International Journal
on Marine Navigation
and Safety of Sea Transportation
Volume 6
Number 4
December 2012
1 INTRODUCTION
Decision-making processes, especially those
occurring in the transport, held responsibility for the
safety of people, equipment and the environment.
Such important decisions should be taken with a
minimum of uncertainty of the decision maker. This
uncertainty may be due to the existence of a number
of factors, such as: the level of training of decision-
makers, the lack of information regarding the
situation of the surrounding area and sometimes an
excess of information provided to decision-makers
simultaneously from multiple sources.
IT systems become more efficient over time
regarding constant increase of computing power
available to standard users. That allows developing
advanced systems which collect and analyze
relevant data to support decisions that minimize the
risk of collision.
These advanced systems may be used in support
decisions on the real ships maneuvering and ship
models and simulators used during the training of
future officers at training centers.
Currently still being developed computational
methods are meuroevolutionary methods which,
thanks to its efficiency are becoming widely used in
many fields of science and technology, such as:
− automation and robotics systems, e.g. control of a
robot arm (Siebel and Sommer 2007);
− designing and diagnostic systems, e.g. mobile
hardware acceleration (Larkin, Kinane and
O'Connor 2006), search hull damage (Kappatos,
Georgoulas, Stylios and Dermatas 2009),
processors design (Ratuszniak 2012), the
detection and evaluation of the risk of breast
cancer (Janghel, Tiwari, Kala and Shukla 2012)
− control systems: for example, a helicopter flight
stabilization (De Nardi, Togelius, Holland and
Lucas 2006);
− decision support systems, e.g. systems applied
artificial intelligence in computer games
(Kenneth, Bryant and Risto 2005).
A large number of positive results of the
implementation of neuroevolutionary methods
obtained in many areas of science encouraged the
author of this paper to undertake research and
Neuroevolutionary Ship Handling System in a
Windy Environment
M. Łącki
Gdynia Maritime University, Gdynia, Poland
ABSTRACT: This paper presents the advanced intelligent ship handling system able to simulate and
demonstrate learning behavior of artificial helmsman which controls model of ship in a windy environment of
restricted water area. Simulated helmsmen are treated as
individuals in population, which through
environmental sensing and evolutionary algorithms learns to perform given task efficiently. The task is: safe
navigation through heavy wind channels. Neuroevolutionary algorithms, which develop artificial neural
networks through evolutionary operations, have been applied in this system.
453
develop his own algorithms intended for use in
maritime transport (Łącki 2010b, a).
In this paper, the extension of the functionality of
navigational decision support systems is proposed.
This solution generates specifications of
maneuvering decisions (rudder angle and propeller
thrust) that maintain a safe ship trajectory computed
in the available water channel. In addition to rudder
angle and propeller thrust this system also includes
information about time of their execution. It is
possible in this system that all maneuvering
decisions may be calculated and presented in real
time for a given ship dynamics in the presence of
certain external disturbances.
1.1 Reinforcement Learning Algorithms
One of the main tasks in machine learning is to
create the advanced systems that can effectively find
a solution of given problem and improve it over
time. Reinforcement learning is a kind of machine
learning, in which an autonomous unit, called a
robot or agent, performs actions in a given
environment. Through interaction and the
observation of the environment (by input signals)
and performing an action he affects this environment
and receives an immediate score called
reinforcement or reward (Figure 1).
Environment
Artificial
Helmsman
Reward
(Reinforcement)
Action
State
vector
Figure 1. Interaction of helmsman with an environment.
The main task of the agent is to take such actions
to adjust the value R which is the sum of the partial
reinforcement as much as possible (1).
=
+
+
(1)
Such abilities are very important for simulating
helmsman behavior in ship maneuvering on
restricted waters.
For simpler layouts learning process can be
performed using classic approach, i.e. Temporal
Difference Reinforcement Learning (Tesauro 1995;
Kaelbling, Littman and Moore 1996) or Artificial
Neural Networks with fixed structures. Dealing with
high-dimensional spaces is a known challenge in
Reinforcement Learning approach (Łącki 2007)
which predicts the long-term reward for taking
actions in different states (Sutton and Barto 1998).
Evolving neural networks with genetic algorithms
has been highly effective in advanced tasks,
particularly those with continuous hidden states
(Kenneth, Bryant and Risto 2005). Neuroevolution
gives an advantage from evolving neural network
topologies along with weights which can effectively
store action values in machine learning tasks. The
main idea of using evolutionary neural networks in
ship handling is based on evolving population of
helmsmen.
