Ocean Engineering 165 (2018) 528–537

Contents lists available at ScienceDirect

Ocean Engineering
journal homepage: www.elsevier.com/locate/oceaneng

On the use of Artificial Neural Networks for the calm water resistance T
prediction of MARAD Systematic Series’ hullforms
Vasiliki Margari, Aphrodite Kanellopoulou, George Zaraphonitis∗
NTUA (National Technical University of Athens), Greece

A R T I C LE I N FO A B S T R A C T

Keywords: The present study investigates the use of Artificial Neural Networks (ANNs) for the resistance prediction of
Artificial Neural Networks hullforms designed according to the MARAD Systematic Series. This series comprises 16 full hullforms, speci-
Multi-layer perceptron fically designed for use as bulk carriers and tankers. Experimental data for the residual resistance coefficient of
MARAD Systematic Series these hulls provided by MARAD in a series of diagrams have been used to train and evaluate a series of neural
Resistance prediction
networks aiming to estimate the residual resistance coefficient of ships designed according to the MARAD Series.
Calm water resistance
The adopted procedure along with the obtained results are presented and discussed.

1. Introduction hullforms, based on the resistance data provided by the MARAD Sys-
tematic Series (Margari, 2017; Margari et al., 2018).
The prediction of a hull's resistance is a problem of great importance Artificial neural networks have recently found application across
for the ship designer, closely related to the design and optimisation of many scientific fields, such as pattern recognition, classification or
the hullform and propeller, the selection of the main engine, the ship's function approximation. They consist of processors (neurons), which
environmental impact and the fuel cost. Resistance predictions are communicate to each other with signals through weighted connections,
traditionally based on tank testing, or nowadays, on software tools mimicking the structure of a biological neural system. Trained ANNs
applying Computational Fluid Dynamics (CFD). However, in the pre- are considered able to provide desired output from a set of input
liminary design stage, resistance predictions are quite often carried out parameters without the need for an exact function or model for the
based on relevant data from systematic series, such as the Series 60 problem, even if the data are noisy. They also offer a number of ad-
Methodical Series of Single-Screw Ships (Todd, 1963), the SSPA Cargo vantages, including: sufficient accuracy of results, flexibility in im-
Liner Series (Williams, 1969), the BSRA Methodical Series (Moor et al., plementing, ability to implicitly detect complex non-linear relation-
1961), or the MARAD Systematic Series of full-form ship models ships between independent variables. These features make ANNs rightly
(Roseman and Seibold, 1987, 1985). The data from the systematic suited for application in a wide range of engineering problems.
series are usually presented in graphical or tabular form, enabling the Examples from the use of ANNs for the prediction of resistance,
manual calculation of a ship's resistance following a relatively simple propulsion, manoeuvering and seakeeping characteristics of ship are
and straightforward procedure. Such a procedure might be perfectly presented in several studies (Cepowski, 2007, 2005; Koushan, 2001;
suitable for the estimation of the resistance curve of one particular Martins and Lobo, 2007; Mason et al., 2005; Ortigosa et al., 2014;
hullform, but when it comes to the systematic optimisation of a ship Pedersen and Larsen, 2009; Yao and Han, 2012).
design, manual calculation procedures are not efficient any more. In There are numerous types of neural networks. For the present study
such cases programmable calculation procedures would be required, the so-called multi-layer perceptron (MLP) networks have been selected
enabling the evaluation of a large number of alternative designs with and systematically trained and evaluated in order to assess their po-
minimal computing time. Regression techniques, utilizing polynomial tential to estimate the resistance of MARAD-type hullforms. Selected
interpolation of the data provided by the systematic series have been alternatives of MLPs and their application for prediction of MARAD
developed by several authors (see e.g. Bojovic, 1997; Radojcic et al., hullforms’ resistance are provided.
2014; Shaher Sabit, 1976, 1972a, 1972b, 1971). The present work in-
vestigates a different approach to this problem, i.e. the use of Artificial
Neural Networks (ANNs) as a tool for the prediction of resistance of full

∗
Corresponding author.
E-mail addresses: [email protected] (V. Margari), [email protected] (A. Kanellopoulou), [email protected] (G. Zaraphonitis).

https://doi.org/10.1016/j.oceaneng.2018.07.035
Received 23 May 2018; Accepted 11 July 2018
Available online 23 July 2018
0029-8018/ © 2018 Elsevier Ltd. All rights reserved.
V. Margari et al. Ocean Engineering 165 (2018) 528–537

