Information System for Autonomous Mobile Robot Interaction

Position Paper

Nadina Battagliotti, Daniela L´opez De Luise and Jin Sung Park

AIGrouo, Palermo University, Mario Bravo 1050, Buenos Aires, Argentine

Keywords:

Mobile Robots, Distributed Information Systems, Distributed Self-organizing Maps, Computational Intelli-

gence.

Abstract:

This paper presents what the authors consider to be a new distributed memory architecture for self-organizing

distributed systems. The work in progress presented here focuses on the representation of a distributedmemory

for multi-robot systems. It states the main characteristics of a shared environment and provides suitable

interfaces to allow autonomous mobile robots to query, publish and exchange relevant information without

requiring a central data storage. The memory is based on a combination of fuzzy systems, distributed Self-

Organizing Maps (SOM), a new speciﬁc design information handshake between robots and a new model for

approach for world-map global administration.

1 INTRODUCTION

This paper describes a proposal to solve the problem

of allocation of different types of services to a set

of heterogeneuos autonomous nobile robots. Multi-

robot solutions have been gaining popularity with the

development of cooperative strategies and enhanced

communication technologies but also due to the new

advances towards the internet of the future. Multi-

robot systems offer various advantages against single-

robot solutions because of the distributed character

that allows for parallel processing and robustness

thanks to the redundancy. Lot of work has been done

in cooperative robot systems, like in localization, for-

mation control, cooperativetransport and sensing, etc.

(Boella G., 2003) (Findler, N.V., 1995) (Findler, N.V.,

2005) (Jennings, N., 1995)(Maza, I., 2005)(Hachour,

O., 2008)(Dillmann, R., 1995)

One important topic when dealing with multi-

robot systems is the knowledge representation, which

can be global (collective awareness of a whole robot

team) or local (awareness of each individual mem-

ber). In centralized systems, the implementation of

a global memory can require lots of resources. Be-

sides, centralized systems are prone to total failure

if the central processing unit fails. Distributed sys-

tems on the other hand are more robust due to the

redundant character of those architectures allowing

an individual unit to get the role of another unit that

has broken down. Nevertheless, a fully distributed

system presents a high complexity and is more dif-

ﬁcult to manage than a centralized one. This brings

along the need for an advanced memory representa-

tion that combines the advantages of both approaches

distributed and centralized. The implementation of

such memory requires not only the deﬁnition of soft-

ware architecture but also the speciﬁcation of hard-

ware requirements regarding communicationbetween

the individual entities in the whole system. The sys-

tems knowledge should also be shared with human

users that can interact with the system via sponta-

neous connection using a smart phone. For scalabil-

ity, an information exchange between systems should

also be made possible. The ﬁeld of Computational

Intelligence offers various concepts and tools to de-

velop a suitable approach to the presented situation.

Following is an overview of the current Soft Comput-

ing alternatives in the ﬁeld.

Neural Networks are used for a variety of appli-

cations from speech processing, vision, optimization,

communications, classiﬁcation activities, control sys-

tems, associative memories, etc (Simencio, E., 1992).

In all the cases the network is made of a set of process-

ing units (neurons) connected to others via links. The

links have different strengths according to the knowl-

edge acquired by the neurons. Every neuron has a

vector of links usually named as weight vector. Most

of Neural Networks require a training set to internal-

ize the patterns of the problem. During the learning

process, the desired outputs are compared to the cur-

376

Battagliotti N., López De Luise D. and Park J..

Information System for Autonomous Mobile Robot Interaction - Position Paper.

DOI: 10.5220/0004222403760382

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 376-382

ISBN: 978-989-8565-38-9

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

rent outputs. If there is any difference between them,

the internal weights are iteratively changed accord-

ing to certain learning algorithms. The process stops

when a convergence criterion is met.

The number of neurons, its disposition into lay-

ers, the way they are interconnected, the approach

used to update weights and the mechanism by which

a neuron forwards its knowledge to another neurons

conform the architecture of the network. The archi-

tecture determines its velocity and precision (Wilson,

W., 1993).

But in certain architectures it is possible to let the

network learn in a competitive fashion. Competitive

learning is a rule based on the idea that only one neu-

ron from a given iteration in a given layer will ﬁre at

a time. These types of networks are known as self-

organizing networks. In this case the current input is

compared to the weight vectors of the neurons and the

closest vectors determine the winning neuron (Koho-

nen, T., 1991).

