Transportation Research Part C 19 (2011) 853–865

Contents lists available at ScienceDirect

Transportation Research Part C

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

Tramp ship routing and scheduling with speed optimization

Inge Norstad a,b, Kjetil Fagerholt a,*, Gilbert Laporte c
a
Department of Industrial Economics and Technology Management, Norwegian University of Science and Technology, Alfred Getz veg 3, NO-7491
Trondheim, Norway
b
Norwegian Marine Technology Research Institute (MARINTEK), POB 4125 Valentinlyst, NO-7450 Trondheim, Norway
c
Canada Research Chair in Distribution Management, HEC Montréal, 3000 chemin de la Côte-Sainte-Catherine, Canada H3T 2A7

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

Article history: Tramp shipping companies are committed to transport a set of contracted cargoes and try
Received 11 January 2010 to derive additional revenue from carrying optional spot cargoes. Traditionally, models for
Received in revised form 30 April 2010 ship routing and scheduling problems are based on fixed speed and a given fuel consump-
Accepted 1 May 2010
tion rate for each ship. However, in real life a ship’s speed is variable within an interval, and
fuel consumption per time unit can be approximated by a cubic function of speed. Here we
present the tramp ship routing and scheduling problem with speed optimization, where
Keywords:
speed on each sailing leg is introduced as a decision variable. We present a multi-start local
Maritime transportation
Speed optimization
search heuristic to solve this problem. To evaluate each move in the local search we have to
Routing determine the optimal speed for each sailing leg of a given ship route. To do this we pro-
Scheduling pose two different algorithms. Extensive computational results show that the solution
method solves problems of realistic size and that taking speed into consideration in tramp
ship routing and scheduling significantly improves the solutions.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction

Seaborne transportation is one of the main freight transportation modes, and the only cost effective option for transpor-
tation of large volumes between the continents. Approximately eight billion tons of goods are carried by sea each year (UNC-
TAD, 2008). The maritime transportation industry has over the last decade experienced an increased awareness of the effect
of fuel usage on operational costs and on environmental emissions. The impact of optimization in transportation planning in
this industry can be significant given the very large volume of goods transported by sea each year.
We study a tramp ship routing and scheduling problem. A shipping company operating in the tramp shipping market will
have some long-term contracts of affreightment (COAs) which specify a number of cargoes that the company is committed to
transport. In addition to the contracted (mandatory) cargoes, there exists a spot market where cargo owners announce their
transportation needs. These are optional cargoes that the shipping company can choose to carry if it finds it profitable and
has sufficient fleet capacity to do so.
Each cargo, both mandatory and optional, consists of a given quantity to be transported from a given loading port to a
given discharge port. For each cargo there is a time window that defines when the cargo can be loaded, and there is often
a time window for the discharge port as well. Each spot cargo has a specified freight income rate that determines the revenue
the shipping company will receive if the cargo is transported. The income derived from carrying the mandatory contract car-
goes can be considered as fixed in this setting as there is no option not to carry them.
A typical tramp shipping company operates a heterogeneous fleet of ships having different load capacities, speeds, cargo
handling equipment, operating costs and physical dimensions (length, beam width, draft, etc.). Due to the different

* Corresponding author.
E-mail addresses: [email protected] (I. Norstad), [email protected] (K. Fagerholt), [email protected] (G. Laporte).

0968-090X/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trc.2010.05.001
854 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865

properties of the ships, there may be compatibility constraints between ships and ports, and between ships and cargoes. For
example, a small ship may not be able to carry a heavy cargo, and a large ship with deep draft may not enter a shallow port.
There are costs associated with visiting a port, dependent on the ship, and there are also fuel costs associated with sailing.
Every ship in the fleet has a service speed that is traditionally used when the shipping company plans its routes and
schedules for the fleet. However, in reality, the ship can sail at other speeds as well. Normally, a ship has a minimum and
a maximum cruising speed which define the range of speeds at which it can actually travel. Fuel consumption, and hence
the cost of sailing a given distance, is strongly dependent on speed. As shown by Ronen (1982), a cubic function provides
a good estimation, within the practical speed range, of the relationship between fuel consumption per time unit and speed
for cargo ships. A quadratic function can be shown to provide a good estimation of the relationship between fuel consump-
tion per distance unit and speed.
Fig. 1 depicts the relationship between speed and fuel consumption for a particular liquefied natural gas carrier with load
capacity of 150,000 m3. The fuel consumption is in tonnes (t) per traveled nautical mile (M) and the speed v is in knots (M/h).
The feasible speed range is between v = 14.1 and v ¼ 22 knots. For this interval the quadratic function
c(v) = 0.0036v2  0.1015v + 0.8848 has shown to provide a good estimate for the relationship between fuel consumption
c(v) and sailing speed v.
The objective in the tramp ship routing and scheduling problem is to maximize profit, i.e. total freight income minus
operating costs. The decisions to be made in the planning are:

1. selecting which spot cargoes to carry,

2. allocating cargoes to ships,
3. determining optimal ship routes and
4. determining optimal ship schedules based on the optimal speed of each ship for each sailing leg.

These decisions are not independent of each other, so in order to find the optimal solution, they must be considered
simultaneously. What is new in this problem description compared to traditional ship routing and scheduling problem for-
mulations, is that we consider speed as a decision variable in the routing problem. This means that the sailing time and costs
between a given pair of ports are not parameters with given values, but variables that depend on the chosen speed.
Note that several authors have also combined the determination of ship routes with speed decisions, albeit in different
contexts from ours. Benford (1981) propose a simple procedure for determining the best mix of ships for a specific transport
service. Perakis (1985) found that the model suggested by Benford (1981) imposed an artificial constraint and was then cor-
rected. Unlike our problem, their formulation has only one speed variable per ship, resulting in each ship using the same
speed during the whole planning period. Perakis and Papadakis (1987a), Perakis and Papadakis (1987b) consider a fleet
deployment problem, where a given amount of cargo is to be transported from one loading port to one destination port with-
in a year. The solution defines the laden and ballast speed for each ship and which ships that should be laid up. Again, each
ship ends up using the same laden and ballast speed during the while planning period. Brown et al. (1987) and Bausch et al.
(1998) both study liquid bulk shipping problems where the speed for each sailing leg is a decision variable. They generate all
feasible ship schedules a priori and select the best combination of schedules by set partitioning. However, there are given
dates for loading and discharging for each cargo, which means that the sailing speed for each leg is implicitly determined
by the distance and time between the port calls on a given candidate route. Papadakis and Perakis (1989) study the problem
of minimum-cost operation of a fleet of ships that has to carry a specific amount of cargo from a set of loadings ports to a set
of unloading ports in a given time period. Unlike our problem there are no time windows within which a given cargo must be
loaded or unloaded, hence the optimal laden and ballast speed will be the same for the entire schedule for each ship. Fag-
erholt (2001) considers a multi-ship pickup and delivery problem similar to ours, but with focus on soft time windows. All

Fig. 1. Fuel consumption characteristics for an LNG carrier.