The artificial neural network is the helmsman's
brain making him capable of observing actual
navigational situation by input signals and choosing
an appropriate action. These input signals are
calculated and encoded from current situation of the
environment.
In every time step the network calculates its
output from signals received on the input layer.
Output signal is then transformed into one of the
available actions influencing helmsman’s
environment. In this case the vessel on route within
the restricted waters is part of the helmsman’s
environment. Main goal of the helmsmen is to
maximize their fitness values. These values are
calculated from helmsmen behavior during
simulation. The best-fitted individuals, which react
properly to wind effect, become parents for next
generation.
2 THE FORCES ACTING ON THE SHIP
External disturbances acting on a seagoing ship are
mainly wind, wave and ocean current. In this paper
the author assumes that the waves in the harbor area
have little effect on ship maneuvering, and this type
of interference is omitted in the system.
As the ship moves forward on the straight path
(assuming there aren’t any significant distorting
external forces) there are the two major forces acting
on her - the force from the propellers and the force
of water resistance. At a constant speed, these forces
are equal, but with the opposite direction.
When the ship is turning the additional forces act
on the rudder and lateral forces from water pressure
appear. During this maneuver, the ship loses a little
of her velocity and the pivot point moves back
toward amidships.
2.1 Effect of wind in the ship handling
Under pressure of wind force, depending of the
ships’ design (location of the superstructure, the
454
deployment of cargo and on-board equipment, etc.)
she tends to deviate from the course, with the wind
or into the wind. The smaller the speed and draft of
the ship, the greater the influence of wind. Of
course, the size of the side surface exposed to wind
is essential to the ships movement.
When ship moves forward the center of effort of
the wind (WP) is generally close to amidships, away
from pivot point (PP). This difference creates a
substantial turning lever between PP and WP thus
making the ship to swing of the bow into the wind
(with the superstructure deployment at stern)
(Figure 2).
Figure 2. Wind effect and turning lever of ship mowing
forward.
For ship moving forward there are defined terms
of relative wind speed V
rw
and angle of attack
γ
rw
as
follows (Fossen 2011):
=
+
(2)
= arctan
,
(3)
where:
=
(4)
=
(5)
where: u, u
w
, v, v
w
are longitudinal and lateral
velocities of ship and wind, respectively.
Wind forces acting on ship are generally
calculated as follows:
=
(
)
(6)
=
(
)
(6)
=
(
)
(6)
where:
−
ρ
air
– air density,
− A
x
– surfaces affected by wind,
− L
0a
– ship’s length,
− C
n
– coefficients calculated from available
characteristics of ships’ model (Figure 3.),
Figure 3. Equations coefficients for relative wind.
3 NEUROEVOLUTION OF AUGMENTING
TOPOLOGIES
Neuroevolution of Augmenting Topologies (NEAT)
method is one of the Topology and Weight Evolving
Artificial Neural Networks (TWEANN’s) method
(Kenneth and Risto 2002b). In this method the
whole population begins evolution with minimal
networks structures and adds nodes and connections
to them over generations, allowing complex
problems to be solved gradually starting from simple
ones.
The modified NEAT method consist four
fundamental rules which deal with challenges that
exist in evolving efficient neural network topology:
455
1 Begin with a minimal structure and add neurons
and connections between them gradually to
discover most efficient solutions throughout
evolution.
2 Cross-over disparate topologies in a meaningful
way by matching up genes with the same
historical markings.
3 Separate each innovative individual into a
different species to protect it disappearing from
the population prematurely.
4 Reduce oversized topologies by removing
neurons and connection between them to provide
and sustain good overall performance of a whole
population of helmsmen.
3.1 Genetic Encoding
Evolving structure requires a flexible genetic
encoding. In order to allow structures to increase
their complexity, their representations must be
dynamic and expandable (Braun and Weisbrod
1993). Each genome in NEAT includes a number of
inputs, neurons and outputs, as well as a list of
connection genes, each of which refers to two nodes
being connected (Figure 4.).
Figure 4. Genotype and phenotype of evolutionary neural
network.
In this approach each connection gene specifies
the output node, the input node, the weight of the
connection, and an innovation number, which allows
finding corresponding genes during crossover.
Connection loopbacks are also allowed, as shown in
figure 4.
3.2 Genetic operations
There are two main genetic operations: cross-over
and mutation. During cross-over two individuals
(parents), exchange their genetic material in purpose
of creating new individual (an offspring). The
system knows exactly which genes match up with
which through innovation numbers. Genes that do
not match are either disjoint or excess, depending on
whether they occur within or outside the range of the
other parent’s innovation numbers.