The Kohonen’s Self-Organizing Maps (SOM) is

probablythe most relevant competitivenetwork archi-

tecture. It has been used successfully in many prob-

lems such as automatic speech recognition (Behme,

H., 1993), clinical voice analysis (Godino-Llorente,

J., 2000) (Hiltunen, T., 1993), etc. Speciﬁcally in the

robotic ﬁeld it is intensively used in many problems

like robot navigation (Hu, H., 2000), cloud classiﬁ-

cation from satellite images (Kilpatrick, D., 1995),

kinematic of a robot arm (Kieffer, S., 1991), adap-

tive controller for autonomous mobile robots (Kim,

Y., 1992), robot motion planning in dynamic envi-

ronments (Knobbe, A., 1995), spatial understand-

ing and temporal correlation for mobile robot (Kr-

ishna, K., 2000), etc. This project uses an exten-

sion of the SOM named Distributed Self-Organizing

Map (DSOM) where NN is trained using decentral-

ized learning. There are many DSOM implementa-

tion approaches(Pascual-Marqui,A., 2001) (Lobo, V.,

1998) and applications in this ﬁeld (?). In (Lobo, V.,

1998) it is applied with the sonar noise to detect the

shape of a navy. In the context of this project, it is also

used for obstacle recognition. But the information

comes from different types of sensors that feeds asyn-

chronously even for the same object. This requires a

redesign of parts of the original DSOM model to let it

aggregate patterns from partial, noisy and incomplete

data of the obstacle.

Regarding the world map, the area is segmented

into portions. Boundaries between those segments

are not sharp but can be fuzzy. Fuzzy logic (FL) has

been successfully used to solve many complex prob-

lems. It is inherently robust, can be modiﬁed and

tweaked easily to improve the system performance

and can control nonlinear systems (Cox, E., 1992)

(Lee, C., 1990). In general, FL is used in many ap-

plications where noise, error and missing data is typ-

ical, mainly in control theory and artiﬁcial intelli-

gence. Among other applications are wireless com-

munications (Erman, M., 2009), velocity induction

for a motor (Kumar, V., 2005), operational meteorol-

ogy (Murtha, J., 1995), data mining and Information

Retrieval(Meunier, B., 2007), etc. This project pro-

poses a new FL usage in the world map segment ad-

ministration. Although it has been used in philoso-

phy (Kosko, B., 1993) and psychology (Didelon, C.,

1991) for evaluating the individual usage and charac-

teristics of its cognitive maps (or interpretative maps),

the approach to manage segments of a robots world-

map is a contribution in the ﬁeld. In the context of this

work, FL is important to guarantee the proper control

of the robot during the migration from one segment

to another in the map. If the services provided by the

robot or its abilities are critical, the fuzzy boundaries

allow the system to take a period of time to adapt the

current activity softly.

This article is organized as follows: Section 2

describes the proposed architecture for a distributed

memory. Later in 3, the proposed information ap-

proach is presented. Section 4 describes the software

that is used in this work. Conclusions and future work

are presented in 5.

2 PROPOSED ARCHITECTURE

2.1 Global Description

Fig. 1 shows the proposed global system architec-

ture. The physical workspace covered by the system

is logically split into segments (S

). Every partition

has a Logical Central Point (LCP) that regulates the

activity of the agents within S. It can also handle a

local world map and deﬁne its boundaries. A team

of autonomous robots is constantly providing services

within each segment. Any robot is able to perform a

set of different services. To do so, it can download

from the LCP the service that it should provide and

the corresponding hardware conﬁguration required to

perform these services. It can also upload its sensory

data to its corresponding LCP.

2.2 Logical Central Point

A Logical Central Point (LCP) is a logical role that

behaves as the manager of a segment S. A LCP can

be any data processing unit, for example a desktop

computer of a processor mounted on a robot. Custom

InformationSystemforAutonomousMobileRobotInteraction-PositionPaper

377

Figure 1: Representation of the proposed global architec-

ture.

services are physically located and associated to one

or more LCP. The activity of any LCP can be:

• Manage shared information with other LCPs us-

ing a speciﬁc LCP-LCP handshake.

