IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO.

4, MAY 2014 1539

Solutions to the Failures and Limitations of Mamdani
Fuzzy Inference in Ship Navigation
Lokukaluge P. Perera, J. P. Carvalho, Member, IEEE, and C. Guedes Soares
AbstractThis paper proposes a methodology for overcoming
Mamdani-type inference failures on a fuzzy-logic-based decision-
making process applied to collision avoidance in ship nav-
igation. The fuzzy inference failures are observed in three
distinct situations: 1) intersected contradictory decision bound-
aries; 2) an improper transition region between the inference
boundaries of nonintersected contradictory decisions; and 3) con-
tradictory decision accumulation under multiple obstacle scenar-
ios. The solutions for overcoming these fuzzy inference failures and
their limitations are also discussed in this paper. The proposed
solutions consist of insertion of smooth transition regions, deter-
mination of the proper size of the smooth transition regions, and
use of multilevel decision/action formulations. Furthermore, this
paper analyzes a decision-making process for ship navigation,
derives input and output fuzzy membership functions (FMFs), for-
mulates an IFTHEN-rule-based fuzzy inference system (FIS),
and presents simulation results that support recovery from rule
inference failures in several contradictory decision boundary
conditions.
Index TermsContradictory decision inference, decision sup-
port systems, Mamdani fuzzy inference failure, maritime trans-
portation, ship collision avoidance.
I. INTRODUCTION
T
HE USE of Mamdani-fuzzy-rule-based inference in
navigation/steering systems has been previously proposed
and extensively applied. However, this solution exhibits several
problems that may become critical when applied in real-world
systems. These problems arise in three different situations that
have been identied in this paper. One of these situations is
very well known, i.e., the problem of failures due to inferred
contradictory decisions. For example, in the presence of a target
located straight ahead on a region covered by two fuzzy rules
Manuscript received May 18, 2013; revised September 7, 2013; accepted
October 23, 2013. Date of publication November 4, 2013; date of current
version May 8, 2014. This work was supported in part by the Foundation
for Science and Technology of Portugal (FCT) through Project Methodology
for Ships Manoeuvrability Tests with Self-Propelled Models under Contract
PTDC/TRA/74332/2006. The work of L. P. Perera was supported by the FCT
through a doctoral fellowship under Contract SFRH/BD/46270/2008. The work
of J. P. Carvalho was supported in part by the FCT under Project PEst-OE/
EEI/LA0021/2013. An earlier version of this paper was presented at the IEEE
Ninth International Conference on Industrial Informatics, Lisbon, Portugal,
July 2011. The review of this paper was coordinated by Dr. M. S. Ahmed.
L. P. Perera and C. Guedes Soares are with the Center for Ma-
rine Technology and Engineering, Instituto Superior Tcnico, University
of Lisbon, Lisbon 1049-001, Portugal (e-mail: prasad.perera@mar.ist.utl.pt;
guedess@mar.ist.utl.pt).
J. P. Carvalho is with the Institute of Systems and Computer Engineering
Research and Development in Lisbon, Instituto Superior Tcnico, University of
Lisbon, Lisbon 1649-004, Portugal (e-mail: joao.carvalho@inesc-id.pt).
Color versions of one or more of the gures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identier 10.1109/TVT.2013.2288306
where one rule states turn left and the other turn right, the
decisions cancel each other out, and the vehicle goes straight
ahead. From here on, we shall refer to this case as Situation I:
intersected contradictory decision boundaries.
Although the other two situations are not commonly ad-
dressed in the literature, they can also result in catastrophic
failures in real-world systems; these distinct situations usually
occur in the presence of stationary/moving multiple targets,
even if this is not a necessary condition, on 2-D navigation/
steering systems. Such situations can occur when an obstacle is
in close range due to the fact that small linear distances traveled
either by a moving obstacle or the own vehicle can correspond
to large angular movements of the obstacle when seen from the
own vehicle. Therefore, the apparent position of the obstacle
may quickly cross through the several angular fuzzy regions
used to infer the decision-making process. For example, in two
consecutive iterations, the moving obstacle can pass from a
region where the correct inferred decision was turn left to a
region where the decision is turn right; as a result, the system
jumps between contradictory decisions. A navigation system
where inertia is a factor (basically, all real-world large-vessel
navigation systems fall in such situations) may not quickly
respond to such decision changes, and the time delay associated
with the system response may cause an erratic, undesirable, and
possibly dangerous behavior. This situation will be referred to
as Situation II: improper transition region between the infer-
ence boundaries of nonintersected contradictory decisions.
The nal situation, Situation III: contradictory decision
accumulation in multiple targets scenarios, occurs when con-
tradictory decisions can cancel each other out due to the pres-
ence of multiple stationary or moving targets. This situation
is commonly observed in what can be referred to as single-
level fuzzy-rule-based decision-making systems, i.e., systems
where fuzzy input and output variables are directly connected
by inference rules. One can theoretically solve this problem by
increasing the rule base size as long as the number of moving
obstacles is limited and known in advance; however, this is
rarely possible or practical since scalability becomes an issue
due to the problem of combinatorial rule explosion. This paper
proposes and describes several requirements that can be applied
to overcoming the given situations and therefore prevent the
inherent system failures. They include the following.
1) The navigation system should be equipped with a mini-
mum of two fuzzy output decisions (e.g., course/heading
change and speed change).
2) There should be proper formulation of decisions in the
input and output linguistic terms and respective fuzzy
0018-9545 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1540 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 4, MAY 2014
membership functions (FMFs), i.e., common boundary
points or regions between contradictory decisions should
be avoided in both input and output FMFs.
3) If a common boundary between regions that indicate
contradictory decisions is unavoidable, then a smooth
transition region should be inserted.
4) When a smooth transition region between common
boundary points or regions of contradictory decisions is
necessary, then its proper size should be discussed to
prevent Situation II, i.e., an improper transition region
between the inference boundaries of nonintersected con-
tradictory decisions.
5) A multilevel decision/action process should be
used to avoid Situation III, i.e., contradictory decision
accumulation.
This paper illustrates the three previously discussed types
of inference rule failures and the solutions to be implemented
on ocean navigation to improve safety by avoiding collision
situations. To overcome the inference rule failures, the insertion
of a smooth transition region between intersected contradic-
tory decision inference boundaries and a geometrical relation
to determine its proper size are proposed. Furthermore, to
overcome the accumulated decision failures, we propose a
multilevel decision-action process composed of a fuzzy-
logic-based parallel decision-making (PDM) module whose
decisions are formulated into sequential actions by a Bayesian-
network-based module.
II. RELATED WORK
A. Decision-Making Processes
Human-type decision-making plays an important role in
modern industrial applications. Therefore, several experimental
platforms and industrial applications are formulated to sim-
ulate human decision-making capabilities. Although human
decision-making capabilities go beyond single-criterion situa-
tions, it is widely accepted that humans have more limitations
in multicriteria decision-making situations. However, humans
have better perceptions of time, direction, speed, shape and
possibility, likelihood, truth, and other attributes of physical and
mental objects when compared with articial computational
technology [1]. Since computers are of great help when dealing
with crisp precise data, whereas humans are better adapted to
deal with perceptions and uncertain data, the integration of
human perceptions into multicriteria computational technology
could be the ultimate objective in next-generation decision-
making processes in machine learning technology.
Decision-making processes, as reviewed in recent litera-
ture, are usually divided into single-criterion decision-making
(SCDM) and multicriteria decision-making (MCDM). MCDM
can be dened as a study of methods and procedures by
which concerns about multiple conicting criteria. That can be
formally incorporated into the management planning process
[2] and can be further divided into individual decision-making
and group decision-making [3]. However, the approaches in
SCDM and MCDM could be further classied into determin-
istic, stochastic, and fuzzy control methods [4][6].
B. Fuzzy-Logic-Based Decision-Making
Fuzzy-logic-based systems, which are formulated on the
human type of thinking [7], are well known for facilitat-
ing a human-friendly environment during a decision-making
process. Hence, several fuzzy-logic-based decision-making
systems have been developed in research and commercial ap-
plications to fulll the technological requirements inspired by
human behavior in decision-making [8]. The human under-
standing of relationships among objects and/or situations in
decision-making are illustrated using various fuzzy functions
in [9] and [10], both based on TakagiSugenoKang-type and
Mamdani-type fuzzy inference systems (FISs).
Decision-making processes formulated to simulate human
decision-making capabilities are extensively inuenced on the
area of autonomous navigation. The main objectives of a nav-
igation system decision-making process are to avoid static and
dynamic targets and to reach the expected nal destination.
Therefore, intelligent decision-making capabilities ultimately
inuence the survivability and success of the autonomous nav-
igation system.
Fuzzy-logic-based approaches for autonomous navigation
systems have been widely considered in several recent studies.
A fuzzy-target-based soft decision for mobile vehicles in dy-
namic environment is proposed in [11], where a navigation tra-
jectory for the nal destination is selected from all possible via
points learned during its navigation. Similarly, a fuzzy-logic-
based algorithm for path selection in autonomous vehicle navi-
gation is proposed in [12]. The drawbacks of the given methods
come from the fact that they were implemented in limited
and constrained environmental models, causing the decision-
making process to be less effective and time-consuming in
larger navigation spaces.
Seraji and Howard present in [13] a fuzzy logic approach
to behavior-based robot navigation on challenging terrain.
Seraji and Bon [14] also presented a study for autonomous
navigation of planetary rovers using a fuzzy logic framework.
These systems consist of multiple behavior capabilities: goal
seeking, terrain traversing, and collision avoidance. Fatmi et al.
propose a similar approach of fuzzy logic for mobile robot
navigation in [15]. Fuzzy-logic-based navigation controllers for
a mobile robot where the input and output FMFs for steering
and speed commands are expressed by linguistic values are
proposed and implemented in [16] and [17]. However, all the
mentioned systems were implemented under limited obstacle
behavior and/or stationary environmental conditions.
Fuzzy decision-making processes in mobile robot navigation
under dynamic environmental conditions are presented in [18]
and [19]. The fuzzy decision-making process in these studies
consists of two algorithms for obstacles avoidance and tar-
get following. The drawback to this approach is that fuzzy
inference rule failures can occur due to Situation I, as described
earlier.
Usually, fuzzy-rule-based decision-making processes are
formalized assuming a single-level system. However, it is
possible to formulate the multiple identities within a sys-
tem using different sets of rules, as proposed in [20], and a
multiple-behavior-based fuzzy control for sonar-based obstacle
PERERA et al.: SOLUTIONS TO FAILURES AND LIMITATIONS OF MAMDANI FUZZY INFERENCE 1541
avoidance of a mobile robot is presented in [21]. Although
these studies facilitated a multiple-behavior-based fuzzy control
approach, they still suffer from the accumulated decision can-
celation situations between the nonintersected-contradictory-
decision inference regions of a single-level fuzzy-logic-based
system (i.e., Situation III), as further discussed in this paper.
C. Fuzzy-Logic-Based Collision Avoidance Decisions
Collision avoidance facilities for an autonomous ocean navi-
gation system with intelligent decision-making capabilities are
considered as a case study in this paper. The applications and
recent developments of autonomous ocean navigation systems
are summarized in [22][26]. However, the law of the sea
has been ignored by most of these studies. Theoretically, all
ocean-going vessels should follow the law of the sea when
trying to avoid collision situations. The current law of the sea
was introduced by the International Maritime Organization in
1972 by the Convention on the International Regulations for
Preventing Collisions at Sea (COLREGs) [27], [28].
Fuzzy-logic-based collision avoidance systems (CASs) for
ship navigation are proposed in [29][31]. However, several
drawbacks are observed in these implementations. The pro-
posed studies are limited to two-vessel collision situations, the
COLREGs rules and regulations and expert knowledge in ocean
navigation are ignored, and the system inputs and outputs are
directly related by fuzzy rules (single-level systems). Therefore,
the systems capabilities to overcome complex multivessel
collision situations are limited.
III. COLLISION AVOIDANCE IN SHIP NAVIGATION
A. Multivessel Collision Situation
A multivessel collision situation is shown in Fig. 1. The
own vessel, i.e., the vessel that is equipped with the pro-
posed CAS, is located at point O(k) (x
o
(k), y
o
(k)). The
target vessels are located at points P
1
(k), P
2
(k), . . . , P
i
(k)
with various navigational trajectories, respectively. The own
vessel trajectory will intercept the target vessel trajectories
around points C
1
(k), C
2
(k), . . . , C
n
(k) at the time instants
T
1
(k), T
2
(k), . . . , T
n
(k), respectively.
The ith target vessel that needs to be avoided is located
at point P
i
(k) (x
i
(k), y
i
(k)). The ith target vessel estimated
relative trajectory P
i
(k)B
i
(k) will intercept the own vessel
domain at the closest point B
i
(k). Therefore, the closest dis-
tance between the vessels in a collision situation is represented
as R
DCPA
(k). The two-vessel collision point is represented
by C
i
(k), with the ith target vessel distance to the collision
point R
ci
(k). The own vessel speed and course conditions
are represented by V
o
(k) and
o
(k), respectively. The ith
target vessel speed and course conditions are represented by
V
i
(k) and
i
(k). The ith target vessel bearing and colli-
sion point relative bearing (CPRB) conditions are represented
by
i
(k) and
ci,o
(k), respectively. The relative course and
speed conditions of the ith target vessel are represented by

