP Heuristic zation problem was solved using Cplex 11.0 and the ILOG Concert
(AMH). technology for building the model from the C] language.
More in details, this method starts including all the yb,s variables We introduced two instance ‘‘difficulty level’’ indexes, called
in F. Then, at any iteration k, the subset YnF of variables to be ex- AverageLoad and PeakLoad, that we computed before solving the
plored is determined, for each box b, as follows: 0–1 linear problem. The AverageLoad is the ratio between the re-
quired train slots and the available ones, defined as:
 by removing from F (i.e., resetting the upper bound to 1) the P
o2O nbo  nlo
variable yb;s0 such that rb;sk1 k1 k1
0 ¼ mini fr b;i g, where r b;s is the Av erageLoad ¼ ð13Þ
g  nw  nlr
reduced cost of yb,s at iteration k  1;
 by inserting in F (i.e., fixing the upper bound to 0) the variable The denominator of (13) is an estimation of the number of available
yb;s00 in YnF such that yb;s00 ¼ 0 at iteration k  1 and slots for serving boxes. This is obtained multiplying g, that is the a
n o
r k1 k1 priori computed average number of boxes that can be carried on a
b;s00 ¼ mini r b;i .
full wagon, nw, the average number of wagons for trains, and nlr,
592 D. Anghinolfi et al. / European Journal of Operational Research 214 (2011) 588–594

the number of different link-train pairs included in all the train se- as t so  U½1; 48 (expressed in hours) and the deadline as
quences found in the pre-analysis for the orders. The numerator of tdo ¼ tso þ d where d  U[96, 120]. In this way, we assume that boxes
(13) indicates the average number of boxes to be served on the dif- must reach their destination from their origin in 4–5 days. Finally,
ferent railway links, since nbo is the number of boxes of order o and Table 2 reports specific data and average difficulty level indexes
nlo is the average number of railway links composing the routes of (including the average number of 0-1 variables) for the groups of
order o. instances.
The PeakLoad characterizes the most critical link-train pair and The computational tests were executed on a 2.8 GHz Pentium 4
is defined as: computer with 2 GB of RAM. For each group we generated 5 in-
P stances: we firstly applied the pre-analysis procedure, then the
maxðl;rÞ2LR o2O nbo ðl; rÞ
PeakLoad ¼ : ð14Þ three solution methods, corresponding to solving Problem 1 with
g  nw Cplex 11.0 and applying the two proposed MIP heuristics (AMH
The denominator in (14) is the estimation of the available slots on a and RANS), again using Cplex 11.0 as MIP solver. For the first meth-
train, whereas the numerator estimates the load of the most critical od, we solved Problem 1 giving a 2 hours time limit to the Cplex
link-train pair, as the maximum number of boxes that must be 11.0 solver. Moreover, we imposed for AMH a time limit of 600 sec-
transported on any railway link by any train. In (14) nbo ðl; rÞ is onds for solving the MIP sub-problems and we fixed the node limit
the average number of boxes of order o associated with link-train for the branch-and-cut exploration to 1000 nodes for the instances
pair (l, r), obtained as the number of train sequences of order o with a number of variables not greater than 30000, whereas we
including the pair (l, r) times the quantity nbo =jS o j, i.e., the average fixed this limit to a single node for larger instances. The maximum
number of boxes of order o for sequences serving o. The difficulty time limit for both the MIP heuristics was set to 1 hour.
in solving an instance should increase for higher values of these In order to evaluate the effectiveness of the AMH and RANS ap-
two indexes; in particular, PeakLoad values larger than 1 should proaches, we report the obtained results in Table 3 together with
characterize instances for which some boxes could be unserved. the ones produced by Cplex. Moreover, the characteristics of each
We built a set of problem instances taking into account some instance are specified, in terms of AverageLoad and PeakLoad in-
realistic data, when available, relevant to the type of boxes and dexes, and number of variables in the 0–1 LP problem. The compu-
wagons, the railway network and train timetables, and randomly tational results for the three methods are reported, as regards the
generating the transportation demand. Moreover, the cost terms value of the objective function and the number of not assigned
P
adopted in this experimental campaign have been provided by lo- boxes (NotAss) that corresponds to b2B v b . Note that Cplex was

