See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221140652

An Online Adaptation Control System Using mnSOM

Conference Paper  in  Lecture Notes in Computer Science · October 2006

DOI: 10.1007/11893028_104 · Source: DBLP

CITATIONS READS
7 28

3 authors:

Shuhei Nishida Kazuo Ishii

19 PUBLICATIONS   62 CITATIONS   
Kyushu Institute of Technology
352 PUBLICATIONS   824 CITATIONS   
SEE PROFILE
SEE PROFILE

Tetsuo Furukawa
Kyushu Institute of Technology
84 PUBLICATIONS   461 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Ship-Hull cleaning Robot Project View project

Vision-Based Behavior Strategy for Tomato Harvesting Robot (III): - Feature point matching for mosaic image
generation expressing cultivation environment- View project

All content following this page was uploaded by Shuhei Nishida on 22 May 2014.

The user has requested enhancement of the downloaded file.

 An Online Adaptation Control System Using mnSOM

Shuhei Nishida1, Kazuo Ishii2, and Tetsuo Furukawa2

1
Mechanical Systems and Environmental Engineering, The University of Kitakyushu,
1-1 Hibikino, Kitakyushu, Fukuoka 808-0135, Japan
[email protected]
2
Brain Science and Engineering, Kyushu Institute of Technology,
2-4 Hibikino, Kitakyushu, Fukuoka 808-0196, Japan
{ishii, furukwa}@brain.kyutech.ac.jp

Abstract. Autonomous Underwater Vehicles (AUVs) are attractive tools to

survey earth science and oceanography, however, there exists a lot of problems
to be solved such as motion control, acquisition of sensor data, decision-
making, navigation without collision, self-localization and so on. In order to re-
alize useful and practical robots, underwater vehicles should take their action by
judging the changing condition from their own sensors and actuators, and are
desirable to make their behavior, because of features caused by the working en-
vironment. We have been investigated the application of brain-inspired tech-
nologies such as Neural Networks (NNs) and Self-Organizing Map (SOM) into
AUVs. A new controller system for AUVs using Modular Network SOM
(mnSOM) proposed by Tokunaga et al. is discussed in this paper. The proposed
system is developed using recurrent NN type mnSOM. The efficiency of the
system is investigated through the simulations.

Keywords: Adaptive Control, mnSOM, Autonomous Underwater Vehicle.

1 Introduction

Autonomous underwater vehicles (AUVs) have great advantages for activities in

deep oceans [1], and are expected as the attractive tool for near future underwater
development or investigation. However, AUVs have various problems which should
be solved for motion control, acquisition of sensors’ information, behavioural deci-
sion, navigation without collision, self-localization and so on. We have been investi-
gating the application of neural network technology into the AUVs focusing on the
capability of neural networks (NNs) such as learning, nonlinear mapping. Consider-
ing the various AUV specific features, several methods proposed using NNs, Self-
Organizing Map [2] and so on [3], [4]. In order to realize the useful and practical
robots which can work in the ocean, underwater vehicles should take their action by
judging the changing condition from their own sensors and actuators, and are desir-
able to make their behaviors with limited efforts of the operators, because of the
features caused by the working environment. Therefore, the AUVs should be
autonomous and have adaptive function to their environment. AUVs have non-liner

I. King et al. (Eds.): ICONIP 2006, Part I, LNCS 4232, pp. 935 – 942, 2006.
© Springer-Verlag Berlin Heidelberg 2006
936 S. Nishida, K. Ishii, and T. Furukawa

coupled dynamics in six degrees of freedom, and the changes of the equipments of
robots have influence on the control system. In the previous adaptive control method
in [3], the information of initial states is getting lost gradually during the process of
adaptation. Therefore, new method which keeps the information of initial state or
previous environment and adapt to new environment is should be developed for
increasing the efficiency of the learning and reducing the learning cost with the use
of the former environmental information which the robot had learned. Human be-
ings are assumed to have a kind of modular architecture about the dynamics and
controller. This modular architecture is called MOdule Selection And Identification
Control (MOSAIC) [5]. This method allows that multiple pairs of dynamics and
controller modules are obtained. The MOSAIC and reinforcement learning are ap-
plied into the task of swinging up a pendulum [6].
In this paper, a new self-organizing decision making system for AUVs using
Modular Network Self-Organizing Map (mnSOM) [7] proposed by Tokunaga et al. is
described. The mnSOM is an extension of the conventional SOM in which each vec-
tor unit is replaced by function modules such as NN, SOM. Several applications are
repotted [8]-[11]. The proposed system is developed using recurrent NN type
mnSOM. The efficiency of the system is investigated through the simulations.