• Manage information of the robots in its segment S

using a speciﬁc LCP-robot handshake.

• Work as a petition center.

• Work as a passive receiver of service requests

from another LCP.

• Manage local storage block to be able to make in-

ferences and command fast actions to the robots

in S.

• Allocate robots to service petitions.

• Request robots to other LCPs.

• Send robots to other segments upon a request from

other LCP.

• Delegate its role to other unit.

• Make a robot to change its service.

• Make a robot to change its hardware conﬁguration

(sending it to a speciﬁc LCP whose segment is

advocated to this tasks).

• Manage a distributed world-map updated with the

information received from the robots in S.

To perform its activity, LCP has a local storage

with information like:

• ID and service of each robot within its segment.

• Number of robots in its corresponding segment S

• Local world map deﬁned as a set of cells, each

representing a physical area.

• The boundaries of S

In this approach, every world map remains private

for its corresponding LCP. Each segments boundary

is deﬁned by:

• Border section

• ID of the neighborthat shares everyborder section

• Border strength

The border strength is deﬁned in terms of a fuzziness

function. For example, in Fig. 1, the neighbors of S1

are:

• S2: with border S1-2 and fuzziness = 0 (i.e. no

fuzziness)

• S4: with border S1-4 and fuzziness = 5 (fuzziness

with member function 5)

• S5: with border S1-5 and fuzziness = 3 (fuzziness

with member function 3)

In this context, the boundary is one or more borders.

And a border is a set of cells of the local world map.

In all cases, fuzziness refers to a number of a pre-

deﬁned member function. Table 1 is an example of

functions that are classical in the ﬁeld.

Table 1: Classical membership functions.

ID Membership formula Shape

0 None None

triangular

trapezium

trapezoid

normalized gaussian

generalized bell

2 sigmoidals

7 ... ...

The membership function shape is very important

to deﬁne the transition behavior when a robot trespass

the boundaries of the current segment. Among others,

it deﬁnes:

• if the Membership Function is Symmetric: the

LCP consumes the same period of time to get

ready to delegate the robot and to release the robot

to the target LCP (see Fig. 2).

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

378

Figure 2: Fuzzy border trespassing.

• if the Membership Function is Asymmetric: the

LCP requires different amount of time to prepare

the robot delegation and to release the robot.

• if there is no Membership Function: the original

LCP delegates the navigation and control of the

robot instantly to the target LCP.

The strength of the border can also be understood

not only as the ability but also as the disposition of

the current LCP to allow a robot to pass-through that

segment of the boundary.

2.3 World Map

A world map is a representation of the internal in-

formation of a segment. As the LCP receives infor-

mation via handshaking with the robots within its as-

signed segment, the current information about obsta-

cles and robots positions in the world map gets con-

stantly updated.

By using the concept of Distributed Self Orga-

nized Maps (DSOM) the LCP is able to recognize

known obstacles from asynchronous and partial in-

formation gathered from the robots in the current seg-

ment. Although by deﬁnition, the SOM is intended to

classify data into known categories (e.g. types of ob-

jects), it was also demonstrated (Lobo, V., 1998) that

DSOMs are able to recognize new objects.

A world map is a small part of the global

workspace, because it is intended to model just the

physical space covered by a given segment S

(and

managed by the corresponding LCP).

2.4 Robot Distribution

Robots are assigned groups to segments handled by

LCPs. Every robot can only belong to a segment.

When a LCP detects an increase in the workload (e.g.

because the robots have many service requirements,

or because the LCP itself has received many queries

from a terminal operator), it can negotiate with the

nearest LCPs to get additional robots.

Conversely, a LCP can receive from its neighbors

a request for borrowing one or more robots from its

segment. Then it will decide whether to accept or not

those requests according to its current own workload.

We assume that each robot is able to navigate

autonomously within its segment with a predeﬁned

collision-avoidance function and that a robot will

never cross the boundaries deﬁned in the LCP world-

map. The LCP will monitor the states of its robots to

improve their collision-avoidance algorithms and so

prevent deadlocks and critical situations.

While executing a service, a robot simultaneously

keeps sampling its environment so its LCP can update

its world map. It also sends a fragment of this infor-

mation to the closest robots in the same area.

