865

1 INTRODUCTION

Restricted water operations (e.g. rivers and ports)

require an autonomous boat to be capable of fast

response to avoid obstacles while maintaining its

course. A well-designed autonomous boat offers

solutions that contribute towards safer waterways

and higher economic yields. Such activities include

periodically patrolling the river for reconnaissance,

environmental monitoring, and transportation of

humans and goods. Currently, autonomous boats rely

heavily on several hardware devices such as global

positioning systems (GPS) [1], distance sensors [2],

radars [3], and cameras [4]. While such hardwares are

common for the construction of an autonomous boat,

several problems exist such as the high price of radar,

GPS multi-pathing which resulted due to high-

density vegetations, and expensive high-quality

cameras which also require frequent maintenance.

This leaves the autonomous boat with the use of

distance sensors, which may offer a robust [5] and

low-cost solution to assist in navigation and obstacle

avoidance tasks, driven by a well-trained artificial

intelligence system.

Fundamentally, an autonomous boat/autonomous

surface vehicle is a subclass of mobile robots.

Conceptually, a mobile robot consists of a control

system, sensors (distance sensors, light sensors, vision

sensors), and actuators (motors, servos) to respond to

its surrounding [6]. An autonomous mobile robot is

capable of sensing its surroundings and acting upon

Neuroevolutionary Autonomous Surface Vehicle

Simulation in Restricted Waters

A.F. Ayob, N.I. Jalal, M.H. Hassri, S.A. Rahman & S. Jamaludin

University of Malaysia, Terengganu, Malaysia

ABSTRACT: Safe, accurate, and predictable autonomous systems in marine vehicles are paramount. An

understanding of an intelligent system fitted inside a ship is critical to ensure an autonomous ship is safe to be

operated. Although the use of artificial intelligence in the design of the road-based vehicle has arrived at the

self-driving level, there exists a significant gap within the research of autonomous ship to operate in restricted

water (riverine and ports). Hence, this article shall discuss the relevant works of literature to set a preliminary

guiding principle for the design of an autonomous ship. We present a simple illustrative framework as a

starting point for ship designers to begin working in a simulated environment, which can be used as a

foundation before the physical autonomous-ships are constructed and tested in a real-world situation. The

framework consists of a virtual 3D environment and a surface vehicle with distance sensors, controlled by a

neuroevolution-based autonomous piloting system. In this work, two scenarios will be presented: navigation in

restricted waters, and obstacle avoidance capability of an autonomous ship. Results show that the resulting

autonomous surface vehicle (ASV) is also capable of performing obstacle avoidance in the test track, albeit not

being trained to do so in the training track. The work demonstrated in this paper is useful to the ship designers

and can be extended for scenario-based planning for autonomous ship design.

http://www.transnav.eu

the International Journal

on Marine Navigation

and Safety of Sea Transportation

Volume 14

Number 4

December 2020

DOI: 10.12716/1001.14.04.11

866

the surrounding based on its mission definition. In

the field of maritime technology, there are vast

potential use-cases for autonomous surface vehicle

robots. While the primary mission of an autonomous

surface vehicle varies (e.g. reconnaissance, patrol,

intercept, environmental monitoring), several

common themes constitute as the sub-tasks of such

autonomous surface vehicles such as manoeuvring

[7]–[10], cruise control [11]–[14] and collision

avoidance [9], [10], [15].

In this paper, we present a concise state of the art

5-years review on the application of

neuroevolutionary methods in maritime technology,

particularly in the ship design discipline. A

fundamental example using state-of-the-art tool is

presented, where an autonomous surface vehicle

(ASV) or an agent is shown to interact with an

uncertain environment (dynamic water response due

to buoyancy, uneven riverbank terrain, potential

collision situation with other floating objects) while

avoiding any collision situation through automated

steering and thruster control.

In this work, autonomous control of a surface

vehicle is designed where an agent demonstrates

river navigation while avoiding obstacles in a

dynamic environment, using artificial neural network

(ANN), as shown in Figure 1. The desirable steering

and throttle control are achieved using ANN in which

the weight and biases of the ANN are found using the

unsupervised machine learning method. The

unsupervised machine learning method used in this

work is reinforcement learning, in which an agent is

rewarded based on the attainment of accumulated

checkpoints number in each simulation run. This is

achieved using Genetic Algorithm (GA), in which for

every simulation runs, the best designs shall be

retained to evolve in a search for the best design until

an arbitrary generation number achieved.