 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865 855

(promising) feasible routes are generated, and the optimal schedule for each route is then found by determining the optimal
speed with respect to minimizing soft time window penalty and operating costs. Eventually, a set partitioning model selects
the best combination of schedules. A drawback with this approach is that it cannot generate optimal solutions to large
unconstrained problems where the number of feasible routes and schedules is too high to allow full enumeration. Ronen
(forthcoming) studies the effect of oil price on the trade-off between reducing sailing speed and increasing the fleet size
for container ships.
The purpose of this paper is threefold. First, we describe a way to incorporate speed decisions in a traditional tramp ship
routing and scheduling problem. Second, we propose a solution method for this problem. Finally, we show that incorporating
speed in ship routing and scheduling can yield significant improvements in profit for the shipping company.
The remainder of this paper is organized as follows. A mathematical formulation of the tramp ship routing and scheduling
problem with speed optimization is given in Section 2. This section also provides a model for optimizing speed along a fixed
single ship route, which arises as an important subproblem. In Section 3 we suggest solution methods, both for the speed opti-
mization problem and the entire tramp ship routing and scheduling problem with speed optimization. Section 4 provides
computational results, both for the subproblem and the global problem. Finally, concluding remarks are given in Section 5.

2. Mathematical model

We start by giving a mathematical formulation of the Tramp Ship Routing and Scheduling Problem with Speed Optimization
(TSRSPSO) in Section 2.1, before we present a formulation of one of its important subproblems, namely the speed optimiza-
tion problem (SOP) along a fixed single ship route.

2.1. The tramp ship routing and scheduling problem with speed optimization

A full mathematical arc flow formulation for the ship routing and scheduling problem with fixed speed is presented in
Christiansen et al. (2007). It is formulated as a pickup and delivery problem with time windows and capacity constraints.
There, as well as in most formulations of maritime transportation problems, the time and cost of sailing between a pair
of ports for a given ship are fixed. Our extensions of that formulation lie in the introduction of new variables for the sailing
speed for each ship and each sailing leg, as well as an adjusted cost function and constraints to incorporate speed as decision
variables.
Let N be the set of cargoes indexed by i. Associated with cargo i is a loading port node i and a discharging port node i + n,
where n is the number of cargoes in N. Let NP ¼ f1; . . . ; ng be the set of loading (pickup) nodes and ND ¼ fn þ 1; . . . ; 2ng,
the set of delivery nodes. Note that several nodes can refer to the same physical port. The problem is defined on a graph
GðNP [ ND ; AÞ, where A ¼ fði; jÞ : i; j 2 NP [ ND ; i – jg. The set of loading nodes NP is partitioned into two subsets, NC
and NO : NC is the set of loading nodes for the mandatory contract cargoes, and NO is the set of loading nodes for the op-
tional cargoes. Let V be the set of ships indexed by k. The graph Gk ðNk ; Ak Þ is the subgraph of G for ship k. Included in Nk
are all the nodes of NP [ ND which correspond to cargoes that ship k can carry. NPk and NDk are the sets of pickup and
delivery nodes, respectively, for ship k. Included in Nk are also o(k), the initial position for k, which can be a port or a point
at sea, and d(k), an artificial destination node. Ship k has a feasible sailing speed interval [v k ; v k ] and a cargo carrying capacity
Qk. Each cargo i has a weight qi and generates a revenue ri. For each node there is a time window [t i ; ti ] which defines when
service at the node must start. Let sik denote the service time at node i when visited by ship k. Let dij be the sailing distance
from node i to node j.
The binary variable xijk is 1 if ship k travels directly from node i to node j, and 0 otherwise. The variable wik is the weight
on board ship k when leaving node i, and the variable tik is the time for start of service for ship k at node i. The variable vijk is
the speed of travel from node i to node j with ship k. The time it takes for ship k to sail along arc (i, j), including the service
time sik at i, is dij/vijk + sik. The non-linear function ck(v), defined in the interval [v k ; v k ] represents the sailing costs per dis-
tance unit for ship k sailing at speed v. The cost of sailing an arc (i, j) with ship k at speed vijk is then dijck(vijk).
The TSRSPSO can now be formulated as follows:
X X X X
maximize r i xijk  dij ck ðv ijk Þxijk ; ð2:1Þ
k2V ði;jÞ2Ak k2V ði;jÞ2Ak

subject to
X X
xijk ¼ 1; i 2 NC ; ð2:2Þ
k2V j2Nk
X X
xijk 6 1; i 2 NO ; ð2:3Þ
k2V j2Nk
X
xoðkÞjk ¼ 1; k 2 V; ð2:4Þ
S
j2NPk fdðkÞg
X X
xijk  xjik ¼ 0; k 2 V; j 2 Nk n foðkÞ; dðkÞg; ð2:5Þ
i2Nk i2Nk
856 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865

