International Journal
on Marine Navigation
and Safety of Sea Transportation
Volume 1
Number 3
September 2007
273
Multiobjective Approach to Weather Routing
J. Szlapczynska
Gdynia Maritime University, Gdynia, Poland
ABSTRACT: The paper presents a weather routing solution for sail-assisted ships. Since the route finding
optimisation process is a multiobjective one, the emphasis is put on possible application of multiobjective
optimisation methods. The paper focuses on two such methods, namely evolutionary algorithms and ranking
methods represented by Fuzzy TOPSIS. In addition, a proposed set of optimization criteria is presented.
Descriptions of assumed ship and sail models as well as exemplary speed characteristic are also provided.
Finally, a proposal of application to a weather routing tool is presented
1 INTRODUCTION
A problem of finding the most suitable vessel route
taking into account changeable weather conditions
and navigational constraints is referred to as a
weather routing optimisation problem. Such a
problem is mostly considered for ocean going ships
where adverse weather conditions may impact both,
often contradictory, economic and security aspects of
voyage.
One of the first approaches to the problem was a
minimum time route planning based on a weather
forecast called an isochrone method. The method
was based on geometrically determined and
recursively defined time fronts, so called isochrones.
Originally proposed by R.W. James (James 1957),
isochrone method was in wide use through decades.
In late seventies based on the original isochrone
method the first computer-aided weather routing
tools were developed. However, along with
computer implementation some problems arose, i.e.
with so called “isochrone loops”. Numerous
improvements to the method were proposed since
early eighties, with (Hagiwara 1989, Spaans 1986,
Wisniewski 1991) among others. Since then several
different approaches to the optimisation problem
was in use, with dynamic programming or genetic
and evolutionary algorithms among others.
Most of recent scientific researches in weather
routing focus on shortening the passage time,
reducing fuel consumption and avoiding severe
weather i.e. tropical cyclones. Nowadays,
evolutionary or genetic algorithms are common
solutions for weather routing services. However, due
to multiobjective nature of weather routing it is
recommended to introduce some state-of-the-art
multiobjective methods to the process of route
finding. It may facilitate the process of reaching a
trade-off between often conflicting economic and
safety criteria sets. Thus, it is proposed to introduce
multiobjective evolutionary algorithms as well as
multiobjective ranking methods to the route finding
process.
The remainder of the paper is organized as
follows: section 2 introduces the basic idea of a
single-objected evolutionary algorithm. Section 3
provides detailed description of multiobjective methods
applicable to weather routing. The description includes
multiobjective evolutionary algorithms (MOEAs) and
multiobjective ranking methods represented by Fuzzy
TOPSIS. Section 4 presents proposed application of
the methods to weather routing. Finally, section 5
summarizes the material presented.
274
2 THE IDEA OF SINGLE-OBJECTED
EVOLUTIONARY ALGORITHMS
Evolutionary algorithms are natural successors of
genetic algorithms. The key difference between them
is in the chromosome structure. Genetic algorithms
assume binary and fixed-length chromosome strings,
whereas the evolutionary ones allow more
complicated chromosome structures. It implies that
original genetic binary operators, namely mutation
and crossover, are substituted by evolutionary
operators specialized to fit given chromosome
structure and the optimisation task. However, the
general idea of both genetic and evolutionary
algorithms remains the same. At first initial
population of individuals is being generated and
evaluated. After modifications by operators designed
to improve algorithm’s convergence some
individuals are selected and a new population is
generated. The process of evaluation, modification
and selection lasts until a termination condition is
met.
Single-objected goal function is utilized to
evaluate the individuals by means of so called fitness
function. The goal function is either equal to the
fitness function or at least is an element of the latter.
In general: the better the individual in terms of goal
function the higher evaluation score it gets. By the
evaluation process future modifications and
selection is executed mainly for a group of “best
fitted” individuals. This way the algorithm
converges to a final set being sufficiently close to the
optimal solution.
Apart from the goal function, constraints are
another important issue in single-objected
optimisation process. In the evolutionary framework
the constraints are met due to specialized operators
assuring that any modified individual remains in the
feasible solution space.
3 MULTIOBJECTIVE METHODS APPLICABLE
TO WEATHER ROUTING
3.1 MultiObjective Evolutionary Algorithms
(MOEAs)
MultiObjective Evolutionary Algorithms (MOEAs)
have been growing in popularity since its inception
in mid-1980s. In general, MOEAs extend the
functionality of regular single-objected evolutionary
algorithms providing a method of dealing with
multiple and often conflicting objectives.
