9th IFAC Conference on Manoeuvring and Control of Marine Craft, 2012

The International Federation of Automatic Control

September 19-21, 2012. Arenzano, Italy

Parallel Simulation Based Predictive Control Scheme With Application To

Approaching Control For Automatic Berthing
Naoki Mizuno*, Hitoshi Kakami**, Tadatsugi Okazaki***

*Nagoya Institute of Technology, Gokiso-cho, Shouwa-ku, Nagoya 466-8555 JAPAN

e-mail: nmizuno@nitech.ac.jp ,Fax: +81-52-735-5342
**Tamano Works, Mitsui Engineering and Shipbuilding, Co. Ltd. 1-1, Tama 3, Tamano
Okayama JAPAN
***Tokyo University of Marine Science and Technology, 2-6-1, Etchujima, Koto-ku,
Tokyo JAPAN

Abstract: This paper presents a parallel simulation based predictive control algorithm for automatic
berthing. The proposed scheme uses many number of control sequences for on-line simulation of
future behaviour of the ship movement. Each control sequence used in simulation generates a predicted
sequence of the ships future course. The predicted courses of the ship are evaluated based on some cost
function suitable for control objective and then, the optimal control signal is applied to the ship. To
simulate the future behaviour of the ship, it is used a non-linear ships dynamic model (called MMG
model) and also the massively parallel computational hardware (called GPU: Graphical Processing Unit).
Application of GPU for simulation (GPGPU) permits directly use of the nonlinear model of the ship for
parallel simulation and makes the control algorithm with highly computational demand feasible.
In order to investigate the effectiveness of the proposed algorithm, computer simulations and actual sea
tests are carried out using a training ship Shioji Maru.
The experimental results show that the proposed method is successfully applied to the approaching
control for automatic berthing.

Key Words: ship control, predictive control, simulation based control, parallel simulation, GPGPU

1. INTRODUCTION system, by using many candidate control sequences. If the

In recent days, many marine systems are still controlled by
simple PID (proportional-integrative-derivative)
algorithms, despite the better performances usually provided
by systems developed based on the modern control theory.
PID controllers can be used to control a wide range of ships
processes, need only rough process models to be easily tuned
and give reasonable control performances. On the other hand it
is clear that PID performances, although satisfactory, could be
improved when dealing with highly nonlinear dynamics and/or
external disturbances.
A Model Based Predictive Control (MBPC) algorithm is
described by using a model to compute the predicted process
outputs. Also, a cost function related to the control
performance of the system is defined, and the control signal is
obtained by means of minimization the cost function. Finally,
the first of these signals is applied to the process [1].
The extension of linear MBPC to nonlinear processes is
straightforward at least conceptually. However, there exist
some difficulties [2]. In these difficulties, the availability of
nonlinear models and the computational complexities are most
important issues in real applications.
For this problem, the basic idea of the proposed algorithm is
the on-line simulation of the future behaviour of the control
future behaviour of the system can be calculated for many
different control sequences during the control sampling period,
we can determine the optimal inputs at each sampling period
by selecting the first input which gives the most desirable
response in the future.
In the last few decades, computer simulation has enabled
more detailed analysis of marine system that could not be
achieved with theoretical or experimental methods. Various
information about the ships behaviour, such as non-linear
response for rudder angle and pitch angle of the propeller, can
be predicted using simulation.
After the control problem has been redefined as the
simulation problem of future behaviour of the marine system,
the optimisation search space of inputs is very large and is not
feasible to be searched by conventional computational
hardware such as single CPU.
However, rapid progress in massively parallel computational
hardware such as GPU makes it possible to perform parallel
simulation for different input sequences simultaneously.
For example, modern GPU has over 1,000 processing
elements in the chip and can be execute over 10,000 different
simulations in one time.

