463
1 INTRODUCTION
The study of autopilot aircraft and vehicles has
entered a period of vigorous development in the
aviationandlandtransportationindustries.Artificial
Intelligence (AI) has begun to play a major role in
theseindustries(Prashanthetal.,2013).There‐search
trend has been moving toward the conception of
robotic vehicles ca
pable of making spontaneous and
effective decisions in demanding or uncertain
situations.Consideringthedepressionof thecurrent
shipping industry, optimizing and adjusting the
industrialstructureiscrucial,incaseswheretheover‐
head of seafarer’s expenditure and crew company
managementaccountforalargepartofthepa
yment.
Unlike unmanned aerial vehicles (UAVs) and un‐
mannedvehicles,unmannedshipswarmsmayhavea
higher superiority for certain ship types such as
regular container ships for certain voyages (Gudelj
andKrcum,2012).
The real demanding circumstance of vehicle to
everything (V2X) scenarios demands the need of
handling fully a
utomatic driving, remote
managementpersonnelinthestateofsurveillance,as
well as any latent hazard requiring the car (where
thereisthepossibilityofthedriverbeinginvolved)to
take their own effective action atthecriticalinstant.
Undoubtedly,thecurrent technicalandlegal aspects
are still facing enormous changes. Many difficult
ies
must be overcome for including fully automatic
navigationinunmannedshipswarms,remotemulti‐
agent monitoring, and efficient transportation of
goods.Ontheotherhand,inspecifictransportareas,
underremotemonitoringandperception,multi‐agent
transportation is technically advanced (Ren, Wei,
2007). Therefore, Autonomous Ship Swarms
Transportation is a very promising field of research
forunmannedships(Sardaetal.,2016).
Conversion Timing of Seafarer’s Decision-making for
Unmanned Ship Navigation
R.L.Zhang&M.Furusho
KobeUniversity,Kobe,Japan
ABSTRACT:The aimof thisstudyistoconstructanunmannedship swarmsmonitoring model toimprove
autonomousdecision‐makingefficiencyandsafetyperformanceofunmannedshipnavigation.Aframeworkis
proposed to determine the relationship between on‐board decision‐making and shore side monitoring, the
processofshipdatadetection,tra
cking,analysisandloss,andtheapplicationofdecision‐makingalgorithm,to
discussthedifferentriskresponsesofspecificunmannedshiptypesundervariouslatenthazardenvironments,
particularly in terms of precise conversion timing in switching over to remote control and full manual
monitoring,toensuresafenavigationwhenthecapabilit
yofautomaticriskresponseinadequate.Thisframe‐
workmakesiteasiertotraindataandtheadjustmentformachinelearningbasedonBayesianriskprediction.It
canbeconcludedthattheautomationlevelcanbeincreasedandtheworkloadofshore‐basedseafarerscanbe
reducedeasily.
http://www.transnav.eu
the International Journal
on Marine Navigation
and Safety of Sea Transportation
Volume 11
Number 3
September 2017
DOI:10.12716/1001.11.03.11
464
The study object of this paper is unmanned
container ship swarms. Container ships, especially
regular container ships, have characteristics such as
standardized operations schedule, high level of
automation in the process of loading and unloading
of containers, and easy remote monitoring of goods
condition.Uponremovalofthebridge,an
unmanned
container ship can expand its packing capacity,
improveenterprisebusinessincome,andreducesea‐
farercosts,inordertofacilitatetheshippingoflarge‐
scale company cargo dispatch, resulting in
significantly improved efficiency (Dubrovsky, 2010).
YangandWangdesignedafullyautomatedloading
andunloadingplatformspecificallyforan
unmanned
container ship. When in a specific terminal wharf,
both sides of the ship can load and unload
simultaneously, thus improving ship loading and
unloading efficiency, increasing wharf benefits and
reducing thestagnationtime of ships in the anchor‐
age,Italsoallowstheoptimizationoftheporttraffic
flow (Yang
and Wang, 2011). Nonetheless, market
factors determine the research prospects of un‐
mannedcontainerships.Thecontainershiptrafficisa
hugebusiness,even considering that a tiny accident
maycauseimmeasurabledisasterandloss.Therefore,
it is necessary to construct the Autonomous Ship
Swarms monitoring model before proceeding with
the technical details. The ship itself has a limited
abilitytoresistrisks,especiallyunderdemandingor
uncertain environment. Therefore, the decision‐
making cycle by itself may not have adequate
capacity to avoid hazards. However, frequently
invokingdecision‐makingresourcesfromtheex‐pert
systemofshorestationmaycauseremote
monitoring
capacity insufficient, and shore‐based monitoring
seafarers may lead to more human errors or other
latent failures. Therefore, it is necessary to maintain
the resource balance between the decision‐making
cycleandtheremotemonitoringofunmannedships,
thereby optimizing the structure of the decision
model.