We evaluate the performance of the

neuroevolutionary ship control using computer

simulation, focusing on the ship’s capability to follow

the river path, while at the same time avoiding

collision with other floating objects such as weather

buoys.

The main contribution of this paper is the state-of-

the-art review of neuroevolutionary ship control and

the experimental results of an autonomous surface

vehicle operating in a highly dynamic environment.

The review of literature shall be elaborated in the next

section, followed by the methodology, experimental

results, discussion, and conclusion.

Figure 1. Neuroevolutionary ship manoeuvring control.

2 RELATED WORKS

Neuroevolutionary-based works in the literature

concerning ship steering or piloting remains lacking

as compared with autonomous cars. This is due to the

high volume of production of cars compared with

ships within consumer markets. Based on the

keyword search (Elsevier’s ScienceDirect) on research

and review articles of ‘autonomous car’ versus

‘autonomous ship’, a stark difference (50%

differences) in terms of volumes of literature can be

seen, as shown in Figure 2. Such preliminary search

has demonstrated a considerable research gap within

this discipline and requires attention as the industry

is moving toward the 4

Industrial Revolution.

Figure 2. Numbers of publications between Autonomous

Cars versus Autonomous Ships spanning across 20 years.

In this section, the related works available in the

literature concerning autonomous ship shall be

reported. The subsection shall be grouped into three

major topics, which are the application of

neuroevolution (evolutionary-based artificial neural

networks), the study of ship motion of autonomous

ships, and finally the safety-related works in

autonomous ship developments.

867

2.1 Neuroevolutionary Works in Ship Steering/Piloting

Autonomous ship berthing study was presented by

[16], [17] where backpropagation (BP) neural network

was incorporated to study the propeller revolution

controller under gust wind disturbances. Through the

use of a feedforward neural network, the study was

extended in [18], [19], [20] to justify the network’s

real-time response within the context of safe distance.

Similar work using ANN controller for automatic

berthing in different ports was presented in [21] and

[22]. A neuroevolutionary neural network

constructed via the use of evolutionary algorithms

was proposed by [23] to perform the task of crossing

or overtaking the ship-based COLREGs rules. This

ensures the ship to manoeuvre safely and efficiently.

Within the same context, [24] compared the use of

direct and indirect encoding for neuroevolutionary-

based ship handling, which uncovered the ability of

indirect encoding to generalize and react to new

states without bias towards the learned patterns,

however noting that direct encoding method may

adapt faster to new sudden changes.

Path following is crucial to maintain the safe

cruise of a ship. The works presented in [25]

demonstrated the capability of a vessel to return the

course-keeping path of a ship on the routes efficiently

using the neuroevolutionary method. Reported by

[26], such an algorithm is independent enough to find

the most effective path by considering a list of

possible solutions.

It can be noted that most of the works presented in

the field of neuroevolutionary ship piloting are

within the scope of simulation works [27]. Ship

simulation is a very powerful method to represent

real-life situations, to train ship operators and

intelligent algorithms while at the same time

examining typical scenarios in real-world situations

[28].

2.2 Motion Study of Autonomous Ships

One of the challenging aspects of control system

development in marine discipline is the highly non-

linear motion of ships, compared with road-based

vehicles. [7] demonstrated the use of artificial

intelligence to steer ships in completing manoeuvring

tasks autonomously. A hybrid of fuzzy logic and

artificial neural network (ANN) to perform autopilot

ship navigation was demonstrated in the work of

[8], while [9] proposed the use of a recursive neural

network to perform tactical circles and zig-zag

motions. An innovative offshore autonomous rescue

vessel was introduced in the work of [11], [12] to

assist large ships in the search-and-rescue mission.

Radial basis function (RBF) neural network was

incorporated by [13] to approximate the unknown

non-linear terms of the non-linear ship course control

system. Via the use of similar techniques, [14]

demonstrated an approach to simulate and control

the ship’s motion which caters to environmental

disturbances. [29] recommended the use of recursive

neural networks (RNN) due to its flexibility in

counter-measuring ship dynamics where they give

useful perspectives for solving the problems in

control theory as well as applications in marine

systems.