2012 IFAC 19 10.3182/20120919-3-IT-2046.00004

IFAC MCMC 2012
September 19-21, 2012. Arenzano, Italy
Based on the above mentioned algorithm and implementation
for modern GPU, we develop a feasible approaching control
system for automatic berthing.
It is very important for a ship's master to draw up a
ship-handling plan before approaching a berth, leaving it,
altering the heading and so on. In order to cope with these
problems, they are required to derive the maximum
maneuverability from the ship. The minimum-time control
technique is one of the rational and effective maneuvering
methods for such tedious problems if the mathematical model
representing a ship's dynamics is available. However, it should
be noted that the model becomes highly non-linear, especially
in the case of low speed and large maneuvering motion. In this
case, the solutions of minimum-time maneuvering problem
can be obtained by solving non-linear two-point boundary
value problems (TPBVP).
In order to take enough account of the non-linearity, Shoji
and Ohtsu [3] have formulated these problems as a non-linear,
two-point boundary-value problem (TPBVP) in the calculus of
variations. The problems have been solved, using the
numerical method called the sequential conjugate gradient
restoration (SCGR) method [4], [5] under various situations:
1) The minimum-time course-alteration problem,
2) The minimum-time stopping problem,
3) The minimum-time parallel deviation problem [6],
4) The minimum-time maneuvering problem with wind
disturbances [7],
5) The minimum-time berthing problems [3], [8].
However, the solutions, unfortunately, do not yield on-line
control laws, because, it took long computational time to
obtain the optimal solution even by using a high-speed
workstation. Moreover, the geometrical and other sea
conditions of maneuvering problem facing to the real ship
handling scenes are various. Thus, it is impossible to prepare
whole minimum time maneuvering solutions to such various
conditions.
For these problems, Okazaki et al., [9] have investigated the
minimum time maneuvering with neural network. In the
proposed method, an off-line trained neural network
interpolates the real-time maneuvering solution from the
available solutions, which has been already solved, but the
control performance varies depend on the disturbances in real
sea conditions. To overcome this problem, Mizuno et al., [10]
have proposed a ships minimum-time maneuvering system
with two types of neural networks. This system contains the
on-line trained neural network, which compensates the
difference between the real ships dynamics and the
mathematical model, and the other errors caused by some
disturbances etc. From the experimental results, the method
has some advantages compared with the previous studies.
There are some papers in which the neural networks are used
in order to compensate the non-linear dynamics of the ship
during the tracking or berthing phase [11]-[14]. Unfortunately
the on-line learning speed of the neural network is rather slow.
This means that it is difficult to obtain the good transient
performance for tracking or berthing in real situations.
From this point of view, Mizuno et al., [15],[16] have
proposed a new design method for ships minimum time
maneuvering system with neural network and non-linear
model based compensator.
20
Moreover, Mizuno et al. [17] have extended the above
method applicable to minimum-time approaching control for
automatic berthing under some disturbances. The extended
system is composed of a neural network based optimal
solution generator and a nonlinear model predictive
compensator. The neural network generates the optimal
solution for real situation by interpolating pre-computed
minimum-time berthing solutions for typical control
conditions. The optimal solutions for the various minimum
time berthing are numerically computed based on the
sophisticated non-linear dynamical model of the ship (MMG
model) and are learned off-line by the neural network for
interpolation. Moreover, the same nonlinear model, which is
used for the computation of the optimal solutions, is used to
simulate the ships future position and heading. Based on the
receding horizon cost function of the predicted control error
caused by some disturbances, the control inputs for minimum
time approaching are modified.
In this paper, we extend the above non-linear model
predictive compensator to achieve more accurate tracking
under some disturbances by parallel simulation based
approach. Based on the parallel simulation, the control input
which achieve the minimum time maneuvering and
disturbance rejection under realistic sea conditions. This is a
new feature of this system.
First, the basic idea of parallel simulation based predictive
control is briefly reviewed. Next, the berthing problems and
the mathematical model of the ships dynamics for parallel
simulation of ship behaviour are explained. In the third part of
this paper, an approaching control system for automatic
berthing is introduced. Finally, computer simulations and
on-line experiments are carried out for a training ship Shioji
Maru (425 gross tonnage). The experimental results show that
the proposed method is successfully applied.
2. BASIC IDEA OF PARALLEL SIMULATION BASED
CONTROL
2.1 Formulation of Simulation Based Control
The purpose of the controller is typically to force the output to
follow the reference signal. If the dynamics of the system is
linear, the control input can be solved using the linear model
of the process.
However, in case of non-linear system, the input cannot be
solved in general case.
For this situation, the simple idea of the solving algorithm is
to compute the following quantities for every sample period:
1) the behaviour of system output over a finite horizon (N);
2) the cost of the objective function for all (prescribed) control
sequences:
u(.) = {u(t), u(t + 1),.., u(t + N )} (1)
3) to choose the first element of control sequence which gives
the best cost.
IFAC MCMC 2012
September 19-21, 2012. Arenzano, Italy