gistic operators. As already pointed out in the Introduction, Met- not able to improve the first trivial solution (no boxes assigned)
rocargo system is at a prototype level and the considered railway for one instance in groups A and D, for three instances in group B
network with fast transfer systems at terminals is not applied and for all the instances in groups C, E and F. In other words, only
yet in Italy. This is the reason why realistic (and not real) data few instances are solved in a satisfactory way by Cplex. In addition,
are used to build the problem instances. We analysed three differ- note that the solver always reached the limit of 7200 seconds.
ent scenarios, described in Table 1, which will be referred to as Moreover, from the results in Table 3 we can clearly appreciate
Small, Medium and Large scenario, characterized by different the effectiveness of both the MIP heuristics that, in half the time
dimensions of the railway network and the available trains. The given to the MIP solver, were able to find solutions that are largely
considered planning horizon is one week for the Small and Medium better. Only for the instance A2 Cplex found a slightly better solu-
scenario, whereas two weeks are considered for the Large one. The tion, whereas the average percentage deviations, obtained as (Heu-
train routes and timetables are supposed to be the same on each ristic Objective  Cplex Objective)/Cplex Objective, are about 67%
day, the available routes for each origin-destination pair range be- for AMH and 77% for RANS.
tween 1 and 3 and each train covers either 2 or 3 railway links. In Table 4 the two MIP heuristics are compared, showing the
We generated six groups of instances, one for the Small sce- percentage deviation of RANS results from AMH, computed as
nario, two for the Medium and three for the Large one. The type (RANS Objective  AMH Objective)/AMH Objective, and the RANS
of boxes was selected among 13 considered types from the discrete deviation in the number of not assigned boxes, computed as RANS
uniform distribution U[1, 13]; the origin and destination terminals NotAss  AMH NotAss. Comparing the two proposed MIP heuristics,
were generated from U[1, nt] where nt is the number of terminals we can observe the prevalence of RANS over AMH: this latter, in
in the network for a scenario; the order release time was chosen fact, produced a better objective solution only for 3 instances over
35 and only in one case it was able to assign one box more than
RANS.
Table 1 The reason of the better performance of RANS with respect to
Dimensions of the three scenarios. AMH could be explained by more carefully analysing the behaviour
of the two methods. AMH iterates the solution of MIP sub-problems
Scenario Number of nodes Number of arcs Number of trains
where a subset of yb,s variables are fixed (bounded to zero); this in
Small 10 34 98
principle was supposed to generate simpler MIP sub-problems from
Medium 13 50 154
Large 13 50 308 whose solutions the method should gain guidance on the most con-
venient assignment decisions. However, the simplification due to

Table 2
Characteristics of the groups of instances.

Group Scenario Orders Boxes per order AverageLoad PeakLoad Variables


A Small 60 U[6, 8] 0.38 2.45 142139
B Medium 80 U[6, 8] 0.38 1.95 218164
C Medium 100 U[8, 10] 0.59 3.47 322087
D Large 100 U[4, 6] 0.18 1.10 437293
E Large 120 U[5, 7] 0.26 1.72 669931
F Large 140 U[6, 8] 0.32 2.05 859769
D. Anghinolfi et al. / European Journal of Operational Research 214 (2011) 588–594 593

Table 3
The computational results with Cplex, AMH and RANS.