2. Artificial neural networks every node is connected with every node of its forward layer or not.
A fundamental feature of ANNs is their ability to learn. For this
2.1. Historical background purpose, they adjust the weights of their connections and threshold
values in order to present a satisfactory behaviour according to a pre-
Artificial neural networks (ANNs) are inspired by the human brain scribed criterion. There are three different learning techniques: su-
functionality and loosely model the way it processes sensory informa- pervised, unsupervised and reinforcement learning. In supervised
tion received from the environment. They consist of parallel-dis- learning, the network is provided with a series of input and output pairs
tributed, interconnected, non-linear processing units, the neurons. in order to find a function that connects them. This procedure is eval-
Information is acquired from their environment via a learning process uated by a cost function, usually the mean square error between actual
and stored in the form of weights in the connections (Haykin, 1999). A and desired output, which needs to be minimized. In unsupervised
simplified model of the biological neuron was initially introduced by learning, the network has to classify the input data, finding a patern
McCulloch and Pitts (1943), as a possible component of a computa- without external information. In reinforcement learning the network is
tional system. This neuron had several inputs, both excitatory and in- in continuous communication with its environment, which rewards or
hibitory,1 and an output, which reflected the state2 of the neuron. In not the input – output mapping that has been achieved, in order to
order for the neuron to be activated, the sum of the excitatory signals minimize a performance indicator.
had to exceed a threshold value T, while inhibitory signals were absent.
In variations of this model, positive and negative stimulations were of 2.3. Feed-forward multi-layer perceptron
the same importance due to the use of adequate linear weights (Haykin,
1999; Russell and Norvig, 1995). One of the most widely used neural networks is the feed-forward,
There were numerous milestones in the field of neural networks multi-layer perceptron (MLP). MLPs are considered particularly sui-
since they were originally introduced, regarding the architecture of the table for function approximation, data classification or pattern re-
networks and their learning process. Hebbian learning is one of the cognition problems; therefore they have been selected to be used in the
oldest learning algorithms, based on the dynamics of biological systems present study. An MLP consists of an input layer, one or more hidden
(Hebb, 1949). According to this hypothesis, when the neurons con- layers and an output layer (Fig. 1). The number of the input and output
nected by a synapse are stimulated repeatedly and simultaneously, this nodes is equal to the dimension of the input and output data respec-
synapse becomes stronger, and the neurons are connected more effec- tively. The signal is transferred in the forward direction from the input
tively. In terms of ANNs, this is accomplished with the increase of sy- to the output layer (Haykin, 1999).
naptic weights (Haykin, 1999; Russell and Norvig, 1995). One of the Every node except for the input layers’ is a neuron with a differ-
first ANNs to be developed was the perceptron, presented by Rosenblatt entiable activation function that generates its output (Haykin, 1999).
(1958). This simple neural network was a binary classifier, based on the Sigmoid transfer functions, such as logistic sigmoid or hyperbolic tan-
McCulloch – Pitts neuron, and had the ability of updating its weights gent sigmoid are usually preferred as activation functions for MLPs,
via a supervised learning method. Minsky and Papert (1969) analysed it because consecutive layers of neurons with non-linear activation
and proved mathematically that a single-layer network is not able to functions enable the network to learn non-linear relationships between
solve more complex problems, such as linear classification in more than input and output data. The linear transfer function is often used for the
two classes. Back-propagation learning algorithm revived the interest neurons of the output layer.
for neural networks in the mid-80s, when it was used in numerous The logistic sigmoid function is given by eq. (1) while its graphical
different problems. This algorithm adapts the synaptic weights in a way representation is illustrated in Fig. 2.
that minimizes the mean square error between the actual and desired
1
output (Haykin, 1999; Rojas, 1996). It was at that time when many yk = f (uk ) =
1 + e−uk (1)
research programs regarding neural networks were initiated. As a re-
sult, their application expanded in more complex and practical pro- The hyperbolic tangent sigmoid function is given by eq. (2) and its
blems. Autonomous drivers, weather or stock value predictions, image graphical representation is illustrated in Fig. 3.
recognition and optimal path selection could be solved with ANNs
1 − e−uk
(Zurada, 1992). yk = f (uk ) =
1 + e−uk (2)

In the above equations, yk is the output of the kth neuron and uk is

2.2. Artificial neural network architectures
the weighted sum of its input synapses.
In MLPs, a popular supervised learning method capable of handling
An artificial neural network consists of a distribution of inter-
complicated learning problems is the back-propagation algorithm (BP),
connected neurons, arranged in layers. Its architecture is defined by the
a gradient descent technique with backward error propagation. This
number of its layers, the kind of neural synapses and the learning al-
algorithm is based on the error correction learning rule and is executed
gorithm used. According to the first characteristic, they may be dis-
in two stages: In the first stage, the input signal is transferred through
tinguished in single-layer and multi-layer networks. The first type of
the layers, in order for the output signal to be derived. During this
networks contains only one layer of neurons as its output layer, while
the later has ‘hidden’ layers between the input and the output layer.
With respect to the kind of their neural synapses, networks may be
characterized as feed-forward and recurrent. The data in feed-forward
networks flow in one direction, from the input layer to the output,
through the hidden layers, if existent. Recurrent networks contain at
least one feedback interconnection, which links the output of a neuron
with the input of another, placed at the same or previous layer. A
network can be also characterized as fully or partially connected if

1
Excitatory signals increase the probability of the neuron to be activated,
whereas the inhibitory ones, in this model, prevent its activation.
2 Fig. 1. Architecture of a multi-layer perceptron with two hidden layers.
i.e. The sum of the neuron's inputs.

529
V. Margari et al. Ocean Engineering 165 (2018) 528–537

2.3.1. Training algorithms

The training algorithm is used to train the neural network to re-
cognize a certain input and map it to the corresponding output. Back-
propagation algorithms can be distinguished into gradient descent,
Newton-Gauss, conjugate gradient and quasi-Newton algorithms. In this
study, three gradient descent algorithms and one quasi-Newton algo-
rithm were utilized:

2.3.1.1. Gradient descent algorithms. They are first order optimisation

algorithms that employ the gradient vector of the cost function to
decide the direction of the training, always pointing towards its
negative value (Snyman, 2005).
In the gradient descent back-propagation algorithm (also known as
steepest descent) the weights and biases are updated in the direction of
the negative gradient of the performance function.
The gradient descent with adaptive learning rate back-propagation al-
gorithm enables changes to the learning rate throughout the training. In
this way, the learning rate adapts to the local complexity of the error
surface. A higher learning rate is accepted, as far as the learning process
Fig. 2. Logistic sigmoid function. stability is ensured; otherwise the learning rate is decreased as neces-
sary (Hagan et al., 2014).
The gradient descent with momentum and adaptive learning rate back-
propagation combines adaptive learning with momentum training. It is
invoked in the same way as the gradient descent with adaptive learning,
with the addition of a momentum coefficient as training parameter,
introduced to help the algorithm escaping from a shallow local
minimum.

2.3.1.2. Quasi-Newton algorithms. Newton's algorithm is a second-order

optimisation algorithm that requires the calculation of a Hessian matrix
containing the second derivatives of the errors. Quasi-Newton
algorithms approximate the Hessian without the need of calculating
second derivatives (Hagan et al., 2014). For small or medium sized
networks, the Levenberg-Marquardt (Levenberg, 1944; Marquardt, 1963)
training algorithm is most frequently used. The update of the weights is
defined as follows:
wm + 1 = wm − (JT J + μI)−1JT e (5)
where J is the Jacobian matrix of the first derivatives of the network's
errors with respect to its weights and biases, e the vector of network's
errors, wm the weights' vector after the mth step and μ a positive
parameter (Beale et al., 2015).
Fig. 3. Hyperbolic tangent sigmoid function.
2.3.2. Learning algorithms
phase, the synaptic weights remain constant. Then, the mean square The learning algorithm defines the way that weights and bias of the
error of the deviation between the actual response of the network and network will be updated during the training. Two different learning
the desired output is calculated. In the second stage, these deviations algorithms were utilized in this study: the gradient descent weight and
are transferred backwards and the synaptic weights change accord- bias and the gradient descent with momentum weight and bias.
ingly, so that the average squared error is minimized. This cost function
is given by eq. (3): 2.3.3. Data sets
The available data for the network training are usually divided into
1 1 1
Eav =
T
∑ E (t ) = T
∑ 2 ∑ (dk (t )− yk (t ))2 three mutually exclusive sets: The training set is used for the update of
T T K (3) weights and biases of the network, in order for the cost function to be
minimized. The validation set is used for the tuning of parameters other
where K is the number of neurons in the output layer, T the number of
than weights and for comparison between different neural networks.
input-output patterns in the training set, dk(t) and yk(t) the target and
The test set is used for the evaluation of the network’s predictive ability.
actual output of the kth neuron (Haykin, 1999). A learning function is
The available data are randomly distributed into the three datasets.
utilized for the adjustment of the weights (e.g. the gradient descent
learning rule known as delta rule). The correction of the weights is
3. MARAD Systematic Series
given by eq. (4):

∂E (T ) The MARAD Systematic Series was developed by the U.S. Maritime

ΔWki = −η Administration (Roseman and Seibold, 1987, 1985), aiming to respond
∂wki (T ) (4)
to the rising interest for full hull form merchant ships and the lack of
where η is the learning rate parameter, which controls the rate of systematic data regarding the performance of these vessels, that could
adaptation of the weights and bias and affects the speed of convergence be used in the design process. As a result, this systematic series is
of the training process (Haykin, 1999). characterized by high block coefficient values (CB). Low length to beam

530
V. Margari et al. Ocean Engineering 165 (2018) 528–537

were obtained from these diagrams, characterized by five variables (Fn,

L/B, B/T, CB, CR).

4. Development of the neural networks

The development of the ANNs for the approximation of the re-

sistance of ships designed according to the MARAD series was carried
out using the MATLAB Neural Networks Toolbox (Beale et al., 2015;
The MathWorks Inc., 2018), which provides tools for designing, im-
plementing, visualizing, and simulating neural networks. The data re-
quired for the networks’ training were obtained from a series of dia-
grams provided in Roseman and Seibold (1987) in the form of pairs of
four-dimensional input vectors and scalar outputs. The input vector
consists of the length to breadth ratio (L/B), the breadth to draft ratio
(B/T), the block coefficient (CB) and the Froude number (Fn), while the
output is the residual resistance coefficient (CR).
As already stated, among the various types of ANNs, multi-layer
perceptron networks (MLPs) have been selected for the present study,
as they are considered particularly suitable for function approximation
problems. In order to identify the MLP that suits better to the problem
at hand, several trials have been conducted with different configura-
tions. Each configuration is determined by the number of hidden layers
used, the number and type of neurons that comprise each one of them,
the training algorithm, learning algorithm and rate (Beale et al., 2015).
In total, 18 different network architectures have been tested. For each
one of them, 16 different networks have been implemented, based on a
combination of four back-propagation training algorithms and four
Fig. 4. Characteristics of the MARAD series hulls (Roseman and Seibold, 1987). different pairs of activation functions. The training algorithms used for
this study are the Levenberg-Marquardt, gradient descent, gradient
ratios (L/B) and high breadth to draft ratios (B/T) were also adopted for descent with adaptive learning rate and gradient descent with mo-
economic and operational reasons (i.e. draft restrictions) respectively. A mentum and adaptive learning rate. Four combinations of three dif-
cylindrical bow shape was selected, as a means of achieving lower re- ferent activation functions have been used, i.e. the hyperbolic tangent
sistance values for full-form hulls, whilst the stern was designed taking sigmoid (herein denoted as T), the pure linear (denoted as P) and the
into account hydrodynamic issues, such as flow separation. Sets of lines logistic sigmoid function (denoted as L). As a result, 288 networks were
were developed for the parent and the offspring hulls as a combination constructed and tested. During this initial stage of trials, the gradient
of three segments; the entrance, the parallel mid-body and two alter- descent weight and bias learning algorithm was used and the learning
native run segments, a short and a long one. rate was set to 0.02. The available data was separated into mutually
Sixteen models were constructed and tank-tested in order to obtain exclusive training (80%) and validation (20%) sets. The validation set
their calm water resistance and manoeuvering characteristics. Their comprises data unknown to the trained network and is usually used for
longitudinal centre of buoyancy (LCB) was placed 2.5% of length for- the comparison and ranking of different network configurations. A
ward of amidships, as a result of preliminary resistance tests. The six- graphical description of this stage of trials is given in Fig. 5.
teen hulls comprising the MARAD series are characterized by their For each one of these 288 networks, the corresponding mean
block coefficient (CB), their beam to draft ratio (B/T) and length to squared error (MSE) and regression value (R) were calculated, based on
beam ratio (L/B) as shown in Fig. 4 (Roseman and Seibold, 1987). the training and validation sets of data. The regression value is an in-
Bare hull resistance tests were performed for the sixteen hulls at full dicator of the relationship between the actual and target outputs and its
load and ballast condition. The obtained results were presented in absolute value ranges between one and zero, suggesting the existence of
diagrams in the form of the residual resistance coefficient (CR), defined an exact linear relationship between outputs and targets or its absence,
as follows: respectively. The performance of these 288 networks was ranked ac-
cording to their mean squared error (MSE) and regression value (R)
R
CR = CTm − CFm = 1
− CFm based on the validation set.
2
ρSυ2 (6) Results from the evaluation of each one of the 18 architectures
tested are presented in Table 1. The second column in the table shows
where RTm is the model's total measured resistance, S is its wetted the number of nodes per layer (input layer – hidden layer(s) – output
surface, υ is its speed, CTm and CFm are its total and frictional resistance layer). The number of nodes of the input and output layer is by default
coefficient respectively and ρ is the mass density of water in the basin. equal to 4 and 1 respectively, corresponding to the dimensions of the
CFm is expressed by the 1975 ITTC formula as a function of Reynold's input and output vectors. The number of neurons and its distribution
number (RN): among the hidden layers varies. The minimum mean squared error
0.075 values and the maximum regression values for each one of the 18 ar-
CF = chitectures obtained with the 16 combinations of training algorithms
(log10 RN − 2)2 (7)
and activation functions, both for the training and validation sets of
The diagrams provided by the MARAD Systematic Series illustrate data, are presented in Table 1.
CR versus L/B or B/T for a series of Froude numbers (Fn), for given The ten networks that presented the lower values of MSE and the
values of block coefficient (CB) and B/T or L/B respectively. Five dia- ten networks with the higher values of R, based on the validation da-
grams for full load and five for ballast condition are included in taset, were selected for further testing. According to the above selection
Roseman and Seibold (1987). Those corresponding to the full load criteria, and due to the observed overlapping between them, six net-
condition were used in this study. A number of 1983 points in total works were eventually identified for further study. The architecture and

531
V. Margari et al. Ocean Engineering 165 (2018) 528–537

Table 1
Tested MLP architectures –1st stage of trials.
Serial numb. Architecture Training set Validation set

minMSE maxR minMSE maxR

1 4-4-1 0.0095 0.9625 0.0086 0.9819

2 4-8-1 0.0079 0.9695 0.0047 0.9694
3 4-12-1 0.0062 0.9750 0.0071 0.9738
4 4-16-1 0.0067 0.9749 0.0046 0.9808
5 4-4-4-1 0.0066 0.9699 0.0084 0.9670
6 4-8-4-1 0.0011 0.9733 0.0068 0.9744
7 4-4-8-1 0.0068 0.9736 0.0068 0.9776
8 4-12-4-1 0.0079 0.9778 0.0060 0.9814
9 4-12-8-1 0.0059 0.9776 0.0052 0.9836
10 4-16-8-1 0.0034 0.9864 0.0047 0.9791
11 4-16-4-1 0.0041 0.9834 0.0062 0.9800
12 4-16-12-1 0.0034 0.9866 0.0047 0.9798
13 4-12-8-4-1 0.0043 0.9832 0.0042 0.9845
14 4-12-4-4-4-1 0.0044 0.9831 0.0064 0.9762
15 4-12-8-4-4-1 0.0050 0.9808 0.0058 0.9822
16 4-16-12-8-4-1 0.0039 0.9845 0.0046 0.9796
17 4-12-12-8-4-1 0.0048 0.9819 0.0049 0.9796
18 4-12-8-8-4-1 0.0045 0.9817 0.0055 0.9598

Table 2
Properties of selected networks.
Code Architecture Training Activation function
number algorithms
Hidden Output layer
layers

2.4.2 4-8-1 LM T P
4.4.1 4-16-1 LM T T
10.4.2 4-16-8-1 LM T P
13.4.1 4-12-8-4-1 LM T T
16.4.2 4-16-12-8-4-1 LM T P
16.4.4 4-16-12-8-4-1 LM L P

(used also during the first stage) and the gradient descent with mo-
mentum weight and bias. As a result, a total of 72 networks were tested
(Fig. 6). The available data was divided into training, validation and
test dataset with ratios 80%, 12% and 8% respectively. For each one of
the six networks listed in Table 2, the minimum MSE and maximum R
values obtained by the 12 combinations of learning rates and learning
algorithms for the training, validation and test data set, are presented in
Table 3.

Fig. 5. Graphic description of 1st set of trials.

the activation functions of these networks are presented in Table 2. The

first digit of their code numbers corresponds to the serial number of
their architecture, the second one to the training algorithm used and the
third one to the combination of activation functions, as described in
Fig. 5.
The most important factor associated with improved MSE and R
values was the training algorithm. The Levenberg-Marquardt training
algorithm presented the best results, in comparison with the other three
training algorithms that were tested. With respect to the pairs of acti-
vation functions, the networks that consisted only of logistic sigmoid
neurons presented the worst performance. As a result, the networks
selected for further study were using the Levenberg-Marquardt algo-
rithm, while the networks with only logistic sigmoid neurons were
excluded.
The objective of the second stage of trials was to determine the most
appropriate learning algorithm and rate. To this end, tests were carried
out using six different learning rates (0.01, 0.02, 0.05, 0.1, 0.15, 0.2)
and two learning algorithms: the gradient descent weight and bias Fig. 6. Graphic description of second set of trials.

532
V. Margari et al. Ocean Engineering 165 (2018) 528–537

Table 3
Obtained MSE & R values of selected networks.
a/a Training set Validation set Test set

min MSE max R min MSE max R min MSE max R

2.4.2 0.0031 0.9875 0.0006 0.9978 0.0003 0.9985

4.4.1 0.0028 0.9886 0.0006 0.9978 0.0004 0.9985
10.4.2 0.0005 0.9977 0.0002 0.9993 0.0001 0.9997
13.4.1 0.0016 0.9981 0.0003 0.9989 0.0003 0.9990
16.4.2 0.0003 0.9989 0.0003 0.9986 0.0001 0.9996
16.4.4 0.0002 0.9991 0.0001 0.9994 0.0001 0.9997

Fig. 9. MSE plot for MLP1 during training.

Fig. 7. MSE vs. 1-R values – second stage of trials.

Fig. 10. Actual output vs. desired for MLP1 after training.

Fig. 8. Architecture of MLP1.

5. Discussion of results

A series of neural networks has been developed, trained and tested

for the prediction of calm water resistance of MARAD-type hullforms.
As already mentioned, this work was conducted in two stages: In the
first stage, the most promising network configurations (i.e. combina-
tions of architecture, training algorithm and pairs of activation func-
tions) were determined. In the second stage, the selected combinations
were tested using 12 pairs of learning algorithms and rates (Fig. 6),
aiming to obtain improved results. The networks constructed in the Fig. 11. Architecture of MLP2.

533
V. Margari et al. Ocean Engineering 165 (2018) 528–537

Table 4
Geometric properties of the three MARAD hullforms.
Ship B E H

L/B 6.0 5.0 6.5

B/T 3.00 3.00 3.00
CB 0.875 0.850 0.850
L [m] 349.9 333.4 372.6
B [m] 58.3 60.6 57.3
T [m] 19.4 20.2 19.1
WSA [m2] 31111 29356 31877

training is presented in Fig. 9. The training stopped when a minimum

value of MSE for the validation dataset was reached. This network was
selected for its low validation MSE, reached in 105th epoch of the
training process. The MSEs based on the training and validation sets are
relatively higher than that based on the test set. After the training
process was completed, scatter diagrams comparing the actual output of
the network with the desired one for all three datasets (i.e. training,
validation, test) were generated (see Fig. 10). Larger deviations be-
tween the actual and the desired output values may be observed in the
Fig. 12. MSE plot for MLP2 during training.
scatter diagram corresponding to the training set. This is an expected
result, since the network uses this set for updating its parameters, while
second stage presented better validation MSE and R values (i.e. MSE validation and test sets are used for the evaluation of its predictive
and R values based on the validation set), compared to those of the first ability. Much smaller deviations are observed in the scatter diagrams
stage: the validation MSE value was reduced by 52.9%–96.9%, while corresponding to the validation and test data sets, which present an
the increase of the R value varied between 0.7% and 3.1%. Since R almost linear relationship between actual and target outputs with R
values were already very high, the actual improvement may be better values over 0.999 for both of them.
observed by looking at the 1-R values, which showed a reduction of The architecture of the second network (denoted as MLP2) is com-
46.7%–98.2% (Fig. 7). The two networks that showed the lowest vali- posed by four hidden layers of 16, 12, 8 and 4 hyperbolic tangent
dation and test MSE values were selected for the prediction of the re- neurons each and one linear output layer (Fig. 11). The same training
sidual resistance coefficient, and subsequently the resistance of hull- algorithm as for MLP1 was used, but the gradient descent with mo-
forms based on the MARAD Systematic Series. mentum bias and weight learning algorithm and a learning rate equal to
The architecture of the first neural network (herein denoted as 0.15 were applied. The training stopped in the 43rd epoch. The lower
MLP1) is illustrated in Fig. 8. It comprises two hidden layers with 16 MSE value was obtained by the test set (Fig. 12). The R values for this
and 8 hyperbolic tangent sigmoid neurons each and a linear output. The network were also very high, with the one corresponding to the test set
Levenberg-Marquardt training algorithm was used, along with the being the higher (Fig. 13).
gradient descent with weight and bias learning algorithm and a The two selected networks were used to evaluate the residual re-
learning rate equal to 0.20. The progression of the MSE value during sistance coefficient of three MARAD hullforms listed in Table 4. It
should be noted that data from these hulls were not included in the
training, validation and test datasets used for the training and evalua-
tion of the networks.
For all three ships the results derived by both networks presented
very small deviations from the target values and between them
(Figs. 14–16).

Fig. 13. Actual output vs. desired for MLP2 after training. Fig. 14. Residual resistance coefficient for ship B.

534
V. Margari et al. Ocean Engineering 165 (2018) 528–537

Table 6
Total resistance prediction for ship E.
MARAD MLP1 MLP2

Fn υ [kn] RT [kN] RT [kN] diff. [%] RT [kN] diff. [%]

0.130 14.0 1832.1 1832.2 0.0 1838.2 −0.3

0.140 15.1 2222.2 2195.5 1.2 2208.9 0.6
0.145 15.6 2455.9 2439.4 0.7 2438.5 0.7
0.150 16.1 2684.7 2677.6 0.3 2697.1 −0.5
0.155 16.7 2905.1 2910.7 −0.2 2904.7 0.0
0.160 17.2 3136.7 3110.6 0.8 3101.9 1.1
0.165 17.8 3379.9 3367.9 0.4 3349.2 0.9
0.170 18.3 3634.8 3642.6 −0.2 3625.4 0.3
0.175 18.8 3986.7 3976.9 0.2 3975.8 0.3
0.180 19.4 4435.6 4420.5 0.3 4407.7 0.6
0.185 19.9 4915.8 4898.9 0.3 4899.7 0.3
0.190 20.5 5444.9 5435.8 0.2 5419.7 0.5

experimental CR values and those estimated by the neural networks for

the three hullforms are compared in Tables 5–7 as well as in
Fig. 15. Residual resistance coefficient for ship E.
Figs. 17–19. The predictions by MLP1 and MLP2 for the three ships
present very low deviations from the experimental values, in the order
Predictions were conducted for Froude numbers up to 0.18, and in of 0.1%–1.8% for both networks.
some cases up to 0.19, which are quite high when compared to Froude
numbers of existing full-form vessels of similar size. However, since the
6. Conclusions
MARAD diagrams are extending up to these values, it was decided to
use the full range of available data.
The present paper investigates the possibility of using artificial
The results from the calculation of the total resistance based on the
neural networks for the prediction of resistance of full ships designed
according to the MARAD hullforms. The data for the training and
evaluation of the networks were obtained from a series of diagrams,
illustrating the residual resistance coefficient (CR) of the sixteen
MARAD hullforms at the full load condition for a range of Froude
numbers.
A large number of Multi-layer perceptron neural networks were
trained and evaluated. Initially, a total number of 288 networks were
developed. In this stage, the learning algorithm and rate were kept
constant, while combinations of different architectures, training algo-
rithms and pairs of activation functions were tested. Based on their
performance on a validation data set, six networks of specific archi-
tecture, training algorithm and pair of activation functions were se-
lected for further study. In the second stage of trials, these six networks
were tested using six learning rates and two learning functions, while
their other characteristics remained constant. Finally, the two networks
that presented the lower mean squared error on the validation dataset
were selected.
In order to illustrate the ability of these two networks to provide
accurate resistance predictions, results have been presented, comparing
their predictions for three ships based on MARAD hullforms with those
Fig. 16. Residual resistance coefficient for ship H. derived from the experimental results. Multi-layer perceptron neural

Table 7
Total resistance prediction for ship H.
Table 5
Total resistance prediction for ship B. MARAD MLP1 MLP2

MARAD MLP1 MLP2 Fn υ [kn] RT [kN] RT [kN] diff. [%] RT [kN] diff. [%]

Fn υ [kn] RT [kN] RT [kN] diff. [%] RT [kN] diff. [%] 0.130 15.3 2012.7 2018.6 −0.3 2012.5 0.0
0.140 16.5 2332.1 2345.6 −0.6 2347.1 −0.6
0.130 14.8 2025.4 2038.2 −0.6 2046.7 −1.1 0.145 17.0 2544.9 2531.2 0.5 2535.5 0.4
0.140 15.9 2486.3 2512.0 −1.0 2526.5 −1.6 0.150 17.6 2756.6 2731.0 0.9 2745.2 0.4
0.145 16.5 2810.3 2841.6 −1.1 2831.6 −0.8 0.155 18.2 2936.0 2941.6 −0.2 2955.4 −0.7
0.150 17.1 3173.3 3230.0 −1.8 3220.5 −1.5 0.160 18.8 3136.2 3155.0 −0.6 3145.4 −0.3
0.155 17.6 3565.7 3600.2 −1.0 3589.5 −0.7 0.165 19.4 3343.8 3366.1 −0.7 3344.9 0.0
0.160 18.2 3904.5 3937.1 −0.8 3946.6 −1.1 0.170 20.0 3558.8 3599.3 −1.1 3591.5 −0.9
0.165 18.8 4294.3 4346.8 −1.2 4367.2 −1.7 0.175 20.6 3964.5 3925.2 1.0 3920.2 1.1
0.170 19.4 4804.3 4874.2 −1.5 4854.5 −1.0 0.180 21.2 4398.9 4388.7 0.2 4360.8 0.9
0.175 19.9 5385.4 5466.6 −1.5 5440.6 −1.0 0.185 21.7 4945.0 4941.4 0.1 4871.8 1.5
0.180 20.5 6027.0 6093.7 −1.1 6073.2 −0.8 0.190 22.3 5552.4 5516.4 0.6 5467.3 1.5

535
V. Margari et al. Ocean Engineering 165 (2018) 528–537

networks proved to be effective in the evaluation of resistance of

MARAD hullforms with geometric characteristics within the limits of
training data, with maximum deviation less than 2%.
Based on the presented results, it may be argued that properly de-
signed and trained multi-layer perceptron neural networks proved to be
rather effective in the evaluation of resistance of MARAD hullforms.
Considering their successful performance in the case of the MARAD
series, they are expected to be equally suitable for the calculation of the
calm water resistance of ships designed according to other systematic
series. In addition, they could be also used in order to provide fast and
accurate predictions based on training sets derived from the systematic
application of CFD calculations for selected hullform types and speed
ranges. If such networks are available and properly trained, they can
provide the basis for systematic hullform optimisations, taking into
account not only resistance predictions but also other important aspects
of a ship's hydrodynamic performance, such as propulsion power,
manoeuvering characteristics and seakeeping responses.

References
Fig. 17. Total resistance prediction for ship B.

Beale, M.H., Hagan, M.T., Demuth, H.B., 2015. Neural Network Toolbox-user’s Guide.
The Mathworks Inc.
Bojovic, P., 1997. Resistance of AMECRC systematic series of high speed displacement
hullforms. In: Proceedings of the IV Symposium on High Speed Marine Vehicles,
Naples, Italy.
Cepowski, T., 2005. Application of statistical methods and artificial neural networks for
approximating ship's roll in beam waves. Pol. Marit. Res. 22–28 No2/2005.
Cepowski, T., 2007. Application of artificial neural networks to approximation and
identification of sea-keeping performance of a bulk Carrier in ballast loading con-
dition. Pol. Marit. Res. 4 (54), 31–39 Vol. 14.
Hagan, M.T., Demuth, H.B., Beale, M.H., De Jesús, O., 2014. Neural Network Design,
second ed. Martin Hagan.
Haykin, S., 1999. Neural Networks – a Comprehensive Foundation, Prentice-Hall, second
ed. Prentice Hall PTR, Upper Saddle River, NJ, USA.
Hebb, D., 1949. The Organization of Behavior. John Willy and Sons Inc, New York.
Koushan, K., 2001. Empirical prediction of ship resistance and wetted surface area using
Artificial Neural Network. In: Proceedings of the Eighth International Symposium on
Practical Design of Ships and Other Floating Structures, vol.1. pp. 501–508.
Levenberg, K., 1944. A method for the solution of certain non-linear problems in least
squares. Q. Appl. Math. 2, 164–168.
Margari, V., 2017. Prediction of Resistance of MARAD Systematic Series' Hullforms Using
Artificial Neural Networks. Diploma Thesis. NTUA, School of Naval Architecture and
Marine Engineering.
Margari, V., Kanellopoulou, A., Zaraphonitis, G., 2018. Prediction of resistance of MARAD
systematic series' hullforms using artificial neural networks. In: Proceeding of the 6th
International Symposium on Ship Operations, Management & Economics (SOME),
SNAME, Athens.
Marquardt, D., 1963. An algorithm for least-squares estimation of nonlinear parameters.
Fig. 18. Total resistance prediction for ship E. J. Soc. Ind. Appl. Math. 11 (No. 2), 431–441.
Martins, P.T., Lobo, V., 2007. Estimating maneuvering and sea-keeping characteristics
with neural networks. In: Proceedings of the IEEE OCEANS 2007, Aberdeen.
Mason, A., Couser, P., Mason, G., Smith, C.R., von Konsky, B.R., 2005. Optimization of
vessel resistance using generic algorithms and artificial neural networks. In:
Proceedings of the COMPIT 2005, Hamburg, Germany, pp. 440–454.
McCulloch, W., Pitts, W., 1943. A logical Calculus of the ideas immanent in nervous
activity. Bull. Math. Biol. 5 (4), 115–133.
Minksy, M., Papert, S., 1969. Perceptrons: an Introduction to Computational Geometry.
MIT Press, Cambridge, Mass.
Moor, D., Parker, M.N., Pattulo, R.N.M., 1961. The BSRA methodical series – an overall
presentation. Trans. RINA 103.
Ortigosa, I., López, R., García, J., 2014. Prediction of total resistance coefficients using
neural networks. J. Marit. Res. 6 (3), 15–26 [S.l.].
Pedersen, B.P., Larsen, J., 2009. Prediction of full-scale propulsion power using artificial
neural networks. In: Proceedings of COMPIT 2009: 8th International Conference on
Computer and IT Applications in the Maritime Industries, Budapest, pp. 537–550.
Radojcic, D., Zgradic, A., Kalajdzic, M., Simic, A., 2014. Resistance prediction for hard
chine hulls in the pre-planning regime. Pol. Marit. Res. 2 (82), 9–26 Vol. 21.
Rojas, R., 1996. Neural Networks – a Systematic Introduction. Springer –Verlag, Berlin.
Roseman, D.P., Seibold, F., 1985. An introduction to the U.S. Maritime administration
systematic series. Mar. Technol. 221, 329–338.
Roseman, D.P., Seibold, F., 1987. The MARAD Systematic Series of Full-Form Ship
Models. The society of Naval Architects and Marine Engineers.
Rosenblatt, F., 1958. The perceptron: a probabilistic model for information storage and
organization in the brain. Psychol. Rev. 65 (6), 386–408 Cornell aeronautical la-
boratory.
Russell, S., Norvig, P., 1995. Artificial Intelligence – a Modern Approach. Prentice Hall,
Eglewood Cliffs, New Jersey.
Fig. 19. Total resistance prediction for ship H. Shaher Sabit, A., 1971. A regression analysis of the resistance results of the BSRA standard
series. Int. Shipbuild. Prog. 3–17.
Shaher Sabit, A., 1972a. An analysis of the series 60 results. Part 2, analysis of propulsion

536
V. Margari et al. Ocean Engineering 165 (2018) 528–537

factors. Int. Shipbuild. Prog. 294–301. Todd, F.H., 1963. Series 60 Methodical Experiments with Models of Single-screw Ships.
Shaher Sabit, A., 1972b. An analysis of the series 60 results. In: Part 1, Analysis of Forms Report 1712. U.S. Government Printing Office.
and Resistance Results. International Shipbuilding Progress, pp. 81–97. Williams, A., 1969. The SSPA Cargo Liner Series Resistance. SSPA Report No 66. Statens
Shaher Sabit, A., 1976. The SSPA Cargo liner series regression analysis of the resistance Skeppsprovningsanstalt, Giitenburg.
and propulsive coefficients. Int. Shipbuild. Prog. 23 (263), 213–217. Yao, J., Han, D., 2012. RBF neural network evaluation model for MDO design of ship.
Snyman, J.A., 2005. Practical Mathematical Optimization. Springer. IPCSIT 47, 309–312.
The MathWorks Inc, 2018. Neural Network Toolbox. [Online] Available at: https:// Zurada, J.M., 1992. Introduction to Artificial Neural Systems. West Publishing Company,
www.mathworks.com/products/neural-network.html/, Accessed date: 19 April St. Paul.
2018.

537