International Journal

on Marine Navigation

and Safety of Sea Transportation

Volume 3

Number 3

September 2009

251

1 INTRODUCTION

An autopilot is defined as a mechanical, electrical or

hydraulic system used to guide a vehicle without as-

sistance from a human being. A ship uses an autopi-

lot for steering during her voyages except when she

navigates in confined waters or maneuvering at port

(COLREGS 72) [8]. A ship’s voyage may last sev-

eral days and a large proportion of it takes place in

the open sea where the autopilot is used almost ex-

clusively. Even though the ship’s bridge, where the

autopilot is located, is always supervised by the of-

ficer on watch (STCW 95) [22], it is necessary to

ensure that the autopilot is a safe and reliable tool in

his / her hands.

Keeping a ship on course is not an easy task since

ships are exposed to severe weather conditions and

are operating in extreme situations. Wind, sea, cur-

rent, etc, are some of the factors affecting a ship’s

deviation from the desired course. An autopilot’s

task is to keep the ship on track, not losing control in

any case and simultaneously minimizing the devia-

tions regardless of cause. To do that, an autopilot

must have the proper configuration so that it would

be able to perform its best according to the situation

at hand. This ideal situation is not easy to achieve

because the weather combinations of wind, sea, cur-

rent, etc, are practically infinite and the same stands

for the ship’s loading conditions which also affect

the final outcome. Moreover, an autopilot device is

designed to work on almost any type of ship, thus its

performance wouldn’t be the same in different hulls.

The actual performance of the device is measured

using parameters like loss of steering, vertical and

angular deviation, extra distance, etc, because they

are closely connected to dangerous situations at sea

or significant losses of fuel and time. Loss of steer-

ing, combined with a generator failure can cause a

serious accident i.e. capsize (Leontopoulos 79) [34],

while vertical deviation from course (Cross Track

Error) leads to unwanted approaches to navigational

dangers. Moreover, extreme angular deviations from

compass settings affect the ability to command, es-

pecially in bad weather (Bowditch 2002) [6]. Final-

ly, an incorrect adjustment increases the total voyage

distance, the fuel consumption, the time delay and

the corresponding costs (Dutton 1958) [11].

Given the above it is very difficult to develop a

method that takes into account all the affecting fac-

tors and being able to maximize the performance on

every ship, under any weather and loading condition.

An ideal situation would be the development of a

customized device able to “understand” its environ-

ment (weather, loading condition and ship’s particu-

lars) and properly adjust itself, responding to any

changes. Even though such a device is not developed

yet, we claim that a pattern able to operate in a simi-

lar way is feasible, provided that a conventional de-

vice will be equipped with some additional features

mentioned below.

Marine Navigation Using Expert System

N. Nikitakos & G. Fikaris

University of the Aegean, Chios, Greece

ABSTRACT: A ship’s autopilot adjustment is a matter of utmost importance since it affects its safety, com-

mand as well as fuel and time efficiency. A number of methods have been developed in order to cope with

this issue usually based on models that simulate the weather conditions and adjust the device accordingly.

Some of them have a considerable degree of success but none dealt with the problem completely. The main

obstacles are the difficulty of simulating the infinite weather and loading conditions and to properly represent

them with mathematical equations or rules. This paper describes a method of selecting the best out of a pre-

existing set of configurations, taking into account any weather situation, loading condition and type of ship.

Moreover, the selected configuration can improve itself during the entire life cycle of the vessel, since it fine

tunes its properties for better results. This approach uses Case Based Reasoning as its core technology and is a

part of a hybrid system that analyses and solves prefixed problems of maritime interest.

252

This pattern is incorporated as an application

within an AI system named POLARIS (POlicy

Leading ARtificial Intelligence System) (Nikitakos

& Fikaris, 2007) [38] able to analyze problems of

maritime interest and propose courses of action for

them. This approach has certain advantages com-

pared to others because it doesn’t deal directly with

the identification and estimation of the parameter

values that constitute a configuration but instead it

presupposes an unlimited number of them already

installed on the device, with known properties that

can be modified according to the user’s wishes.

There is no limit to the number or nature of the pa-

rameters.

The system’s core methodology is CBR (Case

Based Reasoning) which solves current situations –

problems with the assistance of similar cases that

were dealt successfully in the past. These cases are

stored in a case library and retrieved by the system

using the proper indexes. The retrieved cases are

ranked according to the criteria and the system pro-

poses the best solution to solve the current problem.

If necessary, a solution may be adapted to fit a new

situation. When the best solution is proposed a pro-

cedure of fine tuning may begin and last till the solu-

tion meets the pre specified criteria.