X
xidðkÞk ¼ 1; k 2 V; ð2:6Þ
S
i2NDk foðkÞg
X X
xijv  xj;iþn;k ¼ 0; k 2 V; i 2 NPk ; ð2:7Þ
j2Nk j2Nk

tik þ sik þ di;iþn =v i;iþn;k  t iþn;k 6 0; k 2 V; i 2 NPk ; ð2:8Þ

xijk ðt ik þ sik þ dij =v ijk  tjk Þ 6 0; k 2 V; ði; jÞ 2 Ak ; ð2:9Þ
ti 6 t ik 6 t i ; k 2 V; i 2 Nk ; ð2:10Þ
xijk ðwik þ qj  wjk Þ ¼ 0; k 2 V; j 2 NPk ; ði; jÞ 2 Ak ; ð2:11Þ
xi;jþn;k ðwik  qj  wjþn;k Þ ¼ 0; k 2 V; j 2 NPk ; ði; j þ nÞ 2 Ak ; ð2:12Þ
woðkÞk ¼ 0; k 2 V; ð2:13Þ
X X
qi xijk 6 wik 6 Q k xijk ; k 2 V; i 2 NPk ; ð2:14Þ
j2Nk j2Nk
X
0 6 wiþn;k 6 ðQ k  qi Þxiþn;j;k ; k 2 V; i 2 NPk ; ð2:15Þ
j2Nk

v k 6 v ijk 6 v k ; k 2 V; ði; jÞ 2 Ak ; ð2:16Þ

xijk 2 f0; 1g; k 2 V; ði; jÞ 2 Ak : ð2:17Þ
The objective function (2.1) maximizes the profit from operating the fleet. Constraints (2.2) and (2.3) ensure that the con-
tracted cargoes are transported exactly once, and that the spot cargoes are transported at most once, respectively. Con-
straints (2.4)–(2.6) describe the network flow on a route for ship k. Coupling constraints (2.7) ensure that the loading and
discharging nodes for a cargo i must be visited by the same ship. Precedence constraints (2.8) mean that the discharging node
for a cargo must be visited after its loading node. Constraints (2.9) ensure that the time of starting service at a node j must be
greater than or equal to the departure time from the previous node i, plus the sailing time between the nodes. Constraints
(2.10) define the time window in which service must start. Constraints (2.11) and (2.12) handle the relationship between
quantity on board and loading and discharging, respectively, while constraints (2.13) define the initial quantity on board ship
k. Constraints (2.14) and (2.15) ensure that the capacity constraints for the ships are not violated when loading and discharg-
ing, respectively. Constraints (2.16) are lower and upper bounds for the speed variables. Binary restrictions on the flow vari-
ables are given by constraints (2.17).

2.2. The speed optimization problem for a fixed ship route

This section describes a subproblem of the TSRSPSO, which is to determine the optimal speed along a fixed single ship
route. We refer to this subproblem as the Speed Optimization Problem (SOP). It arises when solving the TSRSPSO by column
generation or by local search. For both these methods, it is necessary to determine the profit contribution for each candidate
route. The freight income and port costs for such a route will be fixed, but the fuel costs will depend on speed. In order to
evaluate the correct profit for a given candidate route, one must determine the optimal speed for each leg of the sequence
that will minimize the fuel cost for the route. Other authors, e.g. Sexton and Bodin (1985), Dumas et al. (1990), Ioachim et al.
(1998), have studied a more general problem of schedule optimization on a fixed route, but with different objectives and
constraints.
The problem description is based on the formulation presented by Fagerholt et al. (2010). Given a route, i.e. a sequence of
port calls with time windows that a given ship must sail, the objective is to determine the speed for each leg in the route so
that the total fuel consumption is minimized.
Fig. 2 illustrates the SOP along a given single ship route with four ports calls, or nodes. The node indexed by i = 0 is the
initial position of the ship. The artificial destination node defined in Section 2.1 is not necessary in the case of a given route,
hence it is not included in the figure.
For simplicity, we use the notation of Section 2.1, but, since we only consider a single ship route, the ship index k is re-
moved. The port call nodes along the route are indexed by i = 0, . . . , m. Let the parameter di,i+1 represent the sailing distance
along the arc (i, i + 1), defined for i = 0, . . . , m  1. Each node along the route has a given time window [t i ; ti ] in which service
must start. Let vi,i+1 be the decision variable for speed along arc (i, i + 1), and let v and v be the ship’s minimum and maximum
cruising speeds. Let ti denote the variable for the start of service time at node i. The non-linear cost function c(vi,i+1) describes
the relationship between fuel consumption and sailing speed. The objective is to minimize cost, since we can assume that the
revenue from transporting the cargoes is fixed for the given route. In practice there will also be a specified service time at
each node. This is, however, independent of sailing speed decisions, and hence we do not include this feature in the following
model.
The SOP can now be defined as follows:
X
m1
minimize di;iþ1 cðv i;iþ1 Þ; ð2:18Þ
i¼0
 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865 857

Fig. 2. Example of a single ship route with four port calls.

subject to
tiþ1  t i  di;iþ1 =v i;iþ1 P 0; i ¼ 0; . . . ; m  1; ð2:19Þ
ti 6 ti 6 t i ; i ¼ 1; . . . ; m; ð2:20Þ
v 6 v i;iþ1 6 v ; i ¼ 0; . . . ; m  1: ð2:21Þ
The objective function (2.18) minimizes the fuel consumption for the given route. Constraints (2.19) ensure that the ship
does not start service before it arrives at the node, but waiting is allowed. Constraints (2.20) are the time window con-
straints, while constraints (2.21) ensure that speed on a given arc is within the ship’s feasible speed interval. For the start
node, i = 0, there is no time window, only a start time t0. We note that the objective function is convex and non-linear,
and that constraints (2.19) are also non-linear.

3. Algorithms

In this section we present an algorithm for the TSRSPSO. The model can, in principle, be solved by commercial optimiza-
tion software for non-linear programs, but in practice, real life instances of the problem will be too large to be solved within a
reasonable time. Hence, we suggest a heuristic solution method. We will first introduce two alternative exact algorithms for
the SOP, and we will then describe a heuristic for the TSRSPSO.

3.1. Solution methods for the SOP

The SOP can be solved by applying a non-linear programming (NLP) solver to the mathematical formulation of Section 2.2.
However, even though the typical size of such a problem is quite small (four to 20 nodes), using an NLP solver will be too
time consuming due to the large number of subproblems that need to be solved in the TSRSPSO algorithm. We will therefore
present two alternative solution methods. The first discretizes the possible arrival times for each node along the route. The
second method recursively adjusts the average speed for segments of the route until a feasible and optimal solution has been
identified.