In crossing over operation, the genes with the
same innovation numbers are lined up. The offspring
is then formed in one of three ways:
− In uniform crossover: matching genes are
randomly chosen for the offspring genome, with
all disjoints and excesses from both parents.
− In blended crossover: the connection weights of
matching genes are averaged, disjoints and
excesses are chosen randomly.
− In elite crossover: disjoints and excesses are taken
from more fit parent only, all redundant genes
from less fit parent are discarded. All matching
genes are averaged.
These types of crossover were found to be most
effective in evolution of neural networks in
extensive testing compared to other methods of
crossover (Kenneth and Risto 2002a).
Disabled genes have a chance of being re-enabled
during mutation, allowing networks to make use of
older genes once again.
Evolutionary neural network can keep historic
trails of the origin of every gene in the population,
allowing matching genes to be found and identified
even in different genome structures. Old behaviors
encoded in the pre-existing network structure have a
chance to not to be destroyed and pass their
properties through evolution to the new structures,
thus provide an opportunity to elaborate on these
original behaviors.
Through mutation, the genomes in modified
NEAT will gradually get larger for complex tasks
and lower their size in simpler ones. Genomes of
varying sizes will result, sometimes with different
connections at the same positions. Any crossover
operator must be able to recombine networks with
differing topologies, which can be difficult.
Historical markings represented by innovation
numbers allow NEAT to perform crossover without
analyzing topologies. Genomes of different
organizations and sizes stay compatible throughout
evolution, and the variable-length genome problem
is essentially solved. This methodology allows
NEAT to increase complexity of structure while
different networks still remain compatible.
Additionally different sizes and structures of
networks group their genetic material into species.
456
3.3 Speciation
Speciation of the population assures that individuals
compete primarily within their own niches instead of
competition within the whole population. In this way
topological innovations of neural network are
protected and have time to optimize their structure
before they have to compete with other experienced
agents in the population.
During species assigning process, as described in
(Łącki 2010c), when a new individual appears in
population, its genome shall be assigned to one of
the existing species. If this offspring is structurally
too innovative comparing to any other individuals in
whole population, the new species is created.
In the first step of species reproduction process
the system eliminates the lowest performing
members from the population. In the next step the
offspring replaces eliminated worst individual
(Fig. 5).
Figure 5. Example of reproduction in elitist selection method.
4 SIMULATION RESULTS
The main goal of authors work is to make a system
able to simulate a safe passage of ship moving
through a restricted coastal area in heavy and
variable wind conditions. This goal may be achieved
with Evolutionary Neural Networks.
Figure 6. Sample data signals of ship handling with ENN.
Navigational situation of a moving vessel can be
described in many ways. Most important is to define
proper state vector from abundant range of data
signals (Fig. 6.) and arbitrary determine fitness
function values received by the helmsman.
The main input signals are gathered from data
listed below:
− Ships course over ground,
− Ships angular velocity,
− The ship is on the collision course with an
obstacle,
− Distance to collision,
− The ship is approaching destination,
− Ships angle to destination,
− The ship is heading out of the area,
− Distance to current canal borders,
− Ship is heading on goal,
− Distance to goal,
− Wind velocity,
− Angle of wind.
All the input signals are encoded either binary (0
or 1) or as a real values between 0 and 1. Some of
the input signals may be calculated as multi-criteria
values (Filipowicz, Łącki and Szłapczyńska 2006).
Neural network output values are signals for rudder
angle (δ) [deg] and thrust control [rpm].
Fitness calculation defines helmsman ability to
avoid obstacles and react to wind forces while
sailing toward designated goal. The fitness value of
an individual is calculated from arbitrary set action
values, i.e.:
− -10 when ship is on the collision course (with an
obstacle or shallow waters),
− +10 when she's heading to goal without any
obstacles on course,
− -100 when she hits an obstacle or run aground,
− +100 when ship reaches a goal,
− -100 when she departs from the area in any other
way, etc.
In the simulation of safe passage through
restricted waters there are no moving vessels in the
area (Fig. 7.). In this situation when ship enters a
heavy side wind channel, there is a risk to hit an
obstacle if no action is being made by the helmsman.
Artificial helmsman observes current situation which
is encoded as input signals for his neural network
and calculates the best (in his opinion) rudder angle
(Figure 8).
457
Figure 7. Model of windy coastal environment.
Figure 8. Simulation results of the systems performance in
heavy wind environment.