The application described in this paper includes

the development of a series of diagnostics performed

by the autopilot device in different loading and

weather situations in order to measure the corre-

sponding performances and create a case base out of

them. Thus, when the ship finds herself in a similar

situation, the device will track the case’s characteris-

tics, select the case with the configuration that per-

formed best and steer the ship with it until it detects

another set of conditions. It is important to mention

that the user may choose to measure the perfor-

mance of a given situation again so that the database

would be constantly updated with improved scores.

2 DSS AND CASE BASED REASONING

The literature defines Decision Support Systems

(Raiffa 76) [26] as “interactive computer based sys-

tems that help decision-makers use data and models

to solve ill-structured, unstructured or semi-

structured problems (Goel 92) [15].” The most

popular definitions belong to Gorry & Scott-Morton

(1971) [16], Keen and Scott-Morton (1978) [25] and

Bonczek, Holsapple & Whinston (1981) [5]. DSS

were categorized in seven major categories which

are file drawer systems, data analysis systems, anal-

ysis information systems, accounting and financial

models, representational models, optimization sys-

tems and suggestion systems (Alter 1980) [1]. A

type based categorization defines data driven, model

driven (Knowles 89) [28] and knowledge driven sys-

tems (Dhar & Stein 1997 [9], Holsapple & Whinston

1996) [20]).

Knowledge driven DSS -sometimes called Expert

Systems- incorporate knowledge about a particular

domain, understanding of problem solving and ex-

pertise at solving those problems (Redmond 1992)

[42]. They are also related to data mining techniques

and usually evolve to hybrid systems (Simpson,

1985) [44]. Major components of a DSS are a) the

shell b) the case library c) the knowledge base or the

model and d) the system’s architecture and network

(Sprague and Carlson 1982) [46]. These systems an-

alyze data using symbolic logic, have an explicit

knowledge base and have the ability to explain con-

clusions in an understandable way. Web based DSS

are referring to a computerized system that delivers

decision support information or decision support

tools to a manager or business analyst using a “thin

client” Web browser (Power 2000) [41].

Reasoning is a procedure that draws conclusions

by chaining together generalized rules, usually start-

ing from scratch. However in Case Based Reasoning

new solutions are generated not by chaining, but by

retrieving the most relevant cases from memory and

adapting them to fit new situations (Leake 1996,

2003) [33][32]. A case is a contextualized piece of

knowledge representing an experience that teaches a

lesson fundamental to achieving the goals of the rea-

soner. A case may have different shapes or sizes,

various time horizons and can associate solutions

with problems, outcomes with situations or both. A

case’s main task is to provide a solution to a prob-

lem but it can also provide the necessary context to

assess or understand a situation (Kolodner 93,

Schank 1994) [29] [45]. A case is comprised from

indexes which should be predictive, goal oriented,

abstract and easily recognizable (Birnbaum & Col-

lins 89 [4], Hammond 89 [18]). These indexes must

describe the problem (goals, constraints and situa-

tion), the solution and the outcome. The case base

indexing is organized according to the problem’s re-

quirements and can be checklist based, difference

based (Kolodner 93) [29], similarity and explanation

based (Hammond 87, 89 [17]), etc. The problem /

situation indexes are mainly used for the retrieval,

qualification and ranking of cases while the solution

indexes present the way of action to the user. The

outcome indexes are a part of the evaluation proce-

dure.

The main advantages of CBR are its simplicity,

its capability of incorporating uncertainty and its

plausibility (Kolodner 1993) [29]. Two major clas-

ses of CBR systems have been developing since the

method’s introduction. These are the interpretive and

problem solving CBR systems (Rissland, Kolodner

& Waltz, 1989) [43]. The former use prior cases as

reference points for classifying new situations,

253

whilst the latter use prior cases to suggest solutions

that apply to new circumstances. Another major ad-

vantage of CBR is that because it uses specific epi-

sodes (cases) for reasoning there is no need to de-

velop many rules and thus makes the knowledge

acquisition process –which is vital to AI systems-

very “cheap”. As pointed by “Mark et al, (1996)”

[35] there are some domains that are very suitable

for CBR, while others are not, especially if cases are

unavailable or in hard to use format. The functions

performed by a typical CBR system are recall and

interpretation of past experiences (cases), adaptation

of those cases to fit the new situation, evaluation of

proposed solutions and repair of the “defective” ones

(Kolodner, 1993) [29].

POLARIS is an AI system containing elements of

a DSS since it interacts with the user and helps him

find the best out of a series of alternatives as well as

expert knowledge relevant to the problem’s domain.