computational load for many simulations is still large for

single or a few multi processors. For this situation, we can
perform the parallel simulations using many-core GPU based
hardware for many sets of input candidate. After the parallel
simulations, we can select the best input for control objective
based on the performance index as shown in Fig.3.

State of Process, Candidates of input, Reference

Signal

t0 t1 t2 tf
Simulator Simulator Simulator
Fig.1 Candidate of input sequences

Figure 1 shows the example of input sequences. P. Index P. Index P. Index

The advantages of the above algorithm are as follows:

1) the global minimum of objective function can be obtained
if all input sequence are investigated; Select Input with minimum Cost
2) the algorithm can be easy applied to nonlinear processes;
3) the constraints can easily be implemented.
Fig.3 Parallel Simulation based Control Input Calculation
On the other hands, the drawback of this scheme is very long
computational time. 3. BERTHING PROBLEM AND MATHEMATICAL
To reduce the computational time, first, we reduce the number MODEL OF THE SHIP
of input sequences based on the partial cost function.
2nd stage Predictions 3.1 Minimum-time berthing problem and its solutions
State
Target Course In this paper, we consider the minimum-time berthing
solution as a typical reference
Y O
signal for parallel simulation
time
Target Course based predictive control.
State State Target Course B Figure 4 shows the start point
and terminal point of the
2nd stage Predictions
1st stage Predictions
minimum time berthing
time time problem.
u
State
Target Course
To obtain the minimum-time
berthing solution off-line, the
2nd stage Predictions v
time mathematical model of the
t0 t1 ship is needed. Also, the same
A
mathematical model is used
Fig.2 Reduction of the number of input sequences X

Fig.4 Berthing Maneuvering for parallel simulation of

Figure 2 shows the process of the reduction of input sequence.
In this case, only 3 inputs in first step are selected in
candidate of the combination with the second step of inputs
because the other input in the first step does not give the
better partial cost compared with the selected 3 inputs in the
first step.
2.2 Implementation using parallel hardware
The number of the candidate of the input sequence may
be reduced based on the partial cost function, but the
the future behaviour of the
ship on-line.
In this research, the mathematical model of the training ship
Shioji Maru is used for both calculations.
Figure 5 shows the normalized optimal paths for 250m and
450m approach, ship's headings, rudder angles, ships speeds
and CPP angles in two typical cases.
It should be noted that the time histories are different in each
case because of the non-linearity of the ships synamics..

21
IFAC MCMC 2012
September 19-21, 2012. Arenzano, Italy

Optimal PathPosition Y x u cos v sin (2)

m(u vr ) X H X P X R (3)
y u sin v cos (4)
m(u vr ) X H X P X R 5)
y u sin v cos (6)
Position X
Ships heading Rudder angle m(v ur ) YH YT YR YTh (7)
r (8)
I ZZ r N H N T N R N Th (9)
( * ) (10)