5 REMARKS
Neuroevolution approach to intelligent agents
training tasks can effectively improve learning
process of simulated helmsman behavior in ship
handling (Łącki 2008). Artificial neural networks
based on NEAT increase complexity of considered
model of ship maneuvering in restricted waters.
Implementation of additional disturbances from
wind in neuroevolutionary system allows simulating
complex behavior of the helmsman in the
environments with much larger state space than it
was possible in a classic state machine learning
algorithms (Łącki 2007). Positive simulation results
of maneuvers in variable wind conditions encourage
to add other input signals to the system, like river
currents, which will be included in future research.
REFERENCES
Braun H. and Weisbrod J. (1993). Evolving Feedforward
Neural Networks. International Conference on Artificial
Neural Networks and Genetic Algorithms, Innsbruck,
Springer-Verlag.
De Nardi R., Togelius J., Holland O. E. and Lucas S. M.
(2006). "Evolution of Neural Networks for Helicopter
Control: Why Modularity Matters." Evolutionary
Computation, 2006. CEC 2006. IEEE Congress on: 1799-
1806.
Filipowicz W., Łącki M. and Szłapczyńska J. (2006).
"Multicriteria Decision Support for Vessels Routing."
Archives of Transport 17: 71-83.
Fossen T., I. (2011). Handbook of marine craft hydrodynamics
and motion control, John Wiley & Sons, Ltd.
Janghel R. R., Tiwari R., Kala R. and Shukla A. (2012).
"Breast Cancer Data Prediction by Dimensionality
Reduction Using PCA and Adaptive Neuro Evolution."
IJISSC 3(1): 1-9.
Kaelbling L. P., Littman M. L. and Moore A. W. (1996).
"Reinforcement Learning: A Survey." Journal of Artificial
Intelligence Research cs.AI/9605: 237-285.
Kappatos V. A., Georgoulas G., Stylios C. D. and Dermatas E.
S. (2009). Evolutionary dimensionality reduction for crack
localization in ship structures using a hybrid computational
intelligent approach. Proceedings of the 2009 international
joint conference on Neural Networks. Atlanta, Georgia,
USA, IEEE Press: 1907-1914.
Kenneth O. S., Bryant B. D. and Risto M. (2005). "Real-time
neuroevolution in the NERO video game." IEEE
Transactions on Evolutionary Computation 9(6): 653-668.
Kenneth O. S. and Risto M. (2002a). Efficient evolution of
neural network topologies. Proceedings of the Evolutionary
Computation on 2002. CEC '02. Proceedings of the 2002
Congress - Volume 02, IEEE Computer Society: 1757-
1762.
Kenneth O. S. and Risto M. (2002b). Efficient Reinforcement
Learning Through Evolving Neural Network Topologies.
Proceedings of the Genetic and Evolutionary Computation
Conference, Morgan Kaufmann Publishers Inc.: 569-577.
Larkin D., Kinane A. and O'Connor N. (2006). Towards
hardware acceleration of neuroevolution for multimedia
processing applications on mobile devices. Proceedings of
the 13th international conference on Neural information
processing - Volume Part III. Hong Kong, China, Springer-
Verlag: 1178-1188.
Łącki M. (2007). Machine Learning Algorithms in Decision
Making Support in Ship Handling. TST, Katowice-Ustroń,
WKŁ.
Łącki M. (2008). Neuroevolutionary approach towards ship
handling. TST, Katowice-Ustroń, WKŁ.
Łącki M. (2010a). "Model środowiska wieloagentowego w
neuroewolucyjnym sterowaniu statkiem." Zeszty Naukowe
Akademii Morskiej w Gdyni 67: 31-37.
Łącki M. (2010b). "Wyznaczanie punktów trasy w
neuroewolucyjnym sterowaniu statkiem." Logistyka 6.
Łącki M. (2010c) Speciation of Population in
Neuroevolutionary Ship Handling. TransNav - International
Journal on Marine Navigation and Safety of Sea
Transportation, 4(2), 211-216
Ratuszniak P. (2012). "Processor array design with the use a
genetic algorithm." Lecture Notes in Computer Science
7116.
Siebel N. T. and Sommer G. (2007). "Evolutionary
reinforcement learning of artificial neural networks."
International Journal of Hybrid Intelligent Systems -
Hybridization of Intelligent Systems 4(3): 171-183.
Sutton R. and Barto A. (1998). Reinforcement Learning: An
Introduction, The MIT Press.
Tesauro G. (1995). "Temporal difference learning and TD-
Gammon." Communications of the ACM 38(3): 58-68
458