3.1.1. Discretizing arrival times

This solution method is based on the work of Fagerholt et al. (2010) and Fagerholt (2001). We can discretize the arrival
time within the time window of each node, and solve the problem as a Shortest Path Problem (SPP) on an acyclic graph. Con-
sider the ship route illustrated in Fig. 2 and discretize the arrival times.
Each port call node in the route, except for i = 0, will be replicated into a number of discretized feasible arrival times at
that node. Each of these new replicated nodes will represent a state (i, l), where i is the index of the node in the route, and l
refers to the discretized arrival time til at the node. The variable t01 is the start time for the ship. Each state (i, l) will represent
a node in a graph (see the example in Fig. 3). Every arc ((i, l), (i + 1,l0 )) connects a state corresponding to an arrival time at a
port call with a state corresponding to an arrival time at the next port call on the route. This results in a directed acyclic
graph. Each arc will correspond to a given sailing speed, v ððilÞ;ðiþ1;l0 ÞÞ ¼ di;iþ1 =ðt iþ1;l0  til Þ. Arcs corresponding to speeds outside
the feasible range for the ship and states with no incoming or outgoing arcs are not generated. To each arc is associated a
sailing cost cðilÞ;ðiþ1;l0 Þ that depends non-linearly on the corresponding speed of the arc. The sailing cost is the product of fuel
consumption and fuel price. We note that cðilÞ;ðiþ1;l0 Þ P cðilÞ;ðiþ1;l0 þ1Þ , since the latter allows for slower sailing speed and hence
less fuel consumption. This implies that for the last port call m, the latest arrival time will always be a part of the optimal
solution, and thus all states for the last port call, except the latest arrival (node (4, 3) in Fig. 3), could be removed from the
graph. In practice, however, this could lead to a poor starting position for the following planning period. Although this is not
done here, introducing a penalty cost for late arrival at the last port call is straightforward.
Every path through this graph, from node (0, 1) to one of the states of the last port call node, is equivalent to a feasible ship
schedule. Hence, finding a shortest path through the network is equivalent to finding an optimal arrival time for each port
call. Then we also have determined the optimal speed for each leg of the route. Since the graph is acyclic and generated in a
lexicographic order, the shortest path from the start node to any other node can effectively be computed while generating
the network.
Discretizing speed instead of arrival times is another option which might even seem more intuitive. However, the graph
would then become a tree where the number of nodes will grow exponentially with the number of discretizations and port
calls on the route.
858 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865

3.1.2. A recursive smoothing algorithm

We will now present a second algorithm that also can be used for determining vessel speed for each leg along the route.
The idea behind this method is that since the fuel consumption function for a vessel is convex over the range of feasible
speeds, it should be optimal to keep speed as low as possible and constant. For instance, if a route consists of two sailing
legs, it will be better to keep the speed at 16 knots for both legs, than traveling at reduced speed on the first leg and increased
speed on the second leg, even though the arrival time at the destination would be the same. Of course, the time window of
the node between the two sailing legs may restrict this possibility. This property holds for all fuel consumptions that are
function of speed as long as they are convex and non-decreasing over their defined interval. It is not necessary to know
the coefficients of the function to determine the optimal speeds. This can be summarized in the following proposition.

Proposition 1. Disregarding the time windows, the minimum cost speed v* for a given route with the sequence 0, 1, 2, . . . , m, and a
total sailing distance d, a given starting time t0, and an ending time tm, will be the same for all sailing legs and determined as v* = d/
(tm  t0).

Proof. The proof follows directly from the convex form of the fuel consumption (and hence cost) as a function of speed.
Our recursive smoothing algorithm (RSA) starts by considering the total time available from start to end of the route and
the total distance to travel. Dividing the distance by the available time, we obtain the constant speed v* which is optimal in
the absence of time windows. If none of the time windows is violated, the solution is optimal. If a time window is violated,
the arrival time for that port call is adjusted to the nearest feasible value. Then the route is split at that node into two sub-
routes and the procedure is repeated recursively until all time windows are respected. If the calculated speed in any iteration
is greater than the ship’s maximum speed, the problem is infeasible. The choice of the node on which to make the split can
affect the objective value and the run time. We check all time windows and split the node p with the largest time window
violation d, no matter if the arrival is too early or too late.
Algorithm 1 describes the RSA algorithm. Let the variables ti be the arrival time at node i and vi,i+1 the speed from i to i + 1.
Let s and e be the start and end nodes for the (partial) route, respectively, and di,i+1 the distance from i to i + 1. The time win-
dow for node i is [t i ; ti ]. Before the algorithm is called, the start time and end time of the route must be set to ts = t0, which is
the starting position for the ship, and te ¼ t e , respectively.

Algorithm 1 (RECURSIVE_SMOOTHING_ALGORITHM (s, e)).

d 0
p 0
Pe1
v i¼s di =ðt e  t s Þ
for i s to e do
i i+1
vi v*
ti ti1 + di/vi
if ti  ti > jdj then
d ti  ti
p i
end if
if ti  ti > jdj then
d ti  ti
p i
end if
end for
if d > 0 then
tp tp
RECURSIVE_SMOOTHING_ALGORITHM(s,p)
RECURSIVE_SMOOTHING_ALGORITHM(p,e)
end if
ifd < 0 then
tp tp
RECURSIVE_SMOOTHING_ALGORITHM(s,p)
RECURSIVE_SMOOTHING_ALGORITHM(p,e)
end if
 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865 859

Fig. 3. Graph obtained by discretizing arrival times.

Before calculating the objective value for SOP, all optimal speeds vi that have a value lower than v must be set to v. This
means that the ship will sail at the lowest possible speed and wait for the start of the time window. However, one should
note that the RSA algorithm is based on properties that only apply if the ship has the same fuel consumption function for all
the sailing legs. If fuel consumption also depends on load, for example, then the property does not hold and the RSA cannot
be used. If this is the case, the method using discretized arrival times should be used instead.