In visual perception systems for autonomous

ships, [30] analyzed four artificial intelligent

algorithms and deep learning frameworks on the case

of object detection and tracking, namely; K-Nearest

Neighbor (K-NN), Artificial Neural Network (ANN),

Convolutional Neural Network (CNN) and Deep

Convolutional Neural Network (DCNN). In view of

automatic ship navigation systems with collision

avoidance, [31] proposed the use of potential field

method in terms of its effectiveness and practical

applications. Presented by [32], a Neural Network

Allocation (NNA) was used to provide the

transformation between the desirable generalized

forces as input and the individual thruster command

for the output.

2.3 Safety-Related Works

Convention on the International Regulations for

Preventing Collisions at Sea (COLREGs) is essential

to apply for the manned and unmanned vessel. To

comply with COLREGs rules, [10] presented the

application of neuroevolutionary methods in vessel

overtaking, head-on, and crossing scenarios using

artificial potential field (APF). However, [33] argued

that COLREGs regulation might fail in some instances

when both of the ships decide to take the same

decision, which may lead to ship collision. Therefore

it is suggested that any two vessels; (1) should not be

overtaking when there exists a crossing vessel, and (2)

an overtaking can be done by a faster vessel when

there is no crossing vessel in the area of proximity

[23].

A complete collision avoidance system should

consider certain factors (e.g. ship types, traffic status,

weather, and navigation technologies) [34]. Therefore,

to improve the collision avoidance system, [35]

proposed a reward function as the main component

of Concise Deep Reinforcement Learning Obstacle

Avoidance (CDRLOA) to act as a feedback data to the

system that evaluates the system performance of the

control behaviour at one scalar signal. [36] proposed

the dynamic predictive guidance technique to

address collision avoidance. It is agreed upon that

collision avoidance should meet the following

optimal rules; collision avoidance’s safety and

network availability, and the shortest navigation

distance during collision avoidance [37].

3 NEUROEVOLUTIONARY-BASED SHIP

STEERING/PILOTING

In this section, the methodology for

neuroevolutionary-based ship steering control shall

be presented. The basic concepts of genetic algorithm

shall be discussed, followed by the description of

agents (autonomous surface vehicle, ASV) and finally

the description of both the training track and testing

track incorporated in this work.

868

3.1 Description of Neuroevolutionary-based Ship

Steering Control

An artificial neural network is considered as an

individual in a population generated by Evolutionary

Algorithms (EA) such as Genetic Algorithms (GA).

Each ANN can vary in terms of its structures, or its

weights and biases, which are incorporated into each

of the individual agents within an environment. In

this work, a Conventional Neuroevolution (CNE) is

incorporated, in which the sets of weights inside of a

fixed-structure ANN are evolved using GA [38] [39]

[40]. Such an environment which is being utilized as

a platform to investigate the response of the agent can

be set up in two ways, whether in a physical

environment or virtual simulations. The response of

the agents (e.g. distance to a potential collision, the

velocity of agents, hydrostatic/hydrodynamic

performance, distance to a target, current position of

the agent) can be used as the input (objective

functions, variables, constraints) for the EA, in which

assisting in providing for the next generation of the

population.

The pseudo-code of the CNE-based

neuroevolution case study for an autonomous mobile

robot operating on a water surface can be observed in

Algorithm 1 below. In this case study, the structure of

the ANN is fixed with input with the size of 10

neurons (9 distance sensor inputs and one velocity

input), a hidden layer with the size of 3 neurons, and

an output layer with the size of 2 neurons (turning

angle coefficient, thruster coefficient). Using the fixed

structure of the ANN, an initial set of weights and

biases are generated randomly which are assigned to

each of the agents (P(g=1)). Upon receiving the

neurons, the agent shall complete its task to maximize

the number of checkpoints (reward) by autonomously

piloting itself within a challenging test track. In a

situation where the agent collided with the riverbank

of the track, the agent shall be discarded (penalty) for

that generation (g). The following generation shall

continue where the offspring (Pc(g)) are generated

from the best agent from the previous generation

until the stopping criteria (e.g. number generations)

of the GA is achieved.