When multiobjective problems are being
considered one of the important issues is the
problem of ranking the criteria in terms of their
importance and impact on final result. It is often
designated that a decision maker is a person who
makes such a choice. There are three distinctive
subgroups of MOEA solutions, differentiated by the
way of involving the decision maker:
“a priori” preference, where the decision maker
combines all the objectives into a single scalar
function;
progressive preference, where the decision
making and optimization processes alternate;
“a posteriori” preference, where the resulting set
of Pareto-optimal solutions is presented to the
decision maker who selects the final solution
from the set provided.
This paper is focused entirely on Pareto-based
MOEA solutions with “a posteriori” preference. It is
caused by the fact that decision-making in the
proposal described in the next section is transferred
to the multiobjective ranking method.
One should be aware that “MOEA” term refers
to some algorithmic framework rather than a
specific ready-to-use and universal solution. Thus,
already known MOEA techniques should be applied
prior to building a problem-oriented multiobjective
evolutionary application. Thus, the following
subsections describe core set of basic MOEA
techniques.
3.1.1 Secondary population
Secondary population is an additional population
maintained throughout MOEA execution time,
collecting all Pareto-optimal solutions found so far
during the search process. Its main goal is to
preserve all desirable solutions throughout the
generation process. In accordance with Pareto
notation the secondary population is termed
)(tP
known
,
where t denotes current generation number.
Similarly, a current set of Pareto-optimal solutions
determined at the end of each generation with
respect to the current MOEA generational
population is termed
)(tP
current
. It is assumed, though,
that
)0(
known
P
is an empty set and
known
P
without t
annotation stands for the final set of Pareto optimal
solutions collected before MOEA termination.
Several strategies of secondary population storage
exist. The most obvious and commonly used is the
strategy of adding
)(tP
current
to
)(tP
known
in the end
of each generation t:
)1()()( = tPtPtP
knowncurrentknown
(1)
The set of
)(tP
known
must be periodically checked
against obsolete Pareto solutions as Pareto
optimality should always be evaluated within current
set. The simplest policy does not assume explicit
copying
)(tP
known
solutions back into the next
population. However, other strategies exist where
275
the secondary population participates in a
tournament selecting next generations or is directly
inserted into the next mating population.
3.1.2 Multiobjective ranking
Multiobjective evolutionary approach enforces
that some transformation of the performance vector
into a scalar fitness value is necessary. This
transformation is achieved by means of a
multiobjective ranking, often also referred to as
Pareto ranking. In general, there are several ranking
methods. All these methods are based on an
assumption that preferred Pareto optimal solutions
are ranked the same value whereas other solutions
are assigned some less desirable rank value.
3.1.3 Niching and fitness sharing
The term niching refers to the process of
clustering in either solution space or criterion space.
In this process clusters consist of groups formed by
some individuals selected from the entire population.
Niching is primarily aimed at finding and
maintaining multiple optima. In result, this technique
should assure a good spread of discovered solutions
and prevent MOEA algorithm from being swamped
by solutions with identical fitness. Fitness sharing is
the most popular realization of the niching
technique. It is based on an assumption that
individuals in a particular niche share available
resources. Thus, the more individuals are located in
the vicinity of a certain individual, the more its
fitness value is deteriorated. The vicinity is most
often determined by a distance measure d(i,j) and
specified by niche radius σ
share
. The distance
function d(i,j) operates either in solution space or
criterion space, resulting in appropriate type of
fitness sharing.
3.1.4 Mating restrictions
The idea behind restricted mating is to prevent or
minimize offsprings, so called lethals, created by
recombination of chromosomes from different
niches. Such individuals can lead to degradation of
MOEA performance. To remedy the problem some
restrictions to mating might be introduced providing
a distance metric and a maximum distance value
σ
mate
for which mating is still permitted. The most
popular solution for mating restriction is to introduce
the fitness sharing niche radius σ
share
into the
problem and setting σ
mate
=σ
share
. However, it is
questioned (Van Veldhuizen 2000) whether such
restriction policy is indeed a compulsory MOEA
component, especially when there is no quantitative
evidence of its benefits.