2 Adaptive Control System Using mnSOM

The proposed controller of the robot consists of Recurrent Neural Network (RNN). As
shown in Fig.1, the adaptive controller is realized using RNN-mnSOM. The making
processes of control system have following three steps. (a) Identification of Forward
Model Modules, (b) Adaptation of Controller Modules using the Forward Model
Modules and (c) Implementation of the Control Module to Robot Control.
At process (a) are shown in Fig.1-(a), Forward Model Modules (FMMs) are ac-
quired. Several time series of motion data which represents different dynamics corre-
sponding to the relationship of control signal and states of the robot such that one
module represents an option dynamics property in advance. These time series data are
fed into RNN-mnSOM, and FMMs are obtained.
At process (b) in Fig.1-(b), Controller Modules (CMs) are acquired using the fixed
FMMs which are obtained by process (a). The target states variables are given to CMs
and output data (control signals) calculated in CMs are given to all FMMs. The opti-
mization of CMs is carried out by back-propagation method using the square error
between target states and estimated states of FMMs regarding a FMM and a CM as
one NN.
Figure 1-(c) shows process (c). The condition of robot is determined as the best-
matching module (BMM) by feeding a certain time series data into each FMM. After
the FMM is selected, the output of the CM corresponding to the FMM is given to the
robot.
The adaptive controller using mnSOM is realized according to the processes
(a)-(c).
 An Online Adaptation Control System Using mnSOM 937

Forword Si
Model Sj
Module
(FMM )

FMM

(a) Identification of Forward Model Modules

Target
Controller
Module

Off-line Adaptation
CM

Force
State

FMM

Evaluation

(b) Adaptation of Controller Modules using the Forward Model Modules

Sx Target

FMM Update
CM Adaptation
Best Matching
Module

FMM
CM
On-line
Force Adaptation
State

(c) Implementation of the Control Module to Robot Control

Fig. 1. Learning Processes of an Adaptive Controller System using RNN-mnSOM

3 Simulations
3.1 Forward Model Modules
In order to evaluate the identification capability of RNN-mnSOM, some set of time series
data is prepared by changing the parameter M and C in the following equation of motion.
938 S. Nishida, K. Ishii, and T. Furukawa

F = Mx + Cx x (1)

Here, F means force, x is velocity, x is acceleration of robot. M and C are mass
including added mass and drag coefficient, respectively. The limit-cycle examination
that force is inverted, if absolute of velocity is over 0.2 [m/s] is carried out. Nine sets
of limit-cycle time series are prepared by changing the parameters shown in Table 1.
These time series are measured in 50 [sec], sampling rate is 10[Hz] and are fed into
6x6 lattice modules RNN-mnSOM. Figure 2 shows result of 100,000 times learning.
In Fig.2, each square expresses module, and the color means the distance among the
neighbors. If a square is black, the module has different dynamics as against the
neighbors. And white means similar. In each BMM, time series data is plotted; veloc-
ity is thin line, acceleration is bold one. The BMM for teaching data are located in the
positions shown in Table 2.
The comparison of estimated parameters from input data and FMMs is as shown in
Fig.3. The squares are parameters of input time series, the crosses are estimated pa-
rameters from FMMs using least-square method. Figure 3 shows that the parameters
of FMMs are distributed among input.
In Fig.4, the relationship between acceleration and velocity are plotted. The posi-
tion of each graph corresponds to FMMs. These parallelograms become smaller in
area from left side to right, so that parameter M becomes big. And, gradient becomes
bigger from upper side to lower, therefore parameter C becomes big.

3.2 Controller Model Modules

CMs which are connected to fixed weights FMMs are optimized. Target position is
0.5[m] during 0 ~ 25 [sec] and -0.5 [m] during 25 ~ 50 [sec]. Target velocity is 0.0
[m/s]. And, sampling rate is 10 [Hz].
The result that the iteration of learning is 15,000 times is shown in Fig.5. In each
square which means CM, time series is plotted using CM; the horizontal axis is time.

Fig. 2. A Forward Model Map Obtained from the Time Series of Limit Cycle Simulation Data
 An Online Adaptation Control System Using mnSOM 939

In vertical axis, dash line is position of robot, gray line is target position and solid line
is control force. All CMs follow target so that control corresponding FMMs.