Inst. ID AverageLoad PeakLoad No. of variables Cplex objective Cplex NotAss AMH objective AMH NotAss RANS objective RANS NotAss
A1 0.36 2.22 133251 3391895 3 2386326 2 1389165 1
A2 0.38 2.01 136996 9379194 9 9382326 9 9379687 9
A3 0.41 2.54 163414 424000000 424 3386911 3 1388865 1
A4 0.38 2.99 136600 19403106 19 389755 0 393552 0
A5 0.38 2.49 140435 12395334 12 5388565 5 4391352 4
B1 0.39 2.41 185589 73516161 73 42523507 42 41514833 41
B2 0.38 2.04 197942 50578470 50 24575480 24 17563662 17
B3 0.38 2 242570 563000000 563 14572205 14 10567689 10
B4 0.39 1.9 234597 552000000 552 5560182 5 560189 5
B5 0.38 1.38 230124 578000000 578 4561566 4 553609 5
C1 0.56 3.19 307070 910000000 910 113737485 113 101746837 101
C2 0.58 2.98 301130 901000000 901 129748775 129 102770643 102
C3 0.59 3.81 331078 895000000 895 153680582 153 124704947 124
C4 0.6 3.69 335544 890000000 890 124696449 124 106715182 106
C5 0.6 3.7 335615 895000000 895 132723643 132 101750611 101
D1 0.17 1.41 430006 8705542 8 575054 0 573427 0
D2 0.18 0.98 428932 33617886 33 591723 0 585129 0
D3 0.18 1.07 436317 1712172 1 572312 0 564187 0
D4 0.18 0.88 468674 504000000 504 590978 0 580400 0
D5 0.17 1.15 422538 15638562 15 578566 0 566335 0
E1 0.26 1.76 722530 734000000 734 2879459 2 858728 0
E2 0.26 1.62 732747 707000000 707 5847630 5 835393 0
E3 0.26 1.52 687204 719000000 719 2784957 2 774731 0
E4 0.25 1.95 658116 711000000 711 2819057 2 810395 0
E5 0.25 1.76 549056 720000000 720 796262 0 803483 0
F1 0.31 2.54 864289 981000000 981 6055668 6 1074804 1
F2 0.33 2.08 890972 983000000 983 10094255 10 1087820 1
F3 0.33 2.12 927211 994000000 994 31081102 31 10148714 10
F4 0.3 1.44 730560 998000000 998 14038610 14 1062927 1
F5 0.33 2.06 885812 979000000 979 18107729 18 5122048 5

Table 4
The comparison between AMH and RANS results.
which has 133251 binary variables, we can observe that 94.3% are
xb,s,r,w variables that are never fixed and only 5.7% are yb,s. Therefore
Inst. ID RANS objective % dev. from AMH RANS NotAss dev. from AMH
the number of iterations performed by the AMH in one hour for in-
A1 41,79 1 stance A1 was only 49. With analogous considerations, the AMH
A2 0,03 0 ability to explore the solution space seems to be not very extended,
A3 58,99 2
A4 0,97 0
since changes in variable fixing from one iteration to another one
A5 18,51 1 are rather limited. On the other hand, RANS does not privilege
B1 2,37 1
any subsets of variables for hard fixing (it is a general purpose
B2 28,53 7 MIP heuristic) and it uses an adaptation mechanism to control the
B3 27,48 4 dimension of the MIP sub-problems (i.e., the number of not fixed
B4 89,92 0 variables) in order to optimally solve them most of the times. In
B5 87,86 1
addition, this provides RANS with a good solution space explora-
C1 10,54 12 tion/diversification capability and allows RANS to execute a larger
C2 20,79 27
number of iterations (203 in case of instance A1) that greatly com-
C3 18,85 29
C4 14,42 18 pensate for the use of a random fixing rule.
C5 23,34 31 As already highlighted, the tested instances can be considered
D1 0,28 0 representative of real cases, in particular the largest ones, corre-
D2 1,11 0 sponding to a network with 13 terminals and 50 railway links,
D3 1,42 0 308 trains and an average request of about 1000 boxes to trans-
D4 1,79 0 port. Also the time horizon used for the tested instances can be
D5 2,11 0
considered representative of realistic applications. Of course, the
E1 70,18 2 proposed approach is suitable for an off-line planning, for instance
E2 85,71 5
E3 72,18 2
one day in advance, for which the considered computation times
E4 71,25 2 are acceptable.
E5 0,91 0
F1 82,25 5 6. Conclusions
F2 89,22 9
F3 67,35 21
In this paper we have proposed a planning procedure for serv-
F4 92,43 13
F5 71,71 13 ing some transportation requests (expressed in terms of origin,
destination, release time, deadline, number of boxes, and so on)
by using a railway network with rapid transhipment terminals.
variable fixing was not enough to allow the solution of the MIP sub- In such a system the boxes can change different trains from origin
problems in short times; if, for example, we consider instance A1, to destination and the boxes of a same order can be served in
594 D. Anghinolfi et al. / European Journal of Operational Research 214 (2011) 588–594