3.2 Fuzzy TOPSIS as a multobjective ranking
method
Ranking methods belong to a group of
multiobjective optimisation methods where
preferences of the decision maker are utilized to
build a ranking of alternatives. The ranking is a list
of all possible solutions ordered from the least to the
most desirable one. Given order is achieved by
casting all the objectives into a single-objective goal
function. Preferences are reflected by weight values
assigned to the original objectives in the aggregated
goal function.
Technique for Order Preference by Similarity to
an Ideal Solution (TOPSIS) is a multiobjective
ranking method proposed by Hwang & Yoon
(Hwang et al. 1982). The method is based on a
concept that the best alternative among the available
alternative set is the closest to the best possible
solution and the farthest from the worst possible
solution simultaneously. The best possible solution,
referred to as an ideal one, is defined as a set of the
best attribute values, whereas the worst possible one,
referred to as a negative-ideal solution, is a set of the
worst attribute values. In order to compare the
alternatives and build the output ranking, the
Euclidean distances between each alternative and
both the ideal and the negative-ideal solutions are
calculated. Then the closeness coefficient is
determined to measure the two distances
simultaneously. Sorting in descending order the
coefficient values assigned to the alternatives creates
the final TOPSIS ranking. The alternative with the
highest ranking value is considered as the most
desirable.
Based on the original TOPSIS method, an
extension has been proposed by Chu & Lin (Chu et
al. 2003) providing support for fuzzy criteria and
fuzzy weights both described by triangular fuzzy
values. The new method has already been applied to
navigational problems in (Szlapczynska 2005).
Detailed Fuzzy TOPSIS algorithm differs from
standard TOPSIS one in the following:
each criterion can be either crisp or fuzzy, the
latter means that the criterion is described by a
linguistic variable with triangle fuzzy values;
weight vector is described by a set of triangle
fuzzy values assigned from another linguistic
variable;
decision matrix is converted to a fuzzy decision
matrix;
scaled V matrix is a result of multiplication of
fuzzy weight vector and fuzzy decision matrix;
in order to determine ideal and negative-ideal
solutions first the scaled V matrix is defuzzified,
276
then the standard TOPSIS computations are
continued.
As a result, similarly to original TOPSIS, final
Fuzzy TOPSIS ranking consists of crisp values. The
alternative with the highest ranking value is
considered as the most desirable.
4 WEATHER ROUTING FOR SAIL-ASSISTED
SHIPS AS A MULTIOBJECTIVE
OPTIMISATION PROBLEM
4.1 Model assumed
For a sail-assisted ship separate ship and sail models
are assumed. Ship model is based on a B-470 bulk
carrier. Its basic parameters are shown in Table 1.
Table 1. Basic parameters of the ship model
Value
172 m
22.8 m
9.5 m
14.3 m
15 kn
30,288 t
Sail model presented in Figure 1 is based on
textile winds from “Oceania” ship. There are six
sails forming a palisade. Each sail has 522m
2
sail
surface area.
For given ship and sail models, the speed
characteristic is determined based on algorithm by
Oleksiewicz (Oleksiewicz in prep.). An exemplary
speed characteristic for starboard tack is presented in
Figure 2.
Fig. 1. Sail model
Fig. 2. Exemplary speed characteristic for given ship and sail
model - starboard tack
4.2 Problem definition
Prior to a problem definition a model of ship
movements is assumed as kinematical one with
elements of ship’s dynamics according to possibility
of manoeuvre execution.
The values to be found by the optimisation
process of route finding are route’s waypoints
defined by their geographical coordinates and ship
speed between two consecutive waypoints. The
optimisation criteria are split into two groups,
namely:
economic criteria, such as passage time and fuel
consumption;
safety criteria, represented by traffic intensity and
constraint violation factors.
Thus, the goal function is defined as presented by
equations 2 - 4:
minrifvtfrivtf
traffsafetyfcreconomytrafffcr
= )},(),,({),,,(
(2)
)}(),({),(
__ fcnconsumptiofuelrtimepassagefcreconomy
vftfvtf =
(3)
)}(),({),(
int_int_
rfifrif
violationconstratraffensitytraffictraffsafety
=
(4)
where:
t
r
passage time [h],
v
fc
total fuel consumption [t],
i
traff
traffic intensity factor [/],
277
r penalty function for constraint violation [/].
All limitations to the problem domain in weather
routing are purely navigational. Landmasses that
cannot be crossed constitute the prime constraint.
Even a small violation of the constraint results in a
route unacceptable from navigational standpoint.