The system uses data from old cases to solve new

problems but it also incorporates expert knowledge

from the problem domain. Thus, its type is a mixture

of a data and knowledge driven system strongly de-

pendent on the nature of the application. All these

are significantly affected by the complexity of the

problem and the domain knowledge available. The

system’s architecture follows the CBR procedures

and comprises of the following modules:

− User interface: interacts with the user

− Case Library: contains the old cases

− Knowledge base: contains the expert

knowledge in the form of rules

− Case Retriever: retrieves and ranks the useful

cases

− Solution presentation facility: presents the solu-

tion to the user

− Solution evaluator: evaluates the solution after

the implementation

− Solution adaptor: adapts the solution to fit the

current situation

− Case storage facility: stores new cases to the li-

brary (Moorman & Ram, 1992) [37]

3 ADJUSTMENT METHODS REVIEW

A quick review of the methods used in order to

properly adjust a ship’s autopilot shows that almost

every single AI technology was used by a number of

researchers. Fuzzy logic (Polkinghorne M.N; Burns

R.S 1994 [39], Roberts G.N, Roberts and Sutton

2006 [40]), Neuron Networks (Unar and Murray-

Smith 1999 [48], Jia, X.J Yang and X.R Zhao, 2006

[49]), Optimization techniques (Holzhuter, 1997)

[21], Linear programming (Goheen K.R, Jeffreys

E.R, 1990 [31]), Model Based Reasoning (Honderd

and Winkelman, 1972 [14], Van Amerongen and

Udink ten Cate, 1975 [23], Van Amerongen and Van

Nauta Lemke, 1986 [24]), Self tuning regulators (KJ

Astrom et al, 1977 [27]), Stochastic models (Ohtsu

et al, 1979 [30], Herther et al, 1971 [19]), etc, repre-

sent only a fragment of the work that has been done

in the field.

Most Autopilots are adjusted using the PID con-

troller which calculates a performance variable with

known values and applies the necessary corrective

actions based on the difference between the calculat-

ed and expected value. The controller includes three

parts: The first one responds to the error, the second

applies a correction for the sum of all the errors and

the third responds to the error variation percentage.

PID controllers however cannot perform in non line-

ar systems and their accuracy is very low.

Another interesting work is the one of Unar and

Murray-Smith who developed an artificial neural

network which controls and coordinates a series of

conventional controllers. Each controller is manu-

factured for a specific operational situation of the

vessel. Still, the level of detail is low and the situa-

tion coverage very poor. Moreover, the system’s

cost and maintenance is relatively high. This ap-

proach has some similarities with CBR since each

controller represents a situation, but it is obvious that

the number of controllers is finite and cannot cover

the infinite weather and loading situations.

4 THE AUTOPILOT APPLICATION THEORY

The Autopilot application presupposes a finite num-

ber of configurations available on the device and a

number of known parameters which are adjustable

and affect the configuration significantly. The sys-

tem creates a case library performing a series of tri-

als, assessing each configuration’s performance for a

given situation. The situations and the corresponding

performance values are stored in memory and ideal-

ly some time during the ship’s life cycle there will

be a case for almost every combination of weather

and loading condition.

When the ship’s devices detect a specific weather

situation, and given that the loading data as well as

the ship’s particulars are already stored in memory,

the system retrieves the cases with the best perfor-

mance values from the base. The qualified cases are

ranked and the corresponding configuration is pre-

sented to the user. After the implementation of the

selected configuration the system records the actual

performance and compares it with the expected one.

If the actual performance is not satisfactory the sys-

tem either switches to the second best configuration

or it enables a fine tuning procedure where it per-

forms a sensitivity analysis of every parameter in the

selected configuration aiming to achieve a better

performance. If this is accomplished it stores the

254

new set of parameters and the corresponding per-

formance indicators, thus creating a new configura-

tion for the device.

The Autopilot’s case contains six categories of

indexes which represent the performance criteria

(goals), the weather conditions (situation), the load-

ing condition of the ship (situation), the ship’s par-

ticulars (user characteristics), the configuration used

(solution) and the applied performance criteria (out-

come).

4.1 Performance Indexes

The performance indexes (criteria) mainly cover the

dimensions of danger, ease of command, cost and

time. Three variables have been used. The first one

is the difference of the distance measured by the

ship’s track to bottom from the rhumb line distance

between waypoints (d). The second is the maximum

(dmmax) and mean vertical deviation to course

(XTE) in miles (dm) while the third is the maximum

(ddmax) and mean deviation of the ship’s bow to the

true course in degrees (dd). Loss of steering LS >0,

dmmax>threshold1 and ddmax>threshold2 were set

as hard constraints. The system is able to

automatically calculate all these quantities either

separately with the appropriate sensors either using

the usual bridge electronic equipment (GPS, ECDIS

& ARPA combined) provided that they are

connected to the improved Autopilot device.