Autonomous risk prediction
and autonomous
decision‐makingaregenuineandsignificantpartsof
the unmanned ship swarms monitoring model
(Kirsch,2016).Theyareeffectivewaystodiminishthe
workload of shore‐based seafarers, reduce human
error and improve commercial interests, by
maximizing the risk prediction ability and risk
avoidancedecision‐makingofunmannedships.
Even if the relationship between decision and
maneuveringautomationhasalreadybeenproposed,
ifhighautomationisselectedfortheactionpart,then
designers should resist the temptation of high
automation levels of decision‐making (Parasuraman,
2000). Even the real “noisiness” world always have
somekindofunexpectedsituationemerged
(Endsley
andKiris,1995).
Thepurposeofthispaperistobuildanautomatic
decisioncyclemodeltoimprovethedecision‐making
efficiencyandsafetyperformanceofunmannedships.
According to the previous assumptions, with
unmanned container ship response to various
hazards, calculated training data and parameter
adjustment, the performance of
automatic decision‐
making can be improved. The situation of re‐mote
humaninterventionwillthenbecomelesssignificant.
When a new uncertainty situation appears, original
datatrackingandanalysisandhandlingofdataloss
are required. In this case, the insufficient ability of
automatically respond to facing hazards will cause
themodeltoshifttomanualmonitoringandremote
controltoensuresafenavigation.
2 AUTONOMOUSDECISION‐MAKING
Because the term “automation” has been used in
many different ways, the British Dictionary defines
automationas:
1 The use of methods for controlling industrial
processesautomatically,esp.byelectronicallycon‐
trolledsystems,
oftenreducingmanpower.
2 Theextenttowhichaprocessissocontrolled.
Autonomousdecision‐makingshouldhaveanew
implication, which the authors defined as
automaticity between the different decision‐making
cycles. Although researchers have already described
the conflict between high automation levels and the
automation of decision‐making
(Parasuraman,2000),
their opinions are focused mainly on high levels of
automation, and do not considered the decision‐
making aspect mistake. Moreover,ʺerror‐trappingʺ
showsthatlowerlinkcommunicationautomationcan
allow more action errors. When high automation is
selected for maneuvering, researchers should resist
thetemptationforhighautomationlevels
ofdecision‐
making. Therefore, high levels should be executed
only for low‐risk situation awareness; for all other
situations,thelevelofautomationdecisionshouldnot
exceed the level of the computer suggesting a
preferredalternativetocontroller.
Ontheotherhand,withtheimprovementofma‐
chine learning algorithms,
more unlabeled data can
beused,andmorereliableautomaticdecision‐making
does not require human intervention, considering
mainly the concept ofʺhuman‐centered automationʺ
re‐understanding (Metzger, 2005). As there are two
distinct centers (ship swarms and shore expert
station) for the autonomous decision‐making of
unmannedcontainershipgroups,
andmoredecisions
can be made by the cycle itself before theʺhuman‐
centeredʺisinvolved(Zhang,2016),theenvironment
forautomaticdecision‐makinghasbe‐comeeasier.
2.1 Thelevelsofautonomousdecision‐making
The basis for automatic decision‐making must be
basedonagoodautomaticityclassification(see
Table
1). From low to high, automated carry forward also
showsthedevelopmentofshipautomationdecision‐
making process, which proves that the direction
towardsautomationisinevitable.Itcanbeseenfrom
the table that the lowest level, level 1, is completely
comprisedof manual operations;the secondlevel of
the decision‐making system can provide all the
decisionoptions,butatthislevel,thesystemdoesn’t
make its own decisions (data learning and training
process); The third level can optimize the selection
andreducethepossibledecisionsoutput(perception
process); the fourth level can provide an optimal
decision‐making
program,butstillcannottakeaction
(optimization process); The fifth level of decision‐
465
making is a conversion point, in which the system
usuallyneeds to be agreed with the seafarertotake
action, but this level of decision‐making is able to
provideadecision‐makingoperationbyitself,soany
level of automaticity higher than this can be
considered as completely autonomous
decision‐
making. Note that the decision of the fifth level
accordingtothistableisimportantforthispaper,asit
is also a turning point for the ships own decision‐
making cycle and shore‐based decision‐making. For
decisionstakenatalevelhigherthanthis,suchasthe
sixthlevel,thesystemcantakeactiononitsown,and
only part of the uncertain data is sent to the shore‐
basedseafarerforrecord;forseventhlevelandabove
decision‐making,thesystemcanrecognizethetiming
of conversion by itself, and take action, without
humaninvolvement.