If there is a dynamic obstacle within a segment,

then one or more robots will detect its presence and

forward it to its LCP and closest robots. All this in-

formation is used by the LCP to recognize the type

of obstacle (by using the DSOM). The LCP can also

inform to another neighbor LCP that the obstacle is

about to enter to its area and provide obstacle data

(i.e. velocity, shape, type and last known position).

Upon request, the LCP can reassign a robot a new

service. If the robot hardware features necessary to

provide the new service require human intervention

(for instance to exchange a tool, remove or add a

wheel, add a new hardware module, etc.), the LCP

can command the robot to trespass the boundary of

the current segment and pass to a special maintenance

segment. After the robot has its physical conﬁgu-

ration changed, it updates its internal status, returns

to its former segment and follows the LCP indication

to perform the new service. This can also imply the

robot being transferred to a new segment.

3 INFORMATION APPROACH

Each robot uses a negotiation with its neighbors. Dur-

ing that negotiation, it shares a reduced amount of in-

formation (only with its neighbors). The approach is

similar to that used in the computer networks to solve

the problem of routing information from a source to

target. But in this proposal the data being routed con-

sists of a bunch of information required to acknowl-

edge the features of the surroundings.

Below there is a comparison between the typ-

ical activity of a router and the Distributed robot

Handshaking Approach (DHA)

. Table 2 shows

the features taken from the router approach that we

adapted for robot-to-robothandshaking. While within

a router, data packets are exchanged, in the DHA, this

occurs with sampling data from the robots sensors.

Each robot in the DHA is analog to a node in a routing

It is mentioned in section it is mentioned in section Fu-

ture Work as number 15

InformationSystemforAutonomousMobileRobotInteraction-PositionPaper

379

scheme. The difference is that in the router scheme,

the destination node can be each node in the network,

while in the DHA, the destination node can only be

a neighbor node. Furthermore, contrary to the router

approach, the DHA considers every LCP storing up-

date information about the environment; within the

router approach, a table of routings is stored.

Table 2: Features extracted from router approach for robot-

to-robot handshaking.

Feature Router DHA

Content Data packets Current sampling

changes

Source A node calling

another

The robot that is

sampling

Target The destination

candidates (can

be more than

one)

The robots that are

closest

Feed-

back

Navigation status

of each candidate

Sampling changes of

each closest robots

(when keep being

closer)

Activity Select best desti-

nation using the

feed-back from

every candidate

Select best destina-

tion using the feed-

back from every can-

didate

Storage Table of routings Current cells and sur-

roundings cells

Updating

process

Upon every

network conﬁgu-

ration change

Upon every physical

change in the land

that is being detected

Something similar is implemented between LCP

and robots. The information exchanged is reduced to

allow the LCP to manage a number of robots properly,

as in Table 3. In this case, in the DHA, the source is

given by the LCP itself. Whereas every robot inside

the LCPs segment can communicate with the LCP, in

the handshake, only the robots that are capable to per-

form the task in question are considered targets. In

this handshaking, the information stored is given by

the current set of tickets (queries) and services being

performed by the robot.

The handshake between two LCPs is deﬁned as in

Table 4.

This handshaking is analog to that for robot-to-

robot communication. In this case, only neighbor

LCPs can be set as targets. The information to be ex-

changed here is represented by the global changes in

the environment affecting more than one LCP. Addi-

tionally, information about tickets under negotiation

and about robot sharing is also exchanged.

Table 3: Features extracted from router approach for robot-

to-robot handshaking.

Feature Router DHA

Content Data packets Current world-map

changes of interest

for the robot

Source A node calling

another

The LCP that has to

command a robot

Target The destination

candidates (can

be more than

one)

The robots that are

able to perform the

task

Feed-

back

Navigation status

of each candidate

Current status and

availability

Activity Select best desti-

nation using the

feed-back from

every candidate

Select the best robot

to accomplish the

task and afterwards

monitor it

Storage Table of routings Current tickets and

services being per-

formed by each robot

Updating

process

Upon every

network conﬁgu-

ration change

Upon every new

ticket or service

completion

4 SYSTEM REQUIREMENTS

4.1 Software

The software tools are mostly open, except for the

simulators required to test the functionality:

• Eclipse framework

• Java 2

• svn, subversion system

• Mantis, bug tracker

• LMS simulator

5 CONCLUSIONS AND FUTURE

WORK

There are many pending items for the project. Among

others the following items are to be reﬁned yet:

1. The internal language (numeric or textual) to rep-

resent service features

2. The internal language (numeric or textual) to rep-

resent hardware features

3. The internal language (numeric or textual) to rep-

resent sampling information

4. Parameters to deﬁne the segments:

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

380

Table 4: Features extracted from a router approach for hand-

shaking between LCPs.