i,o
(k) and V
i,o
(k).
Fig. 1. Multivessel collision situation.
B. Collision Avoidance System
A block diagram of the proposed CAS is shown in Fig. 2.
The CAS consists of ve sections: 1) a vessel trafc monitoring
and information system(VTMIS); 2) a collision risk assessment
(CRA) module; 3) a PDM module; 4) a sequential action
formation (SAF) module; and 5) an own vessel control system
(VCS). The inputs into the VTMIS are the position of the own
vessel and the positions of the target vessels.
The VTMIS consists of four sections: 1) a sensor unit; 2) a
vessel detection and tracking (VDT) module; 3) a vessel state
estimation and trajectory prediction (VSETP) module; and
4) an intervessel communication module. The sensor unit (i.e.,
radar and laser applications) acquires the real-time position
data of the target vessels. Then, the target vessels data are
used in the VDT module to identify and to track each target
vessel separately using these data. The VSETP module uses
the collected tracking data to estimate vessel states and to
predict each target vessels trajectory. Finally, the intervessel
communication unit distributes the previous information among
vessels (i.e., own vessel communication unit) and shore-based
centers. Extensive details on the VTMIS have been presented
in [32].
The main objective of the CRA module is to evaluate the
collision risk of each target vessel with respect to the own
vessel navigation. This is achieved by the relative coursespeed
estimation unit and by the time and place until collision es-
timation unit. More details on the CRA module have been
presented in [33][35]. The inputs into the CRA module are the
measured/estimated position data of the own vessel and target
vessels. The output of the CRA module is the time until the
collision situation T
i
(k) of the ith target vessel, which in turn
is the input of the SAF module. The PDM module consists
of a fuzzy-logic-based decision-making process that generates
1542 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 4, MAY 2014
Fig. 2. Block diagram for the CAS.
Fig. 3. Mathematical formulation of a two-vessel collision situation.
parallel collision avoidance decisions D
i
(k), with respect to
each target vessel.
On the next step, the parallel ith decision of collision avoid-
ance D
i
(k) is forwarded from the PDM module to the SAF
module. The main objective of the SAF module is to organize
the parallel decisions made by the PDM module into sequential
actions A
i
(k), while taking into consideration the time until
the collision situation T
i
(k) from the CRA module. The A
i
(k)
actions are divided into course and speed control actions to be
implemented on the course control and speed control subsys-
tems of the own VCS.
The collision avoidance actions A
i
(k) formulated by the
SAF module act as second-level decisions regarding the
fuzzy-logic-based decisions D
i
(k) formulated in the PDM
module. Actions A
i
(k) are designed to overcome contradictory
accumulated rule inference failures (Situation III). The PDM
model is further discussed in Section IV and more details
regarding the fuzzy-Bayesian-based decisionaction formula-
tions in ship navigation can be found in [35].
IV. FUZZY LOGIC APPROACH
The overall design process of the fuzzy-logic-based decision-
making process described in this paper can be resumed in the
following ve steps: 1) identication of fuzzy input and output
system variables; 2) creation of FMFs for each input and output
system variables; 3) formulation of the FIS IFTHEN fuzzy
rules; and 4) defuzzication of the decisions. However, these
input and output system variables are derived by considering a
two-vessel collision situation that is shown in Fig. 3.
PERERA et al.: SOLUTIONS TO FAILURES AND LIMITATIONS OF MAMDANI FUZZY INFERENCE 1543
Fig. 4. Range FMF.
Fig. 5. Bearing FMF.
Fig. 6. Relative course FMF.
Fig. 7. Speed ratio FMF.
A. Fuzzy Membership Functions
The PDM module executes fuzzy inferences in the decision-
making process (see Fig. 2). The module consists of three
main units: 1) a fuzzication unit, 2) a fuzzy-rule unit, and
3) a defuzzication unit. The input fuzzy variables, and the
respective FMFs are range FMF (see Fig. 4), bearing FMF
(see Fig. 5), relative course FMF (see Fig. 6), and speed ratio
FMF (see Fig. 7). Two output fuzzy variables and respective
FMFs are used in the defuzzication unit of the PDM mod-
ule: course change FMF (see Fig. 8) and speed change FMF
(see Fig. 9). The input and output FMFs are normalized and
have trapezoidal shapes. The respective FMFs in Figs. 49
have been reproduced from [35] to improve the readability of
this paper.
The mathematical formulation of a two-vessel collision situa-
tion used to derive the input and output FMFs is shown in Fig. 3
Fig. 8. Course change FMF.
Fig. 9. Speed change FMF.
as previously mentioned. The own vessel navigational space is
divided into three circular regions with radius R
vd
, R
b
, and R
a
.
Radius R
a
represents the approximate range to the target vessel
detection when the own vessel is in a give way situation,
(i.e., when the own vessel has a lower navigation priority) and
should take appropriate actions to avoid collision situations.
Radius R
b
represents the approximate distance to the target
vessel when the own vessel is in a stand on situation (with
the higher priority for navigation) but should take appropriate
actions to avoid the collision situation due to the absence of the
appropriate actions from the target vessel. One should note that
the vessel coming from the starboard side has higher priority
for navigation in accordance with the COLREGs rules and reg-
ulations. The radius R
vd
represents the vessel domain. Dotted
circles separate the regions corresponding with the Range FMF
(see Fig. 4). The R
i
(k) represents the range of the ith target
vessel.
The own vessel navigation domain is divided into ten regions
numbered from I to X (see Fig. 3). Each of these regions
corresponds to one of the ten regions in the Bearing FMF (see
Fig. 5). It is assumed that the target vessel is located within
these ten regions, and the collision avoidance decisions are
taken in accordance with the respective regions. Although eight
regions are generally enough to implement the COLREGs rules
and regulations of ocean navigation as presented in [36], ten
regions are proposed to overcome Situation I rule inference
failures, as further discussed in Section VI.
As presented in Fig. 3, the target vessel position II domain is
divided into eight divisions (from II-a to II-h) of relative course