Table1. Automaticity selection by on‐board decision cycle
andshore‐basedmonitoringcenter
There is a problem that must be noted: when an
unmanned container ship encounters a totally new
situation,automaticitycantransitfromahighlevelto
a lower level; thus, automaticity can change
dynamically according to the ship navigation. The
following discussion of the different target ships for
theobject
ofthedecision‐makingwillusethistable1
asstandardtoillustrate.
Figure1.Differenttargetdetection,tracking,loss,andstateanalysisbyMarkovDecisionProcesses
Table2.AnexampleofMultipleTargetTrackinginaMarkovDecisionProcess
466
2.2 MarkovDecisionProcessfortargetperception
Inpriorresearch(Zhang,2016),theauthorsusedthe
Robot Operating System as a tool, extended the
MarkovDecision‐making (MDM) andsupported the
decision‐making methodologies based on Markov
DecisionProcesses(MDPs)TheaimoftheMDPsisto
provideanaction
set of decision‐making for theon‐
boardcycle.Whenanactionisexecuted,marineareas
whose state changes ac‐cording to a known
probability distribution, are converted to another
state,andtheprobabilitydistributionisrelatedtothe
actions performed. As shown in Figure 1, three
different target ships
can be considered as three
continuously changing MDPs by detecting, tracking,
analyzingandlosing.Accordingtothedataobtained
from the sensors of the ship (unmanned container
ship), in order to observe the navigational status of
thetargetships,thetable2wasobtained.MDP3isa
large‐scaleship,and,
assuch,thecourseofnavigation
isalways withinthe scopeof theown shipscontrol,
andthetargetship can be completelyin accordance
with the navigation rules and decision‐making
procedures for collision avoidance or other actions;
MDP1indicatesthecourseofthesailboat1.Although
thetracking
processisblockedbyalargeobstruction,
theshipitselfcanbebasedontheheadingcourseand
speedbeforedisappearance,inordertospeculatethe
stateofsailboat1,andforthereappeartobeverified;
MDP2illustratetheblockageofthesailboatbyalarge
obstruction,forexample,due
tofishingworkorother
unpredictable factors, which led to the tar‐get
disappearance from the control scope (loss state). In
thissituation,whereinthesystemcannotbebasedon
past data training or experience to get more in‐
formation, a shift to remote control to acquire the
necessaryexpert
supportisthebestchoice.
2.3 Targetperceptionanddecision‐making
Accordingtothe previous scenario, the combination
of the Table 1 and Figure 1 originates Figure 2. Be‐
cause of the size, shape and speed different of the
target ships, divergent detectors may also obtain
different data reliability. The level
of predictability
varies,astheautomaticitylevelisconstantlyunstable.
According to the Table 1, it can be concluded that
there are nine different automaticity levels, in
differentstates,inwhichtargetshipshavetheirown
decisionlevel. It canalso be concludedthatMarkov
Decision Processes can produce decision‐
making
cycles of different automaticity levels decision‐
makingundertheclosedshipitself,thefifthlevelcan
beseenasatimingconversion standard,lower than
fifth level leads to the conversion to shore‐based
remotecontrol,higherthanfifthlevelbelongstothe
on‐boarddecisioncycle.
Figure2.Levelsofautomaticityassessmentforthedecision
ofwhich MDPsshouldconvert toshore‐based monitoring
orkeepdecisionbyon‐boarddecisioncycleitself.
3 ADVERSARIALDECISIONCYCLES
The mainstream technology of auto‐driving
prediction and decision‐making is becoming clearer
nowadays that machine learning is based on both
deeplearning and reinforcementlearning. However,
ma‐chine learning needs large data to achieve high
performance and high reliability. It means that
developers need to install
an automatic driving
equipment in a large number of automatic ships, so
thattheshipsinthecurrentoperationcanproducethe
required amount of data to enrich the decision‐
makingcycleinorderfortheamountofdatatolead
toanimprovementinthedecision‐makingefficiency.