( * TRUD a )
Ships speed CPP angle
( P PTlp )
*
P (11)
P P

Tlp ( P P PTlp TCPP a) Tlp
*

where m and I are the mass and the turning

moment inertia. TRUD , TCPP and a are time
Fig. 5 Typical Solutions for Minimum-time Berthing
constants. The subscripts H , P , R and Th denote
the highly non-linear hydrodynamic force induced by
3.2 Ships motion model
the hull, propeller, rudder and thruster. For detail of
Shioji Maru is equipped with a bow and stern thrusters, the hydrodynamic force, see [7].
besides a rudder and a controllable pitch propeller (CPP). Her
principal dimensions are shown at Table 1 and the coordinate * (t ) sin Dumy (t ) 0 , 0 t 1 (12)
system is shown in Figure 6.
where, Dumy is the dummy variable for the order
Table 1: Principal Dimensions of Shioji Maru
rudder angle.
Length 49.93 m
Breadth 10.00 m 4. PARALLEL SIMULATION BASED PREDICTIVE
Tonnage 425.0 GT CONTROL SYSTEM
Propeller CPP
4.1 Structure of proposed control system
Bow Thruster 2.4 tons
Stern Thruster 1.8 tons
Figure 7 shows the schematic diagram of the proposed control
system.
X

v r

O y

Fig. 6 Ship's coordinate in horizontal plane

The state values are ships position (x, y), ships heading
, surge speed u, sway speed v, yaw rate r, rudder angle
and CPP angle P The control values are order rudder
angle * , order CPP angle P* , notch of bow thruster bt * ,
notch of stern thruster st * . Referencing to this, the
sophisticated mathematical model (called MMG model) is
written by
Fig. 7 Schematic diagram of the proposed control system
In the proposed system, the MMG model based parallel
ships dynamic simulator and the selection mechanism for
optimal rudder angle and the CPP angle which achieve
the best error based cost function as described later.
Although the MMG model is very complicated and the
number of combination of the cadidate rudder angles and CPP
angles are very large, the computational time with GPU is
rather short compared with conventional CPU.

22
IFAC MCMC 2012
September 19-21, 2012. Arenzano, Italy

For example, the computational time for parallel

simulationfor future 40 [sec] ships behaviour with CPU and
GPU (ASUS N10J Notebook computer) are compared and
summalized in Table 2. This indicates that GPU matkes the
parallel simulation based control system feasible for
approaching control system for automatic berthing.

Table 2 Comparison in computational time for parallel

simulation with CPU and GPU

Using the GPU for computational hardware, the parallel

sumulation based predictive control system becomes feasible
in real-time applications. Now, high-level programming
interfaces (APIs), like NVIDIAs Compute Unified Device
Architecture (CUDA) [18] and ATIs Stream technology [19],
make the development of parallel simulation based predictive Fig. 10 A picture of Shioji
controller programming easier. Maru
The algorithm adopted for simulation based on Eqs. (2)-(12)
is Eulers method.
To reduce the future tracking error for the optimal course of
the minimum time maneuvering, we optimize the following
receding horizon cost function J .
N
J | Y ( ) Y ( ) |
M
o s (13)

where, Yo ( ) denotes the optimal deviation of the minimum-time solution

and Ys ( ) is the simulated one of the controlled ship.

Fig. 8 Simulation results of proposed control system with

5 PRELIMINARY EVALUATIONS BY COMPUTER wind
SIMULATIONS

Before implementing the proposed parallel simulation based
predictive control system, computer simulations of proposed
scheme are performed for nonlinear dynamical model of the
Shioji Maru. Figure 8 shows the typical simulation results for
approaching control for automatic berthing with wind
disturbances by using proposed scheme.
The conditions of parallel simulations are also indicated.
6.2 Real Time Control System
The Shioji Maru has her own on-board control system with
GPS position measurement as shown in Fig. 10. However, to
implement the parallel simulation function, we use the
notebook type computer with GPU for real-time controller.

From this result, it can be seen that the influence of the

disturbance is avoided based on the parallel simulations.

6. ACTUAL SEA TEST

In order to evaluate the feasibility

of the proposed parallel simulation
based predictive control system,
the actual approaching tests for
berthing were carried out at sea,
using the Shioji Maru of Tokyo Fig.10 The real time control system aboard the Shioji
University of Marine Science and Fig. 9 Shioji Maru Maru
Technology (Fig. 9).