i,o
(k) (see Fig. 1). These divisions are separated by dotted
lines coinciding with the relative course FMF (see Fig. 6). The
higher collision risk regions, i.e., mid (II-e and II-g) and high
(II-f), are also shown in Fig. 6. Finally, the speed ratio FMFs
(see Fig. 7) are used to describe the speed ratio between the
target vessel and the own vessel, i.e., V
i
(k)/V
o
(k).
1544 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 4, MAY 2014
B. Fuzzication Unit
Fuzzication is the process that transforms crisp inputs into
a collection of the input membership degrees to each of the
fuzzy variable linguistic terms. This process is executed in the
fuzzication unit of the PDMmodule. The inputs fromthe CRA
module: range R
i
(k), bearing
i
(k), relative course
i,o
(k),
and relative speed V
i,o
(k) of the ith target vessel are fuzzied in
this unit with respect to the input FMFs, i.e., range FMF R
i
(k)
(see Fig. 4), bearing FMF
i
(k) (see Fig. 5), relative course
FMF
i,o
(k) (see Fig. 6), and speed ratio FMF V
i
(k)/V
o
(k)
(see Fig. 7). The fuzzied results are transferred to the fuzzy
rules unit. A Mamdani-type rule-based FIS is used in this unit.
The minmax norm is the aggregation operation considered in
this paper. In this norm, the minimum operator is considered
for intersection, and the maximum operator is considered for
the union of two fuzzy sets.
C. Development of the Fuzzy Rules
The fuzzy rules are formulated in accordance with the rules
and regulations of the COLREGs [27] and expert navigational
knowledge to facilitate a regulated prevention of collision and
to eliminate navigation conicts. Tables I and II present a
summary of the CRAs, fuzzy rules, and collision avoidance
decisions used in this paper. The tables have been reproduced
from [35] to improve the readability of this paper. The rst
column in Table I represents bearing
i
(k) (Bear.) of the target
vessel, which is divided into ten regions (IX). The second
column represents the relative course
i,o
(k) (Cou.), divided
into eight regions (a to h) of the target vessel orientations.
The collision risk (Risk) assessment with respect to the relative
course is divided into three sections of low risk (low), medium
risk (mid), and high risk (high). Only high and medium colli-
sion risk situations, where the collision avoidance actions must
be executed, are present in Tables I and II. The target vessel
range R
i
(k) from R
vd
to R
a
and from R
a
to R
b
are presented
in the third and fourth columns, respectively.
The third and fourth columns are further divided into two
subcolumns. The relative speed ratio of V
i
(k)/V
o
(k) is pre-
sented in the rst subcolumn of the third and fourth main
columns. The speed conditions of V
i
/V
o
<, , and > 1 rep-
resent the target vessel speed approximately less than, equal,
and greater than the own vessel speed. Finally, the decisions
that need to be taken to avoid collision situations are presented
in the second subcolumn of the third and fourth columns. The
decisions can be categorized as: course to starboard (
o
> 0);
course to port (
o
< 0); no course change (
o
= 0) increase
speed (V
o
> 0); decrease speed (V
o
< 0); no speed change
(V
o
= 0); and not applicable (NA). A similar organization is
considered for Table II.
D. Defuzzication Unit
The collision avoidance decisions D
i
(k) for each target
vessel are generated by the defuzzication unit, as presented
in Fig. 2. The fuzzy inference results from the fuzzy rule
unit are defuzzied by the output course change FMF (see
TABLE I
CRAS, FUZZY RULES, AND DECISIONS
Fig. 8) and the speed change FMF (see Fig. 9) to obtain
the course change decisions D
i
(k) and the speed change
decisions D
V
i
(k). These decisions will be formulated for
collision avoidance actions in the own vessel navigation.
The defuzzication process uses the centroid method. In this
method, one calculates the center of gravity of the result-
ing fuzzy set and uses its abscissa as the nal result of the
inference.
V. FAILURES IN FUZZY RULE INFERENCE
As introduced in Section I, the rule inference failures in
fuzzy-logic-based navigation/steering systems can occur in
three situations: Situation I (intersected contradictory decision
PERERA et al.: SOLUTIONS TO FAILURES AND LIMITATIONS OF MAMDANI FUZZY INFERENCE 1545
TABLE II
CRAS, FUZZY RULES, AND DECISIONS
boundaries); Situation II (improper transition region between
the inference boundaries of nonintersected contradictory deci-
sions) and Situation III (contradictory decision accumulation in
multiple targets scenarios).
A. Situation I: Intersected Contradictory
Decision Boundaries
Fig. 10 shows a pictorial example of a possible Situation
I failure if one considers that the own vessel should steer
to port when the target vessel is in region VIII and should
steer to starboard when the target vessel is in region I. If the
target vessel is located in the region where the input FMF that
represent regions I and VIII intersect (fuzzy region, see Fig. 10),
Fig. 10. Fuzzy-rule inference failure due to Situation I.
the defuzzied inference result will be No Action due to
the cancelation of the two contradictory decisions (steer to
starboard: o > 0; steer to port: o < 0) when the centroid
defuzzication method is applied. This decision would possibly
lead to a catastrophic system failure. Notice that, although
the contradictory output FMFs do not intersect, the way that
rules are dened generates contradictory decisions due to the
Mamdani inference mechanism.
Solution: Fuzzy Smooth Transition Region Insertion: The
usual solution for solving this problem consists in using a
higher level decision process to override one of the conicting
decisions. However, these conicting decision situations may
not be noticeable in the decision process and almost impossible
to capture the actual positions where these conicts will occur.
Therefore, usually, there is not a simple universal solution;
therefore, one must consider and prepare the decision process
for every single possible rule-based failure case.
Here, a more elegant and simple solution is proposed, which
basically consists of the insertion of a fuzzy smooth transition
region on the boundary of the regions that have contradic-
tory decisions, as presented in Fig. 11. A smooth transition
region X is inserted between the contradictory decision
regions IX and I. As presented in the gure, the decisions in this
smooth region must not contradict any of the decisions in the
original regions, and in addition, the speed of the own vehicle
should be decreased. As a result of these proactive actions,
the relative position of the target vehicle will automatically
be redirected into either one of the original regions IX or I.
As a result, the CAS will end up executing the collision avoid-
ance decisions of the original regions. This solution is further
discussed in Section VI.
1546 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 4, MAY 2014
Fig. 11. Fuzzy-rule inference failure due to Situation II.
B. Situation II: Improper Transition Region Between the
Inference Boundaries of Nonintersected
Contradictory Decisions
Fig. 11 further shows a pictorial example of a possible
Situation II failure. Let us assume that, in this case, a smooth
transition region X exists between regions IX and I and that this
region avoids the problem of intersected contradictory decision
boundaries by stating no course change and speed decrease
while the target is in region X. These actions will make the
target vessel move into one of the nonconict regions I or IX.
However, if the target vessel assumes the relative trajectory
shown in Fig. 11, Situation II failures can occur if region X