Algorithm 1. Pseudo-code for the genetic algorithm in the

neuroevolution case study

1 g = 1;

2 initialize P(g=1);

3 while isNotTerminated() do

4 evaluateFitness P((g));

5 sortFitnes P((g));

6 P(g) = selectIndividual(P(g));

7 P(g) = crossOver(P(g));

8 P(g) = mutation(P(g));

9 Pc(g) =selectParents(P(g));

10 P(g) = Pc(g); // Children population became the

parents for the next generation.

11 g = g + 1;

12 end

The experiment can be described as a highly non-

linear constrained optimization explained as below:

Objective function: Maximization of the number of

checkpoints f(x) = f(x1, … , x10), where variable x1 to x9

is the array of distance measured from the riverbank

to the body of ASV, and variable x10 is the velocity of

the ASV during the time of measurement. Each

member of the population aims to maximize the

number of checkpoints achieved given a prescribed

time in seconds.

Constraint: A boolean data of {0,1} condition in

which the body of the ship should not touch the

riverbank or any floating object on the water. If the

body of the ship collided, the boolean would be set as

{1} which deemed unfit and discarded from the

remaining population of the agents, whereas {0}

means the agent is fit in which the fitness shall be

recorded for further evaluation of the next generation

run.

3.2 Description of Agents (ASV)

Autonomous Surface Vehicle (ASV) is a sub-

classification of Autonomous Mobile Robot, which

operates on the water surface. In this work, the ASV

is modelled using typical dimensions of working-

class Hydrographic Surveying ASV available in the

market [41], shown in Figure 3 and detailed in Table

1. In this work, the ASV is equipped with an array of

9 distance sensors with an interval of 20

, located

from the perpendicular of port-side to the starboard,

as shown in Figure 4.

The simulations conducted in this work

incorporated the commercial mesh-based physics

modelling codes [42] to model the ASV buoyancy

force (Equation (1)) and hydrodynamic force

(Equation (2)) in real-time. The simulation

environment is created within the Unity physics

engine, which suitable for fast physics modelling

(utilizing C# programming language) in the

preliminary design stage. In this work, the ASV shall

only rely on the input gathered from the distance

sensors and its velocity to decide its value of the

thruster coefficient and its rudder angle coefficient.

Figure 3. A typical working-class Hydrographic Surveying

ASV.

869

Figure 4. An array of 9 distance sensors with an interval of

20 degrees is incorporated in the ASV.

Table 1. Principal particulars of a working-class

hydrographic surveying ASV used in this work

_______________________________________________

Particulars Dimension

_______________________________________________

Length 2.4 m

Beam 1.0 m

Draft 0.3 m

Weight 500 kg

Top speed 7 knots

Number of sensors 9 distance sensors

_______________________________________________

F gV



(1)

where,

FB : Buoyant force in N



: Fluid density in kg/m

V : Displaced body of fluid in m

g : Gravity constant, 9.81 m/s



(2)

where,

FD : Drag force in N

cD : Drag coefficient



: Fluid density in kg/m

v : Velocity of ship in m/s

A : Surface area underwater in m

3.3 Description of Tracks (Training and Testing)

The training tracks incorporated in the simulation are

designed to mimic a very tight and challenging

restricted water scenario, as shown in Figure 5. The

training track consists of several characteristics, such

as:

1 one straight cruise

2 four sharp U-turns

3 one sharp 90

corner

4 the width of the riverbank is 10m

5 the dimension of the training track is 60m width

and 100m length

Figure 5. Training track shown from the Top-view, with the

‘checkpoints’ highlighted.

The testing track incorporated in the simulation is

a larger restricted water environment, e.g. an inland

water transport scenario, as shown in Figure 6.

Surrounded by typical objects in ports (buildings,

boats, weather buoys, and deck), the testing track

differs in terms of the size of the river with additional

obstacles to test the generalization capability of the

resulting ANN model. The training track can be

summarized as below:

1 three straight cruises

2 six sharp U-turns

3 two sharp 90

corners

4 the width of the riverbank is 20m

5 the dimension of the training track is 130m width

and 140m length

Figure 6. Testing track shown from the Perspective-view,

with the buoys and boats floating on the water to represent

obstacles.

870

4 RESULTS AND DISCUSSIONS

In this section, the results that demonstrate

neuroevolutionary-based ship steering control shall

be presented. The manoeuvring characteristics of the

best agent shall be discussed concerning the training

track, followed by the observation of the obstacle

avoidance capability shown in the testing track.