different ways (different routes and sequences of trains). The pro- [6] N. Bostel, P. Dejax, Models and algorithms for container allocation problems on
trains in a rapid transshipment shunting yard, Transportation Science 32
posed procedure is composed of a pre-analysis phase and, then, the
(1998) 370–379.
statement and solution of a 0–1 linear programming problem. Two [7] P. Corry, E. Kozan, An assignment model for dynamic load planning of
MIP heuristic approaches have been proposed for solving large intermodal trains, Computers and Operations Research 33 (2006) 1–17.
problem instances. The performed experimental tests have shown [8] J.-F. Cordeau, P. Toth, D. Vigo, A survey of optimization models for train routing
and scheduling, Transportation Science 32 (1998) 380–404.
the effectiveness of the proposed planning approach and its possi- [9] C.A. Yano, A.M. Newman, Scheduling trains and containers with due dates and
ble application to real cases. dynamic arrivals, Transportation Science 35 (2001) 181–191.
Finally, it can be noted that the proposed planning approach can [10] H.N. Newton, C. Barnhart, P.H. Vance, Constructing railroad blocking plans to
minimize handling costs, Transportation Science 32 (1998) 330–345.
be easily integrated in planning procedures of an overall railway [11] S. Sacone, S. Siri. ‘‘A planning approach for freight transportation operations in
network in which rapid transhipment terminals coexist with tradi- railway networks. In Proceedings of the IEEE International Conference on
tional ones. As a matter of fact, the planning procedure proposed in Automation and Logistics, Qingdao, China (2008).
[12] S. Sacone, S. Siri. ‘‘Optimal planning of order-based operations for intermodal
this work considers only terminals with fast transfer equipment freight transportation. In: Proceedings of the International Trade and Freight
and the relevant trains. Other trains and terminals of the same net- Transportation Conference, Cyprus (2008).
work can keep being treated with traditional planning methods. [13] M.L. Fisher, R. Jaikamur, L.N. VanWassenhove, A multiplier adjustment method
for the generalized assignment problem, Management Science 32 (1986)
1095–1103.
Acknowledgments [14] M. Fischetti, A. Lodi, Local branching, Mathematical Programming 98 (2003)
23–47.
[15] E. Danna, E. Rothberg, C. Le Pape, Exploring relaxation induced neighborhoods
This work was partially supported by the Regional Authority of to improve MIP solutions, Mathematical Programming 102 (2005) 71–90.
Liguria in Italy within the project ‘‘ModLog – Modelling and Opti- [16] P. Hansen, N. Mladenović, D. Urosević, Variable neighborhood search and local
branching, Computers and Operations Research 33 (2010) 3034–3045.
mization of Dynamic Logistic Systems’’, developed at the Italian [17] V. Maniezzo, T. Stützle, S. Voß, (Eds.). Matheuristics – Hybridizing
Centre of Excellence for Integrated Logistics and funded by ‘‘Parco Metaheuristics and Mathematical Programming, Annals of Information
Scientifico e Tecnologico della Liguria’’. Systems 10 (2010).
[18] A.R. Clark, Hybrid heuristics for planning lot setups and sizes, Computers and
Industrial Engineering 45 (2003) 545–562.
References [19] A.M. Newman, M. Kuchta, Using aggregation to optimize long-term production
planning at an underground mine, European Journal of Operational Research
[1] Y.M. Bontekoning, C. Macharis, J.J. Trip, Is a new applied transportation 176 (2007) 1205–1218.
research field emerging? – A review of intermodal rail-truck freight transport [20] M.A. Lejeune, A variable neighborhood decomposition search method for
literature, Transportation research part A 38 (2004) 1–34. supply chain management planning problems, European Journal of
[2] T.G. Crainic, K.H. Kim, Intermodal transportation, in: C. Barnhart, G. Laporte Operational Research 175 (2006) 959–976.
(Eds.), Transportation, North Holland, 2007. [21] K.F. Doerner, V. Schmid, Survey: Matheuristics for rich vehicle routing
[3] T.G. Crainic, Long-haul freight transportation, in: R.W. Hall (Ed.), Handbook of problems Hybrid Metaheuristics 2010, Lecture Notes in Computer Science
Transportation Science, Kluwer Academic publishers., 2003, pp. 451–516. 6373 (2010) 206–221.
[4] <http://www.metrocargo.it/index_eng.htm>. [22] J. Lazić, S. Hanafi, N. Mladenović, D. Urosević, Variable neighbourhood
[5] N. Wieberneit, Service network design for freight transportation: A review, OR decomposition search for 0–1 mixed integer programs, Computers and
Spectrum 30 (2008) 77–112. Operations Research 37 (2010) 1055–1067.

You might also like