3.1
Unmannedshipdecisioncycles
Unmanned ship maneuvering cycle is generally di‐
vided into four main sections: sensory detection,
tracking & perception, decision‐making and optimal
action. The model identifies the object from the
environment,performstrackingandriskrecognition,
makes decisions and takes effective action, and the
effectoftheaction
isfedbackintotheenvironmentto
confirm the new position relationship. Consider a
decision cycle as shown in Figure 3. The external
environment also constitutes a part of the cycle,
includingtheshipssurroundingmarineenvironment
and the ships own hardware environment. The
decision‐making stage is the link
to the shore‐based
re‐motecontrolandshipswarms,sohavingdecision‐
makingastheendofthecycle(humandecisioncenter
automation), it is possible to build the de‐pendency
relationship of the big decision‐making cycle. It can
beseeninFigure4.
467
Figure3.Anexampleofunmannedshipmaneuveringcycle
basedonthefour‐stageofhumaninformationprocessing
Figure4.Aflowchartofconversiontimingoptimizationof
decision‐makingbetweenon‐boarddecisioncycleandshore
side monitoring. An adversarial decision cycles approach
hasbeenpresented.
3.2 Autonomousdecisionprocesstraining
There are two main decision‐making cycles of the
unmanned ship: the first one is on‐board decision
cycle, countless monomer decision‐making cycle
constitutes the ship swarms; the second is a large
group of ships and shore‐based cycle. The relation‐
shipbetween thetwo
main sectionsis similarto the
GenerativeAdversarialNets(Goodfellowetal.,2014)
in deep unsupervised learning. Ship swarms and
shore station center belong to two main bodies of
confrontation.Atfirst,thedecision‐makingcycleofa
single ship is very small, and most of the decision‐
making cycle
needs shore support to complete.
However, as unmanned training data increases, due
to the fact that the ships own decision‐making
capacitycan be continuouslyenhanced, high quality
decisions can finally be achieved, being as good as
shore station seafarers would make. In this process,
the workload of shore‐based seafarers
is gradually
reduced, and the ship swarms automatic decision‐
makingcapacityisgraduallyincreased.
4 APPLICATIONEXAMPLE
This autonomous ship swarms transportation model
of human‐automation interaction can be applied to
specific systems in conjunction with a consideration
ofevaluativecriteria,whichwehavediscussedinthis
paper–human
workloadandcostexpenditureofthe
shipcompany,automaticityreliabilityandautomatic
level, the adversarial relationship between the on‐
boarddecisioncycleandshore‐basedmonitoring.To
furtherdemonstratetheapplicationofthismodel,the
authorsbrieflydrawtheoutlineofimageforitsusein
the design of near future
unmanned ship
transportation system, based on the previously
presented study and other researcher’s study
conclusions.
Jansson proposed a vehicle collision avoidance
frameworkbased onstatistical decisionand stochas‐
tic numerical integration (Jansson, 2008). The main
purpose of decision‐making framework is to deal
with the uncertainty of state estimation. Application
of this model suggests the following
recommendations for future swarm ship
transportation automation. Jansson presented a
probabilisticframeworkfordesigningand analyzing
a collision avoidance algorithm, calculate risk for
faultyinterventionandtheconsequencesofdifferent
maneuverings. Jansson’s work was based on Monte
Carlo techniques, where sampling‐resampling
methods are used
to convert sensor readings with
stochasticerrorstoaBayesianrisk.Theauthorsalso
proposed the construction of a reliable decision‐
support system for risk and accident predictions
based on past experience and objective accident‐
probabilitystatisticsusingBayesianNetwork(Zhang,
2016). We dis‐cussed the prospects of an intelligent
decision‐supportsystemtoensurereliablenavigation
safety.Itincludedacasestudywhichcanprovidea
procedure for complete automatic decision‐making
basedonexperienceprobability.