Figure 1: True distance of ship’s track to bottom

The first criterion d was selected because it repre-

sents the extra distance traveled by the ship in a giv-

en part of the journey, so it can be translated to extra

fuel and time that is monetary cost. The true distance

(Figure 1)

is calculated using the formula

∫













λ d

Axis x refers to geographical longitude and axis Y to latitude

The rhumb line distance is k = (ΔΦ)’ secZ

whereas Z : true course, so the first criterion equals:

( )

sec1

φχ

∆−













∫

Similarly the second criterion’s hard constraint is

dmmax = A tan (maxRB) whereas A: ship’s advance

from extreme vertical XTE point E

till the next

point C

where it meets the course again and max

RB: maximum relative bearing to point E while the

mean is calculated as:

pRBAdm

/tan*

∑

where p is the number of selected and calculated

XTE points. This criterion expresses the ship’s mean

XTE from course, thus it’s an indicator of possible

approaches to navigational dangers like shallow wa-

ters, wrecks, etc. The third criterion and the hard

constraint that derives from it are:

pZZdd

/max

∑

−=

and

ZZdd

−= maxmax

Figure 2: Vertical deviation – Max and mean

These express the selection’s performance in

steering or the ship’s “swinging” on either side (Le-

ontopoulos 1979) [34]. Those criteria were com-

bined to measure the negative performance of each

alternative. The analysis

assigned 5 negative points

for each extra mile, 10 points for each XTE mile and

0, 2 for each degree of deviation. Moreover, the two

hard constraint thresholds were set to 0.02 miles /

Beaufort for dmmax and 2 degrees / Beaufort for

ddmax. Dmmax is increased by 20% for each knot

of current with a relative bearing > 45

. These are

default values and are justified after a survey with

experts aiming to assess the severity of each criteri-

on as far as the autopilot device is concerned. If the

user disagrees he / she can intervene and change this

balance by inserting values to the coefficients α, β, γ

assigned to each criterion during the interaction with

The thresholds are for a 65000 DWT Panamax bulk carrier.

For other types of ships the numbers are different, slightly in-

creasing with the tonnage

255

the system. After the normalization the selection’s

negative performance is calculated as follows:

( )

pZZpRBA

ZdaNP

/max/tan*

sec'1













−++













∆−













∑∑

∫













γβ

φχ

4.2 Wheather condition indexes

The weather condition indexes describe the wind,

sea and current. The case includes wind direction

and force, sea direction and force, current direction

and speed as well as swell direction and height. All

directions are expressed in degrees, wind and sea

force in Beauforts, current speed in knots and swell

height in meters. All directions are relative to bow

and current speed is true. A situation is considered

identical when the parameter differences will not ex-

ceed half of a preset allowance in either direction

(+/-). As the case library grows bigger the bounda-

ries can be stricter for better accuracy.

The weather situation is expressed by four major

phenomena which are wind, sea, current and swell.

Sea condition will always be a part of the situation

during the retrieval procedure, while current, swell

and wind can be omitted if there are not any exact

matches. If the phenomenon is to be included in the

case retrieval process, its indexes are analyzed fur-

ther in order to determine their actual importance

and whether they should be included as retrieval cri-

teria. Table 1 shows a strict version of the retrieval

process because the criterion used is the existence of

a Medium importance (M) characterization for the

relative course or the sea force. Relative course has

three importance levels (Low, Medium, and High)

covering 30 degrees from bow and sea force has six

levels (Very Low, Low, Medium, High, Very High,

and Extremely High) each covering two Beauforts.

As seen in table 1, in almost all cases the sea indexes

should be included in the retrieval. The lower part

presents the same data but now the criterion is the

existence of a High importance (H) in any of the two

indexes.

Table 1: Combined importance of sea direction and force (Me-

dium and High Importance Thresholds are set)

The influence of the sea condition parameter is

affected by a lot of things, but since the case refers

to the same ship, the only other factor to be consid-

ered is the loading situation. Sea direction and force

has a much greater impact when the ship is on bal-

last and less when it’s fully loaded. Thus, when the

vessel is on ballast condition more weather combina-

tions should be included. The strict version is used

for ballast condition and the less strict for the fully

loaded condition. Further division i.e. semi loaded

condition can be applied if needed.

Current, swell and wind are represented similarly

in the knowledge base. Importance weights were

assigned to each direction for each one of the three

phenomena. Current was given a scale of 0 – 10

knots ranging from Very low to Very High im-

portance with a pace of 2 knots. The existence of a

Medium importance is the criterion when the ship is

on ballast condition while a value of High im-

portance is necessary when the ship is loaded. Swell

is measured with a scale of 0 – 5 meters ranging

from Very Low to Very high importance while wind

has the same scale as the sea. The thresholds are at

least one Medium importance for the ballast and at

least one High importance for the fully loaded con-

dition.