Feature Router DHA

Content Data packets Current global

changes that affects

more than one LCP.

Inter LCP ticket ne-

gotiation to provide

the services. Robots

sharing (between two

LCPs)t

Source A node calling

another

The LCP that has to

fulﬁll a request

Target The destination

candidates (can

be more than

one)

The neighbor LCPs

Feed-

back

Navigation status

of each candidate

Current status and

availability of the tar-

get LCPs

Activity Select best desti-

nation using the

feed-back from

every candidate

Select the best LCP

to share the task and

assign it

Storage Table of routings Current status and

services

Updating

process

Upon every

network conﬁgu-

ration change

Upon events

• if the boundaries are static or may change its

size and shape

• if the boundaries are fuzzy o sharp

• if the LCP deﬁnes its boundaries automatically

or it needs external information to know its

boundaries

5. Parameters for robot-central point handshake:

• LCP to robot: current service, hardware conﬁg-

uration

• Robot to LCP: sampling data

6. Parameters for the handshake between central

points (LCP).

7. Evaluate whether to implement traditional Koho-

nen’s SOM or Fuzzy c-Means (less computational

complexity).

8. The implementation of the logical central points

in one or more physical processor, and its real lo-

calization in the real world.

9. Deﬁne the approach to balance the resources dis-

tribution (resources are parts of robots, robots,

knowledge, parts of its world-maps, etc.)

10. Deﬁne the concept of closest robots: whether they

are all the robots within certain distance, or the

directly linked conceptually. It must be deﬁned a

new metric to assess that.

11. Deﬁne if a LCP can be used as a petition centers

or it just receives petitions from a central point.

12. Deﬁne if traditional SOM or fuzzy c-means has to

be used.

13. Deﬁne the special segments. The hardware con-

ﬁguration can be changed manually or using

robotics.

14. Deﬁne if there is just one special segment or

many. In the later case, the workload administra-

tion approach.

15. Deﬁne the contention algorithms between robots

to reduce their information storage and amount of

data to be interchanged.

16. Deﬁne the contention algorithms between a robot

and its LCP, to reduce their information storage

and amount of data to be interchanged.

17. Deﬁne the local robot memory updating approach.

18. Deﬁne the number of cells in the world-map that

corresponds to the surroundings term.

19. Deﬁne a hierarchy

20. Deﬁne roles & inter-role communication

21. Create a wiki to share and publish information

about the project, put the publications, etc

22. Evaluate the set of member functions according to

the service and hardware conﬁgurations.

23. Evaluate the shape and size of segments.

24. Evaluate the dynamic of changes in the bound-

aries of the segments.

25. Evaluate the proper set of fuzzy operators to work

with the fuzzy boundaries.

REFERENCES

Behme, H. (1993). Speech processing by hierarchical seg-

ment classiﬁcation. In International Joint Conference

on Neural Networks. IEEE Press.

Boella, G.(2003). Norm governed multiagent systems: The

delegation of control to autonomous agents. In Proc.

of IEEE/WIC IAT. IEEE Press.

Cox, E. (1992). Fuzzy fundamentals. In IEEE Spectrum.

IEEE Press.

Didelon, C., de Ruffray, S. (1991). A world of interstices:

Fuzzy logic approach in the analysis of interpretative

maps. In The Cartographic Journal. British Carto-

graphic Society.

InformationSystemforAutonomousMobileRobotInteraction-PositionPaper

381

Dillmann, R., Kauser, M., Wallner, F., Weckesser, P. (1995).

Priamos: An advanced mobile system for service, in-

spection, and surveillance tasks. In Modelling and

planning for sensor based intellinget robot systems.

H. Bunke, T. Kanade, H. Noltemeirer Eds.

Erman, M., Mohammed, A. (2009). Fuzzy logic applica-

tions in wireless communications. In IFSA-EUSFLAT.