4.1 Training Track Autonomous Manoeuvre

The autonomous manoeuvre of the ASV within the

training track is shown in Figure 7. It can be observed

that the ASV was able to negotiate turns successfully

using the CNE-based ANN. With respect to the

manoeuvring characteristics, three sharp corner

manoeuvres on the North-East of the training track is

highlighted in Figure 8 and highlighted in green

colour, together with its respective zone 1, 2 and 3.

As the ASV cruising from the bottom right (South-

West) of the training track, it can be understood that

the positive angle values of the rudder angle

represent the right turn, while the negative rudder

angle value represents the left turn. In this

experiment, as the ASV cruise via the assistance of

distance sensor (input), the ANN model evaluated

the sensor input to produce the best throttle

percentage and the rudder angle values, as shown in

Figure 9.

It can be observed in Figure 8, accompanied by

Figure 9, as the ASV enters Zone 1, the rudder tends

to turn the ASV to the right to follow the U-turn

curve to the right. The same pattern can be observed

in zone 3. While manoeuvring against zone 2, the

ASV heads to the left to comply with the left U-turn.

In all three zones, it can be observed that the throttle

and velocity value spikes up and down significantly

each time an aggressive manoeuvre is being executed,

which is a desired characteristic for a CNE-based

ANN model.

Figure 7. Overall autonomous manoeuvre at the training

track with the ASV’s path highlighted in black colour.

Figure 8. Autonomous sharp corner manoeuvre,

highlighted in green colour, with the zones labelled as ‘1’,

‘2’ and ‘3’.

Figure 9. Rudder angle, throttle, and velocity responses

while negotiating with sharp corners as highlighted in

green colour.

4.2 Test Track Autonomous Manoeuvre

In the second experiment, the same ANN model

trained in Section A previously was incorporated to

investigate the capability of the model to generalize

the manoeuvring characteristics in an unfamiliar

surrounding using a test track. The top-view of the

test track (Figure 6) is presented in Figure 10,

depicting the successful autonomous manoeuvre of

the ASV using the ANN model. In this section, an

additional challenge vis-à-vis the training track is

incorporated, where six weather buoys were located

on the U-turns and sharp turns, together with two

boats and one deck situated on the North-East of the

training track.

Depicted in Figure 11, two zones were shown;

zone 1 for the boats and deck obstacle, while zone 2 is

for the weather buoy obstacle. It can be observed that

within both zones, two quick manoeuvres to the right

were observed where the ANN is preventing the ASV

from colliding with the boats and weather buoy by

performing quick rudder turn to the negative angle. It

can also be observed that the throttle and velocity

values significantly dropped each time an aggressive

871

manoeuvre is being executed. It is shown in Figure 12

that the ASV is experiencing steady control of the

rudder and speed in between the two zones depicting

that no unsteady manoeuvre is being performed in a

straight direction cruise.

Figure 10. Overall autonomous manoeuvre at the test track

with the ASV’s path highlighted in black colour.

Figure 11. Autonomous obstacle avoidance manoeuvre,

highlighted in green colour.

Figure 12. Rudder angle, throttle, and velocity responses

while negotiating with an obstacle. Obstacle avoidance is

highlighted in green colour.

4.3 Discussions: Comparative Handling Characteristics

between Human and Autonomous System

The performance of the trained ANN model as

evaluated in Section A and Section B has indicated

that the ASV is capable of performing safe

autonomous cruising, while at the same time

avoiding obstacles. Such findings are further

strengthened with the observation that the ASV is

competent to perform safe navigation in an

unfamiliar setting, as evaluated in Section B. While

such capability is very desirable in the realm of

control, it can be hypothesized that such ASV is

capable of performing better than a human operator.

Taking into account that the ASV is equipped with

nine distance sensors (Figure 4), the data of human-

controlled vessel navigating the training track is

recorded (Figure 13) to compare the efficiency of the

autonomous system between both human and the

ASV.

As depicted in Figure 14, a 50m straight-line

cruising operation was captured for 10 seconds are

compared between a human operator and the ASV. It

can be observed that a human operator who relies on

visual perception are more relaxed during the

straight-line course-keeping operation. Whereas, for

an autonomous system equipped with nine distance

sensors, the ASV actively measures the safe distance,

which in turn beneficial for active obstacle avoidance.