4.3 Loading condition and ship particulars indexes

The loading condition indexes include information

about the deadweight, draft, trim, declination, LCG,

VCG, TCG, free surfaces, hogging and sagging. Ad-

ditional indexes include the capacity used, type of

cargo, fuel, ballast, supplies or alternative ones like

hull coefficient proportions, stowage factors, etc.

The ship’s particulars represent the user characteris-

tics and include the basic dimensions, ship’s coeffi-

cients, RPM (sea speed), rudder elements, maneu-

vering characteristics, etc. The loading condition and

the ship particulars indexes are presented in table 2.

256

Table 2: Loading condition and ship particulars indexes

The loading indexes are identified after inter-

views with merchant ship masters and deck officers

with more than adequate experience in the field. Im-

portance weights have been assigned to each one of

them in order to identify those necessary to be in-

cluded as criteria in each retrieval procedure. The

indexes with the highest importance are the DW, d,

δ, dec, VCG, Io, SF and the Ballast percentage. All

others are already covered by them and exist for ac-

curacy reasons. It should be noted that any of these

indexes can be omitted if the user wishes to or if the

case library is not rich enough and cannot retrieve

exact matches. Also, the value boundaries can

change to permit a stricter or a more loose retrieval

in accordance with the needs.

The ship’s particulars indexes describe the user

(ship) characteristics. Even though the Autopilot ap-

plication refers to the same ship the particulars are

inserted in the library in case a possible user compa-

ny decides to integrate the fleet’s libraries to create a

richer one, especially if there are vessels with similar

characteristics. Like every category of indexes and

as the case library grows more indexes can be added

or stricter criteria can be set.

4.4 Solution and outcome indexes

In this application the solution parameters are only

the configuration with the best performance and its

corresponding characteristics. For simplicity reasons

we included two attributes (for demonstration pur-

poses only) which are the angular velocity of the

rudder (AVR) and the rudder angle permitted (RA)

in order to keep the ship on course. The configura-

tions available can be any combination of these val-

ues, thus for AVR values n-2, n and n+2 degrees per

second and RA values of k-5, k and k+5 degrees we

have 9 possible combinations, plus a (n - 4, k – 10)

combination for very calm sea. Finally, the outcome

indexes are the same as the criteria indexes, but their

values will be the actual performance of the configu-

ration during the voyage.

4.5 Case Retrieval

Case retrieval is one of the most important parts of

the system’s reasoning since it is required to select

all the related cases, classify them according to their

utility towards the goals and promoting the most

promising of them. As mentioned in the literature

the proper retrieval requires a degree of similarity

between the new and the retrieved situation. Many

CBR systems use various levels of abstractions in

order to recognise similarities between cases of dif-

ferent domains. There are numerous algorithms used

for the case retrieval strongly dependent on the prob-

lem complexity. Usual serial algorithms are the Flat

memory – serial search enhanced with shallow in-

dexing, case library partitioning or synchronous par-

allel retrieval (Kolodner 93) [29], Shared Featured

Networks (Fischer 87 [13], Michalski and Stepp 83

[36], Cheeseman 88 [7], Quinlan 86), Discrimination

Networks (Feigenbaum 63) [12]) and Redundant

Discrimination Networks (Kolodner 93) while paral-

lel algorithms are Flat Library – Parallel search

(Stanfild and Waltz 81, 88 [47], Simoudis 91, 92,

Domeshek 89, 91 [10]), Hierarchical memory – Par-

allel search (Kolodner 93) [29]. A serial search is

used for the Autopilot application assisted by a case

library partitioning using the sea condition indexes.

Other situation parameters can be used in case the

library grows very big.

When the system detects the cases whose values

fall into the ranges permitted it uses the nearest

neighbour approach (Dasarathy 1991) [3] for each

selected characteristic in order to assess the degree

of situation similarity. This leads to the retrieval of a

set of cases which are ranked according to the crite-

ria. In the Autopilot’s knowledge base the priorities

are safety, command and monetary cost, so the goals

are ordered with this logic: Loss of steering, vertical

deviation, angular deviation and finally difference of

distance. The system rejects any case that violates a

hard constraint and then calculates the negative per-

formance of the remaining cases, proposing the one

with the lowest score to the user.