size is improper, i.e., not wide enough. In such case, the
navigation systemdecisions can be contradictory in consecutive
(or relatively close) instants resulting on an erratic trajectory
and a possible catastrophic failure since the target vessel might
not be avoided. Therefore, the selection of a proper fuzzy
smooth region is an important part of the designing process of
the fuzzy-logic-based system. This solution is further discussed
regarding the fuzzy-logic-based decision-making process of the
CAS in ship navigation in Section VI.
Solution: Proper Sizing of Fuzzy Smooth Transition Regions:
In general, the smaller the size of the smooth transition regions,
the better is the system behavior as it facilitates the system
transition into a better decision-making region. However, since
smaller regions can cause the system to jump between two
contradictory decisions during the avoidance process. To im-
plement two contradictory decisions on the same trajectory,
one must insure that the smooth region is large enough to
avoid contradictory decisions during the avoidance process of a
single target.
As an example, consider a special collision situation ob-
served in this paper regarding the size of the smooth transition
region X (see Fig. 3). This smooth region was necessary due to
the close range intersected contradictory decision boundaries
between regions I and IX. When region X is not properly sized,
Situation II failures might occur if the target vessel transits from
region I into region IX through region X.
To understand the process of designing a proper smooth
region, one should be aware that if no avoidance measures are
taken, when the target vessel is coming from the right end of
a head-on situation (region I), the relative navigation trajectory
converges toward the own vessel domain if there is a collision
risk between vessels, and the relative trajectory diverges from
the target vessel domain when there is no collision risk. Hence,
the size of region X should be designed in such way that the rel-
ative trajectories of the target vessel do not converge from one
region to another region that contains contradictory decisions.
This solution is further discussed regarding the fuzzy-logic-
based decision-making process of the CAS in ship navigation
in Section VI.
C. Situation III: Contradictory Decision Accumulation in
Multiple Targets Scenarios
When observing the ship navigation situation presented in
Fig. 12, it is easy to see that the presence of multiple moving
obstacles can generate catastrophic solutions on a navigation/
steering system since the rules applied in the depicted situation
would be if the target vessel is in region I, then steer to
starboard and if the target vessel is in region IV, then steer to
port. The accumulation of these two contradictory decisions
would once again lead to a no action decision, and the own
vessel would eventually crash into the target vessel in region I.
This situation is categorized as a failure of contradictory deci-
sion accumulation due to multiple stationary or moving targets
and can occur in single-level fuzzy-rule-based navigation
systems.
Solution: Multilevel Decision/Action Formulations: The
decision-making process in single-level fuzzy-rule-based
systems should be associated with a secondary-level decision
process that can automatically overcome rule inference failures
due to multiple stationary or moving targets. In this approach,
one proposes the use of the SAF module (see Fig. 2) as a
secondary-level process whose role is to serialize parallel de-
cisions taken on the PDM module when in the presence of mul-
tiple moving target vessels. This solution is further discussed
with respect to the Bayesian-network-based SAF module of the
CAS in ship navigation in Section VI.
VI. SOLUTIONS TO THE FAILURES
AND ITS LIMITATIONS
Although the three rule inference failures of the FMFs
discussed earlier might rarely occur in ocean navigation, they
cannot be neglected since they usually result in catastrophic
failures. The proposed solutions to overcome fuzzy rule infer-
ence failures in the decision-making process of ocean naviga-
tion are further discussed in the following.
PERERA et al.: SOLUTIONS TO FAILURES AND LIMITATIONS OF MAMDANI FUZZY INFERENCE 1547
Fig. 12. Fuzzy-rule inference failure due to Situation III.
A. Insertion of Fuzzy Smooth Bearing Regions
Consider the mathematical formulation of a two-vessel col-
lision in the ocean navigation situation presented in Fig. 3.
The gure shows the collision avoidance decisions, i.e.,
course change decisions D
i
(k) and speed change decisions
D
V
i
(k), for high-risk collision situations for fuzzy bearing
regions I to X (see Tables I and II), and for the ranges R
vd
to
R
b
, and R
b
to R
a
in each fuzzy bearing region. The decisions
on fuzzy bearing region I are formulated as course to starboard
(
o
> 0) and no speed change (V
o
= 0). In region II, the de-
cisions are course to starboard (
o
> 0) and decrease speed
(V
o
< 0). Therefore, there are no contradictory decisions in
the intersection of both regions.
In region IV, the decisions are course to port (
o
< 0)
and decrease speed (V
o
< 0). Region III was introduced as
a smooth transition region between regions II an IV with
decisions no course change (
o
= 0) and decrease speed
(V
o
< 0). The region was made proper through the use of
Fig. 3, as further discussed in following. In region V, range
R
vd
to R
b
, the decisions are course to port (
o
< 0) and no
speed change (V
o
= 0), and in the range of R
b
to R
a
, the
decisions are no course change (
o
= 0) and no speed change
(V
o
= 0). With the insertion of fuzzy smooth region III, all the
collision avoidance decision transitions in the right half of the
decision space are smooth, and no contradictory decisions exist.
A similar approach was followed in the left half of the decision
space, where regions VI and X were introduced as smoothing
regions.
B. Determination of Proper Sizing of Fuzzy Smooth
Bearing Regions
One should note that this type of fuzzy inference failure is
only observed in the left half of the decision-making process in
own vessel regions XI, X and I, and is due to the fuzzy rules
that were previously dened in accordance with the COLREGs
rules and regulations. As shown in Fig. 3, line O(k)B(k)
separates regions I and X, and line O(k)D(k) separates regions
X and IX. The line O(k)D(k) intercepts the own vessel domain
at point C(k). There is a collision risk if any relative trajectories
of the target vessel starting from region I converge into the own
vessel domain. The straight line B(k)E(k) represents the left-
end relative navigation trajectory starting from region I that
neither converges nor diverges from the own vessel domain.
Hence, all the relative navigational trajectories of the target
vessel in the region I with collision risk stay on the left side
of line B(k)E(k). The line B(k)E(k) intercepts the own vessel
domain at point E(k). Hence, that should be used as a guideline
regarding whether the collision risk increases or decreases in a
collision situation in region I.
To ensure that the relative trajectories starting from region I
never enter region IX (preventing two contradictory decisions
in the same obstacle avoidance process), the line B(k)E(k)
should never cross region IX in the R
b
R
a
range. This is
achieved if point C(k) is always to the left of point E(k).
Hence, the minimum size of region X should be given by
C(k) E(k). These conditions can be formulated into geo-
metrical relationships among regions I, IX, and X. Considering
the triangle of O(k)B(k)E(k) with C(k) E(k), the sine rule
can be written as
R
a
sin(180