3.2. Algorithm for the TSRSPSO

To solve the tramp ship routing and scheduling problem we use a multi-start local search heuristic, following the ideas of
Brønmo et al. (2007). In this heuristic a number of initial solutions are generated by a constructive heuristic. Each solution is
generated partly randomly and partly by a deterministic insertion heuristic. Some cargoes are chosen randomly and each of
them is assigned to a randomly chosen ship. The remaining cargoes are then assigned one by one to the ship that will con-
tribute the most to the objective function. A number of the best initial solutions are then improved by a local search heuristic
by means of intra-route operators, which attempt to improve the route for a single ship, and inter-route operators, which
attempt to improve the solution by moving cargoes between ships.
Since this local search heuristic can be quite time consuming, not all initial solutions are improved, but only a number of
the best ones. There are two local search procedures, a quick one and an extended one. The quick local search explores the
neighborhood defined by three local search operators. In the last stage of the multi-start heuristic an extended local search is
applied. The extended search uses two additional local search operators in addition to the three used in the quick search.
Algorithm 2 gives an overview of the multi-start local search heuristic.

Algorithm 2 (MULTI-START).

Step 1: Generate a initial solutions.

Step 2: Choose the b best initial solutions and use the quick local search to improve them.
Step 3: Choose the c best solutions from Step 2 and improve them by using the extended local search.

Table 1 provides a brief description of the different local search operators. The quick local search uses the first three oper-
ators, while the extended local search uses all five. For more details on the multi-start local search heuristic, we refer to
Brønmo et al. (2007). Whenever an operator is applied, the new solution obtained is evaluated with respect to feasibility
and contribution to the objective function. In the local search heuristic, the maximum speed of the ship must be used in or-
der to check for feasibility. Since the local search operators do not consider the speed decisions, each new ship route in a
solution will have to be evaluated by solving the SOP. This subproblem minimizes voyage cost for a given ship route by
determining the optimal speed for each leg in the route.

Table 1
Local search operators.

Operator Description
1-resequence A cargo i is removed from the schedule of a ship v and
reinserted into the schedule of v in the best possible position
Reassign A cargo i is removed from the schedule of a ship v and inserted
to the ship u that gives the best feasible insertion
2-interchange Cargo i is removed from ship v and cargo j is removed from
ship u. Then i is inserted into the schedule of ship u and
j is assigned to ship v
2-resequence Two cargoes i and j are removed from ship v. First cargo i
is inserted in the schedule for ship v at the best position
then cargo j is reinserted in the best position
3-interchange Cargoes i,j and k are removed from ships v,u and w,
respectively. Cargoes i,j and k are then reinserted into the schedules of ship u,w and v, respectively
860 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865

4. Computational study

We have performed extensive computational experiments in order to measure the effect of introducing speed as decision
variables in the tramp ship routing and scheduling problem. In Section 4.1 we first study the SOP in isolation and compare
the two proposed algorithms with the optimal solution provided by a non-linear mathematical programming solver. Then, in
Section 4.2 we study the complete TSRSPSO.

4.1. Comparing SOP solution methods

Fagerholt et al. (2010) solve a large number of generated problem instances. They compare the shortest path based meth-
od with the optimal solutions provided by a non-linear programming solver using an interior point method. In this paper we
also compare the recursive smoothing algorithm (RSA) with the shortest path based method and with an exact algorithm.

4.1.1. Generation of test problem instances for the SOP

To test the performance of the different solution methods for the speed optimization problem, we have used an instance
generator similar to the one used by Fagerholt et al. (2010). The same ship characteristics, i.e. speed range and fuel consump-
tion function, are used for all the instances. In order to have a varied set of instances, three different parameters are set at
different values. These parameters are:

1. Number of nodes in the route (4, 8, 12 or 16).

2. Amount of slack or waiting in the route (0%, 20% or 40%). This parameter determines the percentage of the port calls
(rounded to the nearest integer) along the route that will have a waiting time for the start of the time window when using
the ship’s service speed. For the nodes chosen to have waiting, the time window will start between one and six (randomly
chosen) days after what would have been the arrival time given sailing at service speed
3. The width of the time window (5 or 10 days).

The combination of four different route lengths, three alternative waiting time settings and two time window widths re-
sults in 24 different cases. Distances between nodes are randomly chosen from a realistic set of port to port distances. For
nodes with waiting time, the start of the time window is given by the amount of waiting time. For the other nodes the start of
the time windows are between 0.25 and 0.75 times the width of the time window earlier than the arrival time given service
speed. Each case is labeled by using the values of the parameters, hence case 8n_10tw_40 refers to a case with eight nodes,
10 day wide time windows and waiting at 40% of the nodes. We have generated 100 instances of each case, resulting in 2400
different instances.

4.1.2. Computational results for the SOP

The 2400 instances were solved by applying the freely available non-linear programming solver IPOPT from COIN-OR, to
the mathematical formulation of Section 2.2. An algorithm for solving the model with discretized arrival times was imple-
mented in C#, as described by Fagerholt et al. (2010). For each instance an acyclic network is generated and a shortest path is
determined. The RSA is also implemented in C#. Table 2 shows the optimality gap for the different algorithms applied on the
24 cases. The entries in Table 2 are average values over 100 instances for each case. SPP x refers to the shortest path based
algorithm with a discretization level of x, whereas RSA refers to the recursive smoothing algorithm.
To compare the response time of the solution methods, we have generated a large number of test instances of four dif-
ferent route lengths and solved them with the different methods. Table 3 shows the average CPU times in milliseconds for
the different algorithms and the different route lengths. Even though the non-linear programming optimization software
solves the instances in less than a second, it is too slow to be used as a move evaluator in the local search based heuristic
for the TSRPSO. The SPP method, on the other hand, is much faster and more practical. We see that the response time for
the SPP grows quickly when the discretization level increases, but it is linearly dependent on route length. When we take
the results from Table 2 into account, we see that SPP5 or SPP10 would probably be sufficient for the evaluation of each route
when solving the TSRPSO. At this level the CPU time is acceptable. The RSA is extremely fast and always provides optimal
solutions. Therefore, for ship routing problems in which each ship has the same fuel consumption function for both laden
and ballast voyages, this algorithm is the obvious choice.
We should recall that the fuel consumption function we are using is just an estimate, and that several factors are not
known at the time of planning, such as weather, that will influence the fuel consumption. Bearing this in mind, one should
not be too focused on the optimality gap, as there is much uncertainty. Also, during the local search heuristic, what is impor-
tant is that the best routing decision should be taken. The final speed decisions can be taken a posteriori. For all practical
purposes, a solution value less than 1% away from the optimal value should be sufficient. A good solution strategy for the
TSRSPSO, if each ship has different fuel consumption functions depending on the ship load, could therefore be to use
SPP5 or SPP10 for the evaluation of each move, and SPP100 or a non-linear programming software for the final solution.
When fuel consumption function per ship only depends on speed, RSA is clearly the best algorithm for the SOP.
 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865 861

Table 2
Average optimality gap (in percent) for different SOP solution methods.