In real life operation, such an autonomous system

might perform abrupt adjustments to maintain its

distance from dangers/obstacles, therefore sacrificing

the comfort of the passengers.

A 90-degree manoeuvre performance comparison

between a human operator and the ASV is shown in

Figure 15. It can be observed that as the vessel made a

90-degree right turn (positive angle), the human

operator carefully adjusted the response of the rudder

accordingly, which resulted in a more relaxed

cornering manoeuvre. As of the ANN model, in order

to compensate with the probability to drift during

cornering manoeuvre, the autonomous system can be

seen as attempting to actively controlling its steering

response to maintain its distance to the vertical bank

of the river.

The comparative assessment discussed above

between a human operator, and the autonomous

system has raised a discussion with regard to the

element of accuracy bias and comfort in vessel

handling. For a human operator, comfort factor in

handling is very important to maintain longer

endurance of work; however, for an autonomous

system, accuracy is more important to reduce the

probability of hitting an obstacle as trained in the

experimental setup. It can be recommended that such

an element of bias and comfort can serve as the

potential works in the future.

872

Figure 13. Straight Line Manoeuvre – Human Operator

view.

Figure 14. Straight Line Manoeuvre - rudder angle

comparison between a Human Operator and Autonomous

System.

Figure 15. 90-Degree Manoeuvre - rudder angle

comparison between a Human Operator and Autonomous

System.

5 CONCLUSIONS

Presented in this work are the state-of-the-art review

and experimental analysis of the use of

neuroevolutionary methods in ship design discipline,

particularly within the scope of autonomous handling

scenarios. Although autonomous vehicles have been

progressing rapidly for the land-based vehicles, the

research for self-driving in restricted waters (riverine

and ports) still possess a significant gap despite its

economic and safety impacts. In this work, an

illustrative example has been presented using a

simplified ship model in a restricted water scenario

(vertical riverbank) which reveals that the

preliminary neuroevolutionary model is not only

good for navigation in restricted water but also

capable for avoiding obstacle within the proximity of

the distance sensors. Using the end-to-end

unsupervised reinforcement learning, the artificial

neural network can predict the best steering angle

and throttle responses while avoiding collision with

other floating objects and riverbanks. Additionally,

the handling performance of the ship has been

compared between a human operator and the

autonomous system (ANN model) The future works

may include the consideration of comfort factor in

ship handling within the ANN training to ensure that

the ASV is not only capable of performing a safe

manoeuvre operation, but also comfortable handling

for human passengers.

ACKNOWLEDGEMENTS

The work presented in this research article is funded by

Universiti Malaysia Terengganu (Research Intensified

Grant Scheme, RIGS, Grant Number: 55192/12) under the

theme of Technology & Engineering (Infrastructure and

Transportation).

REFERENCES

[1] Y. A. Ahmed and K. Hasegawa, “Implementation of

automatic ship berthing using artificial neural network

for free running experiment,” in IFAC Proceedings

Volumes (IFAC-PapersOnline), 2013, doi:

10.3182/20130918-4-JP-3022.00036.

[2] Y. A. Ahmed and K. Hasegawa, “Artificial neural

network based automatic ship berthing combining PD

controlled side thrusters - A combined controller for

final approaching to berth,” in 2014 13th International

Conference on Control Automation Robotics and Vision,

ICARCV 2014, 2014, doi: 10.1109/ICARCV.2014.7064504.

[3] K. Priandana, B. Kusumoputro, and E. T. Rahardjo,

“The design of ISM-band radar antenna for small boat’s

trajectory tracking,” in QiR 2017 - 2017 15th International

Conference on Quality in Research (QiR): International

Symposium on Electrical and Computer Engineering, 2017,

doi: 10.1109/QIR.2017.8168460.

[4] A. Dabrowski, S. Busch, and R. Stelzer, “A Digital

Interface for Imagery and Control of a Navico/Lowrance

Broadband Radar,” in Robotic Sailing, 2011.

[5] T. Wu, Y. Dong, Z. Dong, A. Singa, X. Chen, and Y.

Zhang, “Testing Artiﬁcial Intelligence System Towards

Safety and Robustness: State of the Art,” IAENG Int. J.

Comput. Sci., vol. 47, no. 3, 2020.