IFSA-EUSFLAT.

Findler, N.V. (1995). Multi-agent coordination and cooper-

ation in a distributed dynamic environment with lim-

ited resources. In Artiﬁcial Intelligence in Engineer-

ing. Elsevier.

Findler, N.V. (2005). A robust business resource manage-

ment framework based on a peer-to-peer infrastruc-

ture. In 7th IEEE Int. Conf. on E-Commerce. Kluwer

Academic.

Godino-Llorente, J. (2000). Non supervised neural net ap-

plied to the detection of voice impairment. In IEEE In-

ternational Conference on Acoustics, Speech and Sig-

nal Processing. IEEE Press.

Hachour, O. (2008). Path planning of Autonomous Mobile

robot. In Int. Journal of Systems Appl., Eng. & De-

velop.. University Press. Issue 4(2).

Hiltunen, T., Leinonen, L. (1993). Visualization and classi-

ﬁcation of voice quality with the self-organizing map.

In International Conference on Artiﬁcial Neural Net-

works. S. Gielen and B. Kappen.

Hu, H. (2000). Landmark-based navigation of industrial

mobile robots. University of Portsmouth, Portsmouth,

1st edition.

Jennings, N. (1995). Controlling cooperative problem solv-

ing in industrial multi-agent systems using joint inten-

tions. In Artiﬁcial Intelligence. Elsevier.

Kieffer, S., Morellas, S. (1991). Neural network learning

of the inverse kinematic relationships for a robot arm.

In International Conference on Robotics and Automa-

tion. IEEE Computer Society Press.

Kilpatrick, D. (1995). Unsupervised classiﬁcation of

antarctic satellite imagery using kohonens self-

organizing feature map. In IEEE International Con-

ference on Neural Networks. IEEE Press.

Kim, Y. (1992). Robust adaptive control of an autonomous

mobile robot. In Second International Conference on

Automation, Robotics and Computer Vision. Nanyang

Technol. Univ.

Knobbe, A., Kok, J. (1995). Robot motion planning in un-

known environments using neural networks. In In-

ternational Conference on Artiﬁcial Neural Networks.

NParis : EC2 & Cie.

Kohonen, T. (1991). The self Organizing Map. IEEE Press,

New York, 1st edition.

Kosko, B. (1993). Fuzzy Thinking: The New Science of

Fuzzy Logic. Hyperion, New York, 1st edition.

Krishna, K. (2000). Spatial understanding and temporal

correlation for a mobile robot. In Spatial-Cognition-

and-Computation. Lawrence Erlbaum Associates Inc.

Kumar, V., Joshi, E.R. (2005). Hybrid controller based in-

telligent speed control of induction motor. In Journal

of Theoretical and Applied Information Technology.

Little Lion Scientiﬁc.

Lee, C. (1990). Fuzzy logic in control systems. In IEEE

Trans. on Systems, Man, and Cybernetics. IEEEPress.

Lobo, V., Bandeira, N. (1998). Ship recognition using dis-

tributed self organizing maps. In Neural Networks

Proceedings of the International Conference EANN

98. Systems Engineering Association.

Maza, I., Viguria, A., Ollero, A. (2005). Cooperacin

helicptero autnomo-robot mvil terrestre para aplica-

ciones de bsqueda y localizacin. In Proc. of I Con-

greso Espaol de Informtica. pp 109-116, Thomson,

Granada, Spain, 2005.

Meunier, B. (2007). Real-world fuzzy logic applications in

data mining and information retrievaly. In A Spectrum

of Theoretical and Practical Issues. Springer.

Murtha, J. (1995). Applications of fuzzy logic in opera-

tional meteorology. In Scientiﬁc Services and Pro-

fessional Development Newsletter. Canadian Forces

Weather Service.

Pascual-Marqui, A. , Pascual-Montano, K. (2001).

Smoothly distributed fuzzy c-means: a new self-

organing map. In Pattern Recognition. IEEE Press.

Simencio, E. (1992). Artiﬁcial Neural Networks. IEEE

Press, New York, 1st edition.

Wilson, W. (1993). A comparison of architectural alter-

natives for recurrent networks. In Fourth Australian

Conference on Neural Networks. Sydney University

Electrical Engineering.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

382