Case SPP5 SPP10 SPP20 SPP50 SPP100 RSA

4n_5tw_0 0.418 0.072 0.018 0.003 0.001 0.000
4n_5tw_20 0.180 0.038 0.008 0.001 0.000 0.000
4n_5tw_40 0.112 0.019 0.005 0.001 0.000 0.000
4n_10tw_0 1.082 0.188 0.048 0.007 0.002 0.000
4n_10tw_20 0.526 0.140 0.024 0.004 0.001 0.000
4n_10tw_40 0.261 0.070 0.014 0.002 0.000 0.000
8n_5tw_0 0.838 0.144 0.029 0.004 0.001 0.000
8n_5tw_20 0.640 0.104 0.023 0.003 0.001 0.000
8n_5tw_40 0.253 0.052 0.010 0.002 0.000 0.000
8n_10tw_0 2.413 0.494 0.095 0.014 0.004 0.000
8n_10tw_20 1.413 0.257 0.063 0.009 0.002 0.000
8n_10tw_40 0.715 0.148 0.030 0.005 0.001 0.000
12n_5tw_0 1.078 0.164 0.038 0.006 0.001 0.000
12n_5tw_20 0.563 0.101 0.023 0.003 0.001 0.000
12n_5tw_40 0.261 0.050 0.012 0.002 0.000 0.000
12n_10tw_0 2.910 0.528 0.117 0.017 0.004 0.000
12n_10tw_20 1.799 0.313 0.071 0.011 0.003 0.000
12n_10tw_40 0.965 0.195 0.045 0.007 0.002 0.000
16n_5tw_0 1.176 0.200 0.042 0.006 0.002 0.000
16n_5tw_20 0.597 0.104 0.023 0.003 0.001 0.000
16n_5tw_40 0.327 0.066 0.013 0.002 0.000 0.000
16n_10tw_0 3.545 0.628 0.133 0.020 0.005 0.000
16n_10tw_20 1.963 0.332 0.075 0.011 0.003 0.000
16n_10tw_40 1.081 0.198 0.046 0.006 0.002 0.000
Average 1.047 0.192 0.042 0.006 0.002 0.000

Table 3
Average CPU time (in milliseconds) for different SOP solution methods.

Nodes NLP SPP5 SPP10 SPP20 SPP50 SPP100 RSA

4 332 0.030 0.094 0.29 1.72 6.30 0.0010
10 401 0.078 0.188 0.70 4.12 16.2 0.0050
20 430 0.125 0.375 1.39 8.23 32.2 0.0091
40 577 0.281 0.750 2.81 16.81 64.7 0.0190

4.2. Multi-start local search algorithm

The multi-start local search heuristic was coded in C# and implemented as a solver on the commercial ship routing deci-
sion support system TurboRouter (Fagerholt, 2004; Fagerholt and Lindstad, 2007) from MARINTEK. Brønmo et al. (2007) de-
scribe two different parameter settings for the multi-start heuristic, one with short computational times, and one that
requires more computational time but provides better solutions. In the computational study here we will only use the latter
parameter setting. The computational tests were performed on a PC with Pentium M, 1.8 GHz processor and 512 Mb of RAM
under Windows XP. For each of the test cases we will compare different strategies for solving the problems.

4.2.1. Description of the test problem instances for TSRSPSO

We have used two different sets of problem instances. First, we have nine real cases with data from six different inter-
national bulk shipping companies operating in the tramp shipping market. The first eight are the same instances that have
been studied in Korsvik et al. (2010) and Brønmo et al. (2007). The ships in instances one to six may carry multiple cargoes
simultaneously, while in instances seven to nine all cargoes are full ship loads. Table 4 describes the nine instances. In the
original cases, each ship has a given service speed and a fuel consumption. In order to solve the instances as TSRSPSO, the
service speed is replaced by a minimum and a maximum speed, and the fixed fuel consumption is replaced by a fuel con-
sumption function depending on speed.
In addition to these instances we have developed an instance generator that can generate instances with different char-
acteristics. The fleet is based on real ships and is the same for all the generated problem instances. The fleet has 13 ships with
different load capacities and speed intervals. The cargoes are generated by randomly selecting a loading port and a discharg-
ing port from a set of real ports. The cargo sizes are set randomly between 20% and 100% of the capacity of the largest ship,
hence carrying multiple cargoes simultaneously is possible. Time window widths are chosen randomly between 6 and 14
days, and the freight rates are randomly chosen within an interval representing realistic figures. We have designed four dif-
ferent cases and generated 10 instances of each, resulting in 40 different instances. Table 5 describes the four instance sets.
862 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865

Table 4
Test instance descriptions.

Instance 1 2 3 4 5 6 7 8 9
Planning horizon (days) 23 75 75 40 35 150 20 35 90
# of contract cargoes 18 8 17 12 15 41 28 12 16
# of spot cargoes 0 1 0 2 2 9 2 3 2
# of ships 6 3 6 7 13 13 13 4 6