[6] N. Uddin, “A Two-Wheeled Robot Trajectory Tracking

Control System Design Based on Poles Domination

Approach,” IAENG Int. J. Comput. Sci., vol. 47, no. 2,

2020.

[7] H. Noshahri, T. J. A. De Vries, and J. Van Amerongen,

“Towards Automatic Steering of Underactuated Ships,”

in IFAC-PapersOnLine, 2019, doi:

10.1016/j.ifacol.2019.12.293.

873

[8] S.-D. Lee, C.-Y. Tzeng, and W.-W. Huang, “Ship

Steering Autopilot Based on ANFIS Framework and

Conditional Tuning Scheme,” Mar. Eng. Front., 2013.

[9] L. Moreira and C. Guedes Soares, “Recursive neural

network model of catamaran manoeuvring,” Trans. R.

Inst. Nav. Archit. Part A Int. J. Marit. Eng., 2012, doi:

10.3940/rina.ijme.2012.a3.232.

[10] W. Naeem, S. C. Henrique, and L. Hu, “A Reactive

COLREGs-Compliant Navigation Strategy for

Autonomous Maritime Navigation,” IFAC-

PapersOnLine, 2016, doi: 10.1016/j.ifacol.2016.10.344.

[11] M. Kurowski, H. Korte, and B. P. Lampe, “AGaPaS -

A new approach for search-and-rescue-operations at

sea,” in IFAC Proceedings Volumes (IFAC-PapersOnline),

2012, doi: 10.3182/20120919-3-IT-2046.00013.

[12] M. Kurowski and B. P. Lampe, “AGaPaS: A new

approach for search-and-rescue-operations at sea,” Proc.

Inst. Mech. Eng. Part M J. Eng. Marit. Environ., 2014, doi:

10.1177/1475090213504392.

[13] Q. Zhang, N. Jiang, Y. Hu, and D. Pan, “Design of

Course-Keeping Controller for a Ship Based on

Backstepping and Neural Networks,” Int. J. e-Navigation

Marit. Econ., 2017, doi: 10.1016/j.enavi.2017.06.004.

[14] Y. Wang, S. Chai, and H. D. Nguyen, “Experimental

and numerical study of autopilot using Extended

Kalman Filter trained neural networks for surface

vessels,” Int. J. Nav. Archit. Ocean Eng., 2020, doi:

10.1016/j.ijnaoe.2019.11.004.

[15] M. C. Tsou and C. K. Hsueh, “The study of ship

collision avoidance route planning by ant colony

algorithm,” J. Mar. Sci. Technol., 2010.

[16] Y. A. Ahmed and K. Hasegawa, “Automatic ship

berthing using artificial neural network based on virtual

window concept in wind condition,” in IFAC

Proceedings Volumes (IFAC-PapersOnline), 2012, doi:

10.3182/20120912-3-BG-2031.00059.

[17] K. Hasegawa, “From Ship manoeuvrability,

controllability, captain’s model, traffic model to accident

and tsunami analysis,” in International Workshop on

Nautical Traffic Models, 2013.

[18] Y. A. Ahmed and K. Hasegawa, “Consistently

Trained Artificial Neural Network for Automatic Ship

Berthing Control,” TransNav, Int. J. Mar. Navig. Saf. Sea

Transp., 2015, doi: 10.12716/1001.09.03.15.

[19] M. Łącki, “Reinforcement Learning in Ship

Handling,” TransNav Int. J. Mar. Navig. Saf. Sea

Transp., 2008.

[20] M. Ła̧cki, “Speciation of population in

neuroevolutionary ship handling,” in Marine Navigation

and Safety of Sea Transportation, 2009.

[21] M. Łącki, “Neuroevolutionary Ship Maneuvering

Prediction System,” in Information, Communication and

Environment, 2015.

[22] V. L. Tran and N. Im, “A study on ship automatic

berthing with assistance of auxiliary devices,” Int. J.

Nav. Archit. Ocean Eng., 2012, doi:

10.3744/JNAOE.2012.4.3.199.

[23] M. Łącki, “Neuroevolutionary approach to colregs

ship maneuvers,” TransNav, 2019, doi:

10.12716/1001.13.04.06.

[24] M. Lacki, “Indirect Encoding in Neuroevolutionary

Ship Handling,” TransNav, Int. J. Mar. Navig. Saf. Sea

Transp., 2018, doi: 10.12716/1001.12.01.07.