4.6 Evaluation and adaptation

The evaluation procedure is the comparison of the

actual performance of the configuration used with

the one stored (the best) in the case library. If the

performance is not satisfactory the user has two

choices. One is to select the second best configura-

tion for the specific situation and store it in memory

and the second is to adapt the selected configuration

to fit the new situation. This is done by initiating a

fine tuning procedure (or sensitivity analysis) where

257

the system changes the configuration parameter val-

ues and performs a new series of diagnostics in order

to track the adapted configuration with the best per-

formance. In the Autopilot application the system

assesses the performance of the adapted configura-

tions relatively easy since the parameters are only

two (AVR & RA) and the possible combinations no

more than ten. Of course the configuration

parameters can be much more, with the system’s

processing time increasing exponentially but then,

fuzzy logic classifications can be used to reduce the

processing time. One way to avoid that is to

categorize the configurations in classes and further

examine them if the performance is not adequate.

There is no adaptation procedure in this particular

applcation because the suggestion’s outcome is an

already preset configuration with fixed attributes.

Moreover, instead of trying to modify the reasoning

or re configurate the solution, it is far more

preferable to simply use the second or third best

configuration proposed by the system or re run the

diagnostics with less strict constraints.

5 CASE STUDY

A 65000 DWT bulk carrier was selected for this case

study which is presented based on real voyage data

except the values of the performance criteria, since

such a device is not developed yet. The ship sailed

from Los Angeles (USA) to San Bernardino (Philip-

pines) and performed its diagnostics during a great

circle trip. The ship is loaded with corn and travels

at usual sea speed. We suppose there is an autopilot

on board that has 10 different selections, so the di-

agnostic test will be performed 10 times in each part

of the great circle given that every part has signifi-

cantly different weather conditions. If this ideal situ-

ation occurs a case base of 10 X 11 = 110 cases will

be constructed in a single trip. The great circle data

are shown in table 3. It must be noted that the rest of

the case study is focused to the first way point for

simplicity reasons, since the procedure is similar for

every other part of the voyage. The distance set for

each selection is 10 miles, thus the first test will

cover a total distance of 100 miles.

A general description of the voyage is as follows:

The ship’s draught was 13.3 meters, the cargo holds

were full and the stowage factor was 1.52. There

was no hogging or sagging and the trim was 1 meter

by the stern. The engine’s RPM were 110 and the

ship’s initial stability satisfactory since the GM was

25 centimeters. The ship’s heading was 295 during

the first diagnostic and the wind was NW 6-7. The

sea was NNE moderate to rough and the current 2

knots to the starboard beam.

Table 3: Initial voyage data

The situation was presented with two sets of vari-

ables –weather and loading parameters- and a third

set which is already inserted in memory representing

the ship’s particulars. The variables used for the

weather conditions are the relative directions of sea,

current, wind and swell and are listed in table 4. The

weather situation is identical since the heading and

distance traveled for each test is the same (295, 10),

the wind

, sea and current differences do not exceed

the allowances permitted and there is not any swell.

The loading situation was represented using the car-

go (+/- 25*TPC

% MT), draft (+/- 0, 25 m), SF (+/-

0, 05), % hold capacity (+/- 10%), RPM (+/- 2%),

VCG (+/- 0, 05 m) and trim (+/- 0, 5 m) variables.

The parentheses show the allowances for the loading

situation similarity. The loading situation is shown

in table 5. Table 6 shows the case as it is stored in

the case library.

Table 4: The weather conditions during the diagnostic test

Table 5: The loading condition during the diagnostic test

The (-) declares left (port) from bow

Tons Per Centimeter: the amount of cargo required to alter the

ship’s draft for 1 centimeter

258

Table 6: The situation as it is stored in the base

Table 7 presents a scenario of possible criteria

values measured during the diagnostics. These in-

clude the criteria measuring the performance as well

as the hard constraints with their respective thresh-

olds. The first hard constraint eliminates three selec-

tions since the maximum vertical distance dmmax

exceeds the threshold dmmaxTh. Thus, selections

1.1, 1.2 and 1.3 are no longer considered. Similarly

the second constraint ddmax eliminates the selec-

tions 1.6 and 1.10 since the value must be below the

limit and not equal. The third constraint which re-

quires zero tolerance to steering losses eliminates se-

lections 1.6 and 1.7 as well as 1.2 and 1.3 which

were already excluded. At this point selections 1.4,

1.5, 1.8 and 1.9 remained active and the system cal-

culates their negative performance NP in order to

rank them. Finally, selection 1.8 is proposed as the

best alternative since it has less negative points than

the others.

Table 7: Criteria and hard constraints

An estimation of the potential benefits resulting

from a proper selection is shown comparing the bet-

ter with the worst alternative not taking into account

the hard constraints that exclude it. Those alterna-

tives are 1.8 and 1.2. Criterion d shows that the ship

travels 0, 9 extra miles

in every 10 miles of journey.

This means that during this passage the ship will

travel 6156, 6 * 0, 9 / 10 = 554 extra nautical miles

and will lose 554 / 15 = 37 hours in terms of time.