4.2.2. Computational results

To measure the effect of introducing speed decisions in the ship routing problem, we have compared different strategies
for taking the speed decisions into account. Table 6 shows the results from applying different strategies on test instance
number nine. In this problem there are 18 cargoes and a heterogeneous fleet of six ships, all with a service speed of 17 knots,
and 18 cargoes. In the first column we have solved the problem with a traditional fixed service speed formulation. In the next
column we have taken the solution from the first one and have applied the RSA algorithm a posteriori to each of the six ship-
ping routes. We see that the income is the same, but the fuel consumption is significantly reduced when speed is optimized.
We also see that only 16 of the 18 cargoes are transported. In the third column the problem is solved as a fixed speed prob-
lem, with all ships traveling at the maximal speed of 20 knots. Now, all the 18 cargoes are taken, but due to the constant high
speed, the fuel costs are so high that the profit is less than under the previous strategies. In the fourth column we have opti-
mized the speed a posteriori on the routing solution found in the third column. Finally, in the fifth column we have applied
the RSA to solve the SOP as a move evaluator in the multi-start local search. The income is the same, since all the 18 cargoes
are taken, but because the routing is different, the fuel consumption is 4.3% less and hence the profit is 2.9% higher.
In Table 7 we solve all nine real instances with each of the five different solution strategies. For each instance we report
the gap in percent for both net profit and fuel costs compared with the fixed service speed solution. We see that for all the
instances, applying speed optimization during the evaluation of each move in the local search heuristic provides the best
solution. On average the improvement in profit is 7.0%. The effect of applying speed optimization varies from instance to
instance and is highly dependent on several factors, for example on the relationship between the number of cargoes and fleet
capacity, and on the relationship between fuel price and freight rates. Note that in instance two, solving the problem by
applying a fixed maximum speed increases the profit. This is because speeding up allows the fleet to transport an additional
spot cargo, and hence increases the income. This is also the reason why fuel consumption increases when applying the max-
imum speed with speed optimization and variable speed strategies to instances two and nine.
In Table 8 we solve the 40 generated instances with each of the five different solution strategies. The values in the table
are average results over the 10 instances of each set. We see that on average, over all the 40 instances, there is an 14%
improvement in net profit when applying speed optimization to evaluate each move in the local search, compared to the
fixed service speed strategy.
Note that for instance set D we see an improvement in profit from service speed with speed optimization to maximum
speed with speed optimization, while for sets A to C it is the opposite. In set D, the planning horizon is relatively short with
respect to the number of cargoes. Therefore, on average over the 10 instances, only 66 of the 70 cargoes are taken when using
service speed. By applying maximum speed, the average is 68 cargoes. This leads to an increased income, hence to an
improvement from 5.7% to 8.7% in profit. In instance sets A to C, due to the relatively long planning horizon, it is possible

Table 5
Descriptions of sets of generated problem instances.

Instance set A B C D
Planning horizon (days) 30 60 100 120
# of contract cargoes 0 0 0 0
# of spot cargoes 15 25 50 70
# of ships 13 13 13 13

Table 6
Comparing different strategies for solving instance number nine.

Service speed (17 Service speed + speed Max speed (20 Max speed + speed Variable speed (14–20
knots) opt. knots) opt. knots)
Cargoes 16 16 18 18 18
served
Income ($) 16.102.500 16.102.500 19.072.500 19.072.500 19.072.500
Port costs($) 2.837.050 2.837.050 3.351.450 3.351.450 3.351.450
Fuel costs ($) 5.830.661 5.184.264 10.188.202 6.328.365 6.056.490
Profit ($) 7.434.789 8.081.186 5.532.848 9.392.685 9.664.560
 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865 863

Table 7
Profit and fuel cost gap (in percent) for different strategies compared to fixed service speed

Instance Service speed + speed opt. Max speed Max speed + speed opt. Variable speed
Profit Fuel Profit Fuel Profit Fuel Profit Fuel
1 0.4 2.4 1.6 31.7 1.8 15.3 2.2 21.3
2 0.5 3.1 1.9 53.7 6.4 27.6 7.7 20.1
3 2.0 14.0 7.0 45.6 2.0 14.0 2.0 14.0
4 4.0 24.6 7.9 46.2 4.0 24.4 6.6 38.3
5 1.7 9.7 6.1 35.4 2.0 11.5 2.1 12.4
6 6.8 43.9 6.3 40.1 7.2 45.2 10.9 70.0
7 0.8 10.4 1.9 24.9 1.1 14.2 1.3 18.0
8 0.4 7.9 1.7 31.8 0.4 7.9 0.5 8.8
9 8.7 11.1 25.6 74.7 26.3 8.5 30.0 3.9
Average 2.7 14.1 6.2 42.7 5.7 10.7 7.0 17.6

Table 8
Profit and fuel cost gap (in percent) for different strategies compared to fixed service speed.

Instance set Service speed + speed opt. Max speed Max speed + speed opt. Variable speed
profit fuel profit fuel profit fuel profit fuel
A 12.1 33.1 13.7 43.2 9.3 29.6 16.4 52.4
B 11.5 34.9 13.1 39.8 1.1 3.4 12.6 38.4
C 15.2 58.1 10.2 38.7 5.6 21.5 13.7 51.9
D 5.7 22.1 6.4 48.4 8.7 9.9 13.2 28.9
Average 11.1 37.1 10.9 42.5 6.2 16.1 14.0 42.9

to transport all the cargoes when using the service speed. When applying the maximum speed, the solutions tend to utilize
the ships more unevenly, which leads to a poorer result, compared to the service speed, when the speed optimization is ap-
plied a posteriori. We will now explain why this happens.
In a poor tramp ship market, the number of spot cargoes is relatively low, and a shipping company will have excess fleet
capacity. In this case, using fixed speed will often lead to a solution in which the ships are unevenly utilized. This is because,
when the speed is fixed, fuel consumption is proportional to distance sailed. Using fixed speed when assigning ships to the
cargoes will in practice result in the minimization of the total traveled distance. A posteriori speed optimization will reduce
the fuel cost, but will not change the routing decision. However, if variable speed is taken into account when assigning ships
to the cargoes, the solution often includes a more even distribution of the cargoes among the ships, leading to lower average
speed for the fleet, and hence lower fuel consumption. In the example in Table 9 there are two ships and six cargoes. The first
column shows the optimal cargo assignment when the problem is solved by using maximum speed with a posteriori speed
optimization. The second column shows the solution when solving the SOP for evaluating each move in the local search heu-
ristic. Note that the total traveled distance is shorter in the fixed speed solution, but since the average speed is much lower in
the variable speed solution, the total fuel consumption will be less in the latter.

4.2.3. Computational time

Table 10 shows the computational times associated with the different strategies for solving the TSRSPSO. The values are
average values of all the 10 generated instances of each case. The time spent on applying speed optimization a posteriori is
negligible compared with the time spent on the multi-start local search based heuristic. Hence we do not report response
times for the pure fixed speed strategies, but only the fixed speed with a posteriori speed optimization. The column labeled
Service speed refers to the service speed with a posteriori speed optimization, and Max speed refers to the maximum speed

Table 9
Example of different routing decisions.