23
IFAC MCMC 2012
September 19-21, 2012. Arenzano, Italy
6.3 Experimental results
Figure 11 shows the typical experimental result using
proposed control scheme.
Fig. 11 Experimental results of proposed control system with
wind
In this case, the settings for the controller are same as the
simulation. From these results, it can be concluded that the
proposed parallel simulation based predictive control
system is feasible in actual sea conditions.
7. Conclusion
This paper presented a new practical ships maneuvering
system with parallel simulation based predictive controller. In
the approaching control problem for automatic berthing, the
system gives good tracking performance in real situations.
Moreover, the actual sea trials demonstrate the effectiveness
of the proposed system.
Acknowledgments
24
The authors wish to thank the crewmembers of the Shioji
Maru, for their helpful support in the actual sea tests.
This work is supported in part by Grant-in-Aid for Scientific
Research (c) from the Ministry of Education, Culture, Sports,
Science and Technology of Japan.
References
[1] Camacho E. Bordons, C. (1999) Model Predictive
Control Spriger-Verlag, ISBN 3-540-76241-8, 1999
[2] T. Allgower and A. Zheng (Eds.) Nonlinear Model
Predictive Control, Birkhauser, 2000.
[3] Shoji, K. and K. Ohtsu, Automatic berthing study by
Optimal Control Theory, Proceedings of CAMS92, Genova,
1992, 185-194.
[4] Wu, A.K. and A. Miele, Sequential Conjugate
Gradient-Restoration Algorithm for optimal Control
Problems with Non-Differential Constraints and General
Boundary Conditions, Part 1, Optimal Control Applications
and Method.Vol.1, 1980, pp. 69-88.
[5] Miele, A. and R.R. Iyer, General Technique for Solving
Nonlinear two points Boundary-Value Problems via the
Method of Particular Solutions, Journal of Optimization
Theory and Applications, Vol.5, No.5, 1970, pp. 382-399.
[6] Ohstu, K. and K. Shoji, Minimum time maneuvering of
ships, Proceedings of MCMC94, 1994.
[7] Ohtsu, K. et al., Minimum Time Maneuvering of a ship, with
Wind Disturbances, Control Eng. Practice, Vol.4, No.3, 1996,
385-392.
[8] Okazaki, T. et al., A Study of Minimum Time Berthing
Solutions, Proceedings of MCMC2000, 2000, pp. 135-139.
[9] Okazaki, T. et al., Study on the Minimum Time Maneuvering
with Neural Network, The Journal of Japan Institute of
Navigation, Vol.102, pp.327-337.
[10] Mizuno, N. et al., A Ships Minimum Time Maneuvering
System Using Neural Networks, Proceedings of IECON02,
November 2002.
[11] Yao, Z. et al., "A multivariable Neural controller for
Automatic Ship Berthing", IEEE Control Systems, June 1997.
[12] Yao, Z. et al., "A Neural Network Approach to Ship
Track-Keeping Control", IEEE Journal of engineering,
Vol.21, No.4, October 1996.
[13] Namkyun Im and Kazuhiko Hasegawa, "A Study on
Automatic Berthing Using Parallel Neural Controller", Journal
of the kansai society of naval architects Japan, No.236,
September 2001.
[14] Namkyun Im and Kazuhiko Hasegawa, "A Study on
Automatic Berthing Using Parallel Neural Controller (2nd
Report)", Journal of the kansai society of naval architects
Japan, No.237, March 2002.
[15] Mizuno, N. et al., A Ships Minimum Time Maneuvering
System Using Neural Network and Non-linear Model based
Super Real-Time Simulator, Proceedings of ECC03,
September 2003
[16] Mizuno, N. et al., Minimum Time Ship Maneuvering using
Neural Network and Nonlinear Model Predictive
Compensator, Proceedings of CAMS 04, July 2004.
[17] Mizuno,N. Et al., A Ships Minimum Time Approaching
Control for Automatic Berthing using Neural Network and
Model Predictive Compensator, Proceedings ofMCMC 06,
2006
[18] NVIDIA CUDA Programming Guide Version 2.3, avail-
able at http://developer.download.nvidia.com/compute/
cuda/2_3/toolkit/docs/NVIDIA_CUDA_Programming_Guide_
2.3.pdf..
[19] ATI Close-To-Metal (CTM) Guide Version 1.01, available at
http://ati.amd.com/companyinfo/researcher/documents/
ATI_CTM_Guide.pdf