Moreover the ship will vertically deviate (mean)

from its course 278 meters more and swing about 10

degrees more (mean) if 1.2 is selected. This means

bigger exposure to danger and greater difficulty in

command which in turn wears the hull, engine, pro-

peller, etc. One must not forget the additional wear

and tear of the rudder and engine if a false steering

configuration is set as well as the crew fatigue and

other damages that may result from rolling, pitching,

etc.

6 CONCLUSION – FUTURE RESEARCH

Summarizing the above it is concluded that a way of

selecting the best alternative from a pre existing set

of configurations of an autopilot is possible using

CBR as the core technology. Since such a device is

not yet developed this application is considered con-

ceptional and its main task was to present some ini-

tial thoughts still requiring verification and hard da-

ta. The development of a prototype could give a lot

of answers and test the system’s performance in the

real world.

Apart from that we strongly believe that the mari-

time industry and especially the ship is a very com-

patible environment for CBR and numerous applica-

tions could be developed. In time, an integrated

system able to deal with a number of issues could be

developed and with the accumulation of cases its

performance and learning will constantly improve.

REFERENCES

1. Alter, S. L. (1980), Decision support systems: current

practice and continuing challenges, Reading, Mass.,

Addison-Wesley Pub.

2. Bain, W., (1986), Case Based Reasoning: A computer

model of subjective assessment, PhD diss, Department

of Computer Science, Yale

3. Belur V Dasarathy, (1991), Nearest neighbour,

Norms, Pattern Classification Techniques, IEEE

Computer Society Press

4. Birnbaum, L & Collins G, (1989), Reminding and en-

gineering design themes: A case study in indexing vo-

cabulary, Pensacola, Florida, Morgan Kaufman

5. Bonczek R. H, C. W Holsapple & A. B Whinston.

(1981), Foundations of Decision Support Systems,

New York, Academic Press.