Max speed + speed opt. Variable speed

Ship A Ship B Ship A Ship B
Cargo 1 U U
Cargo 2 U U
Cargo 3 U U
Cargo 4 U U
Cargo 5 U U
Cargo 6 U U
Average speed (knots) 14.0 16.3 14.2 14.0
Distance (M) 32.295 32.869
Fuel cost ($) 628.168 556.774
864 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865

Table 10
Average response times (in seconds) for different strategies used to solve the TSRSPSO.

Instance set # cargoes Service speed + speed opt. Max speed + speed opt. SPP RSA
A 15 8 10 20 11
B 25 18 20 39 22
C 50 282 571 967 599
D 70 340 840 1121 841

with a posteriori speed optimization. The column SPP refers to applying the discretized arrival times method with a discret-
ization level of 10 for each move in the local search and a discretization level of 100 on the final solution. The column labeled
RSA refers to applying the RSA algorithm to evaluate each move in the local search. For all the instances and strategies the
multi-start parameters are the same: 100 initial solutions, 10 quick local searches and five extended searches. When apply-
ing a local search heuristic, it is impossible to predict from the size of the instance how many iterations will be needed to find
a local optimum. Therefore, the computation time varies significantly from instance to instance and from solution method to
solution method. For several of the instances, applying RSA to each local search move resulted in shorter computational time
before an optimum was found. From Table 10 we see that on average solving the TSRSPSO with maximum speed require
more time than with service speed. This is because the neighborhood size defined by the local search operators increases
as a higher speed allows a ship to transport more cargoes. We see that on average over the four cases applying SPP10 re-
quires 63% more time than applying RSA. We also see that applying RSA hardly requires any additional time compared to
the Max speed strategy.

5. Concluding remarks

We have introduced and solved a tramp ship routing and scheduling problem in which speed is a decision variable on the
different sailing legs. Since fuel consumption per distance unit can be approximated by a quadratic function of speed, the
problem becomes non-linear. We have presented two algorithms for the speed optimization problem along a single shipping
route. Discretizing arrival times is quite fast and can be applied to any fuel consumption function. By adjusting the parameter
of discretization level, one can trade off between solution quality and solution time. The recursive smoothing algorithm
(RSA) is even faster, but is only applicable for cases where fuel consumption functions of the ships are not dependent on ship
load, but only on speed. It will, however, work for any convex fuel consumption function. Taking variable speed into consid-
eration significantly improves the profit, partly because increasing speed can make it possible to carry additional spot car-
goes, and partly because reducing speed results in less fuel consumption per distance. Reducing fuel consumption will also
result in reduced environmental emissions, see also the survey of the environmental effects from optimization in land based
transportation by Sbihi and Eglese (2007).

Acknowledgements

This research was carried out with financial support from the DESIMAL project, funded by the Research Council of Norway
and from the Canadian Natural Engineering Research Council under Grant 39682-05. This support is gratefully acknowl-
edged. Thanks are due to the referees for their valuable comments.

References

Bausch, D., Brown, G., Ronen, D., 1998. Scheduling short-term marine transport of bulk products. Maritime Policy and Management 25, 335–348.
Benford, H., 1981. A simple approach to fleet deployment. Maritime Policy & Management 8, 223–228.
Brønmo, G., Christiansen, M., Fagerholt, K., Nygreen, B., 2007. A multi-start local search heuristic for ship scheduling – a computational study. Computers &
Operations Research 34, 884–899.
Brown, G., Graves, G., Ronen, D., 1987. Scheduling ocean transportation of crude oil. Management Science 33, 335–346.
Christiansen, M., Fagerholt, K., Nygreen, B., Ronen, D., 2007. Maritime transportation. In: Barnhart, C., Laporte, G. (Eds.), Transportation. In: Handbooks in
Operations Research and Management Science, vol. 14. North-Holland, Amsterdam, pp. 189–284.
Dumas, Y., Soumis, F., Desrosiers, J., 1990. Optimising the schedule for a fixed vehicle path with convex inconvenience costs. Transportation Science 24, 145–
152.
Fagerholt, K., 2001. Ship scheduling with soft time windows – an optimization based approach. European Journal of Operational Research 131, 559–571.
Fagerholt, K., 2004. A computer-based decision support system for vessel fleet scheduling – experience and future research. Decision Support Systems 37,
35–47.
Fagerholt, K., Laporte, G., Norstad, I., 2010. Reducing fuel emissions by optimizing speed on shipping routes. Journal of the Operational Research Society 61,
523–529.
Fagerholt, K., Lindstad, H., 2007. Turborouter: an interactive optimization-based decision support system for ship routing and scheduling. Maritime
Economics & Logistics 9, 214–233.
Ioachim, I., GTlinas, S., Soumis, F., Desrosiers, J., 1998. A dynamic programming algorithm for the shortest path problem with time windows and linear node
costs. Networks 31, 193–204.
Korsvik, J.E., Fagerholt, K., Laporte, G., 2010. A tabu search heuristic for ship routing and scheduling. Journal of the Operational Research Society 61,
594–603.
Papadakis, N.A., Perakis, A.N., 1989. A nonlinear approach to multiorigin, multidestination fleet deployment problem. Naval Research Logistics 36, 515–528.
Perakis, A.N., 1985. A second look at fleet deployment. Maritime Policy & Management 12, 209–214.
 I. Norstad et al. / Transportation Research Part C 19 (2011) 853–865 865

Perakis, A.N., Papadakis, N.A., 1987a. Fleet deployment models, part 1. Maritime Policy & Management 14, 127–144.
Perakis, A.N., Papadakis, N.A., 1987b. Fleet deployment models part 2. Maritime Policy & Management 14, 145–155.
Ronen, D., 1982. The effect of oil price on the optimal speed of ships. Journal of the Operational Research Society 33, 1035–1040.
Ronen, D., forthcoming. The effect of oil price on containership speed and fleet size. Journal of the Operational Research Society.
Sbihi, A., Eglese, R.W., 2007. Combinatorial optimization and green logistics. 4OR 5, 99–116.
Sexton, T.R., Bodin, L.D., 1985. Optimizing single vehicle many-to-many operations with desired delivery times: I. Scheduling. Transportation Science 19,
378–410.
UNCTAD, 2008. Review of maritime transport, 2008. United Nations, New York and Geneva.