7
)
=
R
vd
sin(
7
)
. (1)
Since E(k)B(k) is a straight line parallel to the Y
o
axis and
angle
1
is symmetric around the Y
o
axis, the angle condition

7
=
1
/2 and (1) becomes

6
= 180 sin
1

R
a
R
vd
sin(
1
/2)

1
2
. (2)
Hence, the minimum size requirement for a proper smooth
transition region X in Fig. 3 is related to the size of the
contradictory decision regions I and IX that motivated the
creation of smooth region X. One should once again note that
if the region is smaller than the derived size, system decisions
can jump between contradictory decisions and/or two contra-
dictory decisions may be implemented on the same navigational
trajectory.
1548 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 4, MAY 2014
Fig. 13. Course and speed collision avoidance actions.
C. Multilevel Decision/Action Formulations in
Ocean Navigation
In the proposed system, the SAF module (see Fig. 2) is
proposed as a second-level process that overcomes failures
due to contradictory decision accumulation in multiple obstacle
scenarios by the PDM module. The main objective of the
SAF module is to transform the parallel collision avoidance
decisions that are generated by the PDM module into sequential
actions that can be executed in the own vessel navigation system
while eliminating the mentioned failures. This can be achieved
by collecting from the PDM module the multiple collision
avoidance decisions, D
i
(k) (D
i
(k), D
V
i
(k)), and eval-
uating them using the time until the collision situation T
i
(k)
fromthe CRAmodule regarding each target vessel. Final results
are arranged as a sequential formation of actions A
i
(k)
(A
i
(k), A
V
i
(k)) involving the course and speed actions at
given time instants T
i
(k) (T
i
(k), T
V
i
(k)). Fig. 13 gives
an example of the accumulated process of sequential course
and speed action execution. D
i
(k) and D
V
i
(k), and A
i
(k)
and A
V
i
(k) represent the course and speed change deci-
sions and actions, respectively. This approach eliminates the
cancelation of contradictory decisions due to decision/action
accumulation.
The SAF module consists of a continuous Bayesian network
that is formulated to update the parallel collision avoidance
decisions into sequential actions that will execute at appropriate
time instants (see Fig. 2). As presented in the gure, the SAF
module consists of four nodes: 1) collision time estimation,
2) collision risk, 3) action delay, and 4) collision avoidance
actions. The inputs of the SAF module are the collision
decisions D
i
(k) and time until the collision situation T
i
(k)
which generated, respectively, by the PDM and CRA modules
(see Section III-B).
The main objective of the time until collision estimation
node is to estimate the time until the collision situation T
i
(k)
between the own vessel and each of the target vessels. The
node collision risk estimation inferences the collision risk with
respect to each target vessel considering the collision time
estimation. The action delay node is designed to formulate the
appropriate time to take collision avoidance actions. The action
delay node, the collision risk estimation node, and the collision
avoidance decisions are used to infer the collision avoidance
action formulation node. The mathematical derivation of the
Fig. 14. Relative trajectories for CPRB 270

to 360

.
Bayesian-network-based SAF module, inferences and their as-
sociated functions are further described in [35].
VII. COMPUTATIONAL SIMULATIONS
The CAS is implemented on the MATLAB software plat-
form. The following values are assigned for the FMF param-
eters of the PDM module.
Range FMF (see Fig. 4): R
vd
1 nmi, R
b
6 nmi, R
A