For a simulation study (Łącki, 2015), the author
assesses the feasibility of this idea.Łącki narrowed
thestudytoasmaller scale.Forecasting
the location
of the target ships and assisting the own ships on
makingthemaneuveringdecisioncanbeusedaspart
of this framework.Łącki using a neuroevolutionary
method, presented a concept of the advanced ship
action prediction system for the simulation of a
learningprocessofanautonomouscontrol
unit.Data
isprovidedtothissystembytheshipsensor,tomake
possible the completion of forecasting data training
tasks as individuals in the population of artificial
neuralnetworks.Theenvironmentalsensingandthe
evolutionary algorithms learn to execute each given
task efficiently. This mathematical model of
maneuvering VLCC tank
ships with the single‐
propeller and single‐rudder was applied to test the
prediction performance of the system (Łącki, 2008).
Artificialneuralnetworksbasedonmodifiedmethods
increase the complexity and performance of the
consideredmodelofshipmaneuveringinrestriction
waters.
468
5 CONCLUSIONS
Thescenariopresentedinthispaper isanautomatic
decision‐making solution in extreme cases where
ships are rarely encountered in such complex
situationswhennavigatinginopenwaters.However,
inordertoimprovethedecision‐makingefficiencyof
theunmannedshipunderremotecontrol,itisneeded
to take into account this kind of situation in the
beginning of the system design. The author first
proposed automaticity stratification, with the data
accumulationoftheon‐boarddecisioncycle,floating
up and down real‐time assessment data training of
the situation encountered. Between the on‐board
decisionand
shoresidemonitoringoftheadversarial
decision cycle model it is demonstrated that the
conversion timing continuously changes and always
tends to on‐board if the decision cycle can make
decisions similar to those of the seafarerʹs (expert
system). This structural model is based on the
Bayesian network machine learning,
with the
advantages of easier data train and parameter
adjustment, easier improvement of the automation
level.Itcanreducetheworkloadofshore‐basedcrew
significantly.
REFERENCES
Endsley, M. R. & Kiris, E. O. 1995. The out‐of‐the‐loop
performance problem and level of control in
automation. Human Factors: The Journal of the Human
FactorsandErgonomicsSociety,37(2):381‐394.
Parasuraman, R., Sheridan T. B. & Wickens C. D. 2000. A
model for types and
levels of human interaction with
automation. IEEE Transactions on systems, man, and
cybernetics‐PartA:SystemsandHumans,30(3):286‐297.
Prashanth,C.R.,SagarT.,BhatN.2013.Obstacledetection
&eliminationofshadowsforanimageprocessingbased
automated vehicle. Advances in Computing,
Communications and Informatics (ICACCI) International
Conferenceon.IEEE,2013:367‐372.
Sarda, E. I., Qu, H., Bertaska, I. R. 2016. Station‐keeping
control of an unmanned surface vehicle exposed to
current and wind disturbances. Ocean Engineering, 127:
305‐324.
HocrafferA.,&NamC.S.2017.Ameta‐analysisofhuman‐
system interfaces in unmanned
aerial vehicle (UAV)
swarmmanagement.AppliedErgonomics,58:66‐80.
Ren W, Beard R W, Atkins E M. 2007. Information
consensus in multivehicle cooperative control. IEEE
ControlSystemsMagazine,129(2):571‐583.
GudeljA.,&KrčumM.2012.TheContainerTransportation
Problem: Model and Solution Methods. First
International Conference
on Traffic and Transport
Engineering.
Yang, C., & Wang, N. 2011. Decision‐making Method for
Berth Allocation Disruption Management in Container
Terminal.OperationsResearchandManagementScience,4:
14.
Dubrovsky, V. A. 2010. Multi‐hulls: new options and
scientific developments. Ships and Offshore Structures,
5(1):81‐92.
Kirsch, A. 2016.
Human‐aware Navigation in Domestic
EnvironmentsUsingHeuristicDecision‐Making.
Metzger,U.&ParasuramanR.2005.Automationinfuture
airtrafficmanagement:Effectsofdecisionaidreliability
oncontrollerperformanceandmentalworkload.Human
Factors: The Journal of the Human Factors and Ergonomics
Society,47(1):35‐49.
Zhang, R. &
Furusho, M. 2016. Constructing a Decision‐
Support System for Safe Ship‐Navigation Using a
Bayesian Network. International Conference on Human‐
Computer Interaction. Springer International Publishing,
616‐628.
Goodfellow, I., Pouget‐Abadie, J., Mirza, M. 2014.
Generative adversarial nets. Advances in Neural
InformationProcessingSystems2672‐2680.