Along with an assumption that land shore does not
change its shape during a route execution this
constraint is assumed static. However, other
navigational constraints exist that do not fall into
category of static ones, namely ice phenomenon and
tropical cyclones. Available information about ice
and cyclones is mostly derived from forecasted, that
is probabilistic, data. Moreover, both ice concentra-
tion as well as a centre of a tropical depression
changes with time. Thus these constraints are
assumed fuzzy dynamic. Last but not least constraint
is determined by the assumed ship’s draught. Given
ship model is able to navigate only through waters
sufficiently deep for assumed draught value.
4.3 Proposed solution for the problem
The solution proposed is based on the optimisation
criteria defined in the previous subsection. It utilizes
two basic multiobjective mechanisms, namely
multiobjective evolutionary algorithm (MOEA) and
multiobjective ranking method - Fuzzy TOPSIS. In
general, the main algorithm is presented in Figure 3.
In the proposed solution the evolutionary framework
is responsible for iterative process of population
development. MOEA-specific techniques, namely
multiobjective ranking, secondary population and
niching extend the framework to achieve a
Pareto-optimal set of routes according to given
criteria set. Mating restriction, as a MOEA technique
of questionable profits, is not utilized in the
proposal.
Initial population of routes will be generated
based on an orthodrome and a route determined by
the adopted isochrone method (Szlapczynska et al. in
prep.). The population should consist of avg. 50
individuals, each being a random mutation of the
basic routes. Also pure orthodrome and isochrone
route should belong to the initial population. In
addition to that, it is worth considering whether
some other routes optimising one of the other criteria
(fuel consumption, vessel traffic intensity, degree of
constraints violation) should be included in the
initial population.
Fig. 3. Main algorithm of the proposed multiobjective weather
routing
The process of fitness assignment will include
multiobjective ranking and niching, necessary
for assumed multiobjective goal function. In the end
of each generation an increase of fitness function
will be determined. The evolutionary process will
be terminated whenever the increase will be
satisfactorily small (smaller than some ε value).
Then, the ranking Fuzzy TOPSIS method will
prepare the output ranking facilitating the final
decision which route to choose.
If the termination condition is not met the
iterative process of population development must
continue. First, dominance between individuals
should be determined. Based on this information the
secondary population will be updated. A new
population will be created by means of selection and
modification processes. Again all individuals in
the population will be assigned their fitness values
and the evolutionary iterations will proceed until the
termination condition is finally met.
5 SUMMARY
The paper describes a possible multiobjective
approach to weather routing. Two families of
multiobjective decision-making methods are
presented, namely multiobjective evolutionary
algorithms and ranking methods. The author
278
describes model assumed and defines problem of
route finding in weather routing. Then a solution to
given model and problem is presented. The proposal
includes an algorithm determining a route satisfying
given optimisation criteria and navigational
constraints. Further details on the proposed
algorithm will be provided as soon as the weather
routing algorithm in finally implemented.
REFERENCES
Chu, T.C. & Lin Y.C. 2003. A Fuzzy TOPSIS Method for
Robot Selection, The International Journal Of Advanced
Manufacturing Technology: pp. 284-290.
Hagiwara, H. 1989. Weather routing of (sail-assisted) motor
vessels, PhD Thesis, Delft: Technical University of Delft.
Hwang, C.L. & Yoon K. 1982. Multiple Attribute Decision
Making - A State-of-the-Art Survey, The Journal of the
Operational Research Society, Vol. 33, No. 3: pp. 289-300.
James, R.W. 1957. Application of wave forecast to marine
navigation, Washington: US Navy Hydrographic Office.
Oleksiewicz, B. (in preparation). Motor-sailer A hybrid
Propulsion for Commercial Vessels. The case study.
Spaans, J.A. 1986. Windship routeing, Delft: Technical
University of Delft.
Szlapczynska, J. 2005. Fuzzy TOPSIS as a multicriteria
decision making method in navigational applications (in
Polish), Works of the Faculty of Navigation, pp. 180-192.
Szlapczynska, J. & Smierzchalski R. (in preparation). Adopted
isochrone method improving ship safety in weather routing
with evolutionary approach.
Wisniewski, B. 1991. Methods of route selection for a sea
going vessel (in Polish), Gdansk: Wydawnictwo Morskie.
Van Veldhuizen, D.A. & Lamont, G.B. 2000. Multiobjective
Evolutionary Algorithms: Analyzing the State-of-the-Art,
Evolutionary Computation, pp.125-147.