[25] M. Łącki, “Ship course-keeping with

neuroevolutionary algorithms,” Zesz. Nauk. Akad.

Morskiej w Szczecinie, 2018, doi: 10.17402/287.

[26] C. Chen, X. Q. Chen, F. Ma, X. J. Zeng, and J. Wang,

“A knowledge-free path planning approach for smart

ships based on reinforcement learning,” Ocean Eng.,

2019, doi: 10.1016/j.oceaneng.2019.106299.

[27] A. Lazarowska, “Research on algorithms for

autonomous navigation of ships,” WMU J. Marit. Aff.,

2019, doi: 10.1007/s13437-019-00172-0.

[28] D. Looije and Y. Koldenhof, “Unmanned ship

simulation with real-time dynamic risk index,” Zesz.

Nauk. Akad. Morskiej w Szczecinie, 2015.

[29] D. A. Oskin, A. A. Dyda, and V. E. Markin, “Neural

network identification of marine ship dynamics,” in

IFAC Proceedings Volumes (IFAC-PapersOnline), 2013, doi:

10.3182/20130918-4-JP-3022.00018.

[30] B. Liu, S. Z. Wang, Z. X. Xie, J. S. Zhao, and M. F. Li,

“Ship recognition and tracking system for intelligent

ship based on deep learning framework,” TransNav,

2019, doi: 10.12716/1001.13.04.01.

[31] Y. Xue, D. Clelland, B. S. Lee, and D. Han,

“Automatic simulation of ship navigation,” Ocean Eng.,

2011, doi: 10.1016/j.oceaneng.2011.10.011.

[32] R. Skulstad, G. Li, H. Zhang, and T. I. Fossen, “A

Neural Network Approach to Control Allocation of

Ships for Dynamic Positioning,” IFAC-PapersOnLine,

2018, doi: 10.1016/j.ifacol.2018.09.481.

[33] L. P. Perera, “Autonomous ship navigation under

deep learning and the challenges in colregs,” in

Proceedings of the International Conference on Offshore

Mechanics and Arctic Engineering - OMAE, 2018, doi:

10.1115/OMAE2018-77672.

[34] T. Statheros, G. Howells, and K. McDonald-Maier,

“Autonomous ship collision avoidance navigation

concepts, technologies and techniques,” J. Navig., 2008,

doi: 10.1017/S037346330700447X.

[35] Y. Cheng and W. Zhang, “Concise deep

reinforcement learning obstacle avoidance for

underactuated unmanned marine vessels,”

Neurocomputing, 2018, doi:

10.1016/j.neucom.2017.06.066.

[36] A. I. Kozynchenko and S. A. Kozynchenko,

“Applying the dynamic predictive guidance to ship

collision avoidance: Crossing case study simulation,”

Ocean Eng., 2018, doi: 10.1016/j.oceaneng.2018.07.012.

[37] X. D. Cheng, Z. Y. Liu, and X. T. Zhang, “Trajectory

optimization for ship collision avoidance system using

genetic algorithm,” in OCEANS 2006 - Asia Pacific, 2006,

doi: 10.1109/OCEANSAP.2006.4393976.

[38] J. D. Schaffer, D. Whitley, and L. J. Eshelman,

“Combinations of genetic algorithms and neural

networks: A survey of the state of the art,” in COGANN

1992 - International Workshop on Combinations of Genetic

Algorithms and Neural Networks, 1992, doi:

10.1109/COGANN.1992.273950.

[39] X. Yao, “Evolving artificial neural networks,” Proc.

IEEE, 1999, doi: 10.1109/5.784219.

[40] D. Floreano, P. Dürr, and C. Mattiussi,

“Neuroevolution: From architectures to learning,”

Evolutionary Intelligence. 2008, doi: 10.1007/s12065-007-

0002-4.

[41] L. Harris, “Unmanned Marine Systems - ASVs, USVs

& Autonomous Boat Control System,” 2020. [Online].

Available:

https://www.unmannedsystemstechnology.com/compa

ny/autonomous-surface-vehicles-ltd/.

[42] A. Rescec, “Dynamic Water Physics 2,” Unity Assets

Store, 2020. [Online]. Available:

https://assetstore.unity.com/packages/tools/physics/dyn

amic-water-physics-2-147990.