6. Bowditch N, (2002), The American practical Naviga-

tor, Maryland, USA, National Imagery and Mapping

Agency

Additional miles travelled if instead of the best, the worst per-

forming configuration is selected by the device

259

7. Cheeseman P, Kelly J, Matthew, Stutz J, Taylor W,

and Freeman, (1988), Autoclass, a Bayesian classifi-

cation system. Proceedings of the Fifth International

Machine Learning Workshop, San Mateo, Morgan

Kaufmann

8. International Maritime Organization, (1972), Conven-

tion on the International Regulations for Preventing

Collisions at Sea, Part B, Rule 9, Narrow Channels

9. Dhar V, Stein R. (1997), Seven methods for trans-

forming corporate data into business intelligence,

Upper Sandle river, NJ, Prentice Hall,

10. Domeshek E and Kolodner J., (1991), Toward a case

based aid for conceptual design, International journal

of Expert Systems, pp. 201 - 220

11. Dutton B, (1958), Navigation and Piloting, Annapolis,

Maryland, United States Naval Institute

12. E. A. Feigenbaum and J. Feldman, (1963), The simu-

lation of natural learning behaviour. Computers and

thoughts, New York, McGraw-Hill

13. Fischer D, (1987), Knowledge Acquisition via incremental

conceptual clustering, Machine Learning Journal, Vol

II, 139 – 172, Springer Netherlands publishers

14. G Honderd, J. E. W Winkelman, (1972), An adaptive

Autopilot for ships, Proceedings 3

Ship Control Sys-

tems Symposium, Bath, UK,

15. Goel A, 1992, Representation of design functions in

experience based design, Intelligent Computer Aided

Design, ed. D Brown, M Waldron, H Yoshikawa,

Amsterdam, North Holland

16. Gorry and Scott Morton. (1971), A framework for

management information systems, Sloan Management

review, 13, pp 56 - 79

17. Hammond K, (1987), Explaining and repairing plans

that fail. Proceedings of IJCAI-87, San Mateo, Mor-

gan Kaufmann

18. Hammond K, (1989), Case Based Planning: Viewing

planning as a memory task, Boston, Academic Press

19. Herther et al, (1971), A fully Automatic Marine Radar

Data Plotter, Journal Inst. Navigation vol 24, pp 43 –

20. Holsapple C. W & A. B. Whinston. (1996), Decision

Support Systems, A knowledge based approach, Min-

neapolis, West publishing company

21. Holzhuter T, (1997), LQG Approach for the High pre-

cision Track Control of ship IEE Proc Control Theory

Application, 44, 121 – 127

22. International Maritime Organization, (1995), Interna-

tional Convention on Standards of Training, Certifi-

cation and Watch keeping for Seafarers (STCW),

Chapter VIII, Watch keeping

23. J Van Amerongen, A. J. U Ten Cate, (1975), Model

reference adaptive autopilots for ships, Automatica

Vol 11, pp 441 – 449, Pergamon press

24. J Van Amerongen, H. R Van Nauta Lemke, (1986)

Recent developments in automatic steering of ships,

Proc. of the meeting of the Royal Institute of Naviga-

tion, Amsterdam, The Netherlands

25. Keen, P. G. W. and Scott Morton M S., (1978), Deci-

sion Support Systems: An Organizational Perspective,

Reading, MA: Addison-Wesley

26. Keeney, Raiffa, (1976), Decisions with multiple objec-

tives, Preferences and value trade offs, Cambridge

University Press

27. KJ Astrom, (1977), Self tuning regulators, NASA

Conference Publication

28. Knowles T W, (1989), Management Science, Building

and using models, pp. 28, Homewood, Illinois, Irwin

pub.

29. Kolodner. J, (1993), Case Based Reasoning, San

Mateo, California, Morgan Kaufmann publ

30. K Ohtsu, M Horigome, G Kitagawa, (1979), A new

ship’s autopilot design through a stochastic model,

Automatica, 15, pp. 255 - 268

31. K. R Goheen, E. R Jeffreys, (1990), System Identifica-

tion of Remotely Operated Vehicle Dynamics, Journal

Offshore Mechanics and Arctic Engineering

32. Leake D, Kolodner J, (2003), Learning through case

analysis, Encyclopaedia of cognitive science, Nature

publishing group, London

33. Leake D. (1996), Case-Based Reasoning: Experienc-

es, Lessons & Future Directions, Menlo Park Califor-

nia, USA, American Association for Artificial Intelli-

gence,

34. Leontopoulos A (1979), Practical Applications of

Ship‘s Stability, Piraeus, Hellenic Educational Center

of Merchant Marine Executives (KESEN)

35. Mark W, E. Simoudis & D. Hinkle, (1996), Case

Based Reasoning, Expectations and results, Menlo

Park, California, AAAI Press

36. Michalski R S and Stepp R E, (1983), Learning from

observation: Conceptual clustering, Machine Learn-

ing: An Artificial Intelligence Approach, VOL I, Los

Altos, California, Morgan Kaufman

37. Moorman K and Ram A, (1992), A case based ap-

proach to reactive control for autonomous robots.

Proceedings of the AAAI Fall Symposium on AI for

Real World Autonomous Robots, Cambridge, Massa-

chusetts, MIT Press

38. Nikitakos N., G. Fikaris, (2007), POLARIS, A Deci-

sion Support System for maritime policy and man-

agement, Proceedings, IMAM 2007, Varna, Bulgaria

39. Polkinghorne M. N. Roberts G. N, Burns R. S, (1994)

The implementation of a fuzzy logic marine autopilot,

Proceedings IEE Control International Conference

VOL II, p. 1572 – 1577, Warwick, UK

40. Roberts G. N and Sutton R, (2006), Advances in Un-

manned Marine Vehicles, IEE Control series, IEE Press

41. Power, D. J. (2000). Web-based and model-driven de-

cision support systems: concepts and issues Proceed-

ings of the Americas Conference on Information Sys-

tems, Long Beach, California.

42. Redmond M A, (1992), Learning by observing and

understanding expert problem solving. Georgia Insti-

tute of technology, College of Computing, Atlanta

43. Rissland E., Kolodner J. & Waltz D., (1989) Case

Based Reasoning from DARPA: Machine learning

program plan, CBR Workshop, Pensacola, Florida,

Morgan Kauffman

44. Simpson R L, (1985), A computer model of case based rea-

soning in problem solving: An investigation in the domain of

dispute mediation. Georgia Institute of technology,

School of Information and computer science, Atlanta

45. Schank, Kass, Riesbeck. (1994), Inside case based ex-

planation, The Institute of the learning sciences,

North-western University, Lawrence Erlbaum, publ.

Hillsdale, New Jersey.

46. Sprague, R. H. and E. D. Carlson (1982), Building ef-

fective decision support systems. Englewood Cliffs,

N.J., Prentice-Hall

47. Stanfild C W and Waltz D, (1987), The memory based

reasoning paradigm, Proceedings AAAI – 87, pp. 576 - 578

48. Unar M. A and Murray Smith D. J, (1999), Automatic

Steering of Ship using Neural Networks, International

Journal of Adaptive Control and Signal Process 13,

203 – 218

49. X. J Yang & X. R Zhao, (2006) Self organizing Neu-

ral Net Control of Ship’s Horizontal Motion, Interna-

tional Symposium on Instrumentation Science and

Technology, Institute of Physics Publishing, Journal

of Physics: Conference series 48, 1284 – 1288