Table 1. Coefficient M and C for Limit-Cycle Motion

Di: (M, C)
D0: (80, 25) D3: (90, 25) D6: (100, 25)
D1: (80, 50) D4: (90, 50) D7: (100, 50)
D2: (80, 100) D5: (90, 100) D8: (100, 100)

Table 2. Best Matching Module for each Data Class

Di is located in (x, y)
D0: (0, 0) D3: (4, 0) D6: (6, 1)
D1: (0, 0) D4: (4, 1) D7: (6, 3)
D2: (2, 6) D5: (4, 6) D8: (6, 6)

3.3 On-Line Adaptive Simulation

The simulation to compare adaptability for unlearned data between adaptive control-
ler proposed by ref. [3] (hereafter, reference system) and this proposed controller are
carried out.
Figure 6 shows transition of evaluation values. In these graphs, horizontal axis is
learning steps and vertical axis is evaluation value on log-scale. Uppers are Forward
Model Error and lowers are controller Error. Solid lines are obtained from the pro-
posed system and dot-lines are from the reference system.

Forward Model Modlue

Input Data
25

50
C
75

100
80 90 100
M

Fig. 3. Forward Model Map Evaluation in M-C Space by the Least Square Method
940 S. Nishida, K. Ishii, and T. Furukawa

Fig. 4. Acceleration-Velocity Relationship Obtained from Limit Cycle Simulation with FMMs

On reference system, the evaluation value of forward model become big at early
stage in learning. And then, it becomes smaller. According to decrease of forward
model error, controller error becomes big. And then, controller adapt to input at
20[sec] in the case of 25, and at 125[sec] in the case of 120.
On proposed system, the forward model module which is expressed given time se-
ries exits. Therefore, at early stage of leaning, proposed system is needed few adapta-
tion. The adaptability of proposed system is better than reference system.

Fig. 5. Acquisition Simulation of 6x6 Controller Network Modules

 An Online Adaptation Control System Using mnSOM 941

Fig. 6. Transition of Evaluation Values

4 Conclusions
The adaptive controller using mnSOM is proposed. The Forward Model Map for
Dynamics identification and the Controller Map are introduced to realize the adaptive
controller system. In the FMMs, the characteristic and interpolations among several
input data are expressed. In the CMMs, suitable controllers corresponding FMMs are
obtained. The efficiency of proposed system will be investigated through the experi-
ments using AUV.

Acknowledgment
This work was supported by a 21st Century Center of Excellence Program, “World of
Brain Computing Interwoven out of Animals and Robots (PI: T. Yamakawa)” granted
in 2003 to Department of Brain Science and Engineering, (Graduate School of Life
Science and Systems Engineering), Kyushu Institute of Technology by Japan Ministry
of Education, Culture, Sports, Science and Technology.

References
1. T. Ura (1989). “Free Swimming Vehicle PTEROA for Deep Sea Survey,” Proc. of
ROV'89, pp.263-268
2. T. Kohonen, (1982), “Self-organized formation of topologically correct feature maps,”
Biological cybernetics, vol. 43, pp.59-69
3. K. Ishii and T. Ura (2000). “An adaptive neural-net controller system for an underwater
vehicle,” Journal of IFAC Control Engineering Practice, Vol. 8, pp.177-184
4. S. Nishida, K. Ishii and T. Ura, (2004), “A Self-Organizing Map Based Navigation System
for an Underwater Robot,” IEEE International Conference on Robotics and Automation,
pp.4466-4471
942 S. Nishida, K. Ishii, and T. Furukawa

5. M. Haruno, D.M. Wolpert and M. Kawato, (2001), “Mosaic: Module selection and identi-
fication for control,” Neural Computation, vol.13 no.10, pp.2201-2220
6. K. Doya, K. Samejima, K. Katagiri and M.Kawato, (2002), “Multiple Model-based Rein-
forement Learning,” Neural Computation, vol.14 pp.1347-1369
7. K. Tokunaga, T. Furukawa and S. Yasui, (2003), “Modular Network SOM: Extension of
SOM to the realm of function space,” WSOM'03, pp.173-178
8. T. Furukawa, K. Tokunaga, K. Moroshita and S. Yasui, (2005), “Modular Network SOM
(mnSOM): From Vector Space to Function Space,” International Joint Conference on
Neural Networks
9. K. Tokunaga, T. Furukawa, (2005), “Nonlinear ASSOM Constituted of Autoassociative
Neural Modules,” 5th Workshop on Self-Organizing Maps
10. T. Furukawa, T. Tokunaga, S. Kaneko, K. Kimotsuki and S. Yasui, (2004), “Generalized
Self-Organizing Maps (mnSOM) for Dealing with Dynamical Systems,” International
Symposium on Nonliner Theory and its Applications, pp.231-234
11. T. Minatohara, T. Furukawa, (2005), “Self-Organizing Adaptive Controllers: Application
to the Inverted Pendulum” 5th Workshop on Self-Organizing Maps, pp.44-48

View publication stats