10 nmi.
Bearing FMF (see Fig. 5):
1
10

,
2
80

,
3
10

4
80

,
5
26

, and
6
26

.
Relative course FMF values (see Fig. 6):
1
5

,
2
5

,
and
3
5

.
Speed ratio FMF (see Fig. 7): f
1
0.8,
2
1.2, and

3
5.
Course change output FMFs (see Fig. 8):
1
10

, and

2
40

.
Speed change output FMF (see Fig. 9):
1
2, and

2
10.
The FMF parameters were derived considering various colli-
sion avoidance situations under the simulated conditions.
A. Proper Fuzzy Smooth Region Insertion Simulations
Figs. 1416 regard a two-vessel collision situation and
present the simulations of the relative trajectories of the target
PERERA et al.: SOLUTIONS TO FAILURES AND LIMITATIONS OF MAMDANI FUZZY INFERENCE 1549
Fig. 15. Relative trajectories for CPRB 180

to 270

.
Fig. 16. Relative trajectories for CPRB 0

to 180

.
Fig. 17. Zoomed view of the relative trajectories for CPRB 270

to 360

.
Fig. 18. Zoomed view of the relative trajectories for CPRB 180

to 270

.
vessel with the CPRB (see Fig. 1)
ci,o
(k), varying from
270

to 360

, 180

to 270

, and 0

to 180

, in 0.1

division
steps. Figs. 1719 represent the zoomed view of the relative
trajectories of the target vessel around the own vessel initial
position with CPRB
ci,o
(k), varying from 270

to 360

, 180

to 270

, and 0

to 180

, also with 0.1

steps, respectively.
The vessel initial speed condition is V
o
/V
i
= 0.5, and the
initial own vessel course and speed are
o
= 0

and V
o
=
12 knots, respectively. The own vessel is initially located at
O(k) = (0 nmi, 0 nmi). The target vessel start positions are
located around the collision point C
i
(k) = (0 nmi, 5 nmi) (see
Fig. 1), with the target vessel startup distance to the collision
point R
ci
(k) = 10 nmi. Constant speed and course conditions
are assumed for the target vessel. Before the insertion of the
fuzzy smooth regions, fuzzy rule inference failures occurred in-
between relative bearing range of 180

to 270

and from 270

to 360

, where the two contradictory decisions were intercepted

and the relative trajectory of the target vessel intercepts the own
vessel at initial position (0 nmi, 0 nmi).
1550 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 4, MAY 2014
Fig. 19. Zoomed view of the relative trajectories for CPRB 0

to 180

.
Fig. 20. Distance to the minimum point approach (CPRB 0

to 90

).
Fig. 21. Distance to the minimum point approach (CPRB 90

to 180

).
The relative minimum distance between the two vessels with
the CPRB,
ci,o
(k), varying from 0

to 90

, 90

to 180

,
180

to 270

, and 270

to 360

, with respect to the own

Fig. 22. Distance to the minimum point approach (CPRB 180

to 270

).
Fig. 23. Distance to the minimum point approach (CPRB 270

to 360

).
vessel, are shown in Figs. 2023. As observed from the gures,
two minimum distance points are noted around the CPRB,
i.e.,
ci,o
(k) = 175

to 180

, and 350

to 360

. However, the
minimum distance around the CPRB, i.e.,
ci,o
(k) = 175

to
180

, can be ignored since contradictory decision boundary

intersections are not present on this region.
Hence, the minimum distance around the CPRB, i.e.,

ci,o
(k) = 350

and 360

, should be further analyzed to observe

any fuzzy rule inference failures in the CAS. The main objective
of this analysis of the minimum distance around the CPRB,
i.e.,
ci,o
(k) = 350

and 360

, is to see whether any fuzzy

rule failure points are hiding around the region. This can be
observed by further zooming the view around the own vessel
initial points and observing the pattern of the relative navigation
trajectory of the target vessel.
The zoomed view of the target vessel relative trajectories
with respect to the CPRB, i.e.,
ci,o
(k) = 350

to 360

, around
the target vessel trajectories startup are presented in Fig. 24,
with the respective trajectories marked as a
1
, a
2
, a
3
, b
1
, b
2
, c
1
,
c
2
, d
1
, d
2
, d
3
, and d
4
. Similarly, the zoomed view of the target
vessel relative trajectories with respect to the CPRB,
ci,o
(k) =
350

to 360

around the own vessel startup point are presented

in Fig. 25, with the respective relative trajectories marked as
a
1
, a
2
, a
3
, b
1
, b
2
, c
1
, c
2
, d
1
, d
2
, d
3
, and d
4
. As presented in
the gure, the own vessel initial position is bounded by the
PERERA et al.: SOLUTIONS TO FAILURES AND LIMITATIONS OF MAMDANI FUZZY INFERENCE 1551
Fig. 24. Zoomed viewof the relative trajectories (CPRB350

to 360

) around
the Target vessel initial position.
Fig. 25. Zoomed viewof the relative trajectories (CPRB350

to 360

) around
the Own vessel initial position.
relative trajectories of b
1
and b
2
that correspond to the minimum
distance between both vessels.
The relative trajectories of c
1
and c
2
are also bounded by
the relative trajectories of b
1
and b
2
, as observed in Fig. 24.
However, relative trajectories c
1
and c
2
shift away from the own
vessel region limited by the relative trajectories of b
1
and b
2
.
Further, the relative trajectories near the trajectory of b
1
, i.e.,
a
1
, a
2
, a
3
, and a
4
, and b
2
, i.e., d
1
, d
2
, d
3
, and d
4
, are also
observed, and all the trajectories shift away from the own vessel
initial position bounded by the relative trajectories of b
1
and b
2
.
Therefore, none of the relative trajectories generated around the
CPRB, i.e.,
ci,o
(k) = 350

to 360

, will converge into the own

vessel initial position that is bounded by the relative trajectories
of b
1
and b
2
. Hence, one can conclude that the system is free
from fuzzy inference failure points.
As presented in the gures, with the introduction of the
proper fuzzy smooth regions, the relative trajectory of the target
vessel no longer intercepts the own vessel. Instead, it shifts
its trajectory from one side of the origin to the other side.
The minimum distance between vessels becomes always above
0.02 nmi. One should note that this distance could be changed
by varying the FMF parameters of the PDM module.
B. Simulation of Overall System Performance
The computational simulations for the overall system, i.e.,
the integrated PDM and SAF modules, on a multivessel colli-
sion situation are presented in Figs. 2632. It is assumed that
the target vessels are moving in constant speed and course and
Fig. 26. Simulations of a multivessel collision situation.
Fig. 27. Simulations of a multivessel collision situation.
Fig. 28. Simulations of a multivessel collision situation.
1552 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 4, MAY 2014
Fig. 29. Simulations of a multivessel collision situation.
Fig. 30. Simulations of a multivessel collision situation.
do not honor any navigational rules and regulations of the sea
to keep the consistency in the collision situation.
In Fig. 26, the own vessel starts navigation from the origin
(0 m, 0 m) and the rst, second and third target vessels start
from positions (1475 m, 1400 m), (9500 m, 1000 m), and
(9220 m, 6500 m), respectively. All startup and nal positions
of the own and target vessels are represented by vessel shape
icons. The CRA is formulated by a Gaussian distribution and
presented in the x = 8000 m axis. Similarly, the collision
avoidance actions for the course and speed changes formulated
by the Gaussian distributions are presented in the x = 6000 m
and x = 4000 m axis, respectively. The scaled time axis
(actual time 5s) is presented in the y-axis, and the scaled
collision risk (%), course actions (%)and speed actions (%) are
presented in the x-axis.
In Fig. 26, the system has observed one possible collision
situation, and the respective collision avoidance actions of no
course change and speed reduction, which are taken to
Fig. 31. Simulations of a multivessel collision situation.
Fig. 32. Simulations of a multivessel collision situation.
avoid the rst target vessel, are presented. Fig. 27 shows the
completion of the rst action segment of collision avoidance.
In the same gure, the second target vessel is detected, and
the respective collision risk and collision avoidance actions of
course to starboard and speed reduction to avoid the second
target vessel are presented. In Fig. 28, the third target vessel
is detected, and the respective collision avoidance actions of
course to the port and no speed change to avoid the third target
vessel are presented. One should note that, in the same gure,
both vessels (i.e., own vessel and rst target vessel) are moving
in a close encounter situation without any collision risk.
Fig. 29 shows the subcompletion of the second action seg-
ment of the collision avoidance actions, consisting in speed
reduction and continuation of course change to starboard
side by the own vessel. In Fig. 30, the own vessel is about
to safely pass the rst target vessel and the completion of the
second action segment of the collision avoidance actions. In
Fig. 31, the own vessel is about to safely pass the second target
vessel and the completion of the third action segment of the
PERERA et al.: SOLUTIONS TO FAILURES AND LIMITATIONS OF MAMDANI FUZZY INFERENCE 1553
collision avoidance. Finally, the completion of all the collision
avoidance actions without collision risk and safe passing of
the second and third target vessel trajectories are presented in
Fig. 32. The own vessel speed, course, position and time values
are shown on top of the respective gures.
VIII. CONCLUSION AND FUTURE WORK
Fuzzy-logic-based systems have been implemented in sev-
eral industrial applications. However, the FMF failures and
limitations in those systems have not been properly identied
in the recent literature. Therefore, this paper has shown how
Mamdani-type fuzzy inference failures observed in ship naviga-
tion can be avoided and what the limitations are of the proposed
solutions. The overall proposed system consists of a fuzzy-
logic-based PDM module and a Bayesian-network-based SAF
module that improves the decision-making process of a CAS
based on COLREGs rules and human expert knowledge in ship
navigation.
As was shown in the presented computational simulations,
successful results were obtained when applying the proposed
solutions of inserting a smooth transition region between
intersected contradictory decision inference boundaries, de-
termining the proper size limitations between nonintersected-
contradictory-decision inference regions and introducing a
secondary-level decision/action formulation module to over-
come single-level rule inference failures in the FMFs.
Although the results are promising, as future work it is
expected that improvements can be obtained by optimizing
the parameters in the FMFs. These parameters are somewhat
related to the navigational characteristics of the own vessel as
well as the navigational sensor capabilities. Therefore, further
studies should be developed to improve the identication of the
parameters of the FMFs. Furthermore, an experimental imple-
mentation of the CAS is shown in [37] and [38], and some pre-
liminary experimental results with the respective videos of the
ship maneuvers can be found at www.youtube.com/thecentec.
To conclude, one should note that, although the proposed
solutions were developed with collision avoidance in ocean
navigation in mind, they can be applied on any Mamdani-type
fuzzy-logic-based navigation system that suffers from similar
rule inference failures.
REFERENCES
[1] L. A. Zadeh, A new direction in AItoward a computational theory of
perceptions, AI Magazine, vol. 22, no. 1, pp. 7384, Jan. 2001.
[2] Umm-e-Habiba and S. Asghar, A survey on multi-criteria decision mak-
ing approaches, in Proc. IEEE Int. Conf. Emerging Technol., 2009,
pp. 321325.
[3] R. C. Kwok, J. Ma, and D. Zhou, Improving group decision making: A
fuzzy logic approach, IEEE Trans. Syst., Man, Cybern. C, Appl. Rev.,
vol. 32, no. 1, pp. 5463, Feb. 2002.
[4] P. Khatun, C. M. Bingham, N. Schoeld, and P. H. Mellor, Application
of fuzzy control algorithms for electric vehicle antilock braking/traction
control systems, IEEE Trans. Veh. Technol., vol. 52, no. 5, pp. 1356
1364, Sep. 2003.
[5] C. Y. Jung, H. Y. Hwang, D. K. Sung, and G. U. Hwang, Enhanced
Markov chain model and throughput analysis of the slotted CSMA/CA
for IEEE 802.15.4 under unsaturated trafc conditions, IEEE Trans. Veh.
Technol., vol. 58, no. 1, pp. 473478, Jan. 2009.
[6] L. Xiang, A new enterprise customer and supplier cooperative system
framework based on multiple criteria decision-making, in Proc. IEEE
Comput. Soc. 3rd Int. Conf. Syst., 2008, pp. 300305.
[7] M. L. Smith, Sensors, appliance control and fuzzy logic, IEEE Trans.
Ind. Appl., vol. 30, no. 2, pp. 305310, Mar./Apr. 1994.
[8] T. L. Hardy, Multi-objective decision-making under uncertainty:
Fuzzy logic method, National Aeronautics and Space Administration,
Washington, DC, USA, NASA Tech. Memo. 106796Comput. Aerosp.
10 Meet., Mar. 1995.
[9] H. J. Rommelfanger, Multicriteria decision making using fuzzy logic,
in Proc. Conf. NAFIPS, Aug. 1998, pp. 360364.
[10] T. Ozen, J. M. Garibaldi, and S. Musikasuwan, Modeling the variation in
human decision making, in Proc. IEEE Annu. Meet. NAFIPS, Jun. 2004,
vol. 2, pp. 617622.
[11] Y. Chen and S. Yasunobu, Fuzzy target based soft decisions for mo-
bile vehicle in dynamic environment, in Proc. 7th World Congr. Intell.
Control Autom., Chongquig, China, Jun. 2008, pp. 800805.
[12] R. A. Hogle and P. P. Bonissone, A fuzzy algorithm for path selection
in autonomous vehicle navigation, in Proc. 23rd Conf. Decision Control,
Las Vegas, NV, USA, Dec. 2004, pp. 898900.
[13] H. Seraji and A. Howard, Behavior-based robot navigation on challeng-
ing terrain: A fuzzy logic approach, IEEE Trans. Robot. Autom., vol. 18,
no. 3, pp. 308321, Jun. 2002.
[14] H. Seraji and B. Bon, Autonomous navigation of planetary rovers: A
fuzzy logic approach, Jet Propulsion Lab., Pasadena, CA, USA, JPL
Internal Document, 1998.
[15] A. Fatmi, A. A. Yahmadi, L. Khriji, and N. Masmoudi, A fuzzy logic
based navigation of a mobile robot, in Proc. 22nd World Academy Sci.,
Eng. Technol., 2006, pp. 169174.
[16] T. E. Mora and E. N. Sanchez, Fuzzy logic-based real-time navigation
controller for a mobile robot, in Proc. IEEE/RSJ Int. Conf. Intell. Robot
Syst., Victoria, BC, Canada, Oct. 1998, pp. 612617.
[17] I. N. D. Silva, F. A. C. Gomide, and W. C. D. Amaral, Navigation
of mobile robots using fuzzy logic controllers, in Proc. AMC IEEE,
Coimbra, Portugal, 1998, pp. 346349.
[18] A. Babalou and N. Seipour, Application of fuzzy decisions making in
mobile robot navigation in dynamic environments, in Proc. Fuzz-IEEE,
Jeju, Korea, Aug. 2009, pp. 877881.
[19] S. M. Raguraman, D. Tamilselvi, and N. Shivakumar, Mobile robot nav-
igation using fuzzy logic controller, in Proc. Int. Conf. Control, Autom.,
Commun. Energy Conserv., Jun. 2009, pp. 15.
[20] M. A. Hersh, Sustainable decision making: The role of decision support
systems, IEEE Trans. Syst., Man, Cybern. C, Appl. Rev., vol. 29, no. 3,
pp. 395407, Aug. 1999.
[21] S. Thongchi and K. Kawamura, Application of fuzzy control to a sonar
based obstacle avoidance mobile robot, in Proc. IEEE Int. Conf. Control
Appl., Anchorage, AK, USA, Sep. 2000, pp. 425430.
[22] T. I. Fossen, Ed., Recent developments in Ship Control Systems Design.
London, U.K.: Sterling, 1999, ser. World Superyacht Review.
[23] K. Ohtsu, Recent development on analysis and control of ships
motions, in Proc. IEEE Int. Conf. Control Appl., Aug. 1999,
pp. 10961103.
[24] L. Moreira, T. I. Fossen, and C. Guedes Soares, Path following control
system for a tanker ship model, Ocean Eng., vol. 34, no. 14/15, pp. 2074
2085, Oct. 2007.
[25] K. D. Do and J. Pan, Robust path-following of underactuated ships:
Theory and experiments on a model ship, Ocean Eng., vol. 33, no. 10,
pp. 13541372, Jul. 2006.
[26] T. Statheros, G. Howells, and K. McDonald-Maier, Autonomous ship
collision avoidance navigation concepts, technologies and techniques,
J. Navigat., vol. 61, no. 1, pp. 129142, Jan. 2008.
[27] Int. Maritime Org., Convention on the international regulations for pre-
venting collisions at sea (COLREGs), 1972. [Online]. Available: http://
www.imo.org/
[28] A. N. Cockcroft and J. N. F. Lameijer, A Guide to The Collision Avoidance
Rules. Burlington, MA, USA: Elsevier, 2001.
[29] K. Hasegawa, Automatic collision avoidance system for ship using fuzzy
control, in Proc. 8th Ship Control Syst. Symp., 1987, pp. 234258.
[30] J. Zhao, M. Tan, W. G. Price, and P. A. Wilson, DCPA simulation model
for automatic collision avoidance decision making systems using fuzzy
sets, in Proc. OCEANS, Brest, France, 1994, vol. 2, pp. 244249.
[31] S. Lee, K. Kwon, and J. Joh, A fuzzy logic for autonomous navigation of
marine vehicles satisfying colreg guidelines, Int. J. Control Autom. Syst.,
vol. 2, no. 2, pp. 171181, Jun. 2004.
[32] L. P. Perera, P. Oliveira, and C. Guedes Soares, Vessel detection, track-
ing, state estimation and navigation trajectory prediction for the vessel
1554 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 63, NO. 4, MAY 2014
trajectory monitoring and information proces, IEEE Trans. Intell. Transp.
Syst., vol. 13, no. 3, pp. 11881200, Sep. 2012.
[33] L. P. Perera and C. Guedes Soares, Detections of potential collision
situations by relative motions of vessels under parameter uncertain-
ties, in Sustainable Maritime Transportation Exploitation Sea Resources,
E. Rizzuto and C. Guedes Soares, Eds. London, U.K.: Taylor & Francis,
2012, pp. 705713.
[34] L. P. Perera and C. Guedes Soares, Vector-product based Collision Esti-
mation and Detection in e-Navigation, in Proc. 9th IFAC Conf. MCMC,
Arenzano, Italy, 2012.
[35] L. P. Perera, J. P. Carvalho, and C. Guedes Soares, Intelligent ocean nav-
igation & Fuzzy-Bayesian decision-action formulation, IEEE J. Ocean.
Eng., vol. 37, no. 2, pp. 204219, Apr. 2012.
[36] L. P. Perera, J. P. Carvalho, and C. Guedes Soares, Fuzzy-logic based
decision making system for collision avoidance of ocean navigation un-
der critical collision conditions, J. Marine Sci. Technol., vol. 16, no. 1,
pp. 8499, Mar. 2011.
[37] L. P. Perera, L. Moreira, F. P. Santos, V. Ferrari, S. Sutulo, and C. Guedes
Soares, A navigation and control platform for real-time manoeuvring of
autonomous ship Models, in Proc. 9th IFAC Conf. MCMC, Arenzano,
Italy, 2012.
[38] L. P. Perera, V. Ferrari, F. P. Santos, M. A. Hinostroza, and C. Guedes
Soares, Experimental results in collisions avoidance of ship manoeu-
vres, presented at the 32nd Int. International Conference on Ocean,
Offshore and Arctic Engineering (OMAE), Nantes, France, Jun. 2013
OMAE2013-11265.
Lokukaluge P. Perera received the B.Sc. and M.Sc.
degrees in mechanical engineering from Oklahoma
State University, Stillwater, OK, USA, in 1999 and
2001, respectively, and the Ph.D. degree in naval ar-
chitecture and marine engineering from the Techni-
cal University of Lisbon, Lisbon, Portugal, in 2012.
He is currently with the Center for Marine Tech-
nology and Engineering, Instituto Superior Tcnico,
University of Lisbon.
Dr. Perera received Doctoral and Postdoctoral
Fellowships from the Foundation for Science and
Technology of Portugal in 2008 and 2012, respectively.
J. P. Carvalho (M12) received the Licentiate,
M.Sc., and Ph.D. degrees from the University of
Lisbon, Lisbon, Portugal, in 1992, 1996, and 2002, re-
spectively, all in electrical and computer engineering.
Since 1991, he has been a Researcher with the
Institute of Systems and Computer Engineering
Research and Development in Lisbon, Instituto Su-
perior Tcnico, University of Lisbon, where until
2008, he was with the Soft Computing Group. He is
currently a Senior Researcher with the Spoken Lan-
guage Systems Laboratory, Institute of Systems and
Computer EngineeringResearch and Development, Lisbon. Since 1998, he has
taught courses on computational intelligence, distributed systems, computer
architectures, and digital circuits. He is currently an Assistant Professor with
the Department of Electrical Engineering and Computation, Instituto Superior
Tcnico, University of Lisbon. He is the co-author of over 70 papers in inter-
national scientic journals, book chapters, and peer-reviewed conferences. His
main research interests include the application of computational intelligence
techniques to solve problems in noncomputing-related areas.
C. Guedes Soares received the M.S. degree in ocean
engineering from Massachusetts Institute of Tech-
nology, Cambridge, MA, USA, in 1976; the Ph.D.
degree from the Norwegian Institute of Science and
Technology, Trondheim, Norway, in 1984; and the
Doctor of Science degree from the Technical Univer-
sity of Lisbon, Lisbon, Portugal, in 1991.
He is currently a Professor of naval architecture
and marine engineering and the President of the
Center for Marine Technology and Engineering, Uni-
versity of Lisbon, which is recognized and funded by
the Portuguese Foundation